diff --git a/app.py b/app.py new file mode 100644 index 0000000000000000000000000000000000000000..40fea624ebb99ce1d5629951da4564125c8eb6a1 --- /dev/null +++ b/app.py @@ -0,0 +1,123 @@ +import argparse +import requests +import logging +import os +import gradio as gr +import numpy as np +import cv2 +import torch +import torch.nn as nn +from PIL import Image +from torchvision import transforms +from timm.data.constants import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD +from timm.data import create_transform +from config import get_config + +from collections import OrderedDict + +import detectron2.utils.comm as comm +from detectron2.checkpoint import DetectionCheckpointer +from detectron2.config import get_cfg +from detectron2.data import MetadataCatalog +from detectron2.engine import DefaultTrainer as Trainer +from detectron2.engine import default_argument_parser, default_setup, hooks, launch +from detectron2.evaluation import ( + CityscapesInstanceEvaluator, + CityscapesSemSegEvaluator, + COCOEvaluator, + COCOPanopticEvaluator, + DatasetEvaluators, + LVISEvaluator, + PascalVOCDetectionEvaluator, + SemSegEvaluator, + verify_results, + FLICKR30KEvaluator, +) +from detectron2.modeling import GeneralizedRCNNWithTTA + +def parse_option(): + parser = argparse.ArgumentParser('RegionCLIP demo script', add_help=False) + parser.add_argument('--config-file', type=str, default="configs/CLIP_fast_rcnn_R_50_C4.yaml", metavar="FILE", help='path to config file', ) + args, unparsed = parser.parse_known_args() + + return args + +def build_transforms(img_size, center_crop=True): + t = [] + if center_crop: + size = int((256 / 224) * img_size) + t.append( + transforms.Resize(size) + ) + t.append( + transforms.CenterCrop(img_size) + ) + else: + t.append( + transforms.Resize(img_size) + ) + t.append(transforms.ToTensor()) + return transforms.Compose(t) + +def setup(args): + """ + Create configs and perform basic setups. + """ + cfg = get_cfg() + cfg.merge_from_file(args.config_file) + cfg.freeze() + default_setup(cfg, args) + return cfg + +''' +build model +''' +args = parse_option() +cfg = setup(args) + +model = Trainer.build_model(cfg) +DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load( + cfg.MODEL.WEIGHTS, resume=False +) +if cfg.MODEL.META_ARCHITECTURE in ['CLIPRCNN', 'CLIPFastRCNN', 'PretrainFastRCNN'] \ + and cfg.MODEL.CLIP.BB_RPN_WEIGHTS is not None\ + and cfg.MODEL.CLIP.CROP_REGION_TYPE == 'RPN': # load 2nd pretrained model + DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR, bb_rpn_weights=True).resume_or_load( + cfg.MODEL.CLIP.BB_RPN_WEIGHTS, resume=False + ) + +''' +build data transform +''' +eval_transforms = build_transforms(800, center_crop=False) +# display_transforms = build_transforms4display(960, center_crop=False) + +def localize_object(image, texts): + print(texts) + img_t = eval_transforms(Image.fromarray(image).convert("RGB")) * 255 + + print(img_t.shape) + model.eval() + with torch.no_grad(): + print(img_t[0][:10, :10]) + res = model(texts, [{"image": img_t}]) + + return res + + +image = gr.inputs.Image() + +gr.Interface( + description="RegionCLIP for Open-Vocabulary Object Detection", + fn=localize_object, + inputs=["image", "text"], + outputs=[ + gr.outputs.Image( + type="pil", + label="grounding results"), + ], + examples=[ + ["./elephants.png", "an elephant"], + ["./apple_with_ipod.jpg", "an apple"], + ], +).launch() diff --git a/config.py b/config.py new file mode 100644 index 0000000000000000000000000000000000000000..f17536ee6d5e9b2f87af6435d2dc6a38d5aa16d9 --- /dev/null +++ b/config.py @@ -0,0 +1,245 @@ +# -------------------------------------------------------- +# Unified Contrastive Learning (UniCL) +# Copyright (c) 2022 Microsoft +# Licensed under The MIT License [see LICENSE for details] +# Written by Jianwei Yang (jianwyan@microsoft.com) +# Based on Swin Transformer written by Zhe Liu +# -------------------------------------------------------- + +import os +import yaml +from yacs.config import CfgNode as CN + +_C = CN() +_C.VERBOSE = False + +# Base config files +_C.BASE = [''] + +# ----------------------------------------------------------------------------- +# Data settings +# ----------------------------------------------------------------------------- +_C.DATA = CN() +# Batch size for a single GPU, could be overwritten by command line argument +_C.DATA.BATCH_SIZE = 128 +# Path to dataset, could be overwritten by command line argument +_C.DATA.DATA_PATH = '' +# Dataset name +_C.DATA.DATASET = 'imagenet' +# Input image size +_C.DATA.IMG_SIZE = 224 +# Interpolation to resize image (random, bilinear, bicubic) +_C.DATA.INTERPOLATION = 'bicubic' +# Use zipped dataset instead of folder dataset +# could be overwritten by command line argument +_C.DATA.ZIP_MODE = False +# Cache Data in Memory, could be overwritten by command line argument +_C.DATA.CACHE_MODE = 'part' +# Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU. +_C.DATA.PIN_MEMORY = True +# Number of data loading threads +_C.DATA.NUM_WORKERS = 8 + +# ----------------------------------------------------------------------------- +# Model settings +# ----------------------------------------------------------------------------- +_C.MODEL = CN() +# Model name +_C.MODEL.NAME = '' +# Checkpoint to resume, could be overwritten by command line argument +_C.MODEL.RESUME = '' +# Number of classes, overwritten in data preparation +_C.MODEL.NUM_CLASSES = 0 +# Label Smoothing +_C.MODEL.LABEL_SMOOTHING = 0.1 +# Whether load pretrained model +_C.MODEL.PRETRAINED = '' +# Projection dimension +_C.MODEL.DIM_PROJECTION = 512 +# Mode specific +_C.MODEL.SPEC = CN(new_allowed=True) +# ----------------------------------------------------------------------------- +# Build Image Encoder +# ----------------------------------------------------------------------------- +_C.MODEL.IMAGE_ENCODER = CN() +# Image encoder type +_C.MODEL.IMAGE_ENCODER.TYPE = 'swin' +# Input image size +_C.MODEL.IMAGE_ENCODER.IMG_SIZE = 224 +# Dropout rate +_C.MODEL.IMAGE_ENCODER.DROP_RATE = 0.0 +# Drop path rate +_C.MODEL.IMAGE_ENCODER.DROP_PATH_RATE = 0.1 + +# Swin Transformer parameters +_C.MODEL.IMAGE_ENCODER.SWIN = CN() +_C.MODEL.IMAGE_ENCODER.SWIN.PATCH_SIZE = 4 +_C.MODEL.IMAGE_ENCODER.SWIN.IN_CHANS = 3 +_C.MODEL.IMAGE_ENCODER.SWIN.EMBED_DIM = 96 +_C.MODEL.IMAGE_ENCODER.SWIN.DEPTHS = [2, 2, 6, 2] +_C.MODEL.IMAGE_ENCODER.SWIN.NUM_HEADS = [3, 6, 12, 24] +_C.MODEL.IMAGE_ENCODER.SWIN.WINDOW_SIZE = 7 +_C.MODEL.IMAGE_ENCODER.SWIN.MLP_RATIO = 4. +_C.MODEL.IMAGE_ENCODER.SWIN.QKV_BIAS = True +_C.MODEL.IMAGE_ENCODER.SWIN.QK_SCALE = None +_C.MODEL.IMAGE_ENCODER.SWIN.APE = False +_C.MODEL.IMAGE_ENCODER.SWIN.PATCH_NORM = True + +# FocalNet parameters +_C.MODEL.IMAGE_ENCODER.FOCAL = CN() +_C.MODEL.IMAGE_ENCODER.FOCAL.PATCH_SIZE = 4 +_C.MODEL.IMAGE_ENCODER.FOCAL.IN_CHANS = 3 +_C.MODEL.IMAGE_ENCODER.FOCAL.EMBED_DIM = 96 +_C.MODEL.IMAGE_ENCODER.FOCAL.DEPTHS = [2, 2, 6, 2] +_C.MODEL.IMAGE_ENCODER.FOCAL.MLP_RATIO = 4. +_C.MODEL.IMAGE_ENCODER.FOCAL.PATCH_NORM = True +_C.MODEL.IMAGE_ENCODER.FOCAL.FOCAL_LEVELS = [2, 2, 2, 2] +_C.MODEL.IMAGE_ENCODER.FOCAL.FOCAL_WINDOWS = [3, 3, 3, 3] +_C.MODEL.IMAGE_ENCODER.FOCAL.FOCAL_FACTORS = [2, 2, 2, 2] +_C.MODEL.IMAGE_ENCODER.FOCAL.USE_CONV_EMBED = False +_C.MODEL.IMAGE_ENCODER.FOCAL.USE_LAYERSCALE = False +_C.MODEL.IMAGE_ENCODER.FOCAL.USE_POSTLN = False + +# ----------------------------------------------------------------------------- +# Build Text Encoder +# ----------------------------------------------------------------------------- +_C.MODEL.TEXT_ENCODER = CN() + +_C.MODEL.TEXT_ENCODER.NAME = 'transformer' +_C.MODEL.TEXT_ENCODER.LOAD_PRETRAINED = False +_C.MODEL.TEXT_ENCODER.PRETRAINED = '' +_C.MODEL.TEXT_ENCODER.TOKENIZER = 'clip' +_C.MODEL.TEXT_ENCODER.CONTEXT_LENGTH = 77 +_C.MODEL.TEXT_ENCODER.WIDTH = 1024 +_C.MODEL.TEXT_ENCODER.HEADS = 16 +_C.MODEL.TEXT_ENCODER.LAYERS = 12 +_C.MODEL.TEXT_ENCODER.AUTOGRESSIVE = True + +# ----------------------------------------------------------------------------- +# Training settings +# ----------------------------------------------------------------------------- +_C.TRAIN = CN() +_C.TRAIN.START_EPOCH = 0 +_C.TRAIN.EPOCHS = 32 +_C.TRAIN.WARMUP_EPOCHS = 5 +_C.TRAIN.WEIGHT_DECAY = 0.1 +_C.TRAIN.BASE_LR = 5e-4 +_C.TRAIN.WARMUP_LR = 5e-7 +_C.TRAIN.MIN_LR = 5e-6 +# Clip gradient norm +_C.TRAIN.CLIP_GRAD = 5.0 +# Auto resume from latest checkpoint +_C.TRAIN.AUTO_RESUME = True +# Gradient accumulation steps +# could be overwritten by command line argument +_C.TRAIN.ACCUMULATION_STEPS = 0 +# Whether to use gradient checkpointing to save memory +# could be overwritten by command line argument +_C.TRAIN.USE_CHECKPOINT = False + +# LR scheduler +_C.TRAIN.LR_SCHEDULER = CN() +_C.TRAIN.LR_SCHEDULER.NAME = 'cosine' +# Epoch interval to decay LR, used in StepLRScheduler +_C.TRAIN.LR_SCHEDULER.DECAY_EPOCHS = 30 +# LR decay rate, used in StepLRScheduler +_C.TRAIN.LR_SCHEDULER.DECAY_RATE = 0.1 + +# Optimizer +_C.TRAIN.OPTIMIZER = CN() +_C.TRAIN.OPTIMIZER.NAME = 'adamw' +# Optimizer Epsilon +_C.TRAIN.OPTIMIZER.EPS = 1e-8 +# Optimizer Betas +_C.TRAIN.OPTIMIZER.BETAS = (0.9, 0.999) +# SGD momentum +_C.TRAIN.OPTIMIZER.MOMENTUM = 0.9 + +# ----------------------------------------------------------------------------- +# Augmentation settings +# ----------------------------------------------------------------------------- +_C.AUG = CN() +# Color jitter factor +_C.AUG.COLOR_JITTER = 0.4 +# Use AutoAugment policy. "v0" or "original" +_C.AUG.AUTO_AUGMENT = 'rand-m9-mstd0.5-inc1' +# Random erase prob +_C.AUG.REPROB = 0.25 +# Random erase mode +_C.AUG.REMODE = 'pixel' +# Random erase count +_C.AUG.RECOUNT = 1 +# Mixup alpha, mixup enabled if > 0 +_C.AUG.MIXUP = 0.8 +# Cutmix alpha, cutmix enabled if > 0 +_C.AUG.CUTMIX = 1.0 +# Cutmix min/max ratio, overrides alpha and enables cutmix if set +_C.AUG.CUTMIX_MINMAX = None +# Probability of performing mixup or cutmix when either/both is enabled +_C.AUG.MIXUP_PROB = 1.0 +# Probability of switching to cutmix when both mixup and cutmix enabled +_C.AUG.MIXUP_SWITCH_PROB = 0.5 +# How to apply mixup/cutmix params. Per "batch", "pair", or "elem" +_C.AUG.MIXUP_MODE = 'batch' + +# ----------------------------------------------------------------------------- +# Testing settings +# ----------------------------------------------------------------------------- +_C.TEST = CN() +# Whether to use center crop when testing +_C.TEST.CROP = True + +# ----------------------------------------------------------------------------- +# Misc +# ----------------------------------------------------------------------------- +# Mixed precision opt level, if O0, no amp is used ('O0', 'O1', 'O2') +# overwritten by command line argument +_C.AMP_OPT_LEVEL = '' +# Path to output folder, overwritten by command line argument +_C.OUTPUT = '' +# Tag of experiment, overwritten by command line argument +_C.TAG = 'default' +# Frequency to save checkpoint +_C.SAVE_FREQ = 1 +# Frequency to logging info +_C.PRINT_FREQ = 100 +# Fixed random seed +_C.SEED = 0 +# Perform evaluation only, overwritten by command line argument +_C.EVAL_MODE = False +# Test throughput only, overwritten by command line argument +_C.THROUGHPUT_MODE = False +# Debug only so that skip dataloader initialization, overwritten by command line argument +_C.DEBUG_MODE = False +# local rank for DistributedDataParallel, given by command line argument +_C.LOCAL_RANK = 0 + + +def _update_config_from_file(config, cfg_file): + config.defrost() + with open(cfg_file, 'r') as f: + yaml_cfg = yaml.load(f, Loader=yaml.FullLoader) + + for cfg in yaml_cfg.setdefault('BASE', ['']): + if cfg: + _update_config_from_file( + config, os.path.join(os.path.dirname(cfg_file), cfg) + ) + print('=> merge config from {}'.format(cfg_file)) + config.merge_from_file(cfg_file) + config.freeze() + + +def update_config(config, args): + _update_config_from_file(config, args.cfg) + config.freeze() + + +def get_config(args): + """Get a yacs CfgNode object with default values.""" + # Return a clone so that the defaults will not be altered + # This is for the "local variable" use pattern + config = _C.clone() + update_config(config, args) + + return config diff --git a/configs/Base-RCNN-C4.yaml b/configs/Base-RCNN-C4.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fbf34a0ea57a587e09997edd94c4012d69d0b6ad --- /dev/null +++ b/configs/Base-RCNN-C4.yaml @@ -0,0 +1,18 @@ +MODEL: + META_ARCHITECTURE: "GeneralizedRCNN" + RPN: + PRE_NMS_TOPK_TEST: 6000 + POST_NMS_TOPK_TEST: 1000 + ROI_HEADS: + NAME: "Res5ROIHeads" +DATASETS: + TRAIN: ("coco_2017_train",) + TEST: ("coco_2017_val",) +SOLVER: + IMS_PER_BATCH: 16 + BASE_LR: 0.02 + STEPS: (60000, 80000) + MAX_ITER: 90000 +INPUT: + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) +VERSION: 2 diff --git a/configs/Base-RCNN-FPN.yaml b/configs/Base-RCNN-FPN.yaml new file mode 100644 index 0000000000000000000000000000000000000000..3e020f2e7b2f26765be317f907126a1556621abf --- /dev/null +++ b/configs/Base-RCNN-FPN.yaml @@ -0,0 +1,42 @@ +MODEL: + META_ARCHITECTURE: "GeneralizedRCNN" + BACKBONE: + NAME: "build_resnet_fpn_backbone" + RESNETS: + OUT_FEATURES: ["res2", "res3", "res4", "res5"] + FPN: + IN_FEATURES: ["res2", "res3", "res4", "res5"] + ANCHOR_GENERATOR: + SIZES: [[32], [64], [128], [256], [512]] # One size for each in feature map + ASPECT_RATIOS: [[0.5, 1.0, 2.0]] # Three aspect ratios (same for all in feature maps) + RPN: + IN_FEATURES: ["p2", "p3", "p4", "p5", "p6"] + PRE_NMS_TOPK_TRAIN: 2000 # Per FPN level + PRE_NMS_TOPK_TEST: 1000 # Per FPN level + # Detectron1 uses 2000 proposals per-batch, + # (See "modeling/rpn/rpn_outputs.py" for details of this legacy issue) + # which is approximately 1000 proposals per-image since the default batch size for FPN is 2. + POST_NMS_TOPK_TRAIN: 1000 + POST_NMS_TOPK_TEST: 1000 + ROI_HEADS: + NAME: "StandardROIHeads" + IN_FEATURES: ["p2", "p3", "p4", "p5"] + ROI_BOX_HEAD: + NAME: "FastRCNNConvFCHead" + NUM_FC: 2 + POOLER_RESOLUTION: 7 + ROI_MASK_HEAD: + NAME: "MaskRCNNConvUpsampleHead" + NUM_CONV: 4 + POOLER_RESOLUTION: 14 +DATASETS: + TRAIN: ("coco_2017_train",) + TEST: ("coco_2017_val",) +SOLVER: + IMS_PER_BATCH: 16 + BASE_LR: 0.02 + STEPS: (60000, 80000) + MAX_ITER: 90000 +INPUT: + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) +VERSION: 2 diff --git a/configs/CLIP_fast_rcnn_R_50_C4.yaml b/configs/CLIP_fast_rcnn_R_50_C4.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5b185ed8bbbbaae27af881da91ab4a3326896e2f --- /dev/null +++ b/configs/CLIP_fast_rcnn_R_50_C4.yaml @@ -0,0 +1,71 @@ +_BASE_: "./Base-RCNN-C4.yaml" +MODEL: + META_ARCHITECTURE: "CLIPFastRCNN" # "CLIPRCNN" # "GeneralizedRCNN" + MASK_ON: False + WEIGHTS: "./model_final.pth" + BACKBONE: + NAME: "build_clip_resnet_backbone" # "build_resnet_fpn_backbone" + FREEZE_AT: 2 + TEXT_BACKBONE: + NAME: "build_clip_language_encoder" + CLIP: + CROP_REGION_TYPE: "RPN" + OFFLINE_RPN_CONFIG: "./configs/mask_rcnn_R_50_FPN_1x.yaml" + USE_TEXT_EMB_CLASSIFIER: True + TEXT_EMB_PATH: "./lvis_1203_cls_emb_notnorm_rn50x4.pth" + NO_BOX_DELTA: True + OFFLINE_RPN_NMS_THRESH: 0.7 + CLSS_TEMP: 0.01 + MULTIPLY_RPN_SCORE: True + TEXT_EMB_DIM: 640 + RESNETS: + DEPTH: 200 + OUT_FEATURES: ["res4"] + NORM: FrozenBN + STEM_OUT_CHANNELS: 64 + RES2_OUT_CHANNELS: 256 + RPN: + HEAD_NAME: StandardRPNHead + IN_FEATURES: ["res4"] + POST_NMS_TOPK_TEST: 1000 + NMS_THRESH: + ROI_HEADS: + NAME: "CLIPRes5ROIHeads" # "Res5ROIHeads" # "StandardROIHeads" + IN_FEATURES: ["res4"] + NUM_CLASSES: 1203 + NMS_THRESH_TEST: 0.3 + SCORE_THRESH_TEST: 0.0 + ROI_BOX_HEAD: + NAME: "" + NUM_FC: 0 + CLS_AGNOSTIC_BBOX_REG: True + POOLER_RESOLUTION: 18 + ROI_MASK_HEAD: + NAME: "MaskRCNNConvUpsampleHead" + NUM_CONV: 0 + POOLER_RESOLUTION: 14 + PIXEL_MEAN: [0.48145466, 0.4578275, 0.40821073] + PIXEL_STD: [0.26862954, 0.26130258, 0.27577711] +INPUT: + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) +DATASETS: + TRAIN: ("lvis_v1_train",) + TEST: ("lvis_v1_val",) +TEST: + DETECTIONS_PER_IMAGE: 300 # LVIS allows up to 300 + EVAL_PERIOD: 25000 +SOLVER: + IMS_PER_BATCH: 16 + BASE_LR: 0.02 + STEPS: (120000, 160000) + MAX_ITER: 180000 # 180000 * 16 / 100000 ~ 28.8 epochs +DATALOADER: + SAMPLER_TRAIN: "RepeatFactorTrainingSampler" + REPEAT_THRESHOLD: 0.001 +INPUT: + MIN_SIZE_TRAIN_SAMPLING: choice + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) + MAX_SIZE_TRAIN: 1333 + MIN_SIZE_TEST: 800 + MAX_SIZE_TEST: 1333 + FORMAT: "RGB" \ No newline at end of file diff --git a/configs/CLIP_fast_rcnn_swin_base_C4.yaml b/configs/CLIP_fast_rcnn_swin_base_C4.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d86f89cfcfc9d526f6039b7d4bb60f260619ed1e --- /dev/null +++ b/configs/CLIP_fast_rcnn_swin_base_C4.yaml @@ -0,0 +1,74 @@ +_BASE_: "./Base-RCNN-C4.yaml" +MODEL: + META_ARCHITECTURE: "CLIPFastRCNN" # "CLIPRCNN" # "GeneralizedRCNN" + BACKBONE: + NAME: "build_clip_swin" # "build_resnet_fpn_backbone" + FREEZE_AT: 2 + TEXT_BACKBONE: + NAME: "build_clip_swin_text_backbone" + SPEC: + EMBED_DIM: 512 + VISION: + PATCH_SIZE: 4 + IN_CHANS: 3 + EMBED_DIM: 128 + DEPTHS: [ 2, 2, 18, 2 ] + NUM_HEADS: [ 4, 8, 16, 32 ] + WINDOW_SIZE: 7 + MLP_RATIO: 4. + QKV_BIAS: True + APE: False + PATCH_NORM: True + DROP_RATE: 0.0 + DROP_PATH_RATE: 0.2 + OUT_FEATURES: ["stage2", "stage3", "stage4", "stage5"] + TEXT: + NAME: 'transformer' + TOKENIZER: clip + CONTEXT_LENGTH: 77 + WIDTH: 512 + HEADS: 8 + LAYERS: 12 + WEIGHTS: "" # "detectron2://ImageNetPretrained/MSRA/R-50.pkl" + MASK_ON: True + RPN: + HEAD_NAME: StandardRPNHead + IN_FEATURES: ["stage4"] + ROI_HEADS: + NAME: "CLIPSwinROIHeads" # "Res5ROIHeads" # "StandardROIHeads" + IN_FEATURES: ["stage4"] + NUM_CLASSES: 1203 + SCORE_THRESH_TEST: 0.0001 + ROI_BOX_HEAD: + NAME: "" + NUM_FC: 0 + POOLER_RESOLUTION: 14 + ROI_MASK_HEAD: + NAME: "MaskRCNNConvUpsampleHead" + NUM_CONV: 0 + POOLER_RESOLUTION: 14 + PIXEL_MEAN: [0.485, 0.456, 0.406] + PIXEL_STD: [0.229, 0.224, 0.225] +INPUT: + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) +DATASETS: + TRAIN: ("lvis_v1_train",) + TEST: ("lvis_v1_val",) +TEST: + DETECTIONS_PER_IMAGE: 300 # LVIS allows up to 300 + EVAL_PERIOD: 25000 +SOLVER: + IMS_PER_BATCH: 16 + BASE_LR: 0.02 + STEPS: (120000, 160000) + MAX_ITER: 180000 # 180000 * 16 / 100000 ~ 28.8 epochs +DATALOADER: + SAMPLER_TRAIN: "RepeatFactorTrainingSampler" + REPEAT_THRESHOLD: 0.001 +INPUT: + MIN_SIZE_TRAIN_SAMPLING: choice + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) + MAX_SIZE_TRAIN: 1333 + MIN_SIZE_TEST: 800 + MAX_SIZE_TEST: 1333 + FORMAT: "RGB" \ No newline at end of file diff --git a/configs/mask_rcnn_CLIP_R_50_C4_1x.yaml b/configs/mask_rcnn_CLIP_R_50_C4_1x.yaml new file mode 100644 index 0000000000000000000000000000000000000000..acb412e2e1485ab1a92ea69fff16240cf7e077c3 --- /dev/null +++ b/configs/mask_rcnn_CLIP_R_50_C4_1x.yaml @@ -0,0 +1,55 @@ +_BASE_: "./Base-RCNN-C4.yaml" +MODEL: + META_ARCHITECTURE: "GeneralizedRCNN" + BACKBONE: + NAME: "build_clip_resnet_backbone" #"build_clip_resnet_fpn_backbone" # "build_resnet_fpn_backbone" + FREEZE_AT: 2 + WEIGHTS: "" # "detectron2://ImageNetPretrained/MSRA/R-50.pkl" + MASK_ON: True + RESNETS: + DEPTH: 50 + OUT_FEATURES: ["res4"] + NORM: FrozenBN + STEM_OUT_CHANNELS: 64 + RES2_OUT_CHANNELS: 256 + RPN: + HEAD_NAME: StandardRPNHead + IN_FEATURES: ["res4"] + ROI_HEADS: + NAME: "CLIPRes5ROIHeads" # "Res5ROIHeads" # "StandardROIHeads" + IN_FEATURES: ["res4"] + NUM_CLASSES: 1203 + SCORE_THRESH_TEST: 0.0001 + ROI_BOX_HEAD: + NAME: "" + NUM_FC: 0 + POOLER_RESOLUTION: 14 + ROI_MASK_HEAD: + NAME: "MaskRCNNConvUpsampleHead" + NUM_CONV: 0 + POOLER_RESOLUTION: 14 + PIXEL_MEAN: [0.48145466, 0.4578275, 0.40821073] # [103.530, 116.280, 123.675] # + PIXEL_STD: [0.26862954, 0.26130258, 0.27577711] # [1.0, 1.0, 1.0] # +INPUT: + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) +DATASETS: + TRAIN: ("lvis_v1_train",) + TEST: ("lvis_v1_val",) +TEST: + DETECTIONS_PER_IMAGE: 300 # LVIS allows up to 300 + EVAL_PERIOD: 25000 +SOLVER: + IMS_PER_BATCH: 16 + BASE_LR: 0.02 + STEPS: (120000, 160000) # (140000,) # + MAX_ITER: 180000 # 180000 * 16 / 100000 ~ 28.8 epochs +DATALOADER: + SAMPLER_TRAIN: "RepeatFactorTrainingSampler" + REPEAT_THRESHOLD: 0.001 +INPUT: + MIN_SIZE_TRAIN_SAMPLING: choice + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) + MAX_SIZE_TRAIN: 1333 + MIN_SIZE_TEST: 800 + MAX_SIZE_TEST: 1333 + FORMAT: "RGB" # "BGR" \ No newline at end of file diff --git a/configs/mask_rcnn_R_50_C4_1x.yaml b/configs/mask_rcnn_R_50_C4_1x.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b0ae6166dee627f78b8f077ef31a1709731d543b --- /dev/null +++ b/configs/mask_rcnn_R_50_C4_1x.yaml @@ -0,0 +1,23 @@ +_BASE_: "./Base-RCNN-C4.yaml" +MODEL: + WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl" + MASK_ON: True + RESNETS: + DEPTH: 50 + ROI_HEADS: + NUM_CLASSES: 1203 + SCORE_THRESH_TEST: 0.0001 +INPUT: + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) +DATASETS: + TRAIN: ("lvis_v1_train",) + TEST: ("lvis_v1_val",) +TEST: + DETECTIONS_PER_IMAGE: 300 # LVIS allows up to 300 + EVAL_PERIOD: 50000 +SOLVER: + STEPS: (120000, 160000) + MAX_ITER: 180000 # 180000 * 16 / 100000 ~ 28.8 epochs +DATALOADER: + SAMPLER_TRAIN: "RepeatFactorTrainingSampler" + REPEAT_THRESHOLD: 0.001 \ No newline at end of file diff --git a/configs/mask_rcnn_R_50_FPN_1x.yaml b/configs/mask_rcnn_R_50_FPN_1x.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8f788784f72d07b156901c3144b8a795be9e66c3 --- /dev/null +++ b/configs/mask_rcnn_R_50_FPN_1x.yaml @@ -0,0 +1,23 @@ +_BASE_: "./Base-RCNN-FPN.yaml" +MODEL: + WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl" + MASK_ON: True + RESNETS: + DEPTH: 50 + ROI_HEADS: + NUM_CLASSES: 1203 + SCORE_THRESH_TEST: 0.0001 +INPUT: + MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) +DATASETS: + TRAIN: ("lvis_v1_train",) + TEST: ("lvis_v1_val",) +TEST: + DETECTIONS_PER_IMAGE: 300 # LVIS allows up to 300 + EVAL_PERIOD: 50000 +SOLVER: + STEPS: (120000, 160000) + MAX_ITER: 180000 # 180000 * 16 / 100000 ~ 28.8 epochs +DATALOADER: + SAMPLER_TRAIN: "RepeatFactorTrainingSampler" + REPEAT_THRESHOLD: 0.001 diff --git a/datasets/README.md b/datasets/README.md new file mode 100644 index 0000000000000000000000000000000000000000..0eb44cc3b23beeb1755ab8d12002d26f13434235 --- /dev/null +++ b/datasets/README.md @@ -0,0 +1,140 @@ +# Use Builtin Datasets + +A dataset can be used by accessing [DatasetCatalog](https://detectron2.readthedocs.io/modules/data.html#detectron2.data.DatasetCatalog) +for its data, or [MetadataCatalog](https://detectron2.readthedocs.io/modules/data.html#detectron2.data.MetadataCatalog) for its metadata (class names, etc). +This document explains how to setup the builtin datasets so they can be used by the above APIs. +[Use Custom Datasets](https://detectron2.readthedocs.io/tutorials/datasets.html) gives a deeper dive on how to use `DatasetCatalog` and `MetadataCatalog`, +and how to add new datasets to them. + +Detectron2 has builtin support for a few datasets. +The datasets are assumed to exist in a directory specified by the environment variable +`DETECTRON2_DATASETS`. +Under this directory, detectron2 will look for datasets in the structure described below, if needed. +``` +$DETECTRON2_DATASETS/ + coco/ + lvis/ + cityscapes/ + VOC20{07,12}/ +``` + +You can set the location for builtin datasets by `export DETECTRON2_DATASETS=/path/to/datasets`. +If left unset, the default is `./datasets` relative to your current working directory. + +The [model zoo](https://github.com/facebookresearch/detectron2/blob/master/MODEL_ZOO.md) +contains configs and models that use these builtin datasets. + +## Expected dataset structure for [COCO instance/keypoint detection](https://cocodataset.org/#download): + +``` +coco/ + annotations/ + instances_{train,val}2017.json + person_keypoints_{train,val}2017.json + {train,val}2017/ + # image files that are mentioned in the corresponding json +``` + +You can use the 2014 version of the dataset as well. + +Some of the builtin tests (`dev/run_*_tests.sh`) uses a tiny version of the COCO dataset, +which you can download with `./datasets/prepare_for_tests.sh`. + +## Expected dataset structure for PanopticFPN: + +Extract panoptic annotations from [COCO website](https://cocodataset.org/#download) +into the following structure: +``` +coco/ + annotations/ + panoptic_{train,val}2017.json + panoptic_{train,val}2017/ # png annotations + panoptic_stuff_{train,val}2017/ # generated by the script mentioned below +``` + +Install panopticapi by: +``` +pip install git+https://github.com/cocodataset/panopticapi.git +``` +Then, run `python datasets/prepare_panoptic_fpn.py`, to extract semantic annotations from panoptic annotations. + +## Expected dataset structure for [LVIS instance segmentation](https://www.lvisdataset.org/dataset): +``` +coco/ + {train,val,test}2017/ +lvis/ + lvis_v0.5_{train,val}.json + lvis_v0.5_image_info_test.json + lvis_v1_{train,val}.json + lvis_v1_image_info_test{,_challenge}.json +``` + +Install lvis-api by: +``` +pip install git+https://github.com/lvis-dataset/lvis-api.git +``` + +To evaluate models trained on the COCO dataset using LVIS annotations, +run `python datasets/prepare_cocofied_lvis.py` to prepare "cocofied" LVIS annotations. + +## Expected dataset structure for [cityscapes](https://www.cityscapes-dataset.com/downloads/): +``` +cityscapes/ + gtFine/ + train/ + aachen/ + color.png, instanceIds.png, labelIds.png, polygons.json, + labelTrainIds.png + ... + val/ + test/ + # below are generated Cityscapes panoptic annotation + cityscapes_panoptic_train.json + cityscapes_panoptic_train/ + cityscapes_panoptic_val.json + cityscapes_panoptic_val/ + cityscapes_panoptic_test.json + cityscapes_panoptic_test/ + leftImg8bit/ + train/ + val/ + test/ +``` +Install cityscapes scripts by: +``` +pip install git+https://github.com/mcordts/cityscapesScripts.git +``` + +Note: to create labelTrainIds.png, first prepare the above structure, then run cityscapesescript with: +``` +CITYSCAPES_DATASET=/path/to/abovementioned/cityscapes python cityscapesscripts/preparation/createTrainIdLabelImgs.py +``` +These files are not needed for instance segmentation. + +Note: to generate Cityscapes panoptic dataset, run cityscapesescript with: +``` +CITYSCAPES_DATASET=/path/to/abovementioned/cityscapes python cityscapesscripts/preparation/createPanopticImgs.py +``` +These files are not needed for semantic and instance segmentation. + +## Expected dataset structure for [Pascal VOC](http://host.robots.ox.ac.uk/pascal/VOC/index.html): +``` +VOC20{07,12}/ + Annotations/ + ImageSets/ + Main/ + trainval.txt + test.txt + # train.txt or val.txt, if you use these splits + JPEGImages/ +``` + +## Expected dataset structure for [ADE20k Scene Parsing](http://sceneparsing.csail.mit.edu/): +``` +ADEChallengeData2016/ + annotations/ + annotations_detectron2/ + images/ + objectInfo150.txt +``` +The directory `annotations_detectron2` is generated by running `python datasets/prepare_ade20k_sem_seg.py`. diff --git a/datasets/custom_images/dog_and_cat.jfif b/datasets/custom_images/dog_and_cat.jfif new file mode 100644 index 0000000000000000000000000000000000000000..73902cca766d976abe368eafaca2a71126831eff Binary files /dev/null and b/datasets/custom_images/dog_and_cat.jfif differ diff --git a/datasets/prepare_ade20k_sem_seg.py b/datasets/prepare_ade20k_sem_seg.py new file mode 100644 index 0000000000000000000000000000000000000000..8b4a58d8f2877544498e328b6d269f23aa1eb59f --- /dev/null +++ b/datasets/prepare_ade20k_sem_seg.py @@ -0,0 +1,26 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. +import numpy as np +import os +from pathlib import Path +import tqdm +from PIL import Image + + +def convert(input, output): + img = np.asarray(Image.open(input)) + assert img.dtype == np.uint8 + img = img - 1 # 0 (ignore) becomes 255. others are shifted by 1 + Image.fromarray(img).save(output) + + +if __name__ == "__main__": + dataset_dir = Path(os.getenv("DETECTRON2_DATASETS", "datasets")) / "ADEChallengeData2016" + for name in ["training", "validation"]: + annotation_dir = dataset_dir / "annotations" / name + output_dir = dataset_dir / "annotations_detectron2" / name + output_dir.mkdir(parents=True, exist_ok=True) + for file in tqdm.tqdm(list(annotation_dir.iterdir())): + output_file = output_dir / file.name + convert(file, output_file) diff --git a/datasets/prepare_cocofied_lvis.py b/datasets/prepare_cocofied_lvis.py new file mode 100644 index 0000000000000000000000000000000000000000..245c88482a9e2405e5a912b5c560aed78a614a13 --- /dev/null +++ b/datasets/prepare_cocofied_lvis.py @@ -0,0 +1,176 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import copy +import json +import os +from collections import defaultdict + +# This mapping is extracted from the official LVIS mapping: +# https://github.com/lvis-dataset/lvis-api/blob/master/data/coco_to_synset.json +COCO_SYNSET_CATEGORIES = [ + {"synset": "person.n.01", "coco_cat_id": 1}, + {"synset": "bicycle.n.01", "coco_cat_id": 2}, + {"synset": "car.n.01", "coco_cat_id": 3}, + {"synset": "motorcycle.n.01", "coco_cat_id": 4}, + {"synset": "airplane.n.01", "coco_cat_id": 5}, + {"synset": "bus.n.01", "coco_cat_id": 6}, + {"synset": "train.n.01", "coco_cat_id": 7}, + {"synset": "truck.n.01", "coco_cat_id": 8}, + {"synset": "boat.n.01", "coco_cat_id": 9}, + {"synset": "traffic_light.n.01", "coco_cat_id": 10}, + {"synset": "fireplug.n.01", "coco_cat_id": 11}, + {"synset": "stop_sign.n.01", "coco_cat_id": 13}, + {"synset": "parking_meter.n.01", "coco_cat_id": 14}, + {"synset": "bench.n.01", "coco_cat_id": 15}, + {"synset": "bird.n.01", "coco_cat_id": 16}, + {"synset": "cat.n.01", "coco_cat_id": 17}, + {"synset": "dog.n.01", "coco_cat_id": 18}, + {"synset": "horse.n.01", "coco_cat_id": 19}, + {"synset": "sheep.n.01", "coco_cat_id": 20}, + {"synset": "beef.n.01", "coco_cat_id": 21}, + {"synset": "elephant.n.01", "coco_cat_id": 22}, + {"synset": "bear.n.01", "coco_cat_id": 23}, + {"synset": "zebra.n.01", "coco_cat_id": 24}, + {"synset": "giraffe.n.01", "coco_cat_id": 25}, + {"synset": "backpack.n.01", "coco_cat_id": 27}, + {"synset": "umbrella.n.01", "coco_cat_id": 28}, + {"synset": "bag.n.04", "coco_cat_id": 31}, + {"synset": "necktie.n.01", "coco_cat_id": 32}, + {"synset": "bag.n.06", "coco_cat_id": 33}, + {"synset": "frisbee.n.01", "coco_cat_id": 34}, + {"synset": "ski.n.01", "coco_cat_id": 35}, + {"synset": "snowboard.n.01", "coco_cat_id": 36}, + {"synset": "ball.n.06", "coco_cat_id": 37}, + {"synset": "kite.n.03", "coco_cat_id": 38}, + {"synset": "baseball_bat.n.01", "coco_cat_id": 39}, + {"synset": "baseball_glove.n.01", "coco_cat_id": 40}, + {"synset": "skateboard.n.01", "coco_cat_id": 41}, + {"synset": "surfboard.n.01", "coco_cat_id": 42}, + {"synset": "tennis_racket.n.01", "coco_cat_id": 43}, + {"synset": "bottle.n.01", "coco_cat_id": 44}, + {"synset": "wineglass.n.01", "coco_cat_id": 46}, + {"synset": "cup.n.01", "coco_cat_id": 47}, + {"synset": "fork.n.01", "coco_cat_id": 48}, + {"synset": "knife.n.01", "coco_cat_id": 49}, + {"synset": "spoon.n.01", "coco_cat_id": 50}, + {"synset": "bowl.n.03", "coco_cat_id": 51}, + {"synset": "banana.n.02", "coco_cat_id": 52}, + {"synset": "apple.n.01", "coco_cat_id": 53}, + {"synset": "sandwich.n.01", "coco_cat_id": 54}, + {"synset": "orange.n.01", "coco_cat_id": 55}, + {"synset": "broccoli.n.01", "coco_cat_id": 56}, + {"synset": "carrot.n.01", "coco_cat_id": 57}, + {"synset": "frank.n.02", "coco_cat_id": 58}, + {"synset": "pizza.n.01", "coco_cat_id": 59}, + {"synset": "doughnut.n.02", "coco_cat_id": 60}, + {"synset": "cake.n.03", "coco_cat_id": 61}, + {"synset": "chair.n.01", "coco_cat_id": 62}, + {"synset": "sofa.n.01", "coco_cat_id": 63}, + {"synset": "pot.n.04", "coco_cat_id": 64}, + {"synset": "bed.n.01", "coco_cat_id": 65}, + {"synset": "dining_table.n.01", "coco_cat_id": 67}, + {"synset": "toilet.n.02", "coco_cat_id": 70}, + {"synset": "television_receiver.n.01", "coco_cat_id": 72}, + {"synset": "laptop.n.01", "coco_cat_id": 73}, + {"synset": "mouse.n.04", "coco_cat_id": 74}, + {"synset": "remote_control.n.01", "coco_cat_id": 75}, + {"synset": "computer_keyboard.n.01", "coco_cat_id": 76}, + {"synset": "cellular_telephone.n.01", "coco_cat_id": 77}, + {"synset": "microwave.n.02", "coco_cat_id": 78}, + {"synset": "oven.n.01", "coco_cat_id": 79}, + {"synset": "toaster.n.02", "coco_cat_id": 80}, + {"synset": "sink.n.01", "coco_cat_id": 81}, + {"synset": "electric_refrigerator.n.01", "coco_cat_id": 82}, + {"synset": "book.n.01", "coco_cat_id": 84}, + {"synset": "clock.n.01", "coco_cat_id": 85}, + {"synset": "vase.n.01", "coco_cat_id": 86}, + {"synset": "scissors.n.01", "coco_cat_id": 87}, + {"synset": "teddy.n.01", "coco_cat_id": 88}, + {"synset": "hand_blower.n.01", "coco_cat_id": 89}, + {"synset": "toothbrush.n.01", "coco_cat_id": 90}, +] + + +def cocofy_lvis(input_filename, output_filename): + """ + Filter LVIS instance segmentation annotations to remove all categories that are not included in + COCO. The new json files can be used to evaluate COCO AP using `lvis-api`. The category ids in + the output json are the incontiguous COCO dataset ids. + + Args: + input_filename (str): path to the LVIS json file. + output_filename (str): path to the COCOfied json file. + """ + + with open(input_filename, "r") as f: + lvis_json = json.load(f) + + lvis_annos = lvis_json.pop("annotations") + cocofied_lvis = copy.deepcopy(lvis_json) + lvis_json["annotations"] = lvis_annos + + # Mapping from lvis cat id to coco cat id via synset + lvis_cat_id_to_synset = {cat["id"]: cat["synset"] for cat in lvis_json["categories"]} + synset_to_coco_cat_id = {x["synset"]: x["coco_cat_id"] for x in COCO_SYNSET_CATEGORIES} + # Synsets that we will keep in the dataset + synsets_to_keep = set(synset_to_coco_cat_id.keys()) + coco_cat_id_with_instances = defaultdict(int) + + new_annos = [] + ann_id = 1 + for ann in lvis_annos: + lvis_cat_id = ann["category_id"] + synset = lvis_cat_id_to_synset[lvis_cat_id] + if synset not in synsets_to_keep: + continue + coco_cat_id = synset_to_coco_cat_id[synset] + new_ann = copy.deepcopy(ann) + new_ann["category_id"] = coco_cat_id + new_ann["id"] = ann_id + ann_id += 1 + new_annos.append(new_ann) + coco_cat_id_with_instances[coco_cat_id] += 1 + cocofied_lvis["annotations"] = new_annos + + for image in cocofied_lvis["images"]: + for key in ["not_exhaustive_category_ids", "neg_category_ids"]: + new_category_list = [] + for lvis_cat_id in image[key]: + synset = lvis_cat_id_to_synset[lvis_cat_id] + if synset not in synsets_to_keep: + continue + coco_cat_id = synset_to_coco_cat_id[synset] + new_category_list.append(coco_cat_id) + coco_cat_id_with_instances[coco_cat_id] += 1 + image[key] = new_category_list + + coco_cat_id_with_instances = set(coco_cat_id_with_instances.keys()) + + new_categories = [] + for cat in lvis_json["categories"]: + synset = cat["synset"] + if synset not in synsets_to_keep: + continue + coco_cat_id = synset_to_coco_cat_id[synset] + if coco_cat_id not in coco_cat_id_with_instances: + continue + new_cat = copy.deepcopy(cat) + new_cat["id"] = coco_cat_id + new_categories.append(new_cat) + cocofied_lvis["categories"] = new_categories + + with open(output_filename, "w") as f: + json.dump(cocofied_lvis, f) + print("{} is COCOfied and stored in {}.".format(input_filename, output_filename)) + + +if __name__ == "__main__": + dataset_dir = os.path.join(os.getenv("DETECTRON2_DATASETS", "datasets"), "lvis") + for s in ["lvis_v0.5_train", "lvis_v0.5_val"]: + print("Start COCOfing {}.".format(s)) + cocofy_lvis( + os.path.join(dataset_dir, "{}.json".format(s)), + os.path.join(dataset_dir, "{}_cocofied.json".format(s)), + ) diff --git a/datasets/prepare_for_tests.sh b/datasets/prepare_for_tests.sh new file mode 100644 index 0000000000000000000000000000000000000000..b22ee8d8e82fa1dd3834b4d98ea5618d5c2ef391 --- /dev/null +++ b/datasets/prepare_for_tests.sh @@ -0,0 +1,22 @@ +#!/bin/bash -e +# Copyright (c) Facebook, Inc. and its affiliates. + +# Download some files needed for running tests. + +cd "${0%/*}" + +BASE=https://dl.fbaipublicfiles.com/detectron2 +mkdir -p coco/annotations + +for anno in instances_val2017_100 \ + person_keypoints_val2017_100 \ + instances_minival2014_100 \ + person_keypoints_minival2014_100; do + + dest=coco/annotations/$anno.json + [[ -s $dest ]] && { + echo "$dest exists. Skipping ..." + } || { + wget $BASE/annotations/coco/$anno.json -O $dest + } +done diff --git a/datasets/prepare_panoptic_fpn.py b/datasets/prepare_panoptic_fpn.py new file mode 100644 index 0000000000000000000000000000000000000000..597d791afab1bcc0013203a66c7fba225065eebe --- /dev/null +++ b/datasets/prepare_panoptic_fpn.py @@ -0,0 +1,116 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import functools +import json +import multiprocessing as mp +import numpy as np +import os +import time +from fvcore.common.download import download +from panopticapi.utils import rgb2id +from PIL import Image + +from detectron2.data.datasets.builtin_meta import COCO_CATEGORIES + + +def _process_panoptic_to_semantic(input_panoptic, output_semantic, segments, id_map): + panoptic = np.asarray(Image.open(input_panoptic), dtype=np.uint32) + panoptic = rgb2id(panoptic) + output = np.zeros_like(panoptic, dtype=np.uint8) + 255 + for seg in segments: + cat_id = seg["category_id"] + new_cat_id = id_map[cat_id] + output[panoptic == seg["id"]] = new_cat_id + Image.fromarray(output).save(output_semantic) + + +def separate_coco_semantic_from_panoptic(panoptic_json, panoptic_root, sem_seg_root, categories): + """ + Create semantic segmentation annotations from panoptic segmentation + annotations, to be used by PanopticFPN. + + It maps all thing categories to class 0, and maps all unlabeled pixels to class 255. + It maps all stuff categories to contiguous ids starting from 1. + + Args: + panoptic_json (str): path to the panoptic json file, in COCO's format. + panoptic_root (str): a directory with panoptic annotation files, in COCO's format. + sem_seg_root (str): a directory to output semantic annotation files + categories (list[dict]): category metadata. Each dict needs to have: + "id": corresponds to the "category_id" in the json annotations + "isthing": 0 or 1 + """ + os.makedirs(sem_seg_root, exist_ok=True) + + stuff_ids = [k["id"] for k in categories if k["isthing"] == 0] + thing_ids = [k["id"] for k in categories if k["isthing"] == 1] + id_map = {} # map from category id to id in the output semantic annotation + assert len(stuff_ids) <= 254 + for i, stuff_id in enumerate(stuff_ids): + id_map[stuff_id] = i + 1 + for thing_id in thing_ids: + id_map[thing_id] = 0 + id_map[0] = 255 + + with open(panoptic_json) as f: + obj = json.load(f) + + pool = mp.Pool(processes=max(mp.cpu_count() // 2, 4)) + + def iter_annotations(): + for anno in obj["annotations"]: + file_name = anno["file_name"] + segments = anno["segments_info"] + input = os.path.join(panoptic_root, file_name) + output = os.path.join(sem_seg_root, file_name) + yield input, output, segments + + print("Start writing to {} ...".format(sem_seg_root)) + start = time.time() + pool.starmap( + functools.partial(_process_panoptic_to_semantic, id_map=id_map), + iter_annotations(), + chunksize=100, + ) + print("Finished. time: {:.2f}s".format(time.time() - start)) + + +if __name__ == "__main__": + dataset_dir = os.path.join(os.getenv("DETECTRON2_DATASETS", "datasets"), "coco") + for s in ["val2017", "train2017"]: + separate_coco_semantic_from_panoptic( + os.path.join(dataset_dir, "annotations/panoptic_{}.json".format(s)), + os.path.join(dataset_dir, "panoptic_{}".format(s)), + os.path.join(dataset_dir, "panoptic_stuff_{}".format(s)), + COCO_CATEGORIES, + ) + + # Prepare val2017_100 for quick testing: + + dest_dir = os.path.join(dataset_dir, "annotations/") + URL_PREFIX = "https://dl.fbaipublicfiles.com/detectron2/" + download(URL_PREFIX + "annotations/coco/panoptic_val2017_100.json", dest_dir) + with open(os.path.join(dest_dir, "panoptic_val2017_100.json")) as f: + obj = json.load(f) + + def link_val100(dir_full, dir_100): + print("Creating " + dir_100 + " ...") + os.makedirs(dir_100, exist_ok=True) + for img in obj["images"]: + basename = os.path.splitext(img["file_name"])[0] + src = os.path.join(dir_full, basename + ".png") + dst = os.path.join(dir_100, basename + ".png") + src = os.path.relpath(src, start=dir_100) + os.symlink(src, dst) + + link_val100( + os.path.join(dataset_dir, "panoptic_val2017"), + os.path.join(dataset_dir, "panoptic_val2017_100"), + ) + + link_val100( + os.path.join(dataset_dir, "panoptic_stuff_val2017"), + os.path.join(dataset_dir, "panoptic_stuff_val2017_100"), + ) diff --git a/detectron2/__init__.py b/detectron2/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a951838f58f8bcf4b2b51a94b2ba31c53e8fe1af --- /dev/null +++ b/detectron2/__init__.py @@ -0,0 +1,10 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +from .utils.env import setup_environment + +setup_environment() + + +# This line will be programatically read/write by setup.py. +# Leave them at the bottom of this file and don't touch them. +__version__ = "0.4" diff --git a/detectron2/__pycache__/__init__.cpython-39.pyc b/detectron2/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5b823f983e2e16fbe0fd6460b4ec7ebe770167e7 Binary files /dev/null and b/detectron2/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/checkpoint/__init__.py b/detectron2/checkpoint/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..99da0469ae7e169d8970e4b642fed3f870076860 --- /dev/null +++ b/detectron2/checkpoint/__init__.py @@ -0,0 +1,10 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. +# File: + + +from . import catalog as _UNUSED # register the handler +from .detection_checkpoint import DetectionCheckpointer +from fvcore.common.checkpoint import Checkpointer, PeriodicCheckpointer + +__all__ = ["Checkpointer", "PeriodicCheckpointer", "DetectionCheckpointer"] diff --git a/detectron2/checkpoint/__pycache__/__init__.cpython-39.pyc b/detectron2/checkpoint/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..915d59c29c850a78d88085eb1ffa6569a8b73e6f Binary files /dev/null and b/detectron2/checkpoint/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/checkpoint/__pycache__/c2_model_loading.cpython-39.pyc b/detectron2/checkpoint/__pycache__/c2_model_loading.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..25b14652b7a6c5d46f59b28e2c6a1dc6bc550fb1 Binary files /dev/null and b/detectron2/checkpoint/__pycache__/c2_model_loading.cpython-39.pyc differ diff --git a/detectron2/checkpoint/__pycache__/catalog.cpython-39.pyc b/detectron2/checkpoint/__pycache__/catalog.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..329cec1fc7c02a651ec5a9702f186ec4e312869d Binary files /dev/null and b/detectron2/checkpoint/__pycache__/catalog.cpython-39.pyc differ diff --git a/detectron2/checkpoint/__pycache__/clip_model_loading.cpython-39.pyc b/detectron2/checkpoint/__pycache__/clip_model_loading.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6ccc48159bd8fcb5c95550346beee03cfb745893 Binary files /dev/null and b/detectron2/checkpoint/__pycache__/clip_model_loading.cpython-39.pyc differ diff --git a/detectron2/checkpoint/__pycache__/detection_checkpoint.cpython-39.pyc b/detectron2/checkpoint/__pycache__/detection_checkpoint.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ab0fb94bbe76fd5323b2b7d458826a01d8498a23 Binary files /dev/null and b/detectron2/checkpoint/__pycache__/detection_checkpoint.cpython-39.pyc differ diff --git a/detectron2/checkpoint/c2_model_loading.py b/detectron2/checkpoint/c2_model_loading.py new file mode 100644 index 0000000000000000000000000000000000000000..8c8d181bd7200bd3fd38446e743f8f16780d6e76 --- /dev/null +++ b/detectron2/checkpoint/c2_model_loading.py @@ -0,0 +1,407 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import logging +import re +from typing import Dict, List +import torch +from tabulate import tabulate + + +def convert_basic_c2_names(original_keys): + """ + Apply some basic name conversion to names in C2 weights. + It only deals with typical backbone models. + + Args: + original_keys (list[str]): + Returns: + list[str]: The same number of strings matching those in original_keys. + """ + layer_keys = copy.deepcopy(original_keys) + layer_keys = [ + {"pred_b": "linear_b", "pred_w": "linear_w"}.get(k, k) for k in layer_keys + ] # some hard-coded mappings + + layer_keys = [k.replace("_", ".") for k in layer_keys] + layer_keys = [re.sub("\\.b$", ".bias", k) for k in layer_keys] + layer_keys = [re.sub("\\.w$", ".weight", k) for k in layer_keys] + # Uniform both bn and gn names to "norm" + layer_keys = [re.sub("bn\\.s$", "norm.weight", k) for k in layer_keys] + layer_keys = [re.sub("bn\\.bias$", "norm.bias", k) for k in layer_keys] + layer_keys = [re.sub("bn\\.rm", "norm.running_mean", k) for k in layer_keys] + layer_keys = [re.sub("bn\\.running.mean$", "norm.running_mean", k) for k in layer_keys] + layer_keys = [re.sub("bn\\.riv$", "norm.running_var", k) for k in layer_keys] + layer_keys = [re.sub("bn\\.running.var$", "norm.running_var", k) for k in layer_keys] + layer_keys = [re.sub("bn\\.gamma$", "norm.weight", k) for k in layer_keys] + layer_keys = [re.sub("bn\\.beta$", "norm.bias", k) for k in layer_keys] + layer_keys = [re.sub("gn\\.s$", "norm.weight", k) for k in layer_keys] + layer_keys = [re.sub("gn\\.bias$", "norm.bias", k) for k in layer_keys] + + # stem + layer_keys = [re.sub("^res\\.conv1\\.norm\\.", "conv1.norm.", k) for k in layer_keys] + # to avoid mis-matching with "conv1" in other components (e.g. detection head) + layer_keys = [re.sub("^conv1\\.", "stem.conv1.", k) for k in layer_keys] + + # layer1-4 is used by torchvision, however we follow the C2 naming strategy (res2-5) + # layer_keys = [re.sub("^res2.", "layer1.", k) for k in layer_keys] + # layer_keys = [re.sub("^res3.", "layer2.", k) for k in layer_keys] + # layer_keys = [re.sub("^res4.", "layer3.", k) for k in layer_keys] + # layer_keys = [re.sub("^res5.", "layer4.", k) for k in layer_keys] + + # blocks + layer_keys = [k.replace(".branch1.", ".shortcut.") for k in layer_keys] + layer_keys = [k.replace(".branch2a.", ".conv1.") for k in layer_keys] + layer_keys = [k.replace(".branch2b.", ".conv2.") for k in layer_keys] + layer_keys = [k.replace(".branch2c.", ".conv3.") for k in layer_keys] + + # DensePose substitutions + layer_keys = [re.sub("^body.conv.fcn", "body_conv_fcn", k) for k in layer_keys] + layer_keys = [k.replace("AnnIndex.lowres", "ann_index_lowres") for k in layer_keys] + layer_keys = [k.replace("Index.UV.lowres", "index_uv_lowres") for k in layer_keys] + layer_keys = [k.replace("U.lowres", "u_lowres") for k in layer_keys] + layer_keys = [k.replace("V.lowres", "v_lowres") for k in layer_keys] + return layer_keys + + +def convert_c2_detectron_names(weights): + """ + Map Caffe2 Detectron weight names to Detectron2 names. + + Args: + weights (dict): name -> tensor + + Returns: + dict: detectron2 names -> tensor + dict: detectron2 names -> C2 names + """ + logger = logging.getLogger(__name__) + logger.info("Renaming Caffe2 weights ......") + original_keys = sorted(weights.keys()) + layer_keys = copy.deepcopy(original_keys) + + layer_keys = convert_basic_c2_names(layer_keys) + + # -------------------------------------------------------------------------- + # RPN hidden representation conv + # -------------------------------------------------------------------------- + # FPN case + # In the C2 model, the RPN hidden layer conv is defined for FPN level 2 and then + # shared for all other levels, hence the appearance of "fpn2" + layer_keys = [ + k.replace("conv.rpn.fpn2", "proposal_generator.rpn_head.conv") for k in layer_keys + ] + # Non-FPN case + layer_keys = [k.replace("conv.rpn", "proposal_generator.rpn_head.conv") for k in layer_keys] + + # -------------------------------------------------------------------------- + # RPN box transformation conv + # -------------------------------------------------------------------------- + # FPN case (see note above about "fpn2") + layer_keys = [ + k.replace("rpn.bbox.pred.fpn2", "proposal_generator.rpn_head.anchor_deltas") + for k in layer_keys + ] + layer_keys = [ + k.replace("rpn.cls.logits.fpn2", "proposal_generator.rpn_head.objectness_logits") + for k in layer_keys + ] + # Non-FPN case + layer_keys = [ + k.replace("rpn.bbox.pred", "proposal_generator.rpn_head.anchor_deltas") for k in layer_keys + ] + layer_keys = [ + k.replace("rpn.cls.logits", "proposal_generator.rpn_head.objectness_logits") + for k in layer_keys + ] + + # -------------------------------------------------------------------------- + # Fast R-CNN box head + # -------------------------------------------------------------------------- + layer_keys = [re.sub("^bbox\\.pred", "bbox_pred", k) for k in layer_keys] + layer_keys = [re.sub("^cls\\.score", "cls_score", k) for k in layer_keys] + layer_keys = [re.sub("^fc6\\.", "box_head.fc1.", k) for k in layer_keys] + layer_keys = [re.sub("^fc7\\.", "box_head.fc2.", k) for k in layer_keys] + # 4conv1fc head tensor names: head_conv1_w, head_conv1_gn_s + layer_keys = [re.sub("^head\\.conv", "box_head.conv", k) for k in layer_keys] + + # -------------------------------------------------------------------------- + # FPN lateral and output convolutions + # -------------------------------------------------------------------------- + def fpn_map(name): + """ + Look for keys with the following patterns: + 1) Starts with "fpn.inner." + Example: "fpn.inner.res2.2.sum.lateral.weight" + Meaning: These are lateral pathway convolutions + 2) Starts with "fpn.res" + Example: "fpn.res2.2.sum.weight" + Meaning: These are FPN output convolutions + """ + splits = name.split(".") + norm = ".norm" if "norm" in splits else "" + if name.startswith("fpn.inner."): + # splits example: ['fpn', 'inner', 'res2', '2', 'sum', 'lateral', 'weight'] + stage = int(splits[2][len("res") :]) + return "fpn_lateral{}{}.{}".format(stage, norm, splits[-1]) + elif name.startswith("fpn.res"): + # splits example: ['fpn', 'res2', '2', 'sum', 'weight'] + stage = int(splits[1][len("res") :]) + return "fpn_output{}{}.{}".format(stage, norm, splits[-1]) + return name + + layer_keys = [fpn_map(k) for k in layer_keys] + + # -------------------------------------------------------------------------- + # Mask R-CNN mask head + # -------------------------------------------------------------------------- + # roi_heads.StandardROIHeads case + layer_keys = [k.replace(".[mask].fcn", "mask_head.mask_fcn") for k in layer_keys] + layer_keys = [re.sub("^\\.mask\\.fcn", "mask_head.mask_fcn", k) for k in layer_keys] + layer_keys = [k.replace("mask.fcn.logits", "mask_head.predictor") for k in layer_keys] + # roi_heads.Res5ROIHeads case + layer_keys = [k.replace("conv5.mask", "mask_head.deconv") for k in layer_keys] + + # -------------------------------------------------------------------------- + # Keypoint R-CNN head + # -------------------------------------------------------------------------- + # interestingly, the keypoint head convs have blob names that are simply "conv_fcnX" + layer_keys = [k.replace("conv.fcn", "roi_heads.keypoint_head.conv_fcn") for k in layer_keys] + layer_keys = [ + k.replace("kps.score.lowres", "roi_heads.keypoint_head.score_lowres") for k in layer_keys + ] + layer_keys = [k.replace("kps.score.", "roi_heads.keypoint_head.score.") for k in layer_keys] + + # -------------------------------------------------------------------------- + # Done with replacements + # -------------------------------------------------------------------------- + assert len(set(layer_keys)) == len(layer_keys) + assert len(original_keys) == len(layer_keys) + + new_weights = {} + new_keys_to_original_keys = {} + for orig, renamed in zip(original_keys, layer_keys): + new_keys_to_original_keys[renamed] = orig + if renamed.startswith("bbox_pred.") or renamed.startswith("mask_head.predictor."): + # remove the meaningless prediction weight for background class + new_start_idx = 4 if renamed.startswith("bbox_pred.") else 1 + new_weights[renamed] = weights[orig][new_start_idx:] + logger.info( + "Remove prediction weight for background class in {}. The shape changes from " + "{} to {}.".format( + renamed, tuple(weights[orig].shape), tuple(new_weights[renamed].shape) + ) + ) + elif renamed.startswith("cls_score."): + # move weights of bg class from original index 0 to last index + logger.info( + "Move classification weights for background class in {} from index 0 to " + "index {}.".format(renamed, weights[orig].shape[0] - 1) + ) + new_weights[renamed] = torch.cat([weights[orig][1:], weights[orig][:1]]) + else: + new_weights[renamed] = weights[orig] + + return new_weights, new_keys_to_original_keys + + +# Note the current matching is not symmetric. +# it assumes model_state_dict will have longer names. +def align_and_update_state_dicts(model_state_dict, ckpt_state_dict, c2_conversion=True): + """ + Match names between the two state-dict, and returns a new chkpt_state_dict with names + converted to match model_state_dict with heuristics. The returned dict can be later + loaded with fvcore checkpointer. + If `c2_conversion==True`, `ckpt_state_dict` is assumed to be a Caffe2 + model and will be renamed at first. + + Strategy: suppose that the models that we will create will have prefixes appended + to each of its keys, for example due to an extra level of nesting that the original + pre-trained weights from ImageNet won't contain. For example, model.state_dict() + might return backbone[0].body.res2.conv1.weight, while the pre-trained model contains + res2.conv1.weight. We thus want to match both parameters together. + For that, we look for each model weight, look among all loaded keys if there is one + that is a suffix of the current weight name, and use it if that's the case. + If multiple matches exist, take the one with longest size + of the corresponding name. For example, for the same model as before, the pretrained + weight file can contain both res2.conv1.weight, as well as conv1.weight. In this case, + we want to match backbone[0].body.conv1.weight to conv1.weight, and + backbone[0].body.res2.conv1.weight to res2.conv1.weight. + """ + model_keys = sorted(model_state_dict.keys()) + if c2_conversion: + ckpt_state_dict, original_keys = convert_c2_detectron_names(ckpt_state_dict) + # original_keys: the name in the original dict (before renaming) + else: + original_keys = {x: x for x in ckpt_state_dict.keys()} + ckpt_keys = sorted(ckpt_state_dict.keys()) + + def match(a, b): + # Matched ckpt_key should be a complete (starts with '.') suffix. + # For example, roi_heads.mesh_head.whatever_conv1 does not match conv1, + # but matches whatever_conv1 or mesh_head.whatever_conv1. + return a == b or a.endswith("." + b) + + # get a matrix of string matches, where each (i, j) entry correspond to the size of the + # ckpt_key string, if it matches + match_matrix = [len(j) if match(i, j) else 0 for i in model_keys for j in ckpt_keys] + match_matrix = torch.as_tensor(match_matrix).view(len(model_keys), len(ckpt_keys)) + # use the matched one with longest size in case of multiple matches + max_match_size, idxs = match_matrix.max(1) + # remove indices that correspond to no-match + idxs[max_match_size == 0] = -1 + + logger = logging.getLogger(__name__) + # matched_pairs (matched checkpoint key --> matched model key) + matched_keys = {} + result_state_dict = {} + for idx_model, idx_ckpt in enumerate(idxs.tolist()): + if idx_ckpt == -1: + continue + key_model = model_keys[idx_model] + key_ckpt = ckpt_keys[idx_ckpt] + value_ckpt = ckpt_state_dict[key_ckpt] + shape_in_model = model_state_dict[key_model].shape + + if shape_in_model != value_ckpt.shape: + logger.warning( + "Shape of {} in checkpoint is {}, while shape of {} in model is {}.".format( + key_ckpt, value_ckpt.shape, key_model, shape_in_model + ) + ) + logger.warning( + "{} will not be loaded. Please double check and see if this is desired.".format( + key_ckpt + ) + ) + continue + + assert key_model not in result_state_dict + result_state_dict[key_model] = value_ckpt + if key_ckpt in matched_keys: # already added to matched_keys + logger.error( + "Ambiguity found for {} in checkpoint!" + "It matches at least two keys in the model ({} and {}).".format( + key_ckpt, key_model, matched_keys[key_ckpt] + ) + ) + raise ValueError("Cannot match one checkpoint key to multiple keys in the model.") + + matched_keys[key_ckpt] = key_model + + # logging: + matched_model_keys = sorted(matched_keys.values()) + if len(matched_model_keys) == 0: + logger.warning("No weights in checkpoint matched with model.") + return ckpt_state_dict + common_prefix = _longest_common_prefix(matched_model_keys) + rev_matched_keys = {v: k for k, v in matched_keys.items()} + original_keys = {k: original_keys[rev_matched_keys[k]] for k in matched_model_keys} + + model_key_groups = _group_keys_by_module(matched_model_keys, original_keys) + table = [] + memo = set() + for key_model in matched_model_keys: + if key_model in memo: + continue + if key_model in model_key_groups: + group = model_key_groups[key_model] + memo |= set(group) + shapes = [tuple(model_state_dict[k].shape) for k in group] + table.append( + ( + _longest_common_prefix([k[len(common_prefix) :] for k in group]) + "*", + _group_str([original_keys[k] for k in group]), + " ".join([str(x).replace(" ", "") for x in shapes]), + ) + ) + else: + key_checkpoint = original_keys[key_model] + shape = str(tuple(model_state_dict[key_model].shape)) + table.append((key_model[len(common_prefix) :], key_checkpoint, shape)) + table_str = tabulate( + table, tablefmt="pipe", headers=["Names in Model", "Names in Checkpoint", "Shapes"] + ) + logger.info( + "Following weights matched with " + + (f"submodule {common_prefix[:-1]}" if common_prefix else "model") + + ":\n" + + table_str + ) + + unmatched_ckpt_keys = [k for k in ckpt_keys if k not in set(matched_keys.keys())] + for k in unmatched_ckpt_keys: + result_state_dict[k] = ckpt_state_dict[k] + return result_state_dict + + +def _group_keys_by_module(keys: List[str], original_names: Dict[str, str]): + """ + Params in the same submodule are grouped together. + + Args: + keys: names of all parameters + original_names: mapping from parameter name to their name in the checkpoint + + Returns: + dict[name -> all other names in the same group] + """ + + def _submodule_name(key): + pos = key.rfind(".") + if pos < 0: + return None + prefix = key[: pos + 1] + return prefix + + all_submodules = [_submodule_name(k) for k in keys] + all_submodules = [x for x in all_submodules if x] + all_submodules = sorted(all_submodules, key=len) + + ret = {} + for prefix in all_submodules: + group = [k for k in keys if k.startswith(prefix)] + if len(group) <= 1: + continue + original_name_lcp = _longest_common_prefix_str([original_names[k] for k in group]) + if len(original_name_lcp) == 0: + # don't group weights if original names don't share prefix + continue + + for k in group: + if k in ret: + continue + ret[k] = group + return ret + + +def _longest_common_prefix(names: List[str]) -> str: + """ + ["abc.zfg", "abc.zef"] -> "abc." + """ + names = [n.split(".") for n in names] + m1, m2 = min(names), max(names) + ret = [a for a, b in zip(m1, m2) if a == b] + ret = ".".join(ret) + "." if len(ret) else "" + return ret + + +def _longest_common_prefix_str(names: List[str]) -> str: + m1, m2 = min(names), max(names) + lcp = [a for a, b in zip(m1, m2) if a == b] + lcp = "".join(lcp) + return lcp + + +def _group_str(names: List[str]) -> str: + """ + Turn "common1", "common2", "common3" into "common{1,2,3}" + """ + lcp = _longest_common_prefix_str(names) + rest = [x[len(lcp) :] for x in names] + rest = "{" + ",".join(rest) + "}" + ret = lcp + rest + + # add some simplification for BN specifically + ret = ret.replace("bn_{beta,running_mean,running_var,gamma}", "bn_*") + ret = ret.replace("bn_beta,bn_running_mean,bn_running_var,bn_gamma", "bn_*") + return ret diff --git a/detectron2/checkpoint/catalog.py b/detectron2/checkpoint/catalog.py new file mode 100644 index 0000000000000000000000000000000000000000..9a85736754a0de4550df96c22f38fc515bd02d71 --- /dev/null +++ b/detectron2/checkpoint/catalog.py @@ -0,0 +1,115 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging + +from detectron2.utils.file_io import PathHandler, PathManager + + +class ModelCatalog(object): + """ + Store mappings from names to third-party models. + """ + + S3_C2_DETECTRON_PREFIX = "https://dl.fbaipublicfiles.com/detectron" + + # MSRA models have STRIDE_IN_1X1=True. False otherwise. + # NOTE: all BN models here have fused BN into an affine layer. + # As a result, you should only load them to a model with "FrozenBN". + # Loading them to a model with regular BN or SyncBN is wrong. + # Even when loaded to FrozenBN, it is still different from affine by an epsilon, + # which should be negligible for training. + # NOTE: all models here uses PIXEL_STD=[1,1,1] + # NOTE: Most of the BN models here are no longer used. We use the + # re-converted pre-trained models under detectron2 model zoo instead. + C2_IMAGENET_MODELS = { + "MSRA/R-50": "ImageNetPretrained/MSRA/R-50.pkl", + "MSRA/R-101": "ImageNetPretrained/MSRA/R-101.pkl", + "FAIR/R-50-GN": "ImageNetPretrained/47261647/R-50-GN.pkl", + "FAIR/R-101-GN": "ImageNetPretrained/47592356/R-101-GN.pkl", + "FAIR/X-101-32x8d": "ImageNetPretrained/20171220/X-101-32x8d.pkl", + "FAIR/X-101-64x4d": "ImageNetPretrained/FBResNeXt/X-101-64x4d.pkl", + "FAIR/X-152-32x8d-IN5k": "ImageNetPretrained/25093814/X-152-32x8d-IN5k.pkl", + } + + C2_DETECTRON_PATH_FORMAT = ( + "{prefix}/{url}/output/train/{dataset}/{type}/model_final.pkl" # noqa B950 + ) + + C2_DATASET_COCO = "coco_2014_train%3Acoco_2014_valminusminival" + C2_DATASET_COCO_KEYPOINTS = "keypoints_coco_2014_train%3Akeypoints_coco_2014_valminusminival" + + # format: {model_name} -> part of the url + C2_DETECTRON_MODELS = { + "35857197/e2e_faster_rcnn_R-50-C4_1x": "35857197/12_2017_baselines/e2e_faster_rcnn_R-50-C4_1x.yaml.01_33_49.iAX0mXvW", # noqa B950 + "35857345/e2e_faster_rcnn_R-50-FPN_1x": "35857345/12_2017_baselines/e2e_faster_rcnn_R-50-FPN_1x.yaml.01_36_30.cUF7QR7I", # noqa B950 + "35857890/e2e_faster_rcnn_R-101-FPN_1x": "35857890/12_2017_baselines/e2e_faster_rcnn_R-101-FPN_1x.yaml.01_38_50.sNxI7sX7", # noqa B950 + "36761737/e2e_faster_rcnn_X-101-32x8d-FPN_1x": "36761737/12_2017_baselines/e2e_faster_rcnn_X-101-32x8d-FPN_1x.yaml.06_31_39.5MIHi1fZ", # noqa B950 + "35858791/e2e_mask_rcnn_R-50-C4_1x": "35858791/12_2017_baselines/e2e_mask_rcnn_R-50-C4_1x.yaml.01_45_57.ZgkA7hPB", # noqa B950 + "35858933/e2e_mask_rcnn_R-50-FPN_1x": "35858933/12_2017_baselines/e2e_mask_rcnn_R-50-FPN_1x.yaml.01_48_14.DzEQe4wC", # noqa B950 + "35861795/e2e_mask_rcnn_R-101-FPN_1x": "35861795/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_1x.yaml.02_31_37.KqyEK4tT", # noqa B950 + "36761843/e2e_mask_rcnn_X-101-32x8d-FPN_1x": "36761843/12_2017_baselines/e2e_mask_rcnn_X-101-32x8d-FPN_1x.yaml.06_35_59.RZotkLKI", # noqa B950 + "48616381/e2e_mask_rcnn_R-50-FPN_2x_gn": "GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q", # noqa B950 + "37697547/e2e_keypoint_rcnn_R-50-FPN_1x": "37697547/12_2017_baselines/e2e_keypoint_rcnn_R-50-FPN_1x.yaml.08_42_54.kdzV35ao", # noqa B950 + "35998355/rpn_R-50-C4_1x": "35998355/12_2017_baselines/rpn_R-50-C4_1x.yaml.08_00_43.njH5oD9L", # noqa B950 + "35998814/rpn_R-50-FPN_1x": "35998814/12_2017_baselines/rpn_R-50-FPN_1x.yaml.08_06_03.Axg0r179", # noqa B950 + "36225147/fast_R-50-FPN_1x": "36225147/12_2017_baselines/fast_rcnn_R-50-FPN_1x.yaml.08_39_09.L3obSdQ2", # noqa B950 + } + + @staticmethod + def get(name): + if name.startswith("Caffe2Detectron/COCO"): + return ModelCatalog._get_c2_detectron_baseline(name) + if name.startswith("ImageNetPretrained/"): + return ModelCatalog._get_c2_imagenet_pretrained(name) + raise RuntimeError("model not present in the catalog: {}".format(name)) + + @staticmethod + def _get_c2_imagenet_pretrained(name): + prefix = ModelCatalog.S3_C2_DETECTRON_PREFIX + name = name[len("ImageNetPretrained/") :] + name = ModelCatalog.C2_IMAGENET_MODELS[name] + url = "/".join([prefix, name]) + return url + + @staticmethod + def _get_c2_detectron_baseline(name): + name = name[len("Caffe2Detectron/COCO/") :] + url = ModelCatalog.C2_DETECTRON_MODELS[name] + if "keypoint_rcnn" in name: + dataset = ModelCatalog.C2_DATASET_COCO_KEYPOINTS + else: + dataset = ModelCatalog.C2_DATASET_COCO + + if "35998355/rpn_R-50-C4_1x" in name: + # this one model is somehow different from others .. + type = "rpn" + else: + type = "generalized_rcnn" + + # Detectron C2 models are stored in the structure defined in `C2_DETECTRON_PATH_FORMAT`. + url = ModelCatalog.C2_DETECTRON_PATH_FORMAT.format( + prefix=ModelCatalog.S3_C2_DETECTRON_PREFIX, url=url, type=type, dataset=dataset + ) + return url + + +class ModelCatalogHandler(PathHandler): + """ + Resolve URL like catalog://. + """ + + PREFIX = "catalog://" + + def _get_supported_prefixes(self): + return [self.PREFIX] + + def _get_local_path(self, path, **kwargs): + logger = logging.getLogger(__name__) + catalog_path = ModelCatalog.get(path[len(self.PREFIX) :]) + logger.info("Catalog entry {} points to {}".format(path, catalog_path)) + return PathManager.get_local_path(catalog_path, **kwargs) + + def _open(self, path, mode="r", **kwargs): + return PathManager.open(self._get_local_path(path), mode, **kwargs) + + +PathManager.register_handler(ModelCatalogHandler()) diff --git a/detectron2/checkpoint/clip_model_loading.py b/detectron2/checkpoint/clip_model_loading.py new file mode 100644 index 0000000000000000000000000000000000000000..12b0fe79aa3d59d72742c33a111fe5d22fb6c725 --- /dev/null +++ b/detectron2/checkpoint/clip_model_loading.py @@ -0,0 +1,415 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import logging +import re +from typing import Dict, List +import torch +from tabulate import tabulate + + +def convert_basic_clip_names(original_keys, add_backbone_prefix=False, use_whole_clip=False, use_fpn_arch=False, regionclip=False): + """ + Apply some basic name conversion to names in CLIP weights. + It only deals with typical backbone models. + + Args: + original_keys (list[str]): + Returns: + list[str]: The same number of strings matching those in original_keys. + """ + layer_keys = copy.deepcopy(original_keys) + + vit = False + for l_k in layer_keys: + if 'visual.transformer' in l_k: + vit = True + + # load pretrained oai clip + if not vit: # resnet + if add_backbone_prefix: # CLIPRCNN or CLIPFastRCNN + if use_whole_clip: # CLIPRCNN + layer_keys = [k.replace("visual.", "clip_backbone.visual.") for k in layer_keys] + else: # CLIPFastRCNN + if use_fpn_arch: # FPN + layer_keys = [k.replace("visual.", "backbone.bottom_up.") for k in layer_keys] + else: # C4 + layer_keys = [k.replace("visual.", "backbone.") for k in layer_keys] + else: # GeneralizedRCNN or ProposalNetwork + #layer_keys = [k.replace("visual.", "backbone.bottom_up.") for k in layer_keys] # + layer_keys = [k.replace("visual.", "") for k in layer_keys] # + #layer_keys = [k.replace("visual.", "backbone.visual.") for k in layer_keys] # + else: # vit + pass + + return layer_keys, vit + + +def convert_clip_names(weights, add_backbone_prefix=False, use_whole_clip=False, use_fpn_arch=False, regionclip=False): + """ + Map CLIP Detectron weight names to Detectron2 names. + + Args: + weights (dict): name -> tensor + + Returns: + dict: detectron2 names -> tensor + dict: detectron2 names -> C2 names + """ + logger = logging.getLogger(__name__) + logger.info("Renaming CLIP weights ......") + original_keys = sorted(weights.keys()) + layer_keys = copy.deepcopy(original_keys) + + layer_keys, use_vit = convert_basic_clip_names(layer_keys, add_backbone_prefix, use_whole_clip, use_fpn_arch, regionclip) + + # -------------------------------------------------------------------------- + # RPN hidden representation conv + # -------------------------------------------------------------------------- + # FPN case + # In the C2 model, the RPN hidden layer conv is defined for FPN level 2 and then + # shared for all other levels, hence the appearance of "fpn2" + layer_keys = [ + k.replace("conv.rpn.fpn2", "proposal_generator.rpn_head.conv") for k in layer_keys + ] + # Non-FPN case + layer_keys = [k.replace("conv.rpn", "proposal_generator.rpn_head.conv") for k in layer_keys] + + # -------------------------------------------------------------------------- + # RPN box transformation conv + # -------------------------------------------------------------------------- + # FPN case (see note above about "fpn2") + layer_keys = [ + k.replace("rpn.bbox.pred.fpn2", "proposal_generator.rpn_head.anchor_deltas") + for k in layer_keys + ] + layer_keys = [ + k.replace("rpn.cls.logits.fpn2", "proposal_generator.rpn_head.objectness_logits") + for k in layer_keys + ] + # Non-FPN case + layer_keys = [ + k.replace("rpn.bbox.pred", "proposal_generator.rpn_head.anchor_deltas") for k in layer_keys + ] + layer_keys = [ + k.replace("rpn.cls.logits", "proposal_generator.rpn_head.objectness_logits") + for k in layer_keys + ] + + # -------------------------------------------------------------------------- + # Fast R-CNN box head + # -------------------------------------------------------------------------- + layer_keys = [re.sub("^bbox\\.pred", "bbox_pred", k) for k in layer_keys] + layer_keys = [re.sub("^cls\\.score", "cls_score", k) for k in layer_keys] + layer_keys = [re.sub("^fc6\\.", "box_head.fc1.", k) for k in layer_keys] + layer_keys = [re.sub("^fc7\\.", "box_head.fc2.", k) for k in layer_keys] + # 4conv1fc head tensor names: head_conv1_w, head_conv1_gn_s + layer_keys = [re.sub("^head\\.conv", "box_head.conv", k) for k in layer_keys] + + # -------------------------------------------------------------------------- + # FPN lateral and output convolutions + # -------------------------------------------------------------------------- + def fpn_map(name): + """ + Look for keys with the following patterns: + 1) Starts with "fpn.inner." + Example: "fpn.inner.res2.2.sum.lateral.weight" + Meaning: These are lateral pathway convolutions + 2) Starts with "fpn.res" + Example: "fpn.res2.2.sum.weight" + Meaning: These are FPN output convolutions + """ + splits = name.split(".") + norm = ".norm" if "norm" in splits else "" + if name.startswith("fpn.inner."): + # splits example: ['fpn', 'inner', 'res2', '2', 'sum', 'lateral', 'weight'] + stage = int(splits[2][len("res") :]) + return "fpn_lateral{}{}.{}".format(stage, norm, splits[-1]) + elif name.startswith("fpn.res"): + # splits example: ['fpn', 'res2', '2', 'sum', 'weight'] + stage = int(splits[1][len("res") :]) + return "fpn_output{}{}.{}".format(stage, norm, splits[-1]) + return name + + layer_keys = [fpn_map(k) for k in layer_keys] + + # -------------------------------------------------------------------------- + # Mask R-CNN mask head + # -------------------------------------------------------------------------- + # roi_heads.StandardROIHeads case + layer_keys = [k.replace(".[mask].fcn", "mask_head.mask_fcn") for k in layer_keys] + layer_keys = [re.sub("^\\.mask\\.fcn", "mask_head.mask_fcn", k) for k in layer_keys] + layer_keys = [k.replace("mask.fcn.logits", "mask_head.predictor") for k in layer_keys] + # roi_heads.Res5ROIHeads case + layer_keys = [k.replace("conv5.mask", "mask_head.deconv") for k in layer_keys] + + # -------------------------------------------------------------------------- + # Keypoint R-CNN head + # -------------------------------------------------------------------------- + # interestingly, the keypoint head convs have blob names that are simply "conv_fcnX" + layer_keys = [k.replace("conv.fcn", "roi_heads.keypoint_head.conv_fcn") for k in layer_keys] + layer_keys = [ + k.replace("kps.score.lowres", "roi_heads.keypoint_head.score_lowres") for k in layer_keys + ] + layer_keys = [k.replace("kps.score.", "roi_heads.keypoint_head.score.") for k in layer_keys] + + # -------------------------------------------------------------------------- + # Done with replacements + # -------------------------------------------------------------------------- + assert len(set(layer_keys)) == len(layer_keys) + assert len(original_keys) == len(layer_keys) + + new_weights = {} + new_keys_to_original_keys = {} + for orig, renamed in zip(original_keys, layer_keys): + new_keys_to_original_keys[renamed] = orig + if renamed.startswith("bbox_pred.") or renamed.startswith("mask_head.predictor."): + # remove the meaningless prediction weight for background class + new_start_idx = 4 if renamed.startswith("bbox_pred.") else 1 + new_weights[renamed] = weights[orig][new_start_idx:] + logger.info( + "Remove prediction weight for background class in {}. The shape changes from " + "{} to {}.".format( + renamed, tuple(weights[orig].shape), tuple(new_weights[renamed].shape) + ) + ) + elif renamed.startswith("cls_score."): + # move weights of bg class from original index 0 to last index + logger.info( + "Move classification weights for background class in {} from index 0 to " + "index {}.".format(renamed, weights[orig].shape[0] - 1) + ) + new_weights[renamed] = torch.cat([weights[orig][1:], weights[orig][:1]]) + else: + new_weights[renamed] = weights[orig] + + return new_weights, new_keys_to_original_keys, use_vit + + +# Note the current matching is not symmetric. +# it assumes model_state_dict will have longer names. +def align_and_update_state_dicts_for_CLIP(model_state_dict, ckpt_state_dict, c2_conversion=True, bb_rpn_weights=False, regionclip=False): + """ + Extended from ./c2_model_loading.py + Match names between the two state-dict, and returns a new chkpt_state_dict with names + converted to match model_state_dict with heuristics. The returned dict can be later + loaded with fvcore checkpointer. + If `c2_conversion==True`, `ckpt_state_dict` is assumed to be a Caffe2 + model and will be renamed at first. + + Strategy: suppose that the models that we will create will have prefixes appended + to each of its keys, for example due to an extra level of nesting that the original + pre-trained weights from ImageNet won't contain. For example, model.state_dict() + might return backbone[0].body.res2.conv1.weight, while the pre-trained model contains + res2.conv1.weight. We thus want to match both parameters together. + For that, we look for each model weight, look among all loaded keys if there is one + that is a suffix of the current weight name, and use it if that's the case. + If multiple matches exist, take the one with longest size + of the corresponding name. For example, for the same model as before, the pretrained + weight file can contain both res2.conv1.weight, as well as conv1.weight. In this case, + we want to match backbone[0].body.conv1.weight to conv1.weight, and + backbone[0].body.res2.conv1.weight to res2.conv1.weight. + """ + model_keys = sorted(model_state_dict.keys()) + use_whole_clip = False # whether use the whole clip (text & visual encoders), typically in CLIPRCNN meta arch + add_backbone_prefix = False # convert to 'backbone.' prefix, typically in CLIPFastRCNN meta arch + use_fpn_arch = False # if use FPN arch then convert to `bottom_up`, typically in CLIPFastRCNN meta arch with FPN backbone + if bb_rpn_weights: # a 2nd pretrained weights to load, for offline backbone & RPN, then convert the ckpt key names and only keep the ones we need + new_ckpt_state_dict = {} + for original_k in ckpt_state_dict: + if 'backbone' in original_k: + new_key = original_k.replace('backbone', 'offline_backbone') + new_ckpt_state_dict[new_key] = ckpt_state_dict[original_k] + if 'proposal_generator' in original_k: + new_key = original_k.replace('proposal_generator', 'offline_proposal_generator') + new_ckpt_state_dict[new_key] = ckpt_state_dict[original_k] + new_ckpt_state_dict['ignore_others'] = torch.tensor([1]) # ignore other model weights (not 'offline_*') in batch_norm.py + ckpt_state_dict = new_ckpt_state_dict + else: # the 1st pretrained weigths to load + for model_key in model_keys: # if use the whole clip, then convert ckpt 'visual.' names to 'clip_backbone.visual.' + if 'clip_backbone' in model_key: + use_whole_clip = True + for model_key in model_keys: # if there are backbone & offline_backbone, then convert the ckpt 'visual.' names to 'backbone.' to avoid ambiguity + if 'offline_backbone' in model_key: + add_backbone_prefix = True + if 'fpn' in model_key: + use_fpn_arch = True + # original_keys: the name in the original dict (before renaming) + ckpt_state_dict, original_keys, use_vit = convert_clip_names(ckpt_state_dict, add_backbone_prefix, use_whole_clip, use_fpn_arch, regionclip) + ckpt_keys = sorted(ckpt_state_dict.keys()) + + def match(a, b): + # Matched ckpt_key should be a complete (starts with '.') suffix. + # For example, roi_heads.mesh_head.whatever_conv1 does not match conv1, + # but matches whatever_conv1 or mesh_head.whatever_conv1. + return a == b or a.endswith("." + b) + + # get a matrix of string matches, where each (i, j) entry correspond to the size of the + # ckpt_key string, if it matches + match_matrix = [len(j) if match(i, j) else 0 for i in model_keys for j in ckpt_keys] + match_matrix = torch.as_tensor(match_matrix).view(len(model_keys), len(ckpt_keys)) + # use the matched one with longest size in case of multiple matches + max_match_size, idxs = match_matrix.max(1) + # remove indices that correspond to no-match + idxs[max_match_size == 0] = -1 + + logger = logging.getLogger(__name__) + # matched_pairs (matched checkpoint key --> matched model key) + matched_keys = {} + result_state_dict = {} + for idx_model, idx_ckpt in enumerate(idxs.tolist()): + if idx_ckpt == -1: + continue + key_model = model_keys[idx_model] + key_ckpt = ckpt_keys[idx_ckpt] + value_ckpt = ckpt_state_dict[key_ckpt] + shape_in_model = model_state_dict[key_model].shape + + if shape_in_model != value_ckpt.shape: + logger.warning( + "Shape of {} in checkpoint is {}, while shape of {} in model is {}.".format( + key_ckpt, value_ckpt.shape, key_model, shape_in_model + ) + ) + logger.warning( + "{} will not be loaded. Please double check and see if this is desired.".format( + key_ckpt + ) + ) + continue + + assert key_model not in result_state_dict + result_state_dict[key_model] = value_ckpt + if key_ckpt in matched_keys: # already added to matched_keys + logger.error( + "Ambiguity found for {} in checkpoint!" + "It matches at least two keys in the model ({} and {}).".format( + key_ckpt, key_model, matched_keys[key_ckpt] + ) + ) + raise ValueError("Cannot match one checkpoint key to multiple keys in the model.") + + matched_keys[key_ckpt] = key_model + + # logging: + matched_model_keys = sorted(matched_keys.values()) + mmk_list = "The following model parameters are loaded from checkpoints:\n" + for mmk in matched_model_keys: + mmk_list += mmk + "\n" + if len(matched_model_keys) == 0: + logger.warning("No weights in checkpoint matched with model.") + return ckpt_state_dict + common_prefix = _longest_common_prefix(matched_model_keys) + rev_matched_keys = {v: k for k, v in matched_keys.items()} + original_keys = {k: original_keys[rev_matched_keys[k]] for k in matched_model_keys} + + model_key_groups = _group_keys_by_module(matched_model_keys, original_keys) + table = [] + memo = set() + for key_model in matched_model_keys: + if key_model in memo: + continue + if key_model in model_key_groups: + group = model_key_groups[key_model] + memo |= set(group) + shapes = [tuple(model_state_dict[k].shape) for k in group] + table.append( + ( + _longest_common_prefix([k[len(common_prefix) :] for k in group]) + "*", + _group_str([original_keys[k] for k in group]), + " ".join([str(x).replace(" ", "") for x in shapes]), + ) + ) + else: + key_checkpoint = original_keys[key_model] + shape = str(tuple(model_state_dict[key_model].shape)) + table.append((key_model[len(common_prefix) :], key_checkpoint, shape)) + table_str = tabulate( + table, tablefmt="pipe", headers=["Names in Model", "Names in Checkpoint", "Shapes"] + ) + if len(table) != 1 and not use_vit: # do this for now; the table function has some bugs when the whole CLIP is loaded + logger.info( + "Following weights matched with " + + (f"submodule {common_prefix[:-1]}" if common_prefix else "model") + + ":\n" + + table_str + ) + else: + logger.info(mmk_list) + + unmatched_ckpt_keys = [k for k in ckpt_keys if k not in set(matched_keys.keys())] + for k in unmatched_ckpt_keys: + result_state_dict[k] = ckpt_state_dict[k] + return result_state_dict + + +def _group_keys_by_module(keys: List[str], original_names: Dict[str, str]): + """ + Params in the same submodule are grouped together. + + Args: + keys: names of all parameters + original_names: mapping from parameter name to their name in the checkpoint + + Returns: + dict[name -> all other names in the same group] + """ + + def _submodule_name(key): + pos = key.rfind(".") + if pos < 0: + return None + prefix = key[: pos + 1] + return prefix + + all_submodules = [_submodule_name(k) for k in keys] + all_submodules = [x for x in all_submodules if x] + all_submodules = sorted(all_submodules, key=len) + + ret = {} + for prefix in all_submodules: + group = [k for k in keys if k.startswith(prefix)] + if len(group) <= 1: + continue + original_name_lcp = _longest_common_prefix_str([original_names[k] for k in group]) + if len(original_name_lcp) == 0: + # don't group weights if original names don't share prefix + continue + + for k in group: + if k in ret: + continue + ret[k] = group + return ret + + +def _longest_common_prefix(names: List[str]) -> str: + """ + ["abc.zfg", "abc.zef"] -> "abc." + """ + names = [n.split(".") for n in names] + m1, m2 = min(names), max(names) + ret = [a for a, b in zip(m1, m2) if a == b] + ret = ".".join(ret) + "." if len(ret) else "" + return ret + + +def _longest_common_prefix_str(names: List[str]) -> str: + m1, m2 = min(names), max(names) + lcp = [a for a, b in zip(m1, m2) if a == b] + lcp = "".join(lcp) + return lcp + + +def _group_str(names: List[str]) -> str: + """ + Turn "common1", "common2", "common3" into "common{1,2,3}" + """ + lcp = _longest_common_prefix_str(names) + rest = [x[len(lcp) :] for x in names] + rest = "{" + ",".join(rest) + "}" + ret = lcp + rest + + # add some simplification for BN specifically + ret = ret.replace("bn_{beta,running_mean,running_var,gamma}", "bn_*") + ret = ret.replace("bn_beta,bn_running_mean,bn_running_var,bn_gamma", "bn_*") + return ret diff --git a/detectron2/checkpoint/detection_checkpoint.py b/detectron2/checkpoint/detection_checkpoint.py new file mode 100644 index 0000000000000000000000000000000000000000..42fbaa5be6f304a799247948b4da5e6b14da2c45 --- /dev/null +++ b/detectron2/checkpoint/detection_checkpoint.py @@ -0,0 +1,134 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import os +import pickle +import torch +from fvcore.common.checkpoint import Checkpointer +from torch.nn.parallel import DistributedDataParallel + +import detectron2.utils.comm as comm +from detectron2.utils.env import TORCH_VERSION +from detectron2.utils.file_io import PathManager + +from .c2_model_loading import align_and_update_state_dicts +from .clip_model_loading import align_and_update_state_dicts_for_CLIP + +class DetectionCheckpointer(Checkpointer): + """ + Same as :class:`Checkpointer`, but is able to: + 1. handle models in detectron & detectron2 model zoo, and apply conversions for legacy models. + 2. correctly load checkpoints that are only available on the master worker + """ + + def __init__(self, model, save_dir="", *, save_to_disk=None, bb_rpn_weights=False, **checkpointables): + is_main_process = comm.is_main_process() + super().__init__( + model, + save_dir, + save_to_disk=is_main_process if save_to_disk is None else save_to_disk, + **checkpointables, + ) + self.path_manager = PathManager + self.bb_rpn_weights = bb_rpn_weights + + def load(self, path, *args, **kwargs): + need_sync = False + + if path and isinstance(self.model, DistributedDataParallel): + logger = logging.getLogger(__name__) + path = self.path_manager.get_local_path(path) + has_file = os.path.isfile(path) + all_has_file = comm.all_gather(has_file) + if not all_has_file[0]: + raise OSError(f"File {path} not found on main worker.") + if not all(all_has_file): + logger.warning( + f"Not all workers can read checkpoint {path}. " + "Training may fail to fully resume." + ) + # TODO: broadcast the checkpoint file contents from main + # worker, and load from it instead. + need_sync = True + if not has_file: + path = None # don't load if not readable + ret = super().load(path, *args, **kwargs) + + if need_sync: + logger.info("Broadcasting model states from main worker ...") + if TORCH_VERSION >= (1, 7): + self.model._sync_params_and_buffers() + return ret + + def _load_file(self, filename): + if filename.endswith(".pkl"): + with PathManager.open(filename, "rb") as f: + data = pickle.load(f, encoding="latin1") + if "model" in data and "__author__" in data: + # file is in Detectron2 model zoo format + self.logger.info("Reading a file from '{}'".format(data["__author__"])) + return data + else: + # assume file is from Caffe2 / Detectron1 model zoo + if "blobs" in data: + # Detection models have "blobs", but ImageNet models don't + data = data["blobs"] + data = {k: v for k, v in data.items() if not k.endswith("_momentum")} + return {"model": data, "__author__": "Caffe2", "matching_heuristics": True} + elif filename.endswith(".pyth"): + # assume file is from pycls; no one else seems to use the ".pyth" extension + with PathManager.open(filename, "rb") as f: + data = torch.load(f) + assert ( + "model_state" in data + ), f"Cannot load .pyth file {filename}; pycls checkpoints must contain 'model_state'." + model_state = { + k: v + for k, v in data["model_state"].items() + if not k.endswith("num_batches_tracked") + } + return {"model": model_state, "__author__": "pycls", "matching_heuristics": True} + elif "OAI_CLIP" in filename: + # assume file is from OpenAI CLIP pre-trained model + loaded = super()._load_file(filename) # load native pth checkpoint + if "model" not in loaded: + loaded = {"model": loaded} + return {"model": loaded["model"], "__author__": "OAI_CLIP", "matching_heuristics": True} + + loaded = super()._load_file(filename) # load native pth checkpoint + if "model" not in loaded: + loaded = {"model": loaded} + return loaded + + def _load_model(self, checkpoint): + # if checkpoint.get("matching_heuristics", False) or self.bb_rpn_weights: + # self._convert_ndarray_to_tensor(checkpoint["model"]) + # # convert weights by name-matching heuristics + # if checkpoint.get("__author__", "NA") == "OAI_CLIP" or self.bb_rpn_weights: # for OAI_CLIP or 2nd ckpt (offline modules) + # checkpoint["model"] = align_and_update_state_dicts_for_CLIP( + # self.model.state_dict(), + # checkpoint["model"], + # bb_rpn_weights=self.bb_rpn_weights, + # ) + # else: # default loading + # checkpoint["model"] = align_and_update_state_dicts( + # self.model.state_dict(), + # checkpoint["model"], + # c2_conversion=checkpoint.get("__author__", None) == "Caffe2", + # ) + # for non-caffe2 models, use standard ways to load it + # if not self.bb_rpn_weights: + # checkpoint = {'model': {'backbone.' + key: val for key, val in checkpoint['model'].items()}} + incompatible = super()._load_model(checkpoint) + del checkpoint # try saving memory + + model_buffers = dict(self.model.named_buffers(recurse=False)) + for k in ["pixel_mean", "pixel_std"]: + # Ignore missing key message about pixel_mean/std. + # Though they may be missing in old checkpoints, they will be correctly + # initialized from config anyway. + if k in model_buffers: + try: + incompatible.missing_keys.remove(k) + except ValueError: + pass + return incompatible \ No newline at end of file diff --git a/detectron2/config/__init__.py b/detectron2/config/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..4e648e632d55c70f160d49630378d202fbde4e45 --- /dev/null +++ b/detectron2/config/__init__.py @@ -0,0 +1,24 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .compat import downgrade_config, upgrade_config +from .config import CfgNode, get_cfg, global_cfg, set_global_cfg, configurable +from .instantiate import instantiate +from .lazy import LazyCall, LazyConfig + +__all__ = [ + "CfgNode", + "get_cfg", + "global_cfg", + "set_global_cfg", + "downgrade_config", + "upgrade_config", + "configurable", + "instantiate", + "LazyCall", + "LazyConfig", +] + + +from detectron2.utils.env import fixup_module_metadata + +fixup_module_metadata(__name__, globals(), __all__) +del fixup_module_metadata diff --git a/detectron2/config/__pycache__/__init__.cpython-39.pyc b/detectron2/config/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c86641eba0ea6c591484cbc85c95d0848ee61fef Binary files /dev/null and b/detectron2/config/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/config/__pycache__/compat.cpython-39.pyc b/detectron2/config/__pycache__/compat.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..98e5fbcd39744275d4786322b03d801f35209214 Binary files /dev/null and b/detectron2/config/__pycache__/compat.cpython-39.pyc differ diff --git a/detectron2/config/__pycache__/config.cpython-39.pyc b/detectron2/config/__pycache__/config.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5c57bdcd149c00431823e32678270e228c2f595a Binary files /dev/null and b/detectron2/config/__pycache__/config.cpython-39.pyc differ diff --git a/detectron2/config/__pycache__/defaults.cpython-39.pyc b/detectron2/config/__pycache__/defaults.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..16804d4e287113840f71707cfd7bad17be516104 Binary files /dev/null and b/detectron2/config/__pycache__/defaults.cpython-39.pyc differ diff --git a/detectron2/config/__pycache__/instantiate.cpython-39.pyc b/detectron2/config/__pycache__/instantiate.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4352512b6c29f0491a57f20fe13e0bbf162e224b Binary files /dev/null and b/detectron2/config/__pycache__/instantiate.cpython-39.pyc differ diff --git a/detectron2/config/__pycache__/lazy.cpython-39.pyc b/detectron2/config/__pycache__/lazy.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6902f9689c260b14502a3ca6327f54842d7b731f Binary files /dev/null and b/detectron2/config/__pycache__/lazy.cpython-39.pyc differ diff --git a/detectron2/config/compat.py b/detectron2/config/compat.py new file mode 100644 index 0000000000000000000000000000000000000000..11a08c439bf14defd880e37a938fab8a08e68eeb --- /dev/null +++ b/detectron2/config/compat.py @@ -0,0 +1,229 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +""" +Backward compatibility of configs. + +Instructions to bump version: ++ It's not needed to bump version if new keys are added. + It's only needed when backward-incompatible changes happen + (i.e., some existing keys disappear, or the meaning of a key changes) ++ To bump version, do the following: + 1. Increment _C.VERSION in defaults.py + 2. Add a converter in this file. + + Each ConverterVX has a function "upgrade" which in-place upgrades config from X-1 to X, + and a function "downgrade" which in-place downgrades config from X to X-1 + + In each function, VERSION is left unchanged. + + Each converter assumes that its input has the relevant keys + (i.e., the input is not a partial config). + 3. Run the tests (test_config.py) to make sure the upgrade & downgrade + functions are consistent. +""" + +import logging +from typing import List, Optional, Tuple + +from .config import CfgNode as CN +from .defaults import _C + +__all__ = ["upgrade_config", "downgrade_config"] + + +def upgrade_config(cfg: CN, to_version: Optional[int] = None) -> CN: + """ + Upgrade a config from its current version to a newer version. + + Args: + cfg (CfgNode): + to_version (int): defaults to the latest version. + """ + cfg = cfg.clone() + if to_version is None: + to_version = _C.VERSION + + assert cfg.VERSION <= to_version, "Cannot upgrade from v{} to v{}!".format( + cfg.VERSION, to_version + ) + for k in range(cfg.VERSION, to_version): + converter = globals()["ConverterV" + str(k + 1)] + converter.upgrade(cfg) + cfg.VERSION = k + 1 + return cfg + + +def downgrade_config(cfg: CN, to_version: int) -> CN: + """ + Downgrade a config from its current version to an older version. + + Args: + cfg (CfgNode): + to_version (int): + + Note: + A general downgrade of arbitrary configs is not always possible due to the + different functionalities in different versions. + The purpose of downgrade is only to recover the defaults in old versions, + allowing it to load an old partial yaml config. + Therefore, the implementation only needs to fill in the default values + in the old version when a general downgrade is not possible. + """ + cfg = cfg.clone() + assert cfg.VERSION >= to_version, "Cannot downgrade from v{} to v{}!".format( + cfg.VERSION, to_version + ) + for k in range(cfg.VERSION, to_version, -1): + converter = globals()["ConverterV" + str(k)] + converter.downgrade(cfg) + cfg.VERSION = k - 1 + return cfg + + +def guess_version(cfg: CN, filename: str) -> int: + """ + Guess the version of a partial config where the VERSION field is not specified. + Returns the version, or the latest if cannot make a guess. + + This makes it easier for users to migrate. + """ + logger = logging.getLogger(__name__) + + def _has(name: str) -> bool: + cur = cfg + for n in name.split("."): + if n not in cur: + return False + cur = cur[n] + return True + + # Most users' partial configs have "MODEL.WEIGHT", so guess on it + ret = None + if _has("MODEL.WEIGHT") or _has("TEST.AUG_ON"): + ret = 1 + + if ret is not None: + logger.warning("Config '{}' has no VERSION. Assuming it to be v{}.".format(filename, ret)) + else: + ret = _C.VERSION + logger.warning( + "Config '{}' has no VERSION. Assuming it to be compatible with latest v{}.".format( + filename, ret + ) + ) + return ret + + +def _rename(cfg: CN, old: str, new: str) -> None: + old_keys = old.split(".") + new_keys = new.split(".") + + def _set(key_seq: List[str], val: str) -> None: + cur = cfg + for k in key_seq[:-1]: + if k not in cur: + cur[k] = CN() + cur = cur[k] + cur[key_seq[-1]] = val + + def _get(key_seq: List[str]) -> CN: + cur = cfg + for k in key_seq: + cur = cur[k] + return cur + + def _del(key_seq: List[str]) -> None: + cur = cfg + for k in key_seq[:-1]: + cur = cur[k] + del cur[key_seq[-1]] + if len(cur) == 0 and len(key_seq) > 1: + _del(key_seq[:-1]) + + _set(new_keys, _get(old_keys)) + _del(old_keys) + + +class _RenameConverter: + """ + A converter that handles simple rename. + """ + + RENAME: List[Tuple[str, str]] = [] # list of tuples of (old name, new name) + + @classmethod + def upgrade(cls, cfg: CN) -> None: + for old, new in cls.RENAME: + _rename(cfg, old, new) + + @classmethod + def downgrade(cls, cfg: CN) -> None: + for old, new in cls.RENAME[::-1]: + _rename(cfg, new, old) + + +class ConverterV1(_RenameConverter): + RENAME = [("MODEL.RPN_HEAD.NAME", "MODEL.RPN.HEAD_NAME")] + + +class ConverterV2(_RenameConverter): + """ + A large bulk of rename, before public release. + """ + + RENAME = [ + ("MODEL.WEIGHT", "MODEL.WEIGHTS"), + ("MODEL.PANOPTIC_FPN.SEMANTIC_LOSS_SCALE", "MODEL.SEM_SEG_HEAD.LOSS_WEIGHT"), + ("MODEL.PANOPTIC_FPN.RPN_LOSS_SCALE", "MODEL.RPN.LOSS_WEIGHT"), + ("MODEL.PANOPTIC_FPN.INSTANCE_LOSS_SCALE", "MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT"), + ("MODEL.PANOPTIC_FPN.COMBINE_ON", "MODEL.PANOPTIC_FPN.COMBINE.ENABLED"), + ( + "MODEL.PANOPTIC_FPN.COMBINE_OVERLAP_THRESHOLD", + "MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH", + ), + ( + "MODEL.PANOPTIC_FPN.COMBINE_STUFF_AREA_LIMIT", + "MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT", + ), + ( + "MODEL.PANOPTIC_FPN.COMBINE_INSTANCES_CONFIDENCE_THRESHOLD", + "MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH", + ), + ("MODEL.ROI_HEADS.SCORE_THRESH", "MODEL.ROI_HEADS.SCORE_THRESH_TEST"), + ("MODEL.ROI_HEADS.NMS", "MODEL.ROI_HEADS.NMS_THRESH_TEST"), + ("MODEL.RETINANET.INFERENCE_SCORE_THRESHOLD", "MODEL.RETINANET.SCORE_THRESH_TEST"), + ("MODEL.RETINANET.INFERENCE_TOPK_CANDIDATES", "MODEL.RETINANET.TOPK_CANDIDATES_TEST"), + ("MODEL.RETINANET.INFERENCE_NMS_THRESHOLD", "MODEL.RETINANET.NMS_THRESH_TEST"), + ("TEST.DETECTIONS_PER_IMG", "TEST.DETECTIONS_PER_IMAGE"), + ("TEST.AUG_ON", "TEST.AUG.ENABLED"), + ("TEST.AUG_MIN_SIZES", "TEST.AUG.MIN_SIZES"), + ("TEST.AUG_MAX_SIZE", "TEST.AUG.MAX_SIZE"), + ("TEST.AUG_FLIP", "TEST.AUG.FLIP"), + ] + + @classmethod + def upgrade(cls, cfg: CN) -> None: + super().upgrade(cfg) + + if cfg.MODEL.META_ARCHITECTURE == "RetinaNet": + _rename( + cfg, "MODEL.RETINANET.ANCHOR_ASPECT_RATIOS", "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS" + ) + _rename(cfg, "MODEL.RETINANET.ANCHOR_SIZES", "MODEL.ANCHOR_GENERATOR.SIZES") + del cfg["MODEL"]["RPN"]["ANCHOR_SIZES"] + del cfg["MODEL"]["RPN"]["ANCHOR_ASPECT_RATIOS"] + else: + _rename(cfg, "MODEL.RPN.ANCHOR_ASPECT_RATIOS", "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS") + _rename(cfg, "MODEL.RPN.ANCHOR_SIZES", "MODEL.ANCHOR_GENERATOR.SIZES") + del cfg["MODEL"]["RETINANET"]["ANCHOR_SIZES"] + del cfg["MODEL"]["RETINANET"]["ANCHOR_ASPECT_RATIOS"] + del cfg["MODEL"]["RETINANET"]["ANCHOR_STRIDES"] + + @classmethod + def downgrade(cls, cfg: CN) -> None: + super().downgrade(cfg) + + _rename(cfg, "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS", "MODEL.RPN.ANCHOR_ASPECT_RATIOS") + _rename(cfg, "MODEL.ANCHOR_GENERATOR.SIZES", "MODEL.RPN.ANCHOR_SIZES") + cfg.MODEL.RETINANET.ANCHOR_ASPECT_RATIOS = cfg.MODEL.RPN.ANCHOR_ASPECT_RATIOS + cfg.MODEL.RETINANET.ANCHOR_SIZES = cfg.MODEL.RPN.ANCHOR_SIZES + cfg.MODEL.RETINANET.ANCHOR_STRIDES = [] # this is not used anywhere in any version diff --git a/detectron2/config/config.py b/detectron2/config/config.py new file mode 100644 index 0000000000000000000000000000000000000000..5c574b4805c7e0e0a0d0aeb9ca49ca51a2f18c44 --- /dev/null +++ b/detectron2/config/config.py @@ -0,0 +1,249 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import functools +import inspect +import logging +from fvcore.common.config import CfgNode as _CfgNode + +from detectron2.utils.file_io import PathManager + + +class CfgNode(_CfgNode): + """ + The same as `fvcore.common.config.CfgNode`, but different in: + + 1. Use unsafe yaml loading by default. + Note that this may lead to arbitrary code execution: you must not + load a config file from untrusted sources before manually inspecting + the content of the file. + 2. Support config versioning. + When attempting to merge an old config, it will convert the old config automatically. + """ + + @classmethod + def _open_cfg(cls, filename): + return PathManager.open(filename, "r") + + # Note that the default value of allow_unsafe is changed to True + def merge_from_file(self, cfg_filename: str, allow_unsafe: bool = True) -> None: + assert PathManager.isfile(cfg_filename), f"Config file '{cfg_filename}' does not exist!" + loaded_cfg = self.load_yaml_with_base(cfg_filename, allow_unsafe=allow_unsafe) + loaded_cfg = type(self)(loaded_cfg) + + # defaults.py needs to import CfgNode + from .defaults import _C + + latest_ver = _C.VERSION + assert ( + latest_ver == self.VERSION + ), "CfgNode.merge_from_file is only allowed on a config object of latest version!" + + logger = logging.getLogger(__name__) + + loaded_ver = loaded_cfg.get("VERSION", None) + if loaded_ver is None: + from .compat import guess_version + + loaded_ver = guess_version(loaded_cfg, cfg_filename) + assert loaded_ver <= self.VERSION, "Cannot merge a v{} config into a v{} config.".format( + loaded_ver, self.VERSION + ) + + if loaded_ver == self.VERSION: + self.merge_from_other_cfg(loaded_cfg) + else: + # compat.py needs to import CfgNode + from .compat import upgrade_config, downgrade_config + + logger.warning( + "Loading an old v{} config file '{}' by automatically upgrading to v{}. " + "See docs/CHANGELOG.md for instructions to update your files.".format( + loaded_ver, cfg_filename, self.VERSION + ) + ) + # To convert, first obtain a full config at an old version + old_self = downgrade_config(self, to_version=loaded_ver) + old_self.merge_from_other_cfg(loaded_cfg) + new_config = upgrade_config(old_self) + self.clear() + self.update(new_config) + + def dump(self, *args, **kwargs): + """ + Returns: + str: a yaml string representation of the config + """ + # to make it show up in docs + return super().dump(*args, **kwargs) + + +global_cfg = CfgNode() + + +def get_cfg() -> CfgNode: + """ + Get a copy of the default config. + + Returns: + a detectron2 CfgNode instance. + """ + from .defaults import _C + + return _C.clone() + + +def set_global_cfg(cfg: CfgNode) -> None: + """ + Let the global config point to the given cfg. + + Assume that the given "cfg" has the key "KEY", after calling + `set_global_cfg(cfg)`, the key can be accessed by: + :: + from detectron2.config import global_cfg + print(global_cfg.KEY) + + By using a hacky global config, you can access these configs anywhere, + without having to pass the config object or the values deep into the code. + This is a hacky feature introduced for quick prototyping / research exploration. + """ + global global_cfg + global_cfg.clear() + global_cfg.update(cfg) + + +def configurable(init_func=None, *, from_config=None): + """ + Decorate a function or a class's __init__ method so that it can be called + with a :class:`CfgNode` object using a :func:`from_config` function that translates + :class:`CfgNode` to arguments. + + Examples: + :: + # Usage 1: Decorator on __init__: + class A: + @configurable + def __init__(self, a, b=2, c=3): + pass + + @classmethod + def from_config(cls, cfg): # 'cfg' must be the first argument + # Returns kwargs to be passed to __init__ + return {"a": cfg.A, "b": cfg.B} + + a1 = A(a=1, b=2) # regular construction + a2 = A(cfg) # construct with a cfg + a3 = A(cfg, b=3, c=4) # construct with extra overwrite + + # Usage 2: Decorator on any function. Needs an extra from_config argument: + @configurable(from_config=lambda cfg: {"a: cfg.A, "b": cfg.B}) + def a_func(a, b=2, c=3): + pass + + a1 = a_func(a=1, b=2) # regular call + a2 = a_func(cfg) # call with a cfg + a3 = a_func(cfg, b=3, c=4) # call with extra overwrite + + Args: + init_func (callable): a class's ``__init__`` method in usage 1. The + class must have a ``from_config`` classmethod which takes `cfg` as + the first argument. + from_config (callable): the from_config function in usage 2. It must take `cfg` + as its first argument. + """ + + if init_func is not None: + assert ( + inspect.isfunction(init_func) + and from_config is None + and init_func.__name__ == "__init__" + ), "Incorrect use of @configurable. Check API documentation for examples." + + @functools.wraps(init_func) + def wrapped(self, *args, **kwargs): + try: + from_config_func = type(self).from_config + except AttributeError as e: + raise AttributeError( + "Class with @configurable must have a 'from_config' classmethod." + ) from e + if not inspect.ismethod(from_config_func): + raise TypeError("Class with @configurable must have a 'from_config' classmethod.") + + if _called_with_cfg(*args, **kwargs): + explicit_args = _get_args_from_config(from_config_func, *args, **kwargs) + init_func(self, **explicit_args) + else: + init_func(self, *args, **kwargs) + + return wrapped + + else: + if from_config is None: + return configurable # @configurable() is made equivalent to @configurable + assert inspect.isfunction( + from_config + ), "from_config argument of configurable must be a function!" + + def wrapper(orig_func): + @functools.wraps(orig_func) + def wrapped(*args, **kwargs): + if _called_with_cfg(*args, **kwargs): + explicit_args = _get_args_from_config(from_config, *args, **kwargs) + return orig_func(**explicit_args) + else: + return orig_func(*args, **kwargs) + + return wrapped + + return wrapper + + +def _get_args_from_config(from_config_func, *args, **kwargs): + """ + Use `from_config` to obtain explicit arguments. + + Returns: + dict: arguments to be used for cls.__init__ + """ + signature = inspect.signature(from_config_func) + if list(signature.parameters.keys())[0] != "cfg": + if inspect.isfunction(from_config_func): + name = from_config_func.__name__ + else: + name = f"{from_config_func.__self__}.from_config" + raise TypeError(f"{name} must take 'cfg' as the first argument!") + support_var_arg = any( + param.kind in [param.VAR_POSITIONAL, param.VAR_KEYWORD] + for param in signature.parameters.values() + ) + if support_var_arg: # forward all arguments to from_config, if from_config accepts them + ret = from_config_func(*args, **kwargs) + else: + # forward supported arguments to from_config + supported_arg_names = set(signature.parameters.keys()) + extra_kwargs = {} + for name in list(kwargs.keys()): + if name not in supported_arg_names: + extra_kwargs[name] = kwargs.pop(name) + ret = from_config_func(*args, **kwargs) + # forward the other arguments to __init__ + ret.update(extra_kwargs) + return ret + + +def _called_with_cfg(*args, **kwargs): + """ + Returns: + bool: whether the arguments contain CfgNode and should be considered + forwarded to from_config. + """ + from omegaconf import DictConfig + + if len(args) and isinstance(args[0], (_CfgNode, DictConfig)): + return True + if isinstance(kwargs.pop("cfg", None), (_CfgNode, DictConfig)): + return True + # `from_config`'s first argument is forced to be "cfg". + # So the above check covers all cases. + return False diff --git a/detectron2/config/defaults.py b/detectron2/config/defaults.py new file mode 100644 index 0000000000000000000000000000000000000000..efca58f222848ea46368a3bee6e5199e7686ce89 --- /dev/null +++ b/detectron2/config/defaults.py @@ -0,0 +1,786 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .config import CfgNode as CN + +# ----------------------------------------------------------------------------- +# Convention about Training / Test specific parameters +# ----------------------------------------------------------------------------- +# Whenever an argument can be either used for training or for testing, the +# corresponding name will be post-fixed by a _TRAIN for a training parameter, +# or _TEST for a test-specific parameter. +# For example, the number of images during training will be +# IMAGES_PER_BATCH_TRAIN, while the number of images for testing will be +# IMAGES_PER_BATCH_TEST + +# ----------------------------------------------------------------------------- +# Config definition +# ----------------------------------------------------------------------------- + +_C = CN() + +# The version number, to upgrade from old configs to new ones if any +# changes happen. It's recommended to keep a VERSION in your config file. +_C.VERSION = 2 + +_C.MODEL = CN() +_C.MODEL.LOAD_PROPOSALS = False +_C.MODEL.MASK_ON = False +_C.MODEL.KEYPOINT_ON = False +_C.MODEL.DEVICE = "cuda" +_C.MODEL.META_ARCHITECTURE = "GeneralizedRCNN" + +# Path (a file path, or URL like detectron2://.., https://..) to a checkpoint file +# to be loaded to the model. You can find available models in the model zoo. +_C.MODEL.WEIGHTS = "" + +# Values to be used for image normalization (BGR order, since INPUT.FORMAT defaults to BGR). +# To train on images of different number of channels, just set different mean & std. +# Default values are the mean pixel value from ImageNet: [103.53, 116.28, 123.675] +_C.MODEL.PIXEL_MEAN = [103.530, 116.280, 123.675] +# When using pre-trained models in Detectron1 or any MSRA models, +# std has been absorbed into its conv1 weights, so the std needs to be set 1. +# Otherwise, you can use [57.375, 57.120, 58.395] (ImageNet std) +_C.MODEL.PIXEL_STD = [1.0, 1.0, 1.0] + + +# ----------------------------------------------------------------------------- +# INPUT +# ----------------------------------------------------------------------------- +_C.INPUT = CN() +# Size of the smallest side of the image during training +_C.INPUT.MIN_SIZE_TRAIN = (800,) +# Sample size of smallest side by choice or random selection from range give by +# INPUT.MIN_SIZE_TRAIN +_C.INPUT.MIN_SIZE_TRAIN_SAMPLING = "choice" +# Maximum size of the side of the image during training +_C.INPUT.MAX_SIZE_TRAIN = 1333 +# Size of the smallest side of the image during testing. Set to zero to disable resize in testing. +_C.INPUT.MIN_SIZE_TEST = 800 +# Maximum size of the side of the image during testing +_C.INPUT.MAX_SIZE_TEST = 1333 +# Mode for flipping images used in data augmentation during training +# choose one of ["horizontal, "vertical", "none"] +_C.INPUT.RANDOM_FLIP = "horizontal" + +# `True` if cropping is used for data augmentation during training +_C.INPUT.CROP = CN({"ENABLED": False}) +# Cropping type. See documentation of `detectron2.data.transforms.RandomCrop` for explanation. +_C.INPUT.CROP.TYPE = "relative_range" +# Size of crop in range (0, 1] if CROP.TYPE is "relative" or "relative_range" and in number of +# pixels if CROP.TYPE is "absolute" +_C.INPUT.CROP.SIZE = [0.9, 0.9] + + +# Whether the model needs RGB, YUV, HSV etc. +# Should be one of the modes defined here, as we use PIL to read the image: +# https://pillow.readthedocs.io/en/stable/handbook/concepts.html#concept-modes +# with BGR being the one exception. One can set image format to BGR, we will +# internally use RGB for conversion and flip the channels over +_C.INPUT.FORMAT = "BGR" +# The ground truth mask format that the model will use. +# Mask R-CNN supports either "polygon" or "bitmask" as ground truth. +_C.INPUT.MASK_FORMAT = "polygon" # alternative: "bitmask" + +################### Text Tokenizer from MSR-CLIP ################## +_C.INPUT.TEXT_TOKENIZER = "openai_bpe" # "bert-base-cased" + +################## Data Augmentation from MSR-CLIP ################## +_C.AUG = CN() +_C.AUG.SCALE = (0.08, 1.0) +_C.AUG.RATIO = (3.0/4.0, 4.0/3.0) +_C.AUG.COLOR_JITTER = [0.4, 0.4, 0.4, 0.1, 0.0] +_C.AUG.GRAY_SCALE = 0.0 +_C.AUG.GAUSSIAN_BLUR = 0.0 +_C.AUG.DROPBLOCK_LAYERS = [3, 4] +_C.AUG.DROPBLOCK_KEEP_PROB = 1.0 +_C.AUG.DROPBLOCK_BLOCK_SIZE = 7 +_C.AUG.MIXUP_PROB = 0.0 +_C.AUG.MIXUP = 0.0 +_C.AUG.MIXCUT = 0.0 +_C.AUG.MIXCUT_MINMAX = [] +_C.AUG.MIXUP_SWITCH_PROB = 0.5 +_C.AUG.MIXUP_MODE = 'batch' +_C.AUG.MIXCUT_AND_MIXUP = False +_C.AUG.INTERPOLATION = 3 +_C.AUG.USE_TIMM = False +_C.AUG.TIMM_AUG = CN(new_allowed=True) +_C.AUG.TIMM_AUG.USE_LOADER = False +_C.AUG.TIMM_AUG.USE_TRANSFORM = False + +_C.AUG.TRAIN = CN() +_C.AUG.TRAIN.IMAGE_SIZE = [224, 224] # width * height, ex: 192 * 256 +_C.AUG.TRAIN.MAX_SIZE = None # the maximum size for longer edge after resizing +_C.AUG.TEST = CN() +_C.AUG.TEST.IMAGE_SIZE = [224, 224] # width * height, ex: 192 * 256 +_C.AUG.TEST.MAX_SIZE = None # the maximum size for longer edge after resizing +_C.AUG.TEST.CENTER_CROP = False +_C.AUG.TEST.INTERPOLATION = 3 + + +# ----------------------------------------------------------------------------- +# Dataset +# ----------------------------------------------------------------------------- +_C.DATASETS = CN() +# List of the dataset names for training. Must be registered in DatasetCatalog +# Samples from these datasets will be merged and used as one dataset. +_C.DATASETS.TRAIN = () +# List of the pre-computed proposal files for training, which must be consistent +# with datasets listed in DATASETS.TRAIN. +_C.DATASETS.PROPOSAL_FILES_TRAIN = () +# Number of top scoring precomputed proposals to keep for training +_C.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TRAIN = 2000 +# List of the dataset names for testing. Must be registered in DatasetCatalog +_C.DATASETS.TEST = () +# List of the pre-computed proposal files for test, which must be consistent +# with datasets listed in DATASETS.TEST. +_C.DATASETS.PROPOSAL_FILES_TEST = () +# Number of top scoring precomputed proposals to keep for test +_C.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TEST = 1000 +################## Data Loading from MSR-CLIP ################## +# List of dataset class names for training +_C.DATASETS.FACTORY_TRAIN = () +# List of dataset folder for training +_C.DATASETS.PATH_TRAIN = () +# List of the dataset names for auxilary training, as present in paths_catalog.py +_C.DATASETS.AUX = () +# List of dataset class names for auxilary training +_C.DATASETS.FACTORY_AUX = () +# List of dataset folder for auxilary training +_C.DATASETS.PATH_AUX = () +# List of dataset class names for testing +_C.DATASETS.FACTORY_TEST = () +# List of dataset folder for testing +_C.DATASETS.PATH_TEST = () +# Labelmap file to convert to tsv or for demo purpose +_C.DATASETS.LABELMAP_FILE = '' +_C.DATASETS.ATTR_LABELMAP_FILE = '' +_C.DATASETS.FILTERED_CLASSIFICATION_DATASETS = '' +# hierarchy file for test time score aggregation (developed on OpenImages) +_C.DATASETS.HIERARCHY_FILE = '' +# List of box extra fields for training/testing +# If given, will not infer from the other cfgs. +_C.DATASETS.BOX_EXTRA_FIELDS = () + +_C.DATASETS.NUM_CLASSES = 0 +_C.DATASETS.ROOT = '' +_C.DATASETS.TRAIN_SET = 'train' +_C.DATASETS.VAL_SET = '' +_C.DATASETS.TEST_SET = 'val' + +# The maximum total input sequence length after WordPiece tokenization +# Sequences longer than this will be truncated, and sequences shorter than this will be padded. +_C.DATASETS.MAX_SEQ_LENGTH = 35 + +# ----------------------------------------------------------------------------- +# DataLoader +# ----------------------------------------------------------------------------- +_C.DATALOADER = CN() +# Number of data loading threads +_C.DATALOADER.NUM_WORKERS = 4 +# If True, each batch should contain only images for which the aspect ratio +# is compatible. This groups portrait images together, and landscape images +# are not batched with portrait images. +_C.DATALOADER.ASPECT_RATIO_GROUPING = True +# Options: TrainingSampler, RepeatFactorTrainingSampler +_C.DATALOADER.SAMPLER_TRAIN = "TrainingSampler" +# Repeat threshold for RepeatFactorTrainingSampler +_C.DATALOADER.REPEAT_THRESHOLD = 0.0 +# Tf True, when working on datasets that have instance annotations, the +# training dataloader will filter out images without associated annotations +_C.DATALOADER.FILTER_EMPTY_ANNOTATIONS = True + +# ---------------------------------------------------------------------------- # +# CLIP options +# ---------------------------------------------------------------------------- # +_C.MODEL.CLIP = CN() + +_C.MODEL.CLIP.CROP_REGION_TYPE = "" # options: "GT", "RPN" +_C.MODEL.CLIP.BB_RPN_WEIGHTS = None # the weights of pretrained MaskRCNN +_C.MODEL.CLIP.IMS_PER_BATCH_TEST = 8 # the #images during inference per batch + +_C.MODEL.CLIP.USE_TEXT_EMB_CLASSIFIER = False # if True, use the CLIP text embedding as the classifier's weights +_C.MODEL.CLIP.TEXT_EMB_PATH = None # "/mnt/output_storage/trained_models/lvis_cls_emb/lvis_1203_cls_emb.pth" +_C.MODEL.CLIP.OFFLINE_RPN_CONFIG = None # option: all configs of pretrained RPN +_C.MODEL.CLIP.NO_BOX_DELTA = False # if True, during inference, no box delta will be applied to region proposals + +_C.MODEL.CLIP.BG_CLS_LOSS_WEIGHT = None # if not None, it is the loss weight for bg regions +_C.MODEL.CLIP.ONLY_SAMPLE_FG_PROPOSALS = False # if True, during training, ignore all bg proposals and only sample fg proposals +_C.MODEL.CLIP.MULTIPLY_RPN_SCORE = False # if True, during inference, multiply RPN scores with classification scores + +_C.MODEL.CLIP.OPENSET_TEST_NUM_CLASSES = None # if an integer, it is #all_cls in test +_C.MODEL.CLIP.OPENSET_TEST_TEXT_EMB_PATH = None # if not None, enables the openset/zero-shot training, the category embeddings during test + +_C.MODEL.CLIP.CLSS_TEMP = None # if None, dot product wo normalization & temperature; if float, normalization plus temperature +_C.MODEL.CLIP.RUN_CVPR_OVR = False # if True, train CVPR OVR model with their text embeddings +_C.MODEL.CLIP.FOCAL_SCALED_LOSS = None # if not None (float value for gamma), apply focal loss scaling idea to standard cross-entropy loss + +_C.MODEL.CLIP.OFFLINE_RPN_NMS_THRESH = None # the threshold of NMS in offline RPN +_C.MODEL.CLIP.PRETRAIN_IMG_TXT_LEVEL = True # if True, pretrain model using image-text level matching +_C.MODEL.CLIP.PRETRAIN_ONLY_EOT = False # if True, use end-of-token emb to match region features, in image-text level matching +_C.MODEL.CLIP.PRETRAIN_RPN_REGIONS = None # if not None, the number of RPN regions per image during pretraining +_C.MODEL.CLIP.PRETRAIN_SAMPLE_REGIONS = None # if not None, the number of regions per image during pretraining after sampling, to avoid overfitting +_C.MODEL.CLIP.GATHER_GPUS = False # if True, gather tensors across GPUS to increase batch size +_C.MODEL.CLIP.GRID_REGIONS = False # if True, use grid boxes to extract grid features, instead of object proposals +_C.MODEL.CLIP.CONCEPT_POOL_EMB = None # if not None, it provides the file path of embs of concept pool and thus enables region-concept matching +_C.MODEL.CLIP.CONCEPT_THRES = None # if not None, the threshold to filter out the regions with low matching score with concept embs, dependent on temp (default: 0.01) + +_C.MODEL.CLIP.OFFLINE_RPN_LSJ_PRETRAINED = False # if True, use large-scale jittering (LSJ) pretrained RPN +_C.MODEL.CLIP.TEACHER_RESNETS_DEPTH = 50 # the type of visual encoder of teacher model, sucha as ResNet 50, 101, 200 (a flag for 50x4) +_C.MODEL.CLIP.TEACHER_CONCEPT_POOL_EMB = None # if not None, it uses the same concept embedding as student model; otherwise, uses a seperate embedding of teacher model +_C.MODEL.CLIP.TEACHER_POOLER_RESOLUTION = 14 # RoIpooling resolution of teacher model + +_C.MODEL.CLIP.TEXT_EMB_DIM = 1024 # the dimension of precomputed class embeddings + +# ---------------------------------------------------------------------------- # +# Backbone options +# ---------------------------------------------------------------------------- # +_C.MODEL.BACKBONE = CN() + +_C.MODEL.BACKBONE.NAME = "build_resnet_backbone" +# Freeze the first several stages so they are not trained. +# There are 5 stages in ResNet. The first is a convolution, and the following +# stages are each group of residual blocks. +_C.MODEL.BACKBONE.FREEZE_AT = 2 + +_C.MODEL.TEXT_BACKBONE = CN() +_C.MODEL.TEXT_BACKBONE.NAME = "build_clip_swin_text_backbone" + + +# ---------------------------------------------------------------------------- # +# FPN options +# ---------------------------------------------------------------------------- # +_C.MODEL.FPN = CN() +# Names of the input feature maps to be used by FPN +# They must have contiguous power of 2 strides +# e.g., ["res2", "res3", "res4", "res5"] +_C.MODEL.FPN.IN_FEATURES = [] +_C.MODEL.FPN.OUT_CHANNELS = 256 + +# Options: "" (no norm), "GN" +_C.MODEL.FPN.NORM = "" + +# Types for fusing the FPN top-down and lateral features. Can be either "sum" or "avg" +_C.MODEL.FPN.FUSE_TYPE = "sum" + + +# ---------------------------------------------------------------------------- # +# Proposal generator options +# ---------------------------------------------------------------------------- # +_C.MODEL.PROPOSAL_GENERATOR = CN() +# Current proposal generators include "RPN", "RRPN" and "PrecomputedProposals" +_C.MODEL.PROPOSAL_GENERATOR.NAME = "RPN" +# Proposal height and width both need to be greater than MIN_SIZE +# (a the scale used during training or inference) +_C.MODEL.PROPOSAL_GENERATOR.MIN_SIZE = 0 + + +# ---------------------------------------------------------------------------- # +# Anchor generator options +# ---------------------------------------------------------------------------- # +_C.MODEL.ANCHOR_GENERATOR = CN() +# The generator can be any name in the ANCHOR_GENERATOR registry +_C.MODEL.ANCHOR_GENERATOR.NAME = "DefaultAnchorGenerator" +# Anchor sizes (i.e. sqrt of area) in absolute pixels w.r.t. the network input. +# Format: list[list[float]]. SIZES[i] specifies the list of sizes to use for +# IN_FEATURES[i]; len(SIZES) must be equal to len(IN_FEATURES) or 1. +# When len(SIZES) == 1, SIZES[0] is used for all IN_FEATURES. +_C.MODEL.ANCHOR_GENERATOR.SIZES = [[32, 64, 128, 256, 512]] +# Anchor aspect ratios. For each area given in `SIZES`, anchors with different aspect +# ratios are generated by an anchor generator. +# Format: list[list[float]]. ASPECT_RATIOS[i] specifies the list of aspect ratios (H/W) +# to use for IN_FEATURES[i]; len(ASPECT_RATIOS) == len(IN_FEATURES) must be true, +# or len(ASPECT_RATIOS) == 1 is true and aspect ratio list ASPECT_RATIOS[0] is used +# for all IN_FEATURES. +_C.MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS = [[0.5, 1.0, 2.0]] +# Anchor angles. +# list[list[float]], the angle in degrees, for each input feature map. +# ANGLES[i] specifies the list of angles for IN_FEATURES[i]. +_C.MODEL.ANCHOR_GENERATOR.ANGLES = [[-90, 0, 90]] +# Relative offset between the center of the first anchor and the top-left corner of the image +# Value has to be in [0, 1). Recommend to use 0.5, which means half stride. +# The value is not expected to affect model accuracy. +_C.MODEL.ANCHOR_GENERATOR.OFFSET = 0.0 + +# ---------------------------------------------------------------------------- # +# RPN options +# ---------------------------------------------------------------------------- # +_C.MODEL.RPN = CN() +_C.MODEL.RPN.HEAD_NAME = "StandardRPNHead" # used by RPN_HEAD_REGISTRY + +# Names of the input feature maps to be used by RPN +# e.g., ["p2", "p3", "p4", "p5", "p6"] for FPN +_C.MODEL.RPN.IN_FEATURES = ["res4"] +# Remove RPN anchors that go outside the image by BOUNDARY_THRESH pixels +# Set to -1 or a large value, e.g. 100000, to disable pruning anchors +_C.MODEL.RPN.BOUNDARY_THRESH = -1 +# IOU overlap ratios [BG_IOU_THRESHOLD, FG_IOU_THRESHOLD] +# Minimum overlap required between an anchor and ground-truth box for the +# (anchor, gt box) pair to be a positive example (IoU >= FG_IOU_THRESHOLD +# ==> positive RPN example: 1) +# Maximum overlap allowed between an anchor and ground-truth box for the +# (anchor, gt box) pair to be a negative examples (IoU < BG_IOU_THRESHOLD +# ==> negative RPN example: 0) +# Anchors with overlap in between (BG_IOU_THRESHOLD <= IoU < FG_IOU_THRESHOLD) +# are ignored (-1) +_C.MODEL.RPN.IOU_THRESHOLDS = [0.3, 0.7] +_C.MODEL.RPN.IOU_LABELS = [0, -1, 1] +# Number of regions per image used to train RPN +_C.MODEL.RPN.BATCH_SIZE_PER_IMAGE = 256 +# Target fraction of foreground (positive) examples per RPN minibatch +_C.MODEL.RPN.POSITIVE_FRACTION = 0.5 +# Options are: "smooth_l1", "giou" +_C.MODEL.RPN.BBOX_REG_LOSS_TYPE = "smooth_l1" +_C.MODEL.RPN.BBOX_REG_LOSS_WEIGHT = 1.0 +# Weights on (dx, dy, dw, dh) for normalizing RPN anchor regression targets +_C.MODEL.RPN.BBOX_REG_WEIGHTS = (1.0, 1.0, 1.0, 1.0) +# The transition point from L1 to L2 loss. Set to 0.0 to make the loss simply L1. +_C.MODEL.RPN.SMOOTH_L1_BETA = 0.0 +_C.MODEL.RPN.LOSS_WEIGHT = 1.0 +# Number of top scoring RPN proposals to keep before applying NMS +# When FPN is used, this is *per FPN level* (not total) +_C.MODEL.RPN.PRE_NMS_TOPK_TRAIN = 12000 +_C.MODEL.RPN.PRE_NMS_TOPK_TEST = 6000 +# Number of top scoring RPN proposals to keep after applying NMS +# When FPN is used, this limit is applied per level and then again to the union +# of proposals from all levels +# NOTE: When FPN is used, the meaning of this config is different from Detectron1. +# It means per-batch topk in Detectron1, but per-image topk here. +# See the "find_top_rpn_proposals" function for details. +_C.MODEL.RPN.POST_NMS_TOPK_TRAIN = 2000 +_C.MODEL.RPN.POST_NMS_TOPK_TEST = 1000 +# NMS threshold used on RPN proposals +_C.MODEL.RPN.NMS_THRESH = 0.7 +# Set this to -1 to use the same number of output channels as input channels. +_C.MODEL.RPN.CONV_DIMS = [-1] + +# ---------------------------------------------------------------------------- # +# ROI HEADS options +# ---------------------------------------------------------------------------- # +_C.MODEL.ROI_HEADS = CN() +_C.MODEL.ROI_HEADS.NAME = "Res5ROIHeads" +# Number of foreground classes +_C.MODEL.ROI_HEADS.NUM_CLASSES = 80 +# Names of the input feature maps to be used by ROI heads +# Currently all heads (box, mask, ...) use the same input feature map list +# e.g., ["p2", "p3", "p4", "p5"] is commonly used for FPN +_C.MODEL.ROI_HEADS.IN_FEATURES = ["res4"] +# IOU overlap ratios [IOU_THRESHOLD] +# Overlap threshold for an RoI to be considered background (if < IOU_THRESHOLD) +# Overlap threshold for an RoI to be considered foreground (if >= IOU_THRESHOLD) +_C.MODEL.ROI_HEADS.IOU_THRESHOLDS = [0.5] +_C.MODEL.ROI_HEADS.IOU_LABELS = [0, 1] +# RoI minibatch size *per image* (number of regions of interest [ROIs]) +# Total number of RoIs per training minibatch = +# ROI_HEADS.BATCH_SIZE_PER_IMAGE * SOLVER.IMS_PER_BATCH +# E.g., a common configuration is: 512 * 16 = 8192 +_C.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 512 +# Target fraction of RoI minibatch that is labeled foreground (i.e. class > 0) +_C.MODEL.ROI_HEADS.POSITIVE_FRACTION = 0.25 + +# Only used on test mode + +# Minimum score threshold (assuming scores in a [0, 1] range); a value chosen to +# balance obtaining high recall with not having too many low precision +# detections that will slow down inference post processing steps (like NMS) +# A default threshold of 0.0 increases AP by ~0.2-0.3 but significantly slows down +# inference. +_C.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.05 +# Overlap threshold used for non-maximum suppression (suppress boxes with +# IoU >= this threshold) +_C.MODEL.ROI_HEADS.NMS_THRESH_TEST = 0.5 +# If True, augment proposals with ground-truth boxes before sampling proposals to +# train ROI heads. +_C.MODEL.ROI_HEADS.PROPOSAL_APPEND_GT = True + +# Use soft NMS instead of standard NMS if set to True +_C.MODEL.ROI_HEADS.SOFT_NMS_ENABLED = False +# See soft NMS paper for definition of these options +_C.MODEL.ROI_HEADS.SOFT_NMS_METHOD = "gaussian" # "linear" +_C.MODEL.ROI_HEADS.SOFT_NMS_SIGMA = 0.5 +# For the linear_threshold we use NMS_THRESH_TEST +_C.MODEL.ROI_HEADS.SOFT_NMS_PRUNE = 0.001 + +# ---------------------------------------------------------------------------- # +# Box Head +# ---------------------------------------------------------------------------- # +_C.MODEL.ROI_BOX_HEAD = CN() +# C4 don't use head name option +# Options for non-C4 models: FastRCNNConvFCHead, +_C.MODEL.ROI_BOX_HEAD.NAME = "" +# Options are: "smooth_l1", "giou" +_C.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_TYPE = "smooth_l1" +# The final scaling coefficient on the box regression loss, used to balance the magnitude of its +# gradients with other losses in the model. See also `MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT`. +_C.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_WEIGHT = 1.0 +# Default weights on (dx, dy, dw, dh) for normalizing bbox regression targets +# These are empirically chosen to approximately lead to unit variance targets +_C.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS = (10.0, 10.0, 5.0, 5.0) +# The transition point from L1 to L2 loss. Set to 0.0 to make the loss simply L1. +_C.MODEL.ROI_BOX_HEAD.SMOOTH_L1_BETA = 0.0 +_C.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION = 14 +_C.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO = 0 +# Type of pooling operation applied to the incoming feature map for each RoI +_C.MODEL.ROI_BOX_HEAD.POOLER_TYPE = "ROIAlignV2" + +_C.MODEL.ROI_BOX_HEAD.NUM_FC = 0 +# Hidden layer dimension for FC layers in the RoI box head +_C.MODEL.ROI_BOX_HEAD.FC_DIM = 1024 +_C.MODEL.ROI_BOX_HEAD.NUM_CONV = 0 +# Channel dimension for Conv layers in the RoI box head +_C.MODEL.ROI_BOX_HEAD.CONV_DIM = 256 +# Normalization method for the convolution layers. +# Options: "" (no norm), "GN", "SyncBN". +_C.MODEL.ROI_BOX_HEAD.NORM = "" +# Whether to use class agnostic for bbox regression +_C.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG = False +# If true, RoI heads use bounding boxes predicted by the box head rather than proposal boxes. +_C.MODEL.ROI_BOX_HEAD.TRAIN_ON_PRED_BOXES = False + +# ---------------------------------------------------------------------------- # +# Cascaded Box Head +# ---------------------------------------------------------------------------- # +_C.MODEL.ROI_BOX_CASCADE_HEAD = CN() +# The number of cascade stages is implicitly defined by the length of the following two configs. +_C.MODEL.ROI_BOX_CASCADE_HEAD.BBOX_REG_WEIGHTS = ( + (10.0, 10.0, 5.0, 5.0), + (20.0, 20.0, 10.0, 10.0), + (30.0, 30.0, 15.0, 15.0), +) +_C.MODEL.ROI_BOX_CASCADE_HEAD.IOUS = (0.5, 0.6, 0.7) + + +# ---------------------------------------------------------------------------- # +# Mask Head +# ---------------------------------------------------------------------------- # +_C.MODEL.ROI_MASK_HEAD = CN() +_C.MODEL.ROI_MASK_HEAD.NAME = "MaskRCNNConvUpsampleHead" +_C.MODEL.ROI_MASK_HEAD.POOLER_RESOLUTION = 14 +_C.MODEL.ROI_MASK_HEAD.POOLER_SAMPLING_RATIO = 0 +_C.MODEL.ROI_MASK_HEAD.NUM_CONV = 0 # The number of convs in the mask head +_C.MODEL.ROI_MASK_HEAD.CONV_DIM = 256 +# Normalization method for the convolution layers. +# Options: "" (no norm), "GN", "SyncBN". +_C.MODEL.ROI_MASK_HEAD.NORM = "" +# Whether to use class agnostic for mask prediction +_C.MODEL.ROI_MASK_HEAD.CLS_AGNOSTIC_MASK = False +# Type of pooling operation applied to the incoming feature map for each RoI +_C.MODEL.ROI_MASK_HEAD.POOLER_TYPE = "ROIAlignV2" + + +# ---------------------------------------------------------------------------- # +# Keypoint Head +# ---------------------------------------------------------------------------- # +_C.MODEL.ROI_KEYPOINT_HEAD = CN() +_C.MODEL.ROI_KEYPOINT_HEAD.NAME = "KRCNNConvDeconvUpsampleHead" +_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_RESOLUTION = 14 +_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_SAMPLING_RATIO = 0 +_C.MODEL.ROI_KEYPOINT_HEAD.CONV_DIMS = tuple(512 for _ in range(8)) +_C.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS = 17 # 17 is the number of keypoints in COCO. + +# Images with too few (or no) keypoints are excluded from training. +_C.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE = 1 +# Normalize by the total number of visible keypoints in the minibatch if True. +# Otherwise, normalize by the total number of keypoints that could ever exist +# in the minibatch. +# The keypoint softmax loss is only calculated on visible keypoints. +# Since the number of visible keypoints can vary significantly between +# minibatches, this has the effect of up-weighting the importance of +# minibatches with few visible keypoints. (Imagine the extreme case of +# only one visible keypoint versus N: in the case of N, each one +# contributes 1/N to the gradient compared to the single keypoint +# determining the gradient direction). Instead, we can normalize the +# loss by the total number of keypoints, if it were the case that all +# keypoints were visible in a full minibatch. (Returning to the example, +# this means that the one visible keypoint contributes as much as each +# of the N keypoints.) +_C.MODEL.ROI_KEYPOINT_HEAD.NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS = True +# Multi-task loss weight to use for keypoints +# Recommended values: +# - use 1.0 if NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS is True +# - use 4.0 if NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS is False +_C.MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT = 1.0 +# Type of pooling operation applied to the incoming feature map for each RoI +_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_TYPE = "ROIAlignV2" + +# ---------------------------------------------------------------------------- # +# Semantic Segmentation Head +# ---------------------------------------------------------------------------- # +_C.MODEL.SEM_SEG_HEAD = CN() +_C.MODEL.SEM_SEG_HEAD.NAME = "SemSegFPNHead" +_C.MODEL.SEM_SEG_HEAD.IN_FEATURES = ["p2", "p3", "p4", "p5"] +# Label in the semantic segmentation ground truth that is ignored, i.e., no loss is calculated for +# the correposnding pixel. +_C.MODEL.SEM_SEG_HEAD.IGNORE_VALUE = 255 +# Number of classes in the semantic segmentation head +_C.MODEL.SEM_SEG_HEAD.NUM_CLASSES = 54 +# Number of channels in the 3x3 convs inside semantic-FPN heads. +_C.MODEL.SEM_SEG_HEAD.CONVS_DIM = 128 +# Outputs from semantic-FPN heads are up-scaled to the COMMON_STRIDE stride. +_C.MODEL.SEM_SEG_HEAD.COMMON_STRIDE = 4 +# Normalization method for the convolution layers. Options: "" (no norm), "GN". +_C.MODEL.SEM_SEG_HEAD.NORM = "GN" +_C.MODEL.SEM_SEG_HEAD.LOSS_WEIGHT = 1.0 + +_C.MODEL.PANOPTIC_FPN = CN() +# Scaling of all losses from instance detection / segmentation head. +_C.MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT = 1.0 + +# options when combining instance & semantic segmentation outputs +_C.MODEL.PANOPTIC_FPN.COMBINE = CN({"ENABLED": True}) # "COMBINE.ENABLED" is deprecated & not used +_C.MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH = 0.5 +_C.MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT = 4096 +_C.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = 0.5 + + +# ---------------------------------------------------------------------------- # +# RetinaNet Head +# ---------------------------------------------------------------------------- # +_C.MODEL.RETINANET = CN() + +# This is the number of foreground classes. +_C.MODEL.RETINANET.NUM_CLASSES = 80 + +_C.MODEL.RETINANET.IN_FEATURES = ["p3", "p4", "p5", "p6", "p7"] + +# Convolutions to use in the cls and bbox tower +# NOTE: this doesn't include the last conv for logits +_C.MODEL.RETINANET.NUM_CONVS = 4 + +# IoU overlap ratio [bg, fg] for labeling anchors. +# Anchors with < bg are labeled negative (0) +# Anchors with >= bg and < fg are ignored (-1) +# Anchors with >= fg are labeled positive (1) +_C.MODEL.RETINANET.IOU_THRESHOLDS = [0.4, 0.5] +_C.MODEL.RETINANET.IOU_LABELS = [0, -1, 1] + +# Prior prob for rare case (i.e. foreground) at the beginning of training. +# This is used to set the bias for the logits layer of the classifier subnet. +# This improves training stability in the case of heavy class imbalance. +_C.MODEL.RETINANET.PRIOR_PROB = 0.01 + +# Inference cls score threshold, only anchors with score > INFERENCE_TH are +# considered for inference (to improve speed) +_C.MODEL.RETINANET.SCORE_THRESH_TEST = 0.05 +# Select topk candidates before NMS +_C.MODEL.RETINANET.TOPK_CANDIDATES_TEST = 1000 +_C.MODEL.RETINANET.NMS_THRESH_TEST = 0.5 + +# Weights on (dx, dy, dw, dh) for normalizing Retinanet anchor regression targets +_C.MODEL.RETINANET.BBOX_REG_WEIGHTS = (1.0, 1.0, 1.0, 1.0) + +# Loss parameters +_C.MODEL.RETINANET.FOCAL_LOSS_GAMMA = 2.0 +_C.MODEL.RETINANET.FOCAL_LOSS_ALPHA = 0.25 +_C.MODEL.RETINANET.SMOOTH_L1_LOSS_BETA = 0.1 +# Options are: "smooth_l1", "giou" +_C.MODEL.RETINANET.BBOX_REG_LOSS_TYPE = "smooth_l1" + +# One of BN, SyncBN, FrozenBN, GN +# Only supports GN until unshared norm is implemented +_C.MODEL.RETINANET.NORM = "" + + +# ---------------------------------------------------------------------------- # +# ResNe[X]t options (ResNets = {ResNet, ResNeXt} +# Note that parts of a resnet may be used for both the backbone and the head +# These options apply to both +# ---------------------------------------------------------------------------- # +_C.MODEL.RESNETS = CN() + +_C.MODEL.RESNETS.DEPTH = 50 +_C.MODEL.RESNETS.OUT_FEATURES = ["res4"] # res4 for C4 backbone, res2..5 for FPN backbone + +# Number of groups to use; 1 ==> ResNet; > 1 ==> ResNeXt +_C.MODEL.RESNETS.NUM_GROUPS = 1 + +# Options: FrozenBN, GN, "SyncBN", "BN" +_C.MODEL.RESNETS.NORM = "FrozenBN" + +# Baseline width of each group. +# Scaling this parameters will scale the width of all bottleneck layers. +_C.MODEL.RESNETS.WIDTH_PER_GROUP = 64 + +# Place the stride 2 conv on the 1x1 filter +# Use True only for the original MSRA ResNet; use False for C2 and Torch models +_C.MODEL.RESNETS.STRIDE_IN_1X1 = True + +# Apply dilation in stage "res5" +_C.MODEL.RESNETS.RES5_DILATION = 1 + +# Output width of res2. Scaling this parameters will scale the width of all 1x1 convs in ResNet +# For R18 and R34, this needs to be set to 64 +_C.MODEL.RESNETS.RES2_OUT_CHANNELS = 256 +_C.MODEL.RESNETS.STEM_OUT_CHANNELS = 64 + +# Apply Deformable Convolution in stages +# Specify if apply deform_conv on Res2, Res3, Res4, Res5 +_C.MODEL.RESNETS.DEFORM_ON_PER_STAGE = [False, False, False, False] +# Use True to use modulated deform_conv (DeformableV2, https://arxiv.org/abs/1811.11168); +# Use False for DeformableV1. +_C.MODEL.RESNETS.DEFORM_MODULATED = False +# Number of groups in deformable conv. +_C.MODEL.RESNETS.DEFORM_NUM_GROUPS = 1 + + +# ---------------------------------------------------------------------------- # +# Swin options +# Note that parts of a resnet may be used for both the backbone and the head +# These options apply to both +# ---------------------------------------------------------------------------- # +_C.MODEL.SPEC = CN() +_C.MODEL.SPEC.EMBED_DIM = 512 + +_C.MODEL.SPEC.VISION = CN() +_C.MODEL.SPEC.VISION.PATCH_SIZE = 4 +_C.MODEL.SPEC.VISION.IN_CHANS = 3 +_C.MODEL.SPEC.VISION.EMBED_DIM = 96 +_C.MODEL.SPEC.VISION.DEPTHS = [2, 2, 6, 2] +_C.MODEL.SPEC.VISION.NUM_HEADS = [3, 6, 12, 24] +_C.MODEL.SPEC.VISION.WINDOW_SIZE = 7 +_C.MODEL.SPEC.VISION.MLP_RATIO = 4. +_C.MODEL.SPEC.VISION.DROP_RATE = .0 +_C.MODEL.SPEC.VISION.ATTN_DROP_RATE = .0 +_C.MODEL.SPEC.VISION.DROP_PATH_RATE = .0 +_C.MODEL.SPEC.VISION.QKV_BIAS = True +_C.MODEL.SPEC.VISION.QK_SCALE = False +_C.MODEL.SPEC.VISION.APE = False +_C.MODEL.SPEC.VISION.PATCH_NORM = True +_C.MODEL.SPEC.VISION.OUT_FEATURES = ["stage2", "stage3", "stage4", "stage5"] + +_C.MODEL.SPEC.TEXT = CN() +_C.MODEL.SPEC.TEXT.NAME = 'transformer' +_C.MODEL.SPEC.TEXT.LOAD_PRETRAINED = False +_C.MODEL.SPEC.TEXT.PRETRAINED = '' +_C.MODEL.SPEC.TEXT.TOKENIZER = 'clip' +_C.MODEL.SPEC.TEXT.CONTEXT_LENGTH = 77 +_C.MODEL.SPEC.TEXT.WIDTH = 512 +_C.MODEL.SPEC.TEXT.HEADS = 8 +_C.MODEL.SPEC.TEXT.LAYERS = 12 +_C.MODEL.SPEC.TEXT.AUTOGRESSIVE = True + +# ---------------------------------------------------------------------------- # +# Solver +# ---------------------------------------------------------------------------- # +_C.SOLVER = CN() + +# See detectron2/solver/build.py for LR scheduler options +_C.SOLVER.LR_SCHEDULER_NAME = "WarmupMultiStepLR" + +_C.SOLVER.MAX_ITER = 40000 + +_C.SOLVER.BASE_LR = 0.001 + +_C.SOLVER.MOMENTUM = 0.9 + +_C.SOLVER.NESTEROV = False + +_C.SOLVER.WEIGHT_DECAY = 0.0001 +# The weight decay that's applied to parameters of normalization layers +# (typically the affine transformation) +_C.SOLVER.WEIGHT_DECAY_NORM = 0.0 + +_C.SOLVER.GAMMA = 0.1 +# The iteration number to decrease learning rate by GAMMA. +_C.SOLVER.STEPS = (30000,) + +_C.SOLVER.WARMUP_FACTOR = 1.0 / 1000 +_C.SOLVER.WARMUP_ITERS = 1000 +_C.SOLVER.WARMUP_METHOD = "linear" + +# Save a checkpoint after every this number of iterations +_C.SOLVER.CHECKPOINT_PERIOD = 5000 + +# Number of images per batch across all machines. This is also the number +# of training images per step (i.e. per iteration). If we use 16 GPUs +# and IMS_PER_BATCH = 32, each GPU will see 2 images per batch. +# May be adjusted automatically if REFERENCE_WORLD_SIZE is set. +_C.SOLVER.IMS_PER_BATCH = 16 + +# The reference number of workers (GPUs) this config is meant to train with. +# It takes no effect when set to 0. +# With a non-zero value, it will be used by DefaultTrainer to compute a desired +# per-worker batch size, and then scale the other related configs (total batch size, +# learning rate, etc) to match the per-worker batch size. +# See documentation of `DefaultTrainer.auto_scale_workers` for details: +_C.SOLVER.REFERENCE_WORLD_SIZE = 0 + +# Detectron v1 (and previous detection code) used a 2x higher LR and 0 WD for +# biases. This is not useful (at least for recent models). You should avoid +# changing these and they exist only to reproduce Detectron v1 training if +# desired. +_C.SOLVER.BIAS_LR_FACTOR = 1.0 +_C.SOLVER.WEIGHT_DECAY_BIAS = _C.SOLVER.WEIGHT_DECAY + +# Gradient clipping +_C.SOLVER.CLIP_GRADIENTS = CN({"ENABLED": False}) +# Type of gradient clipping, currently 2 values are supported: +# - "value": the absolute values of elements of each gradients are clipped +# - "norm": the norm of the gradient for each parameter is clipped thus +# affecting all elements in the parameter +_C.SOLVER.CLIP_GRADIENTS.CLIP_TYPE = "value" +# Maximum absolute value used for clipping gradients +_C.SOLVER.CLIP_GRADIENTS.CLIP_VALUE = 1.0 +# Floating point number p for L-p norm to be used with the "norm" +# gradient clipping type; for L-inf, please specify .inf +_C.SOLVER.CLIP_GRADIENTS.NORM_TYPE = 2.0 + +# Enable automatic mixed precision for training +# Note that this does not change model's inference behavior. +# To use AMP in inference, run inference under autocast() +_C.SOLVER.AMP = CN({"ENABLED": False}) + +# ---------------------------------------------------------------------------- # +# Specific test options +# ---------------------------------------------------------------------------- # +_C.TEST = CN() +# For end-to-end tests to verify the expected accuracy. +# Each item is [task, metric, value, tolerance] +# e.g.: [['bbox', 'AP', 38.5, 0.2]] +_C.TEST.EXPECTED_RESULTS = [] +# The period (in terms of steps) to evaluate the model during training. +# Set to 0 to disable. +_C.TEST.EVAL_PERIOD = 0 +# The sigmas used to calculate keypoint OKS. See http://cocodataset.org/#keypoints-eval +# When empty, it will use the defaults in COCO. +# Otherwise it should be a list[float] with the same length as ROI_KEYPOINT_HEAD.NUM_KEYPOINTS. +_C.TEST.KEYPOINT_OKS_SIGMAS = [] +# Maximum number of detections to return per image during inference (100 is +# based on the limit established for the COCO dataset). +_C.TEST.DETECTIONS_PER_IMAGE = 100 + +_C.TEST.AUG = CN({"ENABLED": False}) +_C.TEST.AUG.MIN_SIZES = (400, 500, 600, 700, 800, 900, 1000, 1100, 1200) +_C.TEST.AUG.MAX_SIZE = 4000 +_C.TEST.AUG.FLIP = True + +_C.TEST.PRECISE_BN = CN({"ENABLED": False}) +_C.TEST.PRECISE_BN.NUM_ITER = 200 + +# ---------------------------------------------------------------------------- # +# Misc options +# ---------------------------------------------------------------------------- # +# Directory where output files are written +_C.OUTPUT_DIR = "./output" +# Set seed to negative to fully randomize everything. +# Set seed to positive to use a fixed seed. Note that a fixed seed increases +# reproducibility but does not guarantee fully deterministic behavior. +# Disabling all parallelism further increases reproducibility. +_C.SEED = -1 +# Benchmark different cudnn algorithms. +# If input images have very different sizes, this option will have large overhead +# for about 10k iterations. It usually hurts total time, but can benefit for certain models. +# If input images have the same or similar sizes, benchmark is often helpful. +_C.CUDNN_BENCHMARK = False +# The period (in terms of steps) for minibatch visualization at train time. +# Set to 0 to disable. +_C.VIS_PERIOD = 0 + +# global config is for quick hack purposes. +# You can set them in command line or config files, +# and access it with: +# +# from detectron2.config import global_cfg +# print(global_cfg.HACK) +# +# Do not commit any configs into it. +_C.GLOBAL = CN() +_C.GLOBAL.HACK = 1.0 diff --git a/detectron2/config/instantiate.py b/detectron2/config/instantiate.py new file mode 100644 index 0000000000000000000000000000000000000000..cbb32e19ea518eee84941b20f58d1054e84d1937 --- /dev/null +++ b/detectron2/config/instantiate.py @@ -0,0 +1,82 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import dataclasses +import logging +from collections import abc +from typing import Any + +from detectron2.utils.registry import _convert_target_to_string, locate + +__all__ = ["dump_dataclass", "instantiate"] + + +def dump_dataclass(obj: Any): + """ + Dump a dataclass recursively into a dict that can be later instantiated. + + Args: + obj: a dataclass object + + Returns: + dict + """ + assert dataclasses.is_dataclass(obj) and not isinstance( + obj, type + ), "dump_dataclass() requires an instance of a dataclass." + ret = {"_target_": _convert_target_to_string(type(obj))} + for f in dataclasses.fields(obj): + v = getattr(obj, f.name) + if dataclasses.is_dataclass(v): + v = dump_dataclass(v) + if isinstance(v, (list, tuple)): + v = [dump_dataclass(x) if dataclasses.is_dataclass(x) else x for x in v] + ret[f.name] = v + return ret + + +def instantiate(cfg): + """ + Recursively instantiate objects defined in dictionaries by + "_target_" and arguments. + + Args: + cfg: a dict-like object with "_target_" that defines the caller, and + other keys that define the arguments + + Returns: + object instantiated by cfg + """ + from omegaconf import ListConfig + + if isinstance(cfg, ListConfig): + lst = [instantiate(x) for x in cfg] + return ListConfig(lst, flags={"allow_objects": True}) + if isinstance(cfg, list): + # Specialize for list, because many classes take + # list[objects] as arguments, such as ResNet, DatasetMapper + return [instantiate(x) for x in cfg] + + if isinstance(cfg, abc.Mapping) and "_target_" in cfg: + # conceptually equivalent to hydra.utils.instantiate(cfg) with _convert_=all, + # but faster: https://github.com/facebookresearch/hydra/issues/1200 + cfg = {k: instantiate(v) for k, v in cfg.items()} + cls = cfg.pop("_target_") + cls = instantiate(cls) + + if isinstance(cls, str): + cls_name = cls + cls = locate(cls_name) + assert cls is not None, cls_name + else: + try: + cls_name = cls.__module__ + "." + cls.__qualname__ + except Exception: + # target could be anything, so the above could fail + cls_name = str(cls) + assert callable(cls), f"_target_ {cls} does not define a callable object" + try: + return cls(**cfg) + except TypeError: + logger = logging.getLogger(__name__) + logger.error(f"Error when instantiating {cls_name}!") + raise + return cfg # return as-is if don't know what to do diff --git a/detectron2/config/lazy.py b/detectron2/config/lazy.py new file mode 100644 index 0000000000000000000000000000000000000000..fc2156c4899ebdd4631489e663742a29ec481b4d --- /dev/null +++ b/detectron2/config/lazy.py @@ -0,0 +1,370 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import ast +import builtins +import importlib +import inspect +import logging +import os +import uuid +from collections import abc +from contextlib import contextmanager +from copy import deepcopy +from typing import List, Tuple, Union +import cloudpickle +import yaml +from omegaconf import DictConfig, ListConfig, OmegaConf + +from detectron2.utils.file_io import PathManager +from detectron2.utils.registry import _convert_target_to_string + +__all__ = ["LazyCall", "LazyConfig"] + + +class LazyCall: + """ + Wrap a callable so that when it's called, the call will not be executed, + but returns a dict that describes the call. + + LazyCall object has to be called with only keyword arguments. Positional + arguments are not yet supported. + + Examples: + :: + from detectron2.config import instantiate, LazyCall + + layer_cfg = LazyCall(nn.Conv2d)(in_channels=32, out_channels=32) + layer_cfg.out_channels = 64 # can edit it afterwards + layer = instantiate(layer_cfg) + """ + + def __init__(self, target): + if not (callable(target) or isinstance(target, (str, abc.Mapping))): + raise TypeError( + "target of LazyCall must be a callable or defines a callable! Got {target}" + ) + self._target = target + + def __call__(self, **kwargs): + kwargs["_target_"] = self._target + return DictConfig(content=kwargs, flags={"allow_objects": True}) + + +def _visit_dict_config(cfg, func): + """ + Apply func recursively to all DictConfig in cfg. + """ + if isinstance(cfg, DictConfig): + func(cfg) + for v in cfg.values(): + _visit_dict_config(v, func) + elif isinstance(cfg, ListConfig): + for v in cfg: + _visit_dict_config(v, func) + + +def _validate_py_syntax(filename): + # see also https://github.com/open-mmlab/mmcv/blob/master/mmcv/utils/config.py + with PathManager.open(filename, "r") as f: + content = f.read() + try: + ast.parse(content) + except SyntaxError as e: + raise SyntaxError(f"Config file {filename} has syntax error!") from e + + +def _cast_to_config(obj): + # if given a dict, return DictConfig instead + if isinstance(obj, dict): + return DictConfig(obj, flags={"allow_objects": True}) + return obj + + +_CFG_PACKAGE_NAME = "detectron2._cfg_loader" +""" +A namespace to put all imported config into. +""" + + +def _random_package_name(filename): + # generate a random package name when loading config files + return _CFG_PACKAGE_NAME + str(uuid.uuid4())[:4] + "." + os.path.basename(filename) + + +@contextmanager +def _patch_import(): + """ + Enhance relative import statements in config files, so that they: + 1. locate files purely based on relative location, regardless of packages. + e.g. you can import file without having __init__ + 2. do not cache modules globally; modifications of module states has no side effect + 3. support other storage system through PathManager + 4. imported dict are turned into omegaconf.DictConfig automatically + """ + old_import = builtins.__import__ + + def find_relative_file(original_file, relative_import_path, level): + cur_file = os.path.dirname(original_file) + for _ in range(level - 1): + cur_file = os.path.dirname(cur_file) + cur_name = relative_import_path.lstrip(".") + for part in cur_name.split("."): + cur_file = os.path.join(cur_file, part) + # NOTE: directory import is not handled. Because then it's unclear + # if such import should produce python module or DictConfig. This can + # be discussed further if needed. + if not cur_file.endswith(".py"): + cur_file += ".py" + if not PathManager.isfile(cur_file): + raise ImportError( + f"Cannot import name {relative_import_path} from " + f"{original_file}: {cur_file} has to exist." + ) + return cur_file + + def new_import(name, globals=None, locals=None, fromlist=(), level=0): + if ( + # Only deal with relative imports inside config files + level != 0 + and globals is not None + and (globals.get("__package__", "") or "").startswith(_CFG_PACKAGE_NAME) + ): + cur_file = find_relative_file(globals["__file__"], name, level) + _validate_py_syntax(cur_file) + spec = importlib.machinery.ModuleSpec( + _random_package_name(cur_file), None, origin=cur_file + ) + module = importlib.util.module_from_spec(spec) + module.__file__ = cur_file + with PathManager.open(cur_file) as f: + content = f.read() + exec(compile(content, cur_file, "exec"), module.__dict__) + for name in fromlist: # turn imported dict into DictConfig automatically + val = _cast_to_config(module.__dict__[name]) + module.__dict__[name] = val + return module + return old_import(name, globals, locals, fromlist=fromlist, level=level) + + builtins.__import__ = new_import + yield new_import + builtins.__import__ = old_import + + +class LazyConfig: + """ + Provid methods to save, load, and overrides an omegaconf config object + which may contain definition of lazily-constructed objects. + """ + + @staticmethod + def load_rel(filename: str, keys: Union[None, str, Tuple[str, ...]] = None): + """ + Similar to :meth:`load()`, but load path relative to the caller's + source file. + + This has the same functionality as a relative import, except that this method + accepts filename as a string, so more characters are allowed in the filename. + """ + caller_frame = inspect.stack()[1] + caller_fname = caller_frame[0].f_code.co_filename + assert caller_fname != "", "load_rel Unable to find caller" + caller_dir = os.path.dirname(caller_fname) + filename = os.path.join(caller_dir, filename) + return LazyConfig.load(filename, keys) + + @staticmethod + def load(filename: str, keys: Union[None, str, Tuple[str, ...]] = None): + """ + Load a config file. + + Args: + filename: absolute path or relative path w.r.t. the current working directory + keys: keys to load and return. If not given, return all keys + (whose values are config objects) in a dict. + """ + has_keys = keys is not None + filename = filename.replace("/./", "/") # redundant + if os.path.splitext(filename)[1] not in [".py", ".yaml", ".yml"]: + raise ValueError(f"Config file {filename} has to be a python or yaml file.") + if filename.endswith(".py"): + _validate_py_syntax(filename) + + with _patch_import(): + # Record the filename + module_namespace = { + "__file__": filename, + "__package__": _random_package_name(filename), + } + with PathManager.open(filename) as f: + content = f.read() + # Compile first with filename to: + # 1. make filename appears in stacktrace + # 2. make load_rel able to find its parent's (possibly remote) location + exec(compile(content, filename, "exec"), module_namespace) + + ret = module_namespace + else: + with PathManager.open(filename) as f: + obj = yaml.unsafe_load(f) + ret = OmegaConf.create(obj, flags={"allow_objects": True}) + + if has_keys: + if isinstance(keys, str): + return _cast_to_config(ret[keys]) + else: + return tuple(_cast_to_config(ret[a]) for a in keys) + else: + if filename.endswith(".py"): + # when not specified, only load those that are config objects + ret = DictConfig( + { + name: _cast_to_config(value) + for name, value in ret.items() + if isinstance(value, (DictConfig, ListConfig, dict)) + and not name.startswith("_") + }, + flags={"allow_objects": True}, + ) + return ret + + @staticmethod + def save(cfg, filename: str): + """ + Args: + cfg: an omegaconf config object + filename: yaml file name to save the config file + """ + logger = logging.getLogger(__name__) + try: + cfg = deepcopy(cfg) + except Exception: + pass + else: + # if it's deep-copyable, then... + def _replace_type_by_name(x): + if "_target_" in x and callable(x._target_): + try: + x._target_ = _convert_target_to_string(x._target_) + except AttributeError: + pass + + # not necessary, but makes yaml looks nicer + _visit_dict_config(cfg, _replace_type_by_name) + + try: + with PathManager.open(filename, "w") as f: + dict = OmegaConf.to_container(cfg, resolve=False) + dumped = yaml.dump(dict, default_flow_style=None, allow_unicode=True, width=9999) + f.write(dumped) + except Exception: + logger.exception("Unable to serialize the config to yaml. Error:") + new_filename = filename + ".pkl" + try: + # retry by pickle + with PathManager.open(new_filename, "wb") as f: + cloudpickle.dump(cfg, f) + logger.warning(f"Config saved using cloudpickle at {new_filename} ...") + except Exception: + pass + + @staticmethod + def apply_overrides(cfg, overrides: List[str]): + """ + In-place override contents of cfg. + + Args: + cfg: an omegaconf config object + overrides: list of strings in the format of "a=b" to override configs. + See https://hydra.cc/docs/next/advanced/override_grammar/basic/ + for syntax. + + Returns: + the cfg object + """ + + def safe_update(cfg, key, value): + parts = key.split(".") + for idx in range(1, len(parts)): + prefix = ".".join(parts[:idx]) + v = OmegaConf.select(cfg, prefix, default=None) + if v is None: + break + if not OmegaConf.is_config(v): + raise KeyError( + f"Trying to update key {key}, but {prefix} " + f"is not a config, but has type {type(v)}." + ) + OmegaConf.update(cfg, key, value, merge=True) + + from hydra.core.override_parser.overrides_parser import OverridesParser + + parser = OverridesParser.create() + overrides = parser.parse_overrides(overrides) + for o in overrides: + key = o.key_or_group + value = o.value() + if o.is_delete(): + # TODO support this + raise NotImplementedError("deletion is not yet a supported override") + safe_update(cfg, key, value) + return cfg + + @staticmethod + def to_py(cfg, prefix: str = "cfg."): + """ + Convert a config object into its equivalent Python code. + + Args: + cfg: an omegaconf config object + prefix: root name for the resulting code (default: "cfg.") + + + Returns: + str of formatted Python code + """ + import black + + cfg = OmegaConf.to_container(cfg, resolve=True) + + def _to_str(obj, prefix=None, inside_call=False): + if prefix is None: + prefix = [] + if isinstance(obj, abc.Mapping) and "_target_" in obj: + # Dict representing a function call + target = _convert_target_to_string(obj.pop("_target_")) + args = [] + for k, v in sorted(obj.items()): + args.append(f"{k}={_to_str(v, inside_call=True)}") + args = ", ".join(args) + call = f"{target}({args})" + return "".join(prefix) + call + elif isinstance(obj, abc.Mapping) and not inside_call: + # Dict that is not inside a call is a list of top-level config objects that we + # render as one object per line with dot separated prefixes + key_list = [] + for k, v in sorted(obj.items()): + if isinstance(v, abc.Mapping) and "_target_" not in v: + key_list.append(_to_str(v, prefix=prefix + [k + "."])) + else: + key = "".join(prefix) + k + key_list.append(f"{key}={_to_str(v)}") + return "\n".join(key_list) + elif isinstance(obj, abc.Mapping): + # Dict that is inside a call is rendered as a regular dict + return ( + "{" + + ",".join( + f"{repr(k)}: {_to_str(v, inside_call=inside_call)}" + for k, v in sorted(obj.items()) + ) + + "}" + ) + elif isinstance(obj, list): + return "[" + ",".join(_to_str(x, inside_call=inside_call) for x in obj) + "]" + else: + return repr(obj) + + py_str = _to_str(cfg, prefix=[prefix]) + try: + return black.format_str(py_str, mode=black.Mode()) + except black.InvalidInput: + return py_str diff --git a/detectron2/data/__init__.py b/detectron2/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..21c83f8cbd7a9388b452372f0444e78a54a33495 --- /dev/null +++ b/detectron2/data/__init__.py @@ -0,0 +1,19 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from . import transforms # isort:skip + +from .build import ( + build_batch_data_loader, + build_detection_test_loader, + build_detection_train_loader, + get_detection_dataset_dicts, + load_proposals_into_dataset, + print_instances_class_histogram, +) +from .catalog import DatasetCatalog, MetadataCatalog, Metadata +from .common import DatasetFromList, MapDataset +from .dataset_mapper import DatasetMapper + +# ensure the builtin datasets are registered +from . import datasets, samplers # isort:skip + +__all__ = [k for k in globals().keys() if not k.startswith("_")] diff --git a/detectron2/data/__pycache__/__init__.cpython-39.pyc b/detectron2/data/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..99f6a47c8993e5ad5648a228fc93cb014e105b79 Binary files /dev/null and b/detectron2/data/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/data/__pycache__/build.cpython-39.pyc b/detectron2/data/__pycache__/build.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b005b9297e76a89e5d27427c291a6823aeac6325 Binary files /dev/null and b/detectron2/data/__pycache__/build.cpython-39.pyc differ diff --git a/detectron2/data/__pycache__/catalog.cpython-39.pyc b/detectron2/data/__pycache__/catalog.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e7695bc870be2f84a52529aeb4b6d20faa574f09 Binary files /dev/null and b/detectron2/data/__pycache__/catalog.cpython-39.pyc differ diff --git a/detectron2/data/__pycache__/clip_build.cpython-39.pyc b/detectron2/data/__pycache__/clip_build.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f3bd633ebe8a6af9861eede650dae430d7a64abd Binary files /dev/null and b/detectron2/data/__pycache__/clip_build.cpython-39.pyc differ diff --git a/detectron2/data/__pycache__/common.cpython-39.pyc b/detectron2/data/__pycache__/common.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9c6a13e52d5d23434a55b81f73d17ca19ff4df97 Binary files /dev/null and b/detectron2/data/__pycache__/common.cpython-39.pyc differ diff --git a/detectron2/data/__pycache__/dataset_mapper.cpython-39.pyc b/detectron2/data/__pycache__/dataset_mapper.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..917e8028f3757afbf90af13796eaaaae09083b68 Binary files /dev/null and b/detectron2/data/__pycache__/dataset_mapper.cpython-39.pyc differ diff --git a/detectron2/data/__pycache__/detection_utils.cpython-39.pyc b/detectron2/data/__pycache__/detection_utils.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7068d4f69d86e2d884b89e83859d336ae0f38be4 Binary files /dev/null and b/detectron2/data/__pycache__/detection_utils.cpython-39.pyc differ diff --git a/detectron2/data/build.py b/detectron2/data/build.py new file mode 100644 index 0000000000000000000000000000000000000000..a1dcfadbd2cc30a0875c4d294e3cabcfa0146a16 --- /dev/null +++ b/detectron2/data/build.py @@ -0,0 +1,536 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import itertools +import logging +import numpy as np +import operator +import pickle +import torch.utils.data +from tabulate import tabulate +from termcolor import colored + +from detectron2.config import configurable +from detectron2.structures import BoxMode +from detectron2.utils.comm import get_world_size +from detectron2.utils.env import seed_all_rng +from detectron2.utils.file_io import PathManager +from detectron2.utils.logger import _log_api_usage, log_first_n + +from .catalog import DatasetCatalog, MetadataCatalog +from .common import AspectRatioGroupedDataset, DatasetFromList, MapDataset +from .dataset_mapper import DatasetMapper +from .detection_utils import check_metadata_consistency +from .samplers import InferenceSampler, RepeatFactorTrainingSampler, TrainingSampler + +from .clip_build import make_clip_dataset + +""" +This file contains the default logic to build a dataloader for training or testing. +""" + +__all__ = [ + "build_batch_data_loader", + "build_detection_train_loader", + "build_detection_test_loader", + "get_detection_dataset_dicts", + "load_proposals_into_dataset", + "print_instances_class_histogram", +] + + +def filter_images_with_only_crowd_annotations(dataset_dicts): + """ + Filter out images with none annotations or only crowd annotations + (i.e., images without non-crowd annotations). + A common training-time preprocessing on COCO dataset. + + Args: + dataset_dicts (list[dict]): annotations in Detectron2 Dataset format. + + Returns: + list[dict]: the same format, but filtered. + """ + num_before = len(dataset_dicts) + + def valid(anns): + for ann in anns: + if ann.get("iscrowd", 0) == 0: + return True + return False + + dataset_dicts = [x for x in dataset_dicts if valid(x["annotations"])] + num_after = len(dataset_dicts) + logger = logging.getLogger(__name__) + logger.info( + "Removed {} images with no usable annotations. {} images left.".format( + num_before - num_after, num_after + ) + ) + return dataset_dicts + + +def filter_images_with_few_keypoints(dataset_dicts, min_keypoints_per_image): + """ + Filter out images with too few number of keypoints. + + Args: + dataset_dicts (list[dict]): annotations in Detectron2 Dataset format. + + Returns: + list[dict]: the same format as dataset_dicts, but filtered. + """ + num_before = len(dataset_dicts) + + def visible_keypoints_in_image(dic): + # Each keypoints field has the format [x1, y1, v1, ...], where v is visibility + annotations = dic["annotations"] + return sum( + (np.array(ann["keypoints"][2::3]) > 0).sum() + for ann in annotations + if "keypoints" in ann + ) + + dataset_dicts = [ + x for x in dataset_dicts if visible_keypoints_in_image(x) >= min_keypoints_per_image + ] + num_after = len(dataset_dicts) + logger = logging.getLogger(__name__) + logger.info( + "Removed {} images with fewer than {} keypoints.".format( + num_before - num_after, min_keypoints_per_image + ) + ) + return dataset_dicts + + +def load_proposals_into_dataset(dataset_dicts, proposal_file): + """ + Load precomputed object proposals into the dataset. + + The proposal file should be a pickled dict with the following keys: + + - "ids": list[int] or list[str], the image ids + - "boxes": list[np.ndarray], each is an Nx4 array of boxes corresponding to the image id + - "objectness_logits": list[np.ndarray], each is an N sized array of objectness scores + corresponding to the boxes. + - "bbox_mode": the BoxMode of the boxes array. Defaults to ``BoxMode.XYXY_ABS``. + + Args: + dataset_dicts (list[dict]): annotations in Detectron2 Dataset format. + proposal_file (str): file path of pre-computed proposals, in pkl format. + + Returns: + list[dict]: the same format as dataset_dicts, but added proposal field. + """ + logger = logging.getLogger(__name__) + logger.info("Loading proposals from: {}".format(proposal_file)) + + with PathManager.open(proposal_file, "rb") as f: + proposals = pickle.load(f, encoding="latin1") + + # Rename the key names in D1 proposal files + rename_keys = {"indexes": "ids", "scores": "objectness_logits"} + for key in rename_keys: + if key in proposals: + proposals[rename_keys[key]] = proposals.pop(key) + + # Fetch the indexes of all proposals that are in the dataset + # Convert image_id to str since they could be int. + img_ids = set({str(record["image_id"]) for record in dataset_dicts}) + id_to_index = {str(id): i for i, id in enumerate(proposals["ids"]) if str(id) in img_ids} + + # Assuming default bbox_mode of precomputed proposals are 'XYXY_ABS' + bbox_mode = BoxMode(proposals["bbox_mode"]) if "bbox_mode" in proposals else BoxMode.XYXY_ABS + + for record in dataset_dicts: + # Get the index of the proposal + i = id_to_index[str(record["image_id"])] + + boxes = proposals["boxes"][i] + objectness_logits = proposals["objectness_logits"][i] + # Sort the proposals in descending order of the scores + inds = objectness_logits.argsort()[::-1] + record["proposal_boxes"] = boxes[inds] + record["proposal_objectness_logits"] = objectness_logits[inds] + record["proposal_bbox_mode"] = bbox_mode + + return dataset_dicts + + +def print_instances_class_histogram(dataset_dicts, class_names): + """ + Args: + dataset_dicts (list[dict]): list of dataset dicts. + class_names (list[str]): list of class names (zero-indexed). + """ + num_classes = len(class_names) + hist_bins = np.arange(num_classes + 1) + histogram = np.zeros((num_classes,), dtype=np.int) + for entry in dataset_dicts: + annos = entry["annotations"] + classes = np.asarray( + [x["category_id"] for x in annos if not x.get("iscrowd", 0)], dtype=np.int + ) + if len(classes): + assert classes.min() >= 0, f"Got an invalid category_id={classes.min()}" + assert ( + classes.max() < num_classes + ), f"Got an invalid category_id={classes.max()} for a dataset of {num_classes} classes" + histogram += np.histogram(classes, bins=hist_bins)[0] + + N_COLS = min(6, len(class_names) * 2) + + def short_name(x): + # make long class names shorter. useful for lvis + if len(x) > 13: + return x[:11] + ".." + return x + + data = list( + itertools.chain(*[[short_name(class_names[i]), int(v)] for i, v in enumerate(histogram)]) + ) + total_num_instances = sum(data[1::2]) + data.extend([None] * (N_COLS - (len(data) % N_COLS))) + if num_classes > 1: + data.extend(["total", total_num_instances]) + data = itertools.zip_longest(*[data[i::N_COLS] for i in range(N_COLS)]) + table = tabulate( + data, + headers=["category", "#instances"] * (N_COLS // 2), + tablefmt="pipe", + numalign="left", + stralign="center", + ) + log_first_n( + logging.INFO, + "Distribution of instances among all {} categories:\n".format(num_classes) + + colored(table, "cyan"), + key="message", + ) + + +def get_detection_dataset_dicts(names, filter_empty=True, min_keypoints=0, proposal_files=None): + """ + Load and prepare dataset dicts for instance detection/segmentation and semantic segmentation. + + Args: + names (str or list[str]): a dataset name or a list of dataset names + filter_empty (bool): whether to filter out images without instance annotations + min_keypoints (int): filter out images with fewer keypoints than + `min_keypoints`. Set to 0 to do nothing. + proposal_files (list[str]): if given, a list of object proposal files + that match each dataset in `names`. + + Returns: + list[dict]: a list of dicts following the standard dataset dict format. + """ + if isinstance(names, str): + names = [names] + assert len(names), names + dataset_dicts = [DatasetCatalog.get(dataset_name) for dataset_name in names] + for dataset_name, dicts in zip(names, dataset_dicts): + assert len(dicts), "Dataset '{}' is empty!".format(dataset_name) + + if proposal_files is not None: + assert len(names) == len(proposal_files) + # load precomputed proposals from proposal files + dataset_dicts = [ + load_proposals_into_dataset(dataset_i_dicts, proposal_file) + for dataset_i_dicts, proposal_file in zip(dataset_dicts, proposal_files) + ] + + dataset_dicts = list(itertools.chain.from_iterable(dataset_dicts)) + + has_instances = "annotations" in dataset_dicts[0] + if filter_empty and has_instances: + dataset_dicts = filter_images_with_only_crowd_annotations(dataset_dicts) + if min_keypoints > 0 and has_instances: + dataset_dicts = filter_images_with_few_keypoints(dataset_dicts, min_keypoints) + + if has_instances: + try: + class_names = MetadataCatalog.get(names[0]).thing_classes + check_metadata_consistency("thing_classes", names) + print_instances_class_histogram(dataset_dicts, class_names) + except AttributeError: # class names are not available for this dataset + pass + + assert len(dataset_dicts), "No valid data found in {}.".format(",".join(names)) + return dataset_dicts + + +def build_batch_data_loader( + dataset, sampler, total_batch_size, *, aspect_ratio_grouping=False, num_workers=0 +): + """ + Build a batched dataloader. The main differences from `torch.utils.data.DataLoader` are: + 1. support aspect ratio grouping options + 2. use no "batch collation", because this is common for detection training + + Args: + dataset (torch.utils.data.Dataset): map-style PyTorch dataset. Can be indexed. + sampler (torch.utils.data.sampler.Sampler): a sampler that produces indices + total_batch_size, aspect_ratio_grouping, num_workers): see + :func:`build_detection_train_loader`. + + Returns: + iterable[list]. Length of each list is the batch size of the current + GPU. Each element in the list comes from the dataset. + """ + world_size = get_world_size() + assert ( + total_batch_size > 0 and total_batch_size % world_size == 0 + ), "Total batch size ({}) must be divisible by the number of gpus ({}).".format( + total_batch_size, world_size + ) + + batch_size = total_batch_size // world_size + if aspect_ratio_grouping: + data_loader = torch.utils.data.DataLoader( + dataset, + sampler=sampler, + num_workers=num_workers, + batch_sampler=None, + collate_fn=operator.itemgetter(0), # don't batch, but yield individual elements + worker_init_fn=worker_init_reset_seed, + ) # yield individual mapped dict + return AspectRatioGroupedDataset(data_loader, batch_size) + else: + batch_sampler = torch.utils.data.sampler.BatchSampler( + sampler, batch_size, drop_last=True + ) # drop_last so the batch always have the same size + return torch.utils.data.DataLoader( + dataset, + num_workers=num_workers, + batch_sampler=batch_sampler, + collate_fn=trivial_batch_collator, + worker_init_fn=worker_init_reset_seed, + ) + + +def _train_loader_from_config(cfg, mapper=None, *, dataset=None, sampler=None): + if 'yfcc100m' in cfg.DATASETS.TRAIN: # dataset, transform/aug., sampler for image-text pairs training + logger = logging.getLogger(__name__) + logger.info("Creating dataset {}".format(cfg.DATASETS.TRAIN)) + datasets, precomputed_tokens, dataset_classes = make_clip_dataset( + cfg, is_train=True, + transforms=None, # for training, we use our own defined transforms + ) + dataset = datasets[0] # during training, a single (possibly concatenated) dataset was returned + if sampler is None: + sampler_name = cfg.DATALOADER.SAMPLER_TRAIN + logger = logging.getLogger(__name__) + logger.info("Using training sampler {}".format(sampler_name)) + if sampler_name == "TrainingSampler": + sampler = TrainingSampler(len(dataset)) + elif sampler_name == "RepeatFactorTrainingSampler": + repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency( + dataset, cfg.DATALOADER.REPEAT_THRESHOLD + ) + sampler = RepeatFactorTrainingSampler(repeat_factors) + else: + raise ValueError("Unknown training sampler: {}".format(sampler_name)) + return { + "dataset": dataset, + "sampler": sampler, + "mapper": None, + "total_batch_size": cfg.SOLVER.IMS_PER_BATCH, + "aspect_ratio_grouping": cfg.DATALOADER.ASPECT_RATIO_GROUPING, + "num_workers": cfg.DATALOADER.NUM_WORKERS, + } + # the following is the default code in Detectron2 + if dataset is None: + dataset = get_detection_dataset_dicts( + cfg.DATASETS.TRAIN, + filter_empty=cfg.DATALOADER.FILTER_EMPTY_ANNOTATIONS, + min_keypoints=cfg.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE + if cfg.MODEL.KEYPOINT_ON + else 0, + proposal_files=cfg.DATASETS.PROPOSAL_FILES_TRAIN if cfg.MODEL.LOAD_PROPOSALS else None, + ) + _log_api_usage("dataset." + cfg.DATASETS.TRAIN[0]) + + if mapper is None: + mapper = DatasetMapper(cfg, True) + + if sampler is None: + sampler_name = cfg.DATALOADER.SAMPLER_TRAIN + logger = logging.getLogger(__name__) + logger.info("Using training sampler {}".format(sampler_name)) + if sampler_name == "TrainingSampler": + sampler = TrainingSampler(len(dataset)) + elif sampler_name == "RepeatFactorTrainingSampler": + repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency( + dataset, cfg.DATALOADER.REPEAT_THRESHOLD + ) + sampler = RepeatFactorTrainingSampler(repeat_factors) + else: + raise ValueError("Unknown training sampler: {}".format(sampler_name)) + + return { + "dataset": dataset, + "sampler": sampler, + "mapper": mapper, + "total_batch_size": cfg.SOLVER.IMS_PER_BATCH, + "aspect_ratio_grouping": cfg.DATALOADER.ASPECT_RATIO_GROUPING, + "num_workers": cfg.DATALOADER.NUM_WORKERS, + } + + +# TODO can allow dataset as an iterable or IterableDataset to make this function more general +@configurable(from_config=_train_loader_from_config) +def build_detection_train_loader( + dataset, *, mapper, sampler=None, total_batch_size, aspect_ratio_grouping=True, num_workers=0 +): + """ + Build a dataloader for object detection with some default features. + This interface is experimental. + + Args: + dataset (list or torch.utils.data.Dataset): a list of dataset dicts, + or a map-style pytorch dataset. They can be obtained by using + :func:`DatasetCatalog.get` or :func:`get_detection_dataset_dicts`. + mapper (callable): a callable which takes a sample (dict) from dataset and + returns the format to be consumed by the model. + When using cfg, the default choice is ``DatasetMapper(cfg, is_train=True)``. + sampler (torch.utils.data.sampler.Sampler or None): a sampler that produces + indices to be applied on ``dataset``. Default to :class:`TrainingSampler`, + which coordinates an infinite random shuffle sequence across all workers. + total_batch_size (int): total batch size across all workers. Batching + simply puts data into a list. + aspect_ratio_grouping (bool): whether to group images with similar + aspect ratio for efficiency. When enabled, it requires each + element in dataset be a dict with keys "width" and "height". + num_workers (int): number of parallel data loading workers + + Returns: + torch.utils.data.DataLoader: + a dataloader. Each output from it is a ``list[mapped_element]`` of length + ``total_batch_size / num_workers``, where ``mapped_element`` is produced + by the ``mapper``. + """ + if isinstance(dataset, list): + dataset = DatasetFromList(dataset, copy=False) + if mapper is not None: + dataset = MapDataset(dataset, mapper) + if sampler is None: + sampler = TrainingSampler(len(dataset)) + assert isinstance(sampler, torch.utils.data.sampler.Sampler) + return build_batch_data_loader( + dataset, + sampler, + total_batch_size, + aspect_ratio_grouping=aspect_ratio_grouping, + num_workers=num_workers, + ) + + +def _test_loader_from_config(cfg, dataset_name, mapper=None): + """ + Uses the given `dataset_name` argument (instead of the names in cfg), because the + standard practice is to evaluate each test set individually (not combining them). + """ + if 'yfcc100m' in cfg.DATASETS.TEST: # dataset, no {transform/aug., sampler for image-text pairs training} + logger = logging.getLogger(__name__) + logger.info("Creating dataset {}".format(cfg.DATASETS.TEST)) + datasets, precomputed_tokens, dataset_classes = make_clip_dataset( + cfg, is_train=False, + transforms=None, # for training, we use our own defined transforms + ) + dataset = datasets[0] # during training, a single (possibly concatenated) dataset was returned + return { + "dataset": dataset, + "mapper": None, + "num_workers": cfg.DATALOADER.NUM_WORKERS, + } + + # the following is the default code in Detectron2 + dataset = get_detection_dataset_dicts( + [dataset_name], + filter_empty=False, + proposal_files=[ + cfg.DATASETS.PROPOSAL_FILES_TEST[list(cfg.DATASETS.TEST).index(dataset_name)] + ] + if cfg.MODEL.LOAD_PROPOSALS + else None, + ) + if mapper is None: + mapper = DatasetMapper(cfg, False) + if cfg.MODEL.META_ARCHITECTURE == 'CLIPRCNN': # speed up when using CLIP in inference + return {"dataset": dataset, "mapper": mapper, "num_workers": cfg.DATALOADER.NUM_WORKERS,\ + "clip_batch_size": cfg.MODEL.CLIP.IMS_PER_BATCH_TEST} + return {"dataset": dataset, "mapper": mapper, "num_workers": cfg.DATALOADER.NUM_WORKERS} + + +@configurable(from_config=_test_loader_from_config) +def build_detection_test_loader(dataset, *, mapper, sampler=None, num_workers=0, clip_batch_size=None): + """ + Similar to `build_detection_train_loader`, but uses a batch size of 1, + and :class:`InferenceSampler`. This sampler coordinates all workers to + produce the exact set of all samples. + This interface is experimental. + + Args: + dataset (list or torch.utils.data.Dataset): a list of dataset dicts, + or a map-style pytorch dataset. They can be obtained by using + :func:`DatasetCatalog.get` or :func:`get_detection_dataset_dicts`. + mapper (callable): a callable which takes a sample (dict) from dataset + and returns the format to be consumed by the model. + When using cfg, the default choice is ``DatasetMapper(cfg, is_train=False)``. + sampler (torch.utils.data.sampler.Sampler or None): a sampler that produces + indices to be applied on ``dataset``. Default to :class:`InferenceSampler`, + which splits the dataset across all workers. + num_workers (int): number of parallel data loading workers + + Returns: + DataLoader: a torch DataLoader, that loads the given detection + dataset, with test-time transformation and batching. + + Examples: + :: + data_loader = build_detection_test_loader( + DatasetRegistry.get("my_test"), + mapper=DatasetMapper(...)) + + # or, instantiate with a CfgNode: + data_loader = build_detection_test_loader(cfg, "my_test") + """ + if isinstance(dataset, list): + dataset = DatasetFromList(dataset, copy=False) + if mapper is not None: + dataset = MapDataset(dataset, mapper) + if sampler is None: + sampler = InferenceSampler(len(dataset)) + + if clip_batch_size: # multiple images per gpu + world_size = get_world_size() + batch_size = clip_batch_size // world_size + batch_sampler = torch.utils.data.sampler.BatchSampler(sampler, batch_size, drop_last=False) + data_loader = torch.utils.data.DataLoader( + dataset, + num_workers=num_workers, + batch_sampler=batch_sampler, + collate_fn=trivial_batch_collator, + ) + return data_loader + # Always use 1 image per worker during inference since this is the + # standard when reporting inference time in papers. + batch_sampler = torch.utils.data.sampler.BatchSampler(sampler, 1, drop_last=False) + data_loader = torch.utils.data.DataLoader( + dataset, + num_workers=num_workers, + batch_sampler=batch_sampler, + collate_fn=trivial_batch_collator, + ) + return data_loader + + +def trivial_batch_collator(batch): + """ + A batch collator that does nothing. + """ + return batch + + +def worker_init_reset_seed(worker_id): + initial_seed = torch.initial_seed() % 2 ** 31 + seed_all_rng(initial_seed + worker_id) diff --git a/detectron2/data/catalog.py b/detectron2/data/catalog.py new file mode 100644 index 0000000000000000000000000000000000000000..45c110c19508f23921b9033cdaf0aa8056f0c125 --- /dev/null +++ b/detectron2/data/catalog.py @@ -0,0 +1,236 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import logging +import types +from collections import UserDict +from typing import List + +from detectron2.utils.logger import log_first_n + +__all__ = ["DatasetCatalog", "MetadataCatalog", "Metadata"] + + +class _DatasetCatalog(UserDict): + """ + A global dictionary that stores information about the datasets and how to obtain them. + + It contains a mapping from strings + (which are names that identify a dataset, e.g. "coco_2014_train") + to a function which parses the dataset and returns the samples in the + format of `list[dict]`. + + The returned dicts should be in Detectron2 Dataset format (See DATASETS.md for details) + if used with the data loader functionalities in `data/build.py,data/detection_transform.py`. + + The purpose of having this catalog is to make it easy to choose + different datasets, by just using the strings in the config. + """ + + def register(self, name, func): + """ + Args: + name (str): the name that identifies a dataset, e.g. "coco_2014_train". + func (callable): a callable which takes no arguments and returns a list of dicts. + It must return the same results if called multiple times. + """ + assert callable(func), "You must register a function with `DatasetCatalog.register`!" + assert name not in self, "Dataset '{}' is already registered!".format(name) + self[name] = func + + def get(self, name): + """ + Call the registered function and return its results. + + Args: + name (str): the name that identifies a dataset, e.g. "coco_2014_train". + + Returns: + list[dict]: dataset annotations. + """ + try: + f = self[name] + except KeyError as e: + raise KeyError( + "Dataset '{}' is not registered! Available datasets are: {}".format( + name, ", ".join(list(self.keys())) + ) + ) from e + return f() + + def list(self) -> List[str]: + """ + List all registered datasets. + + Returns: + list[str] + """ + return list(self.keys()) + + def remove(self, name): + """ + Alias of ``pop``. + """ + self.pop(name) + + def __str__(self): + return "DatasetCatalog(registered datasets: {})".format(", ".join(self.keys())) + + __repr__ = __str__ + + +DatasetCatalog = _DatasetCatalog() +DatasetCatalog.__doc__ = ( + _DatasetCatalog.__doc__ + + """ + .. automethod:: detectron2.data.catalog.DatasetCatalog.register + .. automethod:: detectron2.data.catalog.DatasetCatalog.get +""" +) + + +class Metadata(types.SimpleNamespace): + """ + A class that supports simple attribute setter/getter. + It is intended for storing metadata of a dataset and make it accessible globally. + + Examples: + :: + # somewhere when you load the data: + MetadataCatalog.get("mydataset").thing_classes = ["person", "dog"] + + # somewhere when you print statistics or visualize: + classes = MetadataCatalog.get("mydataset").thing_classes + """ + + # the name of the dataset + # set default to N/A so that `self.name` in the errors will not trigger getattr again + name: str = "N/A" + + _RENAMED = { + "class_names": "thing_classes", + "dataset_id_to_contiguous_id": "thing_dataset_id_to_contiguous_id", + "stuff_class_names": "stuff_classes", + } + + def __getattr__(self, key): + if key in self._RENAMED: + log_first_n( + logging.WARNING, + "Metadata '{}' was renamed to '{}'!".format(key, self._RENAMED[key]), + n=10, + ) + return getattr(self, self._RENAMED[key]) + + # "name" exists in every metadata + if len(self.__dict__) > 1: + raise AttributeError( + "Attribute '{}' does not exist in the metadata of dataset '{}'. Available " + "keys are {}.".format(key, self.name, str(self.__dict__.keys())) + ) + else: + raise AttributeError( + f"Attribute '{key}' does not exist in the metadata of dataset '{self.name}': " + "metadata is empty." + ) + + def __setattr__(self, key, val): + if key in self._RENAMED: + log_first_n( + logging.WARNING, + "Metadata '{}' was renamed to '{}'!".format(key, self._RENAMED[key]), + n=10, + ) + setattr(self, self._RENAMED[key], val) + + # Ensure that metadata of the same name stays consistent + try: + oldval = getattr(self, key) + assert oldval == val, ( + "Attribute '{}' in the metadata of '{}' cannot be set " + "to a different value!\n{} != {}".format(key, self.name, oldval, val) + ) + except AttributeError: + super().__setattr__(key, val) + + def as_dict(self): + """ + Returns all the metadata as a dict. + Note that modifications to the returned dict will not reflect on the Metadata object. + """ + return copy.copy(self.__dict__) + + def set(self, **kwargs): + """ + Set multiple metadata with kwargs. + """ + for k, v in kwargs.items(): + setattr(self, k, v) + return self + + def get(self, key, default=None): + """ + Access an attribute and return its value if exists. + Otherwise return default. + """ + try: + return getattr(self, key) + except AttributeError: + return default + + +class _MetadataCatalog(UserDict): + """ + MetadataCatalog is a global dictionary that provides access to + :class:`Metadata` of a given dataset. + + The metadata associated with a certain name is a singleton: once created, the + metadata will stay alive and will be returned by future calls to ``get(name)``. + + It's like global variables, so don't abuse it. + It's meant for storing knowledge that's constant and shared across the execution + of the program, e.g.: the class names in COCO. + """ + + def get(self, name): + """ + Args: + name (str): name of a dataset (e.g. coco_2014_train). + + Returns: + Metadata: The :class:`Metadata` instance associated with this name, + or create an empty one if none is available. + """ + assert len(name) + r = super().get(name, None) + if r is None: + r = self[name] = Metadata(name=name) + return r + + def list(self): + """ + List all registered metadata. + + Returns: + list[str]: keys (names of datasets) of all registered metadata + """ + return list(self.keys()) + + def remove(self, name): + """ + Alias of ``pop``. + """ + self.pop(name) + + def __str__(self): + return "MetadataCatalog(registered metadata: {})".format(", ".join(self.keys())) + + __repr__ = __str__ + + +MetadataCatalog = _MetadataCatalog() +MetadataCatalog.__doc__ = ( + _MetadataCatalog.__doc__ + + """ + .. automethod:: detectron2.data.catalog.MetadataCatalog.get +""" +) diff --git a/detectron2/data/clip_build.py b/detectron2/data/clip_build.py new file mode 100644 index 0000000000000000000000000000000000000000..bec75db871cd8d66118748aa90fe10d014bdaf89 --- /dev/null +++ b/detectron2/data/clip_build.py @@ -0,0 +1,158 @@ +# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. +import bisect +import copy +import logging +import os +import torch +import torch.utils.data +import torch.distributed +from torch.utils.data.dataset import ConcatDataset + +from .catalog import DatasetCatalog +from .clip_datasets.clip_img_txt_pair_tsv import CLIPImgTxtPairTSVDataset + +from .transforms.build import build_clip_transforms + +def config_tsv_dataset_args(cfg, dataset_file, factory_name=None, is_train=True): + ############### code removecd as tsv_dataset_name = factory_name = "CLIPImgTxtPairTSVDataset" ############## + if factory_name is not None: + tsv_dataset_name = factory_name + + if tsv_dataset_name in ["CLIPImgTxtPairTSVDataset"]: + # no need for extra arguments + args = {} + args['args'] = cfg + args['seq_len'] = cfg.DATASETS.MAX_SEQ_LENGTH # cfg.max_seq_length + + return args, tsv_dataset_name + + +def build_dataset(cfg, transforms, dataset_catalog, is_train=True, is_aux=False): + """ + Arguments: + cfg: config file. + transforms (callable): transforms to apply to each (image, target) sample + dataset_catalog (DatasetCatalog): contains the information on how to construct a dataset. + is_train (bool): whether to setup the dataset for training or testing + """ + + dataset_list = (cfg.DATASETS.TRAIN if not is_aux else cfg.DATASETS.AUX) if is_train else cfg.DATASETS.TEST + factory_list = (cfg.DATASETS.FACTORY_TRAIN if not is_aux else cfg.DATASETS.FACTORY_AUX) if is_train else cfg.DATASETS.FACTORY_TEST + path_list = (cfg.DATASETS.PATH_TRAIN if not is_aux else cfg.DATASETS.PATH_AUX) if is_train else cfg.DATASETS.PATH_TEST + + if not isinstance(dataset_list, (list, tuple)): + raise RuntimeError( + "dataset_list should be a list of strings, got {}".format(dataset_list)) + if not isinstance(factory_list, (list, tuple)): + raise RuntimeError( + "factory_list should be a list of strings, got {}".format(factory_list)) + datasets = [] + target_offset = 0 + for i, dataset_name in enumerate(dataset_list): + factory_name = factory_list[i] if i < len(factory_list) else None + + if factory_name == "CLIPImgTxtPairTSVDataset": + dataset_names_merged = dataset_name.split('+') + path_lists_merged = path_list[i].split('+') + + assert len(dataset_names_merged) == len(path_lists_merged), "number of datasets must match that of dataset paths" + + image_tsv_list = [] + text_tsv_list = [] + dataset_name_list = [] + map_files = [] + max_num_tsv = 20 # maximum tsv files to load within a given folder + + for dname, dpath in zip(dataset_names_merged, path_lists_merged): + args, tsv_dataset_name = config_tsv_dataset_args( + cfg, dataset_name, factory_name, is_train + ) + factory = CLIPImgTxtPairTSVDataset if tsv_dataset_name in ["CLIPImgTxtPairTSVDataset"] else None + prev_len = len(image_tsv_list) + + isFile = os.path.isfile(dpath) + if isFile: + dpath_listed_files = [os.path.basename(dpath)] + dpath = os.path.dirname(dpath) + else: + dpath_listed_files = sorted(os.listdir(dpath)) + + for filename in dpath_listed_files: + if ("images" in filename or "image" in filename or "img" in filename) and filename.endswith(".tsv"): + image_tsv_list.append(os.path.join(dpath, filename)) + if "images" in filename: # "images" - "text" + text_tsv_list.append(os.path.join(dpath, filename.replace("images", "text"))) + elif "image" in filename: # "image"-"text" + text_tsv_list.append(os.path.join(dpath, filename.replace("image", "text"))) + elif "img" in filename: # "img"-"caption" + text_tsv_list.append(os.path.join(dpath, filename.replace("img", "caption"))) + if len(image_tsv_list) - prev_len == max_num_tsv: + break + dataset_name_list += [dname] * (len(image_tsv_list) - prev_len) + + if dname == "imagenet22k": + map_files += [os.path.join(dpath, 'darknet_data_imagenet.labels.list')] * (len(image_tsv_list) - prev_len) + else: + map_files += [None] * (len(image_tsv_list) - prev_len) + + assert len(image_tsv_list) == len(text_tsv_list), \ + "the number image tsv files must be equal to that of text tsv files, otherwise check your data!" + + args["image_tsv_file"] = image_tsv_list + args["text_tsv_file"] = text_tsv_list + args["dataset_name"] = dataset_name_list + args["map_file"] = map_files + args["filtered_datasets"] = cfg.DATASETS.FILTERED_CLASSIFICATION_DATASETS + assert len(image_tsv_list) == len(text_tsv_list) == len(dataset_name_list) == len(map_files) + + print("number of image tsv files: ", len(image_tsv_list)) + print("number of text tsv fies: ", len(text_tsv_list)) + + args["is_train"] = is_train + args["transforms"] = transforms + args["target_offset"] = target_offset + if "bpe" in cfg.INPUT.TEXT_TOKENIZER: + from detectron2.data.datasets.clip_prompt_utils import SimpleTokenizer as _Tokenizer + tokenizer = _Tokenizer() + args["tokenizer_type"] = "bpe" + args["tokenizer"] = tokenizer + # make dataset from factory + dataset = factory(**args) + datasets.append(dataset) + + precomputed_tokens = {} + dataset_classes = {} + for dataset in datasets: + if hasattr(dataset, "input_ids_all_classes"): + precomputed_tokens["imagenet"] = \ + [dataset.input_ids_all_classes, dataset.input_mask_all_classes, dataset.segment_ids_all_classes] + if hasattr(dataset, "classnames"): + if isinstance(dataset.classnames, dict): + dataset_classes.update(dataset.classnames) + else: + dataset_classes[dataset.dataset_name] = dataset.classnames + + # for testing, return a list of datasets + if not is_train: + return datasets, precomputed_tokens, dataset_classes + + if len(datasets) == 0: + return None, None, None + + # for training, concatenate all datasets into a single one + dataset = datasets[0] + if len(datasets) > 1: + dataset = ConcatDataset(datasets) + return [dataset], precomputed_tokens, dataset_classes + + +def make_clip_dataset(cfg, is_train=True, is_aux=False, transforms=None): + if transforms is None: + transforms = build_clip_transforms(cfg, is_train) + print("data transforms: ") + print(transforms) + datasets, precomputed_tokens, dataset_classes = build_dataset(cfg, transforms, DatasetCatalog, is_train, is_aux) + + if not datasets: + return None, None, None + return datasets, precomputed_tokens, dataset_classes \ No newline at end of file diff --git a/detectron2/data/clip_datasets/__init__.py b/detectron2/data/clip_datasets/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/detectron2/data/clip_datasets/__pycache__/__init__.cpython-39.pyc b/detectron2/data/clip_datasets/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..09fdcd58dd9d60629ce3f6e8719f29ab1a93dcb7 Binary files /dev/null and b/detectron2/data/clip_datasets/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/data/clip_datasets/__pycache__/clip_img_txt_pair_tsv.cpython-39.pyc b/detectron2/data/clip_datasets/__pycache__/clip_img_txt_pair_tsv.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2ca3298cd1e9363fe8c0b84c4926cedb753a9750 Binary files /dev/null and b/detectron2/data/clip_datasets/__pycache__/clip_img_txt_pair_tsv.cpython-39.pyc differ diff --git a/detectron2/data/clip_datasets/__pycache__/clip_prompt_engineering.cpython-39.pyc b/detectron2/data/clip_datasets/__pycache__/clip_prompt_engineering.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b20b5cd2693d91df28023019233052b5dfd3a086 Binary files /dev/null and b/detectron2/data/clip_datasets/__pycache__/clip_prompt_engineering.cpython-39.pyc differ diff --git a/detectron2/data/clip_datasets/__pycache__/oscar_tsv.cpython-39.pyc b/detectron2/data/clip_datasets/__pycache__/oscar_tsv.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ffd5c1eb04031e7a25d9e4bcdda67c559ebf75ab Binary files /dev/null and b/detectron2/data/clip_datasets/__pycache__/oscar_tsv.cpython-39.pyc differ diff --git a/detectron2/data/clip_datasets/bpe_simple_vocab_16e6.txt.gz b/detectron2/data/clip_datasets/bpe_simple_vocab_16e6.txt.gz new file mode 100644 index 0000000000000000000000000000000000000000..36a15856e00a06a9fbed8cdd34d2393fea4a3113 --- /dev/null +++ b/detectron2/data/clip_datasets/bpe_simple_vocab_16e6.txt.gz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a +size 1356917 diff --git a/detectron2/data/clip_datasets/clip_img_txt_pair_tsv.py b/detectron2/data/clip_datasets/clip_img_txt_pair_tsv.py new file mode 100644 index 0000000000000000000000000000000000000000..49487b81952459b11d37684fb2e6b9fefead0d9f --- /dev/null +++ b/detectron2/data/clip_datasets/clip_img_txt_pair_tsv.py @@ -0,0 +1,602 @@ +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import os +from io import BytesIO +import json +import logging +import base64 +import threading +import random +import numpy as np +from typing import Callable, List, Tuple, Union +from PIL import Image +from PIL import ImageFile +ImageFile.LOAD_TRUNCATED_IMAGES = True +import torch +import torch.utils.data as data +from .oscar_tsv import InputExample, convert_example_to_features +from detectron2.structures.tsv_file import TSVFile, CompositeTSVFile +from detectron2.data.clip_datasets.clip_prompt_engineering import get_prompt_templates, prompt_engineering +#import spacy + +def pre_fetch(tsv_filename: str): + logging.info('Pre-loading %s ...' % tsv_filename) + with open(tsv_filename, 'r'): + logging.info('Pre-loading %s ended.' % tsv_filename) + +class CLIPImgTxtPairTSVDataset(data.Dataset): + """ + This class is intended for encapsulating Image/Text pair data for contrastive learning described in + the following paper, + "Learning Transferable Visual Models From Natural Language Supervision" (a.k.a CLIP) + Specifically, it is used to accomadate the tsv data format from Azure Cognition Service Group. + """ + def __init__(self, + image_tsv_file: Union[str, List[str]], + text_tsv_file: Union[str, List[str]], + transforms: Callable = None, + tokenizer: Callable = None, + seq_len = 0, context_length = 77, target_offset=0, + args = None, + dataset_name = "", + tokenizer_type = "bert", + is_train = True, + map_file = None, + filtered_datasets = ''): + self.args = args + self.is_train = is_train + self.dataset_names = dataset_name + self.tokenizer_type = tokenizer_type + self.target_offset = target_offset + self.seq_len = seq_len + + self.transforms = transforms + self.tokenizer = tokenizer + self._chunk_sizes = None + self.context_length = context_length + + self.prompt_templates = get_prompt_templates() # [:2] + self.spacy_nlp = None # spacy.load('en_core_web_sm') + + self.class_selector = None + # self.class_selector = list(self.label2idx.keys()) if self.label2idx else None + + self.label2idx = {} + self.idx2label = {} + self.classnames = {} + self.dataset_target_offsets = {}; offset = 0 + + self.num_classes = sum([len(val) for val in self.classnames.values()]) + + self.filtered_classnames = [] + + if isinstance(image_tsv_file, str) and isinstance(text_tsv_file, str): + # single tsv file + if ( + os.path.splitext(image_tsv_file)[1].lower() == '.tsv' + and os.path.splitext(text_tsv_file)[1].lower() == '.tsv' + ): + self.image_tsv_file = TSVFile(image_tsv_file, if_generate_lineidx=True) + self.text_tsv_file = TSVFile(text_tsv_file, if_generate_lineidx=True) + # multiple tsv files specified in a text file + elif ( + os.path.splitext(image_tsv_file)[1].lower() == '.txt' + and os.path.splitext(text_tsv_file)[1].lower() == '.txt' + ): + self.image_tsv_file = CompositeTSVFile(image_tsv_file) + self.text_tsv_file = CompositeTSVFile(text_tsv_file) + self._chunk_sizes = self.image_tsv_file.get_chunk_size() + else: + raise ValueError("Invalid input! Please check the tsv filenames.") + # multiple tsv files specified in a list + elif ( + isinstance(image_tsv_file, list) + and isinstance(text_tsv_file, list) + ): + assert len(image_tsv_file) == len(text_tsv_file), \ + "Inconsistent number of Image/Text tsv files!" + assert len(image_tsv_file) == len(text_tsv_file), \ + "Inconsistent number of Image/Text tsv files!" + self.image_tsv_path = image_tsv_file + self.text_tsv_path = text_tsv_file + self.image_tsv_file = CompositeTSVFile(image_tsv_file, class_selector=self.class_selector) + self.text_tsv_file = CompositeTSVFile(text_tsv_file, class_selector=self.class_selector) + self._chunk_sizes = self.image_tsv_file.get_chunk_size() + self._accumulated_chunk_sizes = np.cumsum(self._chunk_sizes).tolist() + else: + raise ValueError("Invalid input! Please check the tsv filenames.") + + assert len(self.image_tsv_file) == len(self.text_tsv_file), \ + "Inconsistent size of Image/Text ({}/{}) data!".format( + len(self.image_tsv_file), len(self.text_tsv_file) + ) + + def get_chunk_sizes(self): + return self._chunk_sizes + + def get_class_boundaries(self): + # The samples of each class are organized class-by-class. + # _class_boundaries stores the lower- and upper-bound of each class. + return self.image_tsv_file.get_class_boundaries() + + def _load_map(self, map_file: str): + if not map_file: + return None + + label2idx = {} + with open(map_file) as f: + for line in f: + items = line.strip().split('\t') + label2idx[items[0]] = int(items[1]) + + return label2idx + + def _load_darknet_map(self, map_file): + if not map_file: + return None + + label2idx = {} + with open(map_file) as f: + linenum = 0 + for l in f: + item = l.strip() + label2idx[item] = linenum + linenum += 1 + + return label2idx + + def _pre_tokenize(self): + """ + pre-tokenize class names + """ + input_ids_all = [] + input_masks_all = [] + segment_ids_all = [] + for k in range(len(self.classnames["imagenet"])): + cur_id = 0; img_id = 0 + scale = 1.0 + + v = self.classnames["imagenet"].label_to_name(k) + if isinstance(v, str): + vs = [v] + elif isinstance(v, list): + vs = v + t1s = [] + t2s = [] + for v in vs: + for pt in self.prompt_templates: + t1s.append(prompt_engineering(v, template=pt)) + t2s.append("") + input_ids = [] + input_masks = [] + segment_ids = [] + is_next_labels = [0] * len(t1s) + is_img_matchs = [1] * len(t1s) + img_feat_len = 0 + for t1, t2, is_next_label, is_img_match in zip(t1s, t2s, is_next_labels, is_img_matchs): + if self.tokenizer_type == "bert": + # tokenize + tokens_a = self.tokenizer.tokenize(t1) + tokens_b = None + + # combine to one sample + cur_example = InputExample(guid=cur_id, tokens_a=tokens_a, + tokens_b=tokens_b, is_next=is_next_label, + img_id=img_id, is_img_match=is_img_match) + + # transform sample to features + cur_features = convert_example_to_features(self.args, cur_example, + self.seq_len, self.tokenizer, + img_feat_len) + + input_ids.append(torch.tensor(cur_features.input_ids, dtype=torch.long)) + input_masks.append(torch.tensor(cur_features.input_mask, dtype=torch.long)) + segment_ids.append(torch.tensor(cur_features.segment_ids, dtype=torch.long)) + + elif self.tokenizer_type == "bpe": + tokens_a = t1; tokens_b = None + # combine to one sample + cur_example = InputExample(guid=cur_id, tokens_a=tokens_a, + tokens_b=tokens_b, is_next=is_next_label, + img_id=img_id, is_img_match=is_img_match) + + # transform sample to features + cur_features = convert_example_to_features_bpe(self.args, cur_example, + self.seq_len, self.tokenizer, + img_feat_len) + + input_ids.append(torch.tensor(cur_features.input_ids, dtype=torch.long)) + input_masks.append(torch.tensor(cur_features.input_mask, dtype=torch.long)) + segment_ids.append(torch.tensor(cur_features.segment_ids, dtype=torch.long)) + + else: + raise NotImplementedError + input_ids_all.append(torch.stack(input_ids, 0)) + input_masks_all.append(torch.stack(input_masks, 0)) + segment_ids_all.append(torch.stack(segment_ids, 0)) + + self.input_ids_all_classes = torch.stack(input_ids_all, 0) + self.input_mask_all_classes = torch.stack(input_masks_all, 0) + self.segment_ids_all_classes = torch.stack(segment_ids_all, 0) + + def _online_tokenize(self, text): + + # random select a prompt template + temp_idx = np.random.randint(len(self.prompt_templates)) + pt = self.prompt_templates[temp_idx] + + names = text.split(";") + num_names = np.random.randint(len(names)) + 1 + names_sampled = random.sample(names, num_names) + text = ", ".join(names_sampled) + + t1 = prompt_engineering(text, template=pt) + + cur_id = 0; img_id = 0; scale = 1.0 + is_next_label = 0; is_img_match = 1 + img_feat_len = 0 + + if self.tokenizer_type == "bert": + # tokenize + tokens_a = self.tokenizer.tokenize(t1) + tokens_b = None + + # combine to one sample + cur_example = InputExample(guid=cur_id, tokens_a=tokens_a, + tokens_b=tokens_b, is_next=is_next_label, + img_id=img_id, is_img_match=is_img_match) + + # transform sample to features + cur_features = convert_example_to_features(self.args, cur_example, + self.context_length, self.tokenizer, + img_feat_len) + + + elif self.tokenizer_type == "bpe": + tokens_a = t1; tokens_b = None + # combine to one sample + cur_example = InputExample(guid=cur_id, tokens_a=tokens_a, + tokens_b=tokens_b, is_next=is_next_label, + img_id=img_id, is_img_match=is_img_match) + + # transform sample to features + cur_features = convert_example_to_features_bpe(self.args, cur_example, + self.context_length, self.tokenizer, + img_feat_len) + + return torch.tensor(cur_features.input_ids, dtype=torch.long), \ + torch.tensor(cur_features.input_mask, dtype=torch.long), \ + torch.tensor(cur_features.segment_ids, dtype=torch.long) + + def get_dataset_name(self, index): + """ + get dataset name according to index + """ + assert index < self._accumulated_chunk_sizes[-1], "index must in the range of accumulated data size" + for k, boundary in enumerate(self._accumulated_chunk_sizes): + if index < boundary: + return self.dataset_names[k], k + + def get_target_offset(self, dataset_name): + return self.dataset_target_offsets[dataset_name] + + def get_img_label_pair(self, items_image, index): + dataset_name, chunk_id = self.get_dataset_name(index) + target_offset = self.get_target_offset(dataset_name) + _, target, img = self._decode_data(items_image, dataset_name) + + if self.transforms: + img = self.transforms(img) + + if target == -1: + input_ids, input_mask, segment_ids = \ + self._online_tokenize("uncovered image") + else: + classname = self.classnames[dataset_name].labels2names[self.idx2label[dataset_name][target]] + if classname in self.filtered_classnames: + # we filter these classnames for training + target = -1 + input_ids, input_mask, segment_ids = \ + self._online_tokenize("uncovered image") + else: + input_ids, input_mask, segment_ids = \ + self._online_tokenize(classname) + target += target_offset + return img, \ + input_ids, \ + input_mask, \ + segment_ids, \ + torch.LongTensor([target]), \ + dataset_name + + def get_img_txt_pair(self, items_image, items_text, index): + dataset_name, chunk_id = self.get_dataset_name(index) + assert items_text[0] == items_image[0], \ + 'keys do not match for image ({}) and text ({}) for {} at chunk {}-{}'.format( + len(items_text[0]), len(items_image[0]), dataset_name, chunk_id, self.image_tsv_path[chunk_id] + ) + + img = self._decode_image(items_image, dataset_name) + # print("index {}, chunk id {}, name {}".format(index, chunk_id, self.image_tsv_path[chunk_id])) + # raise TypeError("cannot decode current item") + img_width, img_height = img.size # img_height, img_width = np.array(img).shape + + txts = self._decode_text(items_text) + if self.spacy_nlp is not None: + np_input_ids, np_input_masks, np_segment_ids = self.create_phrase_text(txts) + + if self.transforms: + img = self.transforms(img) + + if isinstance(txts, str): + input_ids, input_masks, segment_ids = \ + convert_txt_to_tokens_bpe(txts, self.tokenizer, self.context_length) + all_str2id_links = [] + elif isinstance(txts, list): + input_ids = [] + input_masks = [] + segment_ids = [] + all_str2id_links = [] + for txt in txts: + input_id, input_mask, segment_id, str2id_links = \ + convert_txt_to_tokens_bpe(txt, self.tokenizer, self.context_length, return_link=True) + input_ids += input_id + input_masks += input_mask + segment_ids += segment_id + all_str2id_links += [str2id_links] + scale = 1.0 + img_id = 0 + + if self.spacy_nlp is not None: + return img, \ + torch.tensor(input_ids).long().view(-1), \ + torch.tensor(input_masks).long().view(-1), \ + torch.tensor(segment_ids).long().view(-1), \ + torch.LongTensor([1e5]), \ + dataset_name, \ + torch.tensor(np_input_ids).long().view(-1), \ + torch.tensor(np_input_masks).long().view(-1), \ + torch.tensor(np_segment_ids).long().view(-1) + else: + return img, \ + torch.tensor(input_ids).long().view(-1), \ + torch.tensor(input_masks).long().view(-1), \ + torch.tensor(segment_ids).long().view(-1), \ + torch.LongTensor([1e5]), \ + (dataset_name, items_text[0], (img_height, img_width), all_str2id_links) # dataset name, image id, image height&width, links bet string and tokenized texts + + def create_phrase_text(self, txt_list): + """ Use NLP tool to detect noun phrases in captions, fill each identified phrase into a random prompt to create a sentence, + and convert each sentence to bpe tokens + """ + if isinstance(txt_list, str): + txt_list = [txt_list] + # detect noun phrase + noun_phrase = [] + for txt in txt_list: + doc = self.spacy_nlp(txt.lower()) + this_text = [nc.text for nc in doc.noun_chunks] + this_text = [nc.replace('a ', '').replace('the ', '') for nc in this_text] + noun_phrase.extend(this_text) + noun_phrase = list(set(noun_phrase)) + # fill each phrase into a random prompt + text_list = [] + pts = random.sample(self.prompt_templates, len(noun_phrase)) + for i, np in enumerate(noun_phrase): + text_list.append(prompt_engineering(np, pts[i])) + # convert string into bpe tokens + input_ids = [] + input_masks = [] + segment_ids = [] + for txt in text_list: + input_id, input_mask, segment_id = \ + convert_txt_to_tokens_bpe(txt, self.tokenizer, self.context_length) + input_ids += input_id + input_masks += input_mask + segment_ids += segment_id + return input_ids, input_masks, segment_ids + + def __getitem__(self, index: Union[int, Tuple[int, int]]): + if isinstance(index, tuple): + items_image = self.image_tsv_file[index[0]] + items_text = self.text_tsv_file[index[0]] + if index[1] >= 0: + tsv_filename = self.image_tsv_file.file_list[index[1]] + + # Python threads are not truly parallel. Spawn a new process instead. + # logging.info('Pre-loading %s ...' % tsv_filename) + # os.system('cat ' + tsv_filename + ' > /dev/null &') + x = threading.Thread( + target=pre_fetch, args=(tsv_filename,), daemon=True + ) + x.start() + curr_index = index[0] + else: + items_image = self.image_tsv_file[index] + items_text = self.text_tsv_file[index] + curr_index = index + + # NOTE: since we duplicate image tsv to text tsv for image-label data, + # we can determine whether the current instance is an image-label pair or + # a image-text pair data based on whether items_image is identical to items_text or not. + if items_image == items_text: + return self.get_img_label_pair(items_image, curr_index) + else: + return self.get_img_txt_pair(items_image, items_text, curr_index) + + def _decode_image(self, items: Tuple[str, str], dataset_name=""): + key = items[0] + image = Image.open(BytesIO(base64.b64decode(items[1]))).convert('RGB') + return image + + def _decode_text(self, items: Tuple[str, Union[str, dict]]): + key = items[0] + text = '' + if isinstance(items[1], str): + try: + str_dict = json.loads(items[1]) + # in this dict, it may contain either "tags" or "captions" or both + keys = [key for key in str_dict.keys()] + selected_key = random.sample(keys, 1)[0] + if selected_key == "captions": + # if this is a caption, we sample a caption + captions = str_dict[selected_key] + text = captions[:5] + # text = random.sample(captions, 1)[0] + elif selected_key == "tags": + # for tags, we randomly disorder it + tags = str_dict[selected_key] + tag_words = tags.split(' ') + random.shuffle(tag_words) + tags_shuffled = " ".join(tag_words) + # add prompt template + pt = random.sample(self.prompt_templates, 1)[0] + text = prompt_engineering(tags_shuffled, pt) + except: + text = items[1] + elif isinstance(items[1], dict): + assert 'captions' in items[1], '"captions" does not in {}'.format(items[1]) + captions = items[1]['captions'] + if isinstance(captions, list): + text = random.choice(captions) + elif isinstance(captions, str): + text = captions + else: + raise ValueError('captions should be str or list') + + return text + + def _decode_data(self, items, dataset_name): + key = items[0] + label = self._get_label(items[1], dataset_name) + try: + image = Image.open(BytesIO(base64.b64decode(items[2]))) + except: + return None + + return key, label, image.convert('RGB') + + def _get_label(self, item, dataset_name): + if not self.label2idx[dataset_name]: + return int(item) + + if item in self.label2idx[dataset_name]: + return self.label2idx[dataset_name][item] + + label = json.loads(item)[0]['class'] + if label in self.label2idx[dataset_name]: + return self.label2idx[dataset_name][label] + else: + return -1 + + def __len__(self): + return len(self.image_tsv_file) + +def convert_txt_to_tokens_bpe(text, tokenizer, context_length, return_link=False): + + sot_token = tokenizer.encoder["<|startoftext|>"] + eot_token = tokenizer.encoder["<|endoftext|>"] + if return_link: + bpe_tokens, str2id_links = tokenizer.encode(text, return_link=return_link) + str2id_links = [["<|startoftext|>", [sot_token]]] + str2id_links + [["<|endoftext|>", [eot_token]]] + else: + bpe_tokens = tokenizer.encode(text, return_link=return_link) + input_ids = [sot_token] + bpe_tokens + [eot_token] + + if len(input_ids) > context_length: + input_ids = input_ids[:context_length] + segment_ids = [0] * len(input_ids) + lm_label_ids = [-1] * len(input_ids) + + # The mask has 1 for real tokens and 0 for padding tokens. Only real tokens are attended to. + input_mask = [1] * len(input_ids) + + # Zero-pad up to the sequence length. + while len(input_ids) < context_length: + input_ids.append(0) + input_mask.append(0) + segment_ids.append(0) + lm_label_ids.append(-1) + + assert len(input_ids) == context_length + assert len(input_mask) == context_length + assert len(segment_ids) == context_length + assert len(lm_label_ids) == context_length + + if return_link: + return input_ids, input_mask, segment_ids, str2id_links + return input_ids, input_mask, segment_ids + +def convert_example_to_features_bpe(args, example, max_seq_length, tokenizer, + img_feat_len, context_length=77): + """ + Convert a raw sample (pair of sentences as tokenized strings) into a proper training sample with + IDs, LM labels, input_mask, CLS and SEP tokens etc. + :param args: parameter settings + :param img_feat_len: lens of actual img features + :param example: InputExample, containing sentence input as strings and is_next label + :param max_seq_length: int, maximum length of sequence. + :param tokenizer: Tokenizer + :return: InputFeatures, containing all inputs and labels of one sample as IDs (as used for model training) + """ + # we do not consider tokens_b for now in original CLIP + text = example.tokens_a + assert isinstance(text, str) + + sot_token = tokenizer.encoder["<|startoftext|>"] + eot_token = tokenizer.encoder["<|endoftext|>"] + input_ids = [sot_token] + tokenizer.encode(text) + [eot_token] + + if len(input_ids) > context_length: + input_ids = input_ids[:context_length] + segment_ids = [0] * len(input_ids) + lm_label_ids = [-1] * len(input_ids) + + # The mask has 1 for real tokens and 0 for padding tokens. Only real tokens are attended to. + input_mask = [1] * len(input_ids) + + # Zero-pad up to the sequence length. + while len(input_ids) < context_length: + input_ids.append(0) + input_mask.append(0) + segment_ids.append(0) + lm_label_ids.append(-1) + + assert len(input_ids) == context_length + assert len(input_mask) == context_length + assert len(segment_ids) == context_length + assert len(lm_label_ids) == context_length + + if example.guid < 1: + logging.info("*** Example ***") + logging.info("guid: %s" % example.guid) + logging.info("input_ids: %s" % " ".join([str(x) for x in input_ids])) + logging.info("input_mask: %s" % " ".join([str(x) for x in input_mask])) + logging.info("segment_ids: %s" % " ".join([str(x) for x in segment_ids])) + logging.info("LM label: %s " % lm_label_ids) + logging.info("Is next sentence label: %s " % example.is_next) + + features = InputFeatures(input_ids=input_ids, + input_mask=input_mask, + segment_ids=segment_ids, + lm_label_ids=lm_label_ids, + is_next=example.is_next, + img_feat_len=img_feat_len, + is_img_match=example.is_img_match) + return features + +class InputFeatures(object): + """A single set of features of data.""" + + def __init__(self, input_ids, input_mask, segment_ids, is_next, + lm_label_ids, img_feat_len, is_img_match): + self.input_ids = input_ids + self.input_mask = input_mask + self.segment_ids = segment_ids + self.is_next = is_next + self.lm_label_ids = lm_label_ids + + self.img_feat_len = img_feat_len + self.is_img_match = is_img_match \ No newline at end of file diff --git a/detectron2/data/clip_datasets/clip_prompt_engineering.py b/detectron2/data/clip_datasets/clip_prompt_engineering.py new file mode 100644 index 0000000000000000000000000000000000000000..bba36e22156ac19c1789d0c9132c03f4e5752555 --- /dev/null +++ b/detectron2/data/clip_datasets/clip_prompt_engineering.py @@ -0,0 +1,381 @@ +import gzip +import html +import os +from functools import lru_cache + +import ftfy +import regex as re +import torch +import numpy as np +from typing import Union, List + +# https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py +@lru_cache() +def default_bpe(): + return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz") + + +@lru_cache() +def bytes_to_unicode(): + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a signficant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1)) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8+n) + n += 1 + cs = [chr(n) for n in cs] + return dict(zip(bs, cs)) + + +def get_pairs(word): + """Return set of symbol pairs in a word. + Word is represented as tuple of symbols (symbols being variable-length strings). + """ + pairs = set() + prev_char = word[0] + for char in word[1:]: + pairs.add((prev_char, char)) + prev_char = char + return pairs + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r'\s+', ' ', text) + text = text.strip() + return text + + +class SimpleTokenizer(object): + def __init__(self, bpe_path: str = default_bpe()): + self.byte_encoder = bytes_to_unicode() + self.byte_decoder = {v: k for k, v in self.byte_encoder.items()} + merges = gzip.open(bpe_path).read().decode("utf-8").split('\n') + merges = merges[1:49152-256-2+1] + merges = [tuple(merge.split()) for merge in merges] + vocab = list(bytes_to_unicode().values()) + vocab = vocab + [v+'' for v in vocab] + self.vocab = vocab + for merge in merges: + vocab.append(''.join(merge)) + vocab.extend(['<|startoftext|>', '<|endoftext|>']) + self.encoder = dict(zip(vocab, range(len(vocab)))) + self.decoder = {v: k for k, v in self.encoder.items()} + self.bpe_ranks = dict(zip(merges, range(len(merges)))) + self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'} + self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE) + + def bpe(self, token): + if token in self.cache: + return self.cache[token] + word = tuple(token[:-1]) + ( token[-1] + '',) + pairs = get_pairs(word) + + if not pairs: + return token+'' + + while True: + bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf'))) + if bigram not in self.bpe_ranks: + break + first, second = bigram + new_word = [] + i = 0 + while i < len(word): + try: + j = word.index(first, i) + new_word.extend(word[i:j]) + i = j + except: + new_word.extend(word[i:]) + break + + if word[i] == first and i < len(word)-1 and word[i+1] == second: + new_word.append(first+second) + i += 2 + else: + new_word.append(word[i]) + i += 1 + new_word = tuple(new_word) + word = new_word + if len(word) == 1: + break + else: + pairs = get_pairs(word) + word = ' '.join(word) + self.cache[token] = word + return word + + def encode(self, text): + bpe_tokens = [] + text = whitespace_clean(basic_clean(text)).lower() + for token in re.findall(self.pat, text): + token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8')) + bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' ')) + return bpe_tokens + + def decode(self, tokens): + text = ''.join([self.decoder[token] for token in tokens]) + text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('', ' ') + return text + + +# https://github.com/openai/CLIP/blob/main/clip/clip.py +#_tokenizer = SimpleTokenizer() + +def tokenize(texts: Union[str, List[str]], context_length: int = 77): + if isinstance(texts, str): + texts = [texts] + + sot_token = _tokenizer.encoder["<|startoftext|>"] + eot_token = _tokenizer.encoder["<|endoftext|>"] + all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts] + result = torch.zeros(len(all_tokens), context_length, dtype=torch.long) + + for i, tokens in enumerate(all_tokens): + if len(tokens) > context_length: + raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}") + result[i, :len(tokens)] = torch.tensor(tokens) + + return result + + +# prompt_engineering.py +def get_prompt_templates(): + # prompt_templates = [ + # 'There is a {} in the scene.', + # 'There is the {} in the scene.', + # 'a photo of a {} in the scene.', + # 'a photo of the {} in the scene.', + # 'a photo of one {} in the scene.', + + # 'itap of a {}.', + # 'itap of my {}.', # itap: I took a picture of + # 'itap of the {}.', + # 'a photo of a {}.', + # 'a photo of my {}.', + # 'a photo of the {}.', + # 'a photo of one {}.', + # 'a photo of many {}.', + + # 'a good photo of a {}.', + # 'a good photo of the {}.', + # 'a bad photo of a {}.', + # 'a bad photo of the {}.', + # 'a photo of a nice {}.', + # 'a photo of the nice {}.', + # 'a photo of a cool {}.', + # 'a photo of the cool {}.', + # 'a photo of a weird {}.', + # 'a photo of the weird {}.', + + # 'a photo of a small {}.', + # 'a photo of the small {}.', + # 'a photo of a large {}.', + # 'a photo of the large {}.', + + # 'a photo of a clean {}.', + # 'a photo of the clean {}.', + # 'a photo of a dirty {}.', + # 'a photo of the dirty {}.', + + # 'a bright photo of a {}.', + # 'a bright photo of the {}.', + # 'a dark photo of a {}.', + # 'a dark photo of the {}.', + + # 'a photo of a hard to see {}.', + # 'a photo of the hard to see {}.', + # 'a low resolution photo of a {}.', + # 'a low resolution photo of the {}.', + # 'a cropped photo of a {}.', + # 'a cropped photo of the {}.', + # 'a close-up photo of a {}.', + # 'a close-up photo of the {}.', + # 'a jpeg corrupted photo of a {}.', + # 'a jpeg corrupted photo of the {}.', + # 'a blurry photo of a {}.', + # 'a blurry photo of the {}.', + # 'a pixelated photo of a {}.', + # 'a pixelated photo of the {}.', + + # 'a black and white photo of the {}.', + # 'a black and white photo of a {}.', + + # 'a plastic {}.', + # 'the plastic {}.', + + # 'a toy {}.', + # 'the toy {}.', + # 'a plushie {}.', + # 'the plushie {}.', + # 'a cartoon {}.', + # 'the cartoon {}.', + + # 'an embroidered {}.', + # 'the embroidered {}.', + + # 'a painting of the {}.', + # 'a painting of a {}.', + # ] + + prompt_templates = [ + '{}.', + 'a photo of a {}.', + 'a bad photo of a {}.', + 'a photo of many {}.', + 'a sculpture of a {}.', + 'a photo of the hard to see {}.', + 'a low resolution photo of the {}.', + 'a rendering of a {}.', + 'graffiti of a {}.', + 'a bad photo of the {}.', + 'a cropped photo of the {}.', + 'a tattoo of a {}.', + 'the embroidered {}.', + 'a photo of a hard to see {}.', + 'a bright photo of a {}.', + 'a photo of a clean {}.', + 'a photo of a dirty {}.', + 'a dark photo of the {}.', + 'a drawing of a {}.', + 'a photo of my {}.', + 'the plastic {}.', + 'a photo of the cool {}.', + 'a close-up photo of a {}.', + 'a black and white photo of the {}.', + 'a painting of the {}.', + 'a painting of a {}.', + 'a pixelated photo of the {}.', + 'a sculpture of the {}.', + 'a bright photo of the {}.', + 'a cropped photo of a {}.', + 'a plastic {}.', + 'a photo of the dirty {}.', + 'a jpeg corrupted photo of a {}.', + 'a blurry photo of the {}.', + 'a photo of the {}.', + 'a good photo of the {}.', + 'a rendering of the {}.', + 'a {} in a video game.', + 'a photo of one {}.', + 'a doodle of a {}.', + 'a close-up photo of the {}.', + 'the origami {}.', + 'the {} in a video game.', + 'a sketch of a {}.', + 'a doodle of the {}.', + 'a origami {}.', + 'a low resolution photo of a {}.', + 'the toy {}.', + 'a rendition of the {}.', + 'a photo of the clean {}.', + 'a photo of a large {}.', + 'a rendition of a {}.', + 'a photo of a nice {}.', + 'a photo of a weird {}.', + 'a blurry photo of a {}.', + 'a cartoon {}.', + 'art of a {}.', + 'a sketch of the {}.', + 'a embroidered {}.', + 'a pixelated photo of a {}.', + 'itap of the {}.', + 'a jpeg corrupted photo of the {}.', + 'a good photo of a {}.', + 'a plushie {}.', + 'a photo of the nice {}.', + 'a photo of the small {}.', + 'a photo of the weird {}.', + 'the cartoon {}.', + 'art of the {}.', + 'a drawing of the {}.', + 'a photo of the large {}.', + 'a black and white photo of a {}.', + 'the plushie {}.', + 'a dark photo of a {}.', + 'itap of a {}.', + 'graffiti of the {}.', + 'a toy {}.', + 'itap of my {}.', + 'a photo of a cool {}.', + 'a photo of a small {}.', + 'a tattoo of the {}.', + ] + return prompt_templates + +def prompt_engineering(classnames, template=""): + return template.replace('{}', classnames.replace(',', '').replace('+', ' ')) + +# clip_img_tsv.py +def convert_example_to_features_bpe(text, tokenizer, sot_token, eot_token, context_length=77): + """ + Convert a raw sample (pair of sentences as tokenized strings) into a proper training sample. + :param tokenizer: Tokenizer + :return: List, a list containing token id, padded by 0 + """ + assert isinstance(text, str) + input_ids = [sot_token] + tokenizer.encode(text) + [eot_token] + if len(input_ids) > context_length: + input_ids = input_ids[:context_length] + input_ids = np.array(input_ids) + + pad_input_ids = np.zeros(context_length) + pad_input_ids[:input_ids.shape[0]] = input_ids + + return pad_input_ids + +def pre_tokenize(class_names): + """ + pre-tokenize class names + :param class_names: List, a list of class names + :param tokenizer: Tokenizer, SimpleTokenizer() + :return: Tensor, containing all prompts for all classes, [#cls, #prompts, context_length] + """ + # tokenizer + tokenizer = SimpleTokenizer() + sot_token = tokenizer.encoder["<|startoftext|>"] + eot_token = tokenizer.encoder["<|endoftext|>"] + + # prompt engineering + prompt_templates = get_prompt_templates() + input_ids_all = [] + for k in range(len(class_names)): + v = class_names[k] + if isinstance(v, str): + vs = [v] + elif isinstance(v, list): + vs = v + t1s = [] + for v in vs: + for pt in prompt_templates: + t1s.append(prompt_engineering(v, template=pt)) + input_ids = [] + for t1 in t1s: + this_input_ids = convert_example_to_features_bpe(t1, tokenizer, sot_token, eot_token) + input_ids.append(torch.tensor(this_input_ids, dtype=torch.long)) + + input_ids_all.append(torch.stack(input_ids, 0)) + + input_ids_all_classes = torch.stack(input_ids_all, 0) + return input_ids_all_classes + + +if __name__ == "__main__": + flatten_input_ids = pre_tokenize() diff --git a/detectron2/data/clip_datasets/oscar_tsv.py b/detectron2/data/clip_datasets/oscar_tsv.py new file mode 100644 index 0000000000000000000000000000000000000000..c2dfce34f399ef2efc44e13d45f13dc9fba66970 --- /dev/null +++ b/detectron2/data/clip_datasets/oscar_tsv.py @@ -0,0 +1,218 @@ +import logging +import random + +class InputExample(object): + """A single training/test example for the language model.""" + + def __init__(self, guid, tokens_a, tokens_b=None, is_next=None, + lm_labels=None, img_id=None, is_img_match=None, + img_label=None): + """Constructs a InputExample. + + Args: + guid: Unique id for the example. + tokens_a: string. The untokenized text of the first sequence. For single + sequence tasks, only this sequence must be specified. + tokens_b: (Optional) string. The untokenized text of the second sequence. + Only must be specified for sequence pair tasks. + """ + self.guid = guid + self.tokens_a = tokens_a + self.tokens_b = tokens_b + self.is_next = is_next # nextSentence + self.lm_labels = lm_labels # masked words for language model + + self.img_id = img_id + self.is_img_match = is_img_match + self.img_label = img_label + +class InputFeatures(object): + """A single set of features of data.""" + + def __init__(self, input_ids, input_mask, segment_ids, is_next, + lm_label_ids, img_feat_len, is_img_match): + self.input_ids = input_ids + self.input_mask = input_mask + self.segment_ids = segment_ids + self.is_next = is_next + self.lm_label_ids = lm_label_ids + + self.img_feat_len = img_feat_len + self.is_img_match = is_img_match + + +def random_word(tokens, tokenizer): + """ + Masking some random tokens for Language Model task with probabilities as in the original BERT paper. + :param tokens: list of str, tokenized sentence. + :param tokenizer: Tokenizer, object used for tokenization (we need it's vocab here) + :return: (list of str, list of int), masked tokens and related labels for LM prediction + """ + output_label = [] + + for i, token in enumerate(tokens): + prob = random.random() + # mask token with 15% probability + if prob < 0.15: + prob /= 0.15 + + # 80% randomly change token to mask token + if prob < 0.8: + tokens[i] = "[MASK]" + + # 10% randomly change token to random token + elif prob < 0.9: + tokens[i] = random.choice(list(tokenizer.vocab.items()))[0] + + # -> rest 10% randomly keep current token + + # append current token to output (we will predict these later) + try: + output_label.append(tokenizer.vocab[token]) + except KeyError: + # For unknown words (should not occur with BPE vocab) + output_label.append(tokenizer.vocab["[UNK]"]) + logging.warning( + "Cannot find token '{}' in vocab. Using [UNK] insetad".format( + token)) + else: + # no masking token (will be ignored by loss function later) + output_label.append(-1) + + return tokens, output_label + + +def convert_example_to_features(args, example, max_seq_length, tokenizer, + img_feat_len): + """ + Convert a raw sample (pair of sentences as tokenized strings) into a proper training sample with + IDs, LM labels, input_mask, CLS and SEP tokens etc. + :param args: parameter settings + :param img_feat_len: lens of actual img features + :param example: InputExample, containing sentence input as strings and is_next label + :param max_seq_length: int, maximum length of sequence. + :param tokenizer: Tokenizer + :return: InputFeatures, containing all inputs and labels of one sample as IDs (as used for model training) + """ + + tokens_a = example.tokens_a + tokens_b = None + if example.tokens_b: + tokens_b = example.tokens_b + # Modifies `tokens_a` and `tokens_b` in place so that the total + # length is less than the specified length. + # Account for [CLS], [SEP], [SEP] with "- 3" + _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3) + else: + if len(tokens_a) > max_seq_length - 2: + tokens_a = tokens_a[:(max_seq_length - 2)] + + tokens_a, t1_label = random_word(tokens_a, tokenizer) + if tokens_b: + tokens_b, t2_label = random_word(tokens_b, tokenizer) + + # concatenate lm labels and account for CLS, SEP, SEP + if tokens_b: + lm_label_ids = ([-1] + t1_label + [-1] + t2_label + [-1]) + else: + lm_label_ids = ([-1] + t1_label + [-1]) + + # The convention in BERT is: + # (a) For sequence pairs: + # tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP] + # type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1 + # (b) For single sequences: + # tokens: [CLS] the dog is hairy . [SEP] + # type_ids: 0 0 0 0 0 0 0 + # + # Where "type_ids" are used to indicate whether this is the first + # sequence or the second sequence. The embedding vectors for `type=0` and + # `type=1` were learned during pre-training and are added to the wordpiece + # embedding vector (and position vector). This is not *strictly* necessary + # since the [SEP] token unambigiously separates the sequences, but it makes + # it easier for the model to learn the concept of sequences. + # + # For classification tasks, the first vector (corresponding to [CLS]) is + # used as as the "sentence vector". Note that this only makes sense because + # the entire model is fine-tuned. + tokens = [] + segment_ids = [] + tokens.append("[CLS]") + segment_ids.append(0) + for token in tokens_a: + tokens.append(token) + segment_ids.append(0) + tokens.append("[SEP]") + segment_ids.append(0) + + if tokens_b: + assert len(tokens_b) > 0 + for token in tokens_b: + tokens.append(token) + segment_ids.append(1) + tokens.append("[SEP]") + segment_ids.append(1) + + input_ids = tokenizer.convert_tokens_to_ids(tokens) + + # The mask has 1 for real tokens and 0 for padding tokens. Only real tokens are attended to. + input_mask = [1] * len(input_ids) + + # Zero-pad up to the sequence length. + while len(input_ids) < max_seq_length: + input_ids.append(0) + input_mask.append(0) + segment_ids.append(0) + lm_label_ids.append(-1) + + assert len(input_ids) == max_seq_length + assert len(input_mask) == max_seq_length + assert len(segment_ids) == max_seq_length + assert len(lm_label_ids) == max_seq_length + + # image features + if args.max_img_seq_length > 0: + if img_feat_len > args.max_img_seq_length: + input_mask = input_mask + [1] * img_feat_len + else: + input_mask = input_mask + [1] * img_feat_len + pad_img_feat_len = args.max_img_seq_length - img_feat_len + input_mask = input_mask + ([0] * pad_img_feat_len) + + lm_label_ids = lm_label_ids + [-1] * args.max_img_seq_length + + if example.guid < 1: + logging.info("*** Example ***") + logging.info("guid: %s" % example.guid) + logging.info("tokens: %s" % " ".join([str(x) for x in tokens])) + logging.info("input_ids: %s" % " ".join([str(x) for x in input_ids])) + logging.info("input_mask: %s" % " ".join([str(x) for x in input_mask])) + logging.info("segment_ids: %s" % " ".join([str(x) for x in segment_ids])) + logging.info("LM label: %s " % lm_label_ids) + logging.info("Is next sentence label: %s " % example.is_next) + + features = InputFeatures(input_ids=input_ids, + input_mask=input_mask, + segment_ids=segment_ids, + lm_label_ids=lm_label_ids, + is_next=example.is_next, + img_feat_len=img_feat_len, + is_img_match=example.is_img_match) + return features + + +def _truncate_seq_pair(tokens_a, tokens_b, max_length): + """Truncates a sequence pair in place to the maximum length.""" + + # This is a simple heuristic which will always truncate the longer sequence + # one token at a time. This makes more sense than truncating an equal percent + # of tokens from each, since if one sequence is very short then each token + # that's truncated likely contains more information than a longer sequence. + while True: + total_length = len(tokens_a) + len(tokens_b) + if total_length <= max_length: + break + if len(tokens_a) > len(tokens_b): + tokens_a.pop() + else: + tokens_b.pop() \ No newline at end of file diff --git a/detectron2/data/common.py b/detectron2/data/common.py new file mode 100644 index 0000000000000000000000000000000000000000..ef7d97c2860b63d7ff4686f2e86f00fe6e181a35 --- /dev/null +++ b/detectron2/data/common.py @@ -0,0 +1,186 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import itertools +import logging +import numpy as np +import pickle +import random +import torch.utils.data as data +from torch.utils.data.sampler import Sampler + +from detectron2.utils.serialize import PicklableWrapper + +__all__ = ["MapDataset", "DatasetFromList", "AspectRatioGroupedDataset", "ToIterableDataset"] + + +class MapDataset(data.Dataset): + """ + Map a function over the elements in a dataset. + + Args: + dataset: a dataset where map function is applied. + map_func: a callable which maps the element in dataset. map_func is + responsible for error handling, when error happens, it needs to + return None so the MapDataset will randomly use other + elements from the dataset. + """ + + def __init__(self, dataset, map_func): + self._dataset = dataset + self._map_func = PicklableWrapper(map_func) # wrap so that a lambda will work + + self._rng = random.Random(42) + self._fallback_candidates = set(range(len(dataset))) + + def __len__(self): + return len(self._dataset) + + def __getitem__(self, idx): + retry_count = 0 + cur_idx = int(idx) + + while True: + data = self._map_func(self._dataset[cur_idx]) + if data is not None: + self._fallback_candidates.add(cur_idx) + return data + + # _map_func fails for this idx, use a random new index from the pool + retry_count += 1 + self._fallback_candidates.discard(cur_idx) + cur_idx = self._rng.sample(self._fallback_candidates, k=1)[0] + + if retry_count >= 3: + logger = logging.getLogger(__name__) + logger.warning( + "Failed to apply `_map_func` for idx: {}, retry count: {}".format( + idx, retry_count + ) + ) + + +class DatasetFromList(data.Dataset): + """ + Wrap a list to a torch Dataset. It produces elements of the list as data. + """ + + def __init__(self, lst: list, copy: bool = True, serialize: bool = True): + """ + Args: + lst (list): a list which contains elements to produce. + copy (bool): whether to deepcopy the element when producing it, + so that the result can be modified in place without affecting the + source in the list. + serialize (bool): whether to hold memory using serialized objects, when + enabled, data loader workers can use shared RAM from master + process instead of making a copy. + """ + self._lst = lst + self._copy = copy + self._serialize = serialize + + def _serialize(data): + buffer = pickle.dumps(data, protocol=-1) + return np.frombuffer(buffer, dtype=np.uint8) + + if self._serialize: + logger = logging.getLogger(__name__) + logger.info( + "Serializing {} elements to byte tensors and concatenating them all ...".format( + len(self._lst) + ) + ) + self._lst = [_serialize(x) for x in self._lst] + self._addr = np.asarray([len(x) for x in self._lst], dtype=np.int64) + self._addr = np.cumsum(self._addr) + self._lst = np.concatenate(self._lst) + logger.info("Serialized dataset takes {:.2f} MiB".format(len(self._lst) / 1024 ** 2)) + + def __len__(self): + if self._serialize: + return len(self._addr) + else: + return len(self._lst) + + def __getitem__(self, idx): + if self._serialize: + start_addr = 0 if idx == 0 else self._addr[idx - 1].item() + end_addr = self._addr[idx].item() + bytes = memoryview(self._lst[start_addr:end_addr]) + return pickle.loads(bytes) + elif self._copy: + return copy.deepcopy(self._lst[idx]) + else: + return self._lst[idx] + + +class ToIterableDataset(data.IterableDataset): + """ + Convert an old indices-based (also called map-style) dataset + to an iterable-style dataset. + """ + + def __init__(self, dataset, sampler): + """ + Args: + dataset (torch.utils.data.Dataset): an old-style dataset with ``__getitem__`` + sampler (torch.utils.data.sampler.Sampler): a cheap iterable that produces indices + to be applied on ``dataset``. + """ + assert not isinstance(dataset, data.IterableDataset), dataset + assert isinstance(sampler, Sampler), sampler + self.dataset = dataset + self.sampler = sampler + + def __iter__(self): + worker_info = data.get_worker_info() + if worker_info is None or worker_info.num_workers == 1: + for idx in self.sampler: + yield self.dataset[idx] + else: + # With map-style dataset, `DataLoader(dataset, sampler)` runs the + # sampler in main process only. But `DataLoader(ToIterableDataset(dataset, sampler))` + # will run sampler in every of the N worker and only keep 1/N of the ids on each + # worker. The assumption is that sampler is cheap to iterate and it's fine to discard + # ids in workers. + for idx in itertools.islice( + self.sampler, worker_info.id, None, worker_info.num_workers + ): + yield self.dataset[idx] + + +class AspectRatioGroupedDataset(data.IterableDataset): + """ + Batch data that have similar aspect ratio together. + In this implementation, images whose aspect ratio < (or >) 1 will + be batched together. + This improves training speed because the images then need less padding + to form a batch. + + It assumes the underlying dataset produces dicts with "width" and "height" keys. + It will then produce a list of original dicts with length = batch_size, + all with similar aspect ratios. + """ + + def __init__(self, dataset, batch_size): + """ + Args: + dataset: an iterable. Each element must be a dict with keys + "width" and "height", which will be used to batch data. + batch_size (int): + """ + self.dataset = dataset + self.batch_size = batch_size + self._buckets = [[] for _ in range(2)] + # Hard-coded two aspect ratio groups: w > h and w < h. + # Can add support for more aspect ratio groups, but doesn't seem useful + + def __iter__(self): + for d in self.dataset: + w, h = d["width"], d["height"] + bucket_id = 0 if w > h else 1 + bucket = self._buckets[bucket_id] + bucket.append(d) + if len(bucket) == self.batch_size: + yield bucket[:] + del bucket[:] diff --git a/detectron2/data/dataset_mapper.py b/detectron2/data/dataset_mapper.py new file mode 100644 index 0000000000000000000000000000000000000000..5e03ea2f428a271fcc85de1d97a17a8914a8978a --- /dev/null +++ b/detectron2/data/dataset_mapper.py @@ -0,0 +1,214 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import logging +import numpy as np +from typing import List, Optional, Union +import torch + +from detectron2.config import configurable + +from . import detection_utils as utils +from . import transforms as T + +""" +This file contains the default mapping that's applied to "dataset dicts". +""" + +__all__ = ["DatasetMapper"] + + +class DatasetMapper: + """ + A callable which takes a dataset dict in Detectron2 Dataset format, + and map it into a format used by the model. + + This is the default callable to be used to map your dataset dict into training data. + You may need to follow it to implement your own one for customized logic, + such as a different way to read or transform images. + See :doc:`/tutorials/data_loading` for details. + + The callable currently does the following: + + 1. Read the image from "file_name" + 2. Applies cropping/geometric transforms to the image and annotations + 3. Prepare data and annotations to Tensor and :class:`Instances` + """ + + @configurable + def __init__( + self, + is_train: bool, + *, + augmentations: List[Union[T.Augmentation, T.Transform]], + image_format: str, + use_instance_mask: bool = False, + use_keypoint: bool = False, + instance_mask_format: str = "polygon", + keypoint_hflip_indices: Optional[np.ndarray] = None, + precomputed_proposal_topk: Optional[int] = None, + recompute_boxes: bool = False, + filter_open_cls: bool = False, + clip_crop: bool = False, + ): + """ + NOTE: this interface is experimental. + + Args: + is_train: whether it's used in training or inference + augmentations: a list of augmentations or deterministic transforms to apply + image_format: an image format supported by :func:`detection_utils.read_image`. + use_instance_mask: whether to process instance segmentation annotations, if available + use_keypoint: whether to process keypoint annotations if available + instance_mask_format: one of "polygon" or "bitmask". Process instance segmentation + masks into this format. + keypoint_hflip_indices: see :func:`detection_utils.create_keypoint_hflip_indices` + precomputed_proposal_topk: if given, will load pre-computed + proposals from dataset_dict and keep the top k proposals for each image. + recompute_boxes: whether to overwrite bounding box annotations + by computing tight bounding boxes from instance mask annotations. + filter_open_cls: open-set setting, filter the open-set categories during training + clip_crop: the mode that directly use CLIP on cropped image regions + """ + if recompute_boxes: + assert use_instance_mask, "recompute_boxes requires instance masks" + # fmt: off + self.is_train = is_train + self.augmentations = T.AugmentationList(augmentations) + self.image_format = image_format + self.use_instance_mask = use_instance_mask + self.instance_mask_format = instance_mask_format + self.use_keypoint = use_keypoint + self.keypoint_hflip_indices = keypoint_hflip_indices + self.proposal_topk = precomputed_proposal_topk + self.recompute_boxes = recompute_boxes + self.filter_open_cls = filter_open_cls + self.clip_crop = clip_crop + # fmt: on + logger = logging.getLogger(__name__) + mode = "training" if is_train else "inference" + logger.info(f"[DatasetMapper] Augmentations used in {mode}: {augmentations}") + + @classmethod + def from_config(cls, cfg, is_train: bool = True): + augs = utils.build_augmentation(cfg, is_train) + if cfg.INPUT.CROP.ENABLED and is_train: + augs.insert(0, T.RandomCrop(cfg.INPUT.CROP.TYPE, cfg.INPUT.CROP.SIZE)) + recompute_boxes = cfg.MODEL.MASK_ON + else: + recompute_boxes = False + + ret = { + "is_train": is_train, + "augmentations": augs, + "image_format": cfg.INPUT.FORMAT, + "use_instance_mask": cfg.MODEL.MASK_ON, + "instance_mask_format": cfg.INPUT.MASK_FORMAT, + "use_keypoint": cfg.MODEL.KEYPOINT_ON, + "recompute_boxes": recompute_boxes, + } + + if cfg.MODEL.KEYPOINT_ON: + ret["keypoint_hflip_indices"] = utils.create_keypoint_hflip_indices(cfg.DATASETS.TRAIN) + + if cfg.MODEL.LOAD_PROPOSALS: + ret["precomputed_proposal_topk"] = ( + cfg.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TRAIN + if is_train + else cfg.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TEST + ) + # open-set setting, filter the open-set categories during training + # filter_open_cls = cfg.SOLVER.IMS_PER_BATCH < 10 # debug + # if filter_open_cls: + # ret["filter_open_cls"] = True + # CLIP inference on cropped image regions + if cfg.MODEL.META_ARCHITECTURE in ["CLIPRCNN", "CLIPFastRCNN"]: + ret["clip_crop"] = True + return ret + + def __call__(self, dataset_dict): + """ + Args: + dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format. + + Returns: + dict: a format that builtin models in detectron2 accept + """ + dataset_dict = copy.deepcopy(dataset_dict) # it will be modified by code below + # USER: Write your own image loading if it's not from a file + image = utils.read_image(dataset_dict["file_name"], format=self.image_format) + utils.check_image_size(dataset_dict, image) + + # USER: Remove if you don't do semantic/panoptic segmentation. + if "sem_seg_file_name" in dataset_dict: + sem_seg_gt = utils.read_image(dataset_dict.pop("sem_seg_file_name"), "L").squeeze(2) + else: + sem_seg_gt = None + + aug_input = T.AugInput(image, sem_seg=sem_seg_gt) + transforms = self.augmentations(aug_input) + # if self.clip_crop: # load original images into CLIP model, without resizing + # pass + # else: + image, sem_seg_gt = aug_input.image, aug_input.sem_seg + + image_shape = image.shape[:2] # h, w + # Pytorch's dataloader is efficient on torch.Tensor due to shared-memory, + # but not efficient on large generic data structures due to the use of pickle & mp.Queue. + # Therefore it's important to use torch.Tensor. + dataset_dict["image"] = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1))) + if sem_seg_gt is not None: + dataset_dict["sem_seg"] = torch.as_tensor(sem_seg_gt.astype("long")) + + # USER: Remove if you don't use pre-computed proposals. + # Most users would not need this feature. + if self.proposal_topk is not None: + utils.transform_proposals( + dataset_dict, image_shape, transforms, proposal_topk=self.proposal_topk + ) + + if not self.is_train: + if self.clip_crop: # still load the GT annotations + pass + else: + # USER: Modify this if you want to keep them for some reason. + dataset_dict.pop("annotations", None) + dataset_dict.pop("sem_seg_file_name", None) + return dataset_dict + + if "annotations" in dataset_dict: + # if self.filter_open_cls: # filter categories for open-set training + # obj_annos = dataset_dict['annotations'] + # clean_obj_annos = [obj_anno for obj_anno in obj_annos if obj_anno['frequency'] != 'r'] # filter rare classes + # if len(clean_obj_annos) == 0: # empty annotation + # print("\n\nImage {} has no annotation after filtering open-set classes!\n\n".format(dataset_dict['image_id'])) + # clean_obj_annos = obj_annos[0] # keep one for compatability, fix it later + # dataset_dict['annotations'] = clean_obj_annos + + # USER: Modify this if you want to keep them for some reason. + for anno in dataset_dict["annotations"]: + if not self.use_instance_mask: + anno.pop("segmentation", None) + if not self.use_keypoint: + anno.pop("keypoints", None) + + # USER: Implement additional transformations if you have other types of data + annos = [ + utils.transform_instance_annotations( + obj, transforms, image_shape, keypoint_hflip_indices=self.keypoint_hflip_indices + ) + for obj in dataset_dict.pop("annotations") + if obj.get("iscrowd", 0) == 0 + ] + instances = utils.annotations_to_instances( + annos, image_shape, mask_format=self.instance_mask_format + ) + + # After transforms such as cropping are applied, the bounding box may no longer + # tightly bound the object. As an example, imagine a triangle object + # [(0,0), (2,0), (0,2)] cropped by a box [(1,0),(2,2)] (XYXY format). The tight + # bounding box of the cropped triangle should be [(1,0),(2,1)], which is not equal to + # the intersection of original bounding box and the cropping box. + if self.recompute_boxes: + instances.gt_boxes = instances.gt_masks.get_bounding_boxes() + dataset_dict["instances"] = utils.filter_empty_instances(instances) + return dataset_dict diff --git a/detectron2/data/datasets/README.md b/detectron2/data/datasets/README.md new file mode 100644 index 0000000000000000000000000000000000000000..9fb3e4f7afec17137c95c78be6ef06d520ec8032 --- /dev/null +++ b/detectron2/data/datasets/README.md @@ -0,0 +1,9 @@ + + +### Common Datasets + +The dataset implemented here do not need to load the data into the final format. +It should provide the minimal data structure needed to use the dataset, so it can be very efficient. + +For example, for an image dataset, just provide the file names and labels, but don't read the images. +Let the downstream decide how to read. diff --git a/detectron2/data/datasets/__init__.py b/detectron2/data/datasets/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..dbd92e8e2e1295d73e28f1eb2ed2368f368849a3 --- /dev/null +++ b/detectron2/data/datasets/__init__.py @@ -0,0 +1,9 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .coco import load_coco_json, load_sem_seg, register_coco_instances +from .coco_panoptic import register_coco_panoptic, register_coco_panoptic_separated +from .lvis import load_lvis_json, register_lvis_instances, get_lvis_instances_meta +from .pascal_voc import load_voc_instances, register_pascal_voc +from . import builtin as _builtin # ensure the builtin datasets are registered + + +__all__ = [k for k in globals().keys() if not k.startswith("_")] diff --git a/detectron2/data/datasets/__pycache__/__init__.cpython-39.pyc b/detectron2/data/datasets/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2d4796162b092a43721fe69e4ba17f6968a313b3 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/builtin.cpython-39.pyc b/detectron2/data/datasets/__pycache__/builtin.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..32a84b1aac477f61f93f51a442674196d357a364 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/builtin.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/builtin_meta.cpython-39.pyc b/detectron2/data/datasets/__pycache__/builtin_meta.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5940593750785fd36644ec7359e1ad6e088271fe Binary files /dev/null and b/detectron2/data/datasets/__pycache__/builtin_meta.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/cityscapes.cpython-39.pyc b/detectron2/data/datasets/__pycache__/cityscapes.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..51f13421f492ac16436ac9c2db093a479e1d8491 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/cityscapes.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/cityscapes_panoptic.cpython-39.pyc b/detectron2/data/datasets/__pycache__/cityscapes_panoptic.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..50c55a5f123b4a8a3c18eccadcc77e3a2cd00099 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/cityscapes_panoptic.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/clip_prompt_utils.cpython-39.pyc b/detectron2/data/datasets/__pycache__/clip_prompt_utils.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..adfb59eb263901ea0f0731f8767a3b4e02320383 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/clip_prompt_utils.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/coco.cpython-39.pyc b/detectron2/data/datasets/__pycache__/coco.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..757cadb9ad435037586fe628fc5ea9ac0525a9a1 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/coco.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/coco_panoptic.cpython-39.pyc b/detectron2/data/datasets/__pycache__/coco_panoptic.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9efc4023b601892d1ba35d9ff2f4befe99bf0ad7 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/coco_panoptic.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/coco_zeroshot_categories.cpython-39.pyc b/detectron2/data/datasets/__pycache__/coco_zeroshot_categories.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9a4abd7eb4c345cf6a6cefee5a8c14a0111b651a Binary files /dev/null and b/detectron2/data/datasets/__pycache__/coco_zeroshot_categories.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/lvis.cpython-39.pyc b/detectron2/data/datasets/__pycache__/lvis.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..67a6dd3672b6a2ed1b4f782b8c5adbb8ec33bf75 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/lvis.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/lvis_v0_5_categories.cpython-39.pyc b/detectron2/data/datasets/__pycache__/lvis_v0_5_categories.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..85b4cdc2836e6a112a795443d87041ace60508de Binary files /dev/null and b/detectron2/data/datasets/__pycache__/lvis_v0_5_categories.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/lvis_v1_categories.cpython-39.pyc b/detectron2/data/datasets/__pycache__/lvis_v1_categories.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d0b5e9aee46633b2ff9c5d547210002fda2b1a46 Binary files /dev/null and b/detectron2/data/datasets/__pycache__/lvis_v1_categories.cpython-39.pyc differ diff --git a/detectron2/data/datasets/__pycache__/pascal_voc.cpython-39.pyc b/detectron2/data/datasets/__pycache__/pascal_voc.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0b260ea0a801956ba688bfb3b42ff905bc0a21fe Binary files /dev/null and b/detectron2/data/datasets/__pycache__/pascal_voc.cpython-39.pyc differ diff --git a/detectron2/data/datasets/bpe_simple_vocab_16e6.txt.gz b/detectron2/data/datasets/bpe_simple_vocab_16e6.txt.gz new file mode 100644 index 0000000000000000000000000000000000000000..36a15856e00a06a9fbed8cdd34d2393fea4a3113 --- /dev/null +++ b/detectron2/data/datasets/bpe_simple_vocab_16e6.txt.gz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a +size 1356917 diff --git a/detectron2/data/datasets/builtin.py b/detectron2/data/datasets/builtin.py new file mode 100644 index 0000000000000000000000000000000000000000..9eabfeecc0024017b25adccea02ac8b919236af1 --- /dev/null +++ b/detectron2/data/datasets/builtin.py @@ -0,0 +1,302 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + + +""" +This file registers pre-defined datasets at hard-coded paths, and their metadata. + +We hard-code metadata for common datasets. This will enable: +1. Consistency check when loading the datasets +2. Use models on these standard datasets directly and run demos, + without having to download the dataset annotations + +We hard-code some paths to the dataset that's assumed to +exist in "./datasets/". + +Users SHOULD NOT use this file to create new dataset / metadata for new dataset. +To add new dataset, refer to the tutorial "docs/DATASETS.md". +""" + +import os + +from detectron2.data import DatasetCatalog, MetadataCatalog + +from .builtin_meta import ADE20K_SEM_SEG_CATEGORIES, _get_builtin_metadata +from .cityscapes import load_cityscapes_instances, load_cityscapes_semantic +from .cityscapes_panoptic import register_all_cityscapes_panoptic +from .coco import load_sem_seg, register_coco_instances +from .coco_panoptic import register_coco_panoptic, register_coco_panoptic_separated +from .lvis import get_lvis_instances_meta, register_lvis_instances +from .pascal_voc import register_pascal_voc + +# ==== Predefined datasets and splits for COCO ========== + +_PREDEFINED_SPLITS_COCO = {} +_PREDEFINED_SPLITS_COCO["coco"] = { + "coco_2014_train": ("coco/train2014", "coco/annotations/instances_train2014.json"), + "coco_2014_val": ("coco/val2014", "coco/annotations/instances_val2014.json"), + "coco_2014_minival": ("coco/val2014", "coco/annotations/instances_minival2014.json"), + "coco_2014_minival_100": ("coco/val2014", "coco/annotations/instances_minival2014_100.json"), + "coco_2014_valminusminival": ( + "coco/val2014", + "coco/annotations/instances_valminusminival2014.json", + ), + "coco_2017_train": ("coco/train2017", "coco/annotations/instances_train2017.json"), + "coco_2017_val": ("coco/val2017", "coco/annotations/instances_val2017.json"), + "coco_2017_test": ("coco/test2017", "coco/annotations/image_info_test2017.json"), + "coco_2017_test-dev": ("coco/test2017", "coco/annotations/image_info_test-dev2017.json"), + "coco_2017_val_100": ("coco/val2017", "coco/annotations/instances_val2017_100.json"), +} +_PREDEFINED_SPLITS_COCO["coco_ovd"] = { + "coco_2017_ovd_all_train": ("coco/train2017", "coco/annotations/ovd_ins_train2017_all.json"), + "coco_2017_ovd_b_train": ("coco/train2017", "coco/annotations/ovd_ins_train2017_b.json"), + "coco_2017_ovd_t_train": ("coco/train2017", "coco/annotations/ovd_ins_train2017_t.json"), + "coco_2017_ovd_all_test": ("coco/val2017", "coco/annotations/ovd_ins_val2017_all.json"), + "coco_2017_ovd_b_test": ("coco/val2017", "coco/annotations/ovd_ins_val2017_b.json"), + "coco_2017_ovd_t_test": ("coco/val2017", "coco/annotations/ovd_ins_val2017_t.json"), +} + +# zeroshot inference of grounding evaluation +_PREDEFINED_SPLITS_FLICKR30K = {} +_PREDEFINED_SPLITS_FLICKR30K["yfcc100m"] = { + "flickr30k_train": ('flickr30k_images', "flickr30k_anno", "split/train.txt"), + "flickr30k_val": ('flickr30k_images', "flickr30k_anno", "split/val.txt"), + "flickr30k_test": ('flickr30k_images', "flickr30k_anno", "split/test.txt"), +} + +_PREDEFINED_SPLITS_COCO["coco_person"] = { + "keypoints_coco_2014_train": ( + "coco/train2014", + "coco/annotations/person_keypoints_train2014.json", + ), + "keypoints_coco_2014_val": ("coco/val2014", "coco/annotations/person_keypoints_val2014.json"), + "keypoints_coco_2014_minival": ( + "coco/val2014", + "coco/annotations/person_keypoints_minival2014.json", + ), + "keypoints_coco_2014_valminusminival": ( + "coco/val2014", + "coco/annotations/person_keypoints_valminusminival2014.json", + ), + "keypoints_coco_2014_minival_100": ( + "coco/val2014", + "coco/annotations/person_keypoints_minival2014_100.json", + ), + "keypoints_coco_2017_train": ( + "coco/train2017", + "coco/annotations/person_keypoints_train2017.json", + ), + "keypoints_coco_2017_val": ("coco/val2017", "coco/annotations/person_keypoints_val2017.json"), + "keypoints_coco_2017_val_100": ( + "coco/val2017", + "coco/annotations/person_keypoints_val2017_100.json", + ), +} + + +_PREDEFINED_SPLITS_COCO_PANOPTIC = { + "coco_2017_train_panoptic": ( + # This is the original panoptic annotation directory + "coco/panoptic_train2017", + "coco/annotations/panoptic_train2017.json", + # This directory contains semantic annotations that are + # converted from panoptic annotations. + # It is used by PanopticFPN. + # You can use the script at detectron2/datasets/prepare_panoptic_fpn.py + # to create these directories. + "coco/panoptic_stuff_train2017", + ), + "coco_2017_val_panoptic": ( + "coco/panoptic_val2017", + "coco/annotations/panoptic_val2017.json", + "coco/panoptic_stuff_val2017", + ), + "coco_2017_val_100_panoptic": ( + "coco/panoptic_val2017_100", + "coco/annotations/panoptic_val2017_100.json", + "coco/panoptic_stuff_val2017_100", + ), +} + + +def register_all_coco(root): + for dataset_name, splits_per_dataset in _PREDEFINED_SPLITS_COCO.items(): + if dataset_name == 'coco_ovd': # for zero-shot split + for key, (image_root, json_file) in splits_per_dataset.items(): + # Assume pre-defined datasets live in `./datasets`. + register_coco_instances( + key, + {}, # empty metadata, it will be overwritten in load_coco_json() function + os.path.join(root, json_file) if "://" not in json_file else json_file, + os.path.join(root, image_root), + ) + else: # default splits + for key, (image_root, json_file) in splits_per_dataset.items(): + # Assume pre-defined datasets live in `./datasets`. + register_coco_instances( + key, + _get_builtin_metadata(dataset_name), + os.path.join(root, json_file) if "://" not in json_file else json_file, + os.path.join(root, image_root), + ) + + for ( + prefix, + (panoptic_root, panoptic_json, semantic_root), + ) in _PREDEFINED_SPLITS_COCO_PANOPTIC.items(): + prefix_instances = prefix[: -len("_panoptic")] + instances_meta = MetadataCatalog.get(prefix_instances) + image_root, instances_json = instances_meta.image_root, instances_meta.json_file + # The "separated" version of COCO panoptic segmentation dataset, + # e.g. used by Panoptic FPN + register_coco_panoptic_separated( + prefix, + _get_builtin_metadata("coco_panoptic_separated"), + image_root, + os.path.join(root, panoptic_root), + os.path.join(root, panoptic_json), + os.path.join(root, semantic_root), + instances_json, + ) + # The "standard" version of COCO panoptic segmentation dataset, + # e.g. used by Panoptic-DeepLab + register_coco_panoptic( + prefix, + _get_builtin_metadata("coco_panoptic_standard"), + image_root, + os.path.join(root, panoptic_root), + os.path.join(root, panoptic_json), + instances_json, + ) + +# ==== Predefined datasets and splits for LVIS ========== + +def register_all_flickr30k(): + MetadataCatalog.get('yfcc100m').set(evaluator_type="flickr30k") + + +# ==== Predefined datasets and splits for LVIS ========== + + +_PREDEFINED_SPLITS_LVIS = { + "lvis_v1": { + "lvis_v1_train": ("coco/", "lvis/lvis_v1_train.json"), + "lvis_v1_val": ("coco/", "lvis/lvis_v1_val.json"), + "lvis_v1_test_dev": ("coco/", "lvis/lvis_v1_image_info_test_dev.json"), + "lvis_v1_test_challenge": ("coco/", "lvis/lvis_v1_image_info_test_challenge.json"), + }, + "lvis_v1_zeroshot": { + "lvis_v1_train_zeroshot": ("coco/", "lvis/lvis_v1_train.json"), + "lvis_v1_val_zeroshot": ("coco/", "lvis/lvis_v1_val.json"), + "lvis_v1_test_dev_zeroshot": ("coco/", "lvis/lvis_v1_image_info_test_dev.json"), + "lvis_v1_test_challenge_zeroshot": ("coco/", "lvis/lvis_v1_image_info_test_challenge.json"), + }, + "lvis_v0.5": { + "lvis_v0.5_train": ("coco/", "lvis/lvis_v0.5_train.json"), + "lvis_v0.5_val": ("coco/", "lvis/lvis_v0.5_val.json"), + "lvis_v0.5_val_rand_100": ("coco/", "lvis/lvis_v0.5_val_rand_100.json"), + "lvis_v0.5_test": ("coco/", "lvis/lvis_v0.5_image_info_test.json"), + }, + "lvis_v0.5_cocofied": { + "lvis_v0.5_train_cocofied": ("coco/", "lvis/lvis_v0.5_train_cocofied.json"), + "lvis_v0.5_val_cocofied": ("coco/", "lvis/lvis_v0.5_val_cocofied.json"), + }, +} + + +def register_all_lvis(root): + for dataset_name, splits_per_dataset in _PREDEFINED_SPLITS_LVIS.items(): + for key, (image_root, json_file) in splits_per_dataset.items(): + register_lvis_instances( + key, + get_lvis_instances_meta(dataset_name), + os.path.join(root, json_file) if "://" not in json_file else json_file, + os.path.join(root, image_root), + ) + + +# ==== Predefined splits for raw cityscapes images =========== +_RAW_CITYSCAPES_SPLITS = { + "cityscapes_fine_{task}_train": ("cityscapes/leftImg8bit/train/", "cityscapes/gtFine/train/"), + "cityscapes_fine_{task}_val": ("cityscapes/leftImg8bit/val/", "cityscapes/gtFine/val/"), + "cityscapes_fine_{task}_test": ("cityscapes/leftImg8bit/test/", "cityscapes/gtFine/test/"), +} + + +def register_all_cityscapes(root): + for key, (image_dir, gt_dir) in _RAW_CITYSCAPES_SPLITS.items(): + meta = _get_builtin_metadata("cityscapes") + image_dir = os.path.join(root, image_dir) + gt_dir = os.path.join(root, gt_dir) + + inst_key = key.format(task="instance_seg") + DatasetCatalog.register( + inst_key, + lambda x=image_dir, y=gt_dir: load_cityscapes_instances( + x, y, from_json=True, to_polygons=True + ), + ) + MetadataCatalog.get(inst_key).set( + image_dir=image_dir, gt_dir=gt_dir, evaluator_type="cityscapes_instance", **meta + ) + + sem_key = key.format(task="sem_seg") + DatasetCatalog.register( + sem_key, lambda x=image_dir, y=gt_dir: load_cityscapes_semantic(x, y) + ) + MetadataCatalog.get(sem_key).set( + image_dir=image_dir, + gt_dir=gt_dir, + evaluator_type="cityscapes_sem_seg", + ignore_label=255, + **meta, + ) + + +# ==== Predefined splits for PASCAL VOC =========== +def register_all_pascal_voc(root): + SPLITS = [ + ("voc_2007_trainval", "VOC2007", "trainval"), + ("voc_2007_train", "VOC2007", "train"), + ("voc_2007_val", "VOC2007", "val"), + ("voc_2007_test", "VOC2007", "test"), + ("voc_2012_trainval", "VOC2012", "trainval"), + ("voc_2012_train", "VOC2012", "train"), + ("voc_2012_val", "VOC2012", "val"), + ] + for name, dirname, split in SPLITS: + year = 2007 if "2007" in name else 2012 + register_pascal_voc(name, os.path.join(root, dirname), split, year) + MetadataCatalog.get(name).evaluator_type = "pascal_voc" + + +def register_all_ade20k(root): + root = os.path.join(root, "ADEChallengeData2016") + for name, dirname in [("train", "training"), ("val", "validation")]: + image_dir = os.path.join(root, "images", dirname) + gt_dir = os.path.join(root, "annotations_detectron2", dirname) + name = f"ade20k_sem_seg_{name}" + DatasetCatalog.register( + name, lambda x=image_dir, y=gt_dir: load_sem_seg(y, x, gt_ext="png", image_ext="jpg") + ) + MetadataCatalog.get(name).set( + stuff_classes=ADE20K_SEM_SEG_CATEGORIES[:], + image_root=image_dir, + sem_seg_root=gt_dir, + evaluator_type="sem_seg", + ignore_label=255, + ) + + +# True for open source; +# Internally at fb, we register them elsewhere +if __name__.endswith(".builtin"): + # Assume pre-defined datasets live in `./datasets`. + _root = os.getenv("DETECTRON2_DATASETS", "datasets") + register_all_coco(_root) + register_all_lvis(_root) + register_all_cityscapes(_root) + register_all_cityscapes_panoptic(_root) + register_all_pascal_voc(_root) + register_all_ade20k(_root) + register_all_flickr30k() diff --git a/detectron2/data/datasets/builtin_meta.py b/detectron2/data/datasets/builtin_meta.py new file mode 100644 index 0000000000000000000000000000000000000000..68b325f4c0326196736d39d6214898e38576e3c0 --- /dev/null +++ b/detectron2/data/datasets/builtin_meta.py @@ -0,0 +1,560 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +""" +Note: +For your custom dataset, there is no need to hard-code metadata anywhere in the code. +For example, for COCO-format dataset, metadata will be obtained automatically +when calling `load_coco_json`. For other dataset, metadata may also be obtained in other ways +during loading. + +However, we hard-coded metadata for a few common dataset here. +The only goal is to allow users who don't have these dataset to use pre-trained models. +Users don't have to download a COCO json (which contains metadata), in order to visualize a +COCO model (with correct class names and colors). +""" +# meta data for 65-48-17 zeroshot split of COCO +COCO_OVD_CATEGORIES = { + 'target': [ + {"color": [106, 0, 228], "isthing": 1, "id": 5, "name": "airplane"}, + {"color": [0, 60, 100], "isthing": 1, "id": 6, "name": "bus"}, + {"color": [255, 77, 255], "isthing": 1, "id": 17, "name": "cat"}, + {"color": [0, 226, 252], "isthing": 1, "id": 18, "name": "dog"}, + {"color": [120, 166, 157], "isthing": 1, "id": 21, "name": "cow"}, + {"color": [110, 76, 0], "isthing": 1, "id": 22, "name": "elephant"}, + {"color": [0, 125, 92], "isthing": 1, "id": 28, "name": "umbrella"}, + {"color": [188, 208, 182], "isthing": 1, "id": 32, "name": "tie"}, + {"color": [133, 129, 255], "isthing": 1, "id": 36, "name": "snowboard"}, + {"color": [134, 134, 103], "isthing": 1, "id": 41, "name": "skateboard"}, + {"color": [109, 63, 54], "isthing": 1, "id": 47, "name": "cup"}, + {"color": [151, 0, 95], "isthing": 1, "id": 49, "name": "knife"}, + {"color": [147, 186, 208], "isthing": 1, "id": 61, "name": "cake"}, + {"color": [3, 95, 161], "isthing": 1, "id": 63, "name": "couch"}, + {"color": [166, 74, 118], "isthing": 1, "id": 76, "name": "keyboard"}, + {"color": [127, 167, 115], "isthing": 1, "id": 81, "name": "sink"}, + {"color": [128, 76, 255], "isthing": 1, "id": 87, "name": "scissors"}, + ], + 'base': [ + {"color": [220, 20, 60], "isthing": 1, "id": 1, "name": "person"}, + {"color": [119, 11, 32], "isthing": 1, "id": 2, "name": "bicycle"}, + {"color": [0, 0, 142], "isthing": 1, "id": 3, "name": "car"}, + {"color": [0, 0, 230], "isthing": 1, "id": 4, "name": "motorcycle"}, + {"color": [0, 80, 100], "isthing": 1, "id": 7, "name": "train"}, + {"color": [0, 0, 70], "isthing": 1, "id": 8, "name": "truck"}, + {"color": [0, 0, 192], "isthing": 1, "id": 9, "name": "boat"}, + {"color": [250, 0, 30], "isthing": 1, "id": 15, "name": "bench"}, + {"color": [165, 42, 42], "isthing": 1, "id": 16, "name": "bird"}, + {"color": [182, 182, 255], "isthing": 1, "id": 19, "name": "horse"}, + {"color": [0, 82, 0], "isthing": 1, "id": 20, "name": "sheep"}, + {"color": [174, 57, 255], "isthing": 1, "id": 23, "name": "bear"}, + {"color": [199, 100, 0], "isthing": 1, "id": 24, "name": "zebra"}, + {"color": [72, 0, 118], "isthing": 1, "id": 25, "name": "giraffe"}, + {"color": [255, 179, 240], "isthing": 1, "id": 27, "name": "backpack"}, + {"color": [209, 0, 151], "isthing": 1, "id": 31, "name": "handbag"}, + {"color": [0, 220, 176], "isthing": 1, "id": 33, "name": "suitcase"}, + {"color": [255, 99, 164], "isthing": 1, "id": 34, "name": "frisbee"}, + {"color": [92, 0, 73], "isthing": 1, "id": 35, "name": "skis"}, + {"color": [0, 228, 0], "isthing": 1, "id": 38, "name": "kite"}, + {"color": [145, 148, 174], "isthing": 1, "id": 42, "name": "surfboard"}, + {"color": [197, 226, 255], "isthing": 1, "id": 44, "name": "bottle"}, + {"color": [207, 138, 255], "isthing": 1, "id": 48, "name": "fork"}, + {"color": [9, 80, 61], "isthing": 1, "id": 50, "name": "spoon"}, + {"color": [84, 105, 51], "isthing": 1, "id": 51, "name": "bowl"}, + {"color": [74, 65, 105], "isthing": 1, "id": 52, "name": "banana"}, + {"color": [166, 196, 102], "isthing": 1, "id": 53, "name": "apple"}, + {"color": [208, 195, 210], "isthing": 1, "id": 54, "name": "sandwich"}, + {"color": [255, 109, 65], "isthing": 1, "id": 55, "name": "orange"}, + {"color": [0, 143, 149], "isthing": 1, "id": 56, "name": "broccoli"}, + {"color": [179, 0, 194], "isthing": 1, "id": 57, "name": "carrot"}, + {"color": [5, 121, 0], "isthing": 1, "id": 59, "name": "pizza"}, + {"color": [227, 255, 205], "isthing": 1, "id": 60, "name": "donut"}, + {"color": [153, 69, 1], "isthing": 1, "id": 62, "name": "chair"}, + {"color": [119, 0, 170], "isthing": 1, "id": 65, "name": "bed"}, + {"color": [0, 165, 120], "isthing": 1, "id": 70, "name": "toilet"}, + {"color": [183, 130, 88], "isthing": 1, "id": 72, "name": "tv"}, + {"color": [95, 32, 0], "isthing": 1, "id": 73, "name": "laptop"}, + {"color": [130, 114, 135], "isthing": 1, "id": 74, "name": "mouse"}, + {"color": [110, 129, 133], "isthing": 1, "id": 75, "name": "remote"}, + {"color": [79, 210, 114], "isthing": 1, "id": 78, "name": "microwave"}, + {"color": [178, 90, 62], "isthing": 1, "id": 79, "name": "oven"}, + {"color": [65, 70, 15], "isthing": 1, "id": 80, "name": "toaster"}, + {"color": [59, 105, 106], "isthing": 1, "id": 82, "name": "refrigerator"}, + {"color": [142, 108, 45], "isthing": 1, "id": 84, "name": "book"}, + {"color": [196, 172, 0], "isthing": 1, "id": 85, "name": "clock"}, + {"color": [95, 54, 80], "isthing": 1, "id": 86, "name": "vase"}, + {"color": [191, 162, 208], "isthing": 1, "id": 90, "name": "toothbrush"}, + ], + 'all': [ + {"color": [220, 20, 60], "isthing": 1, "id": 1, "name": "person"}, + {"color": [119, 11, 32], "isthing": 1, "id": 2, "name": "bicycle"}, + {"color": [0, 0, 142], "isthing": 1, "id": 3, "name": "car"}, + {"color": [0, 0, 230], "isthing": 1, "id": 4, "name": "motorcycle"}, + {"color": [106, 0, 228], "isthing": 1, "id": 5, "name": "airplane"}, + {"color": [0, 60, 100], "isthing": 1, "id": 6, "name": "bus"}, + {"color": [0, 80, 100], "isthing": 1, "id": 7, "name": "train"}, + {"color": [0, 0, 70], "isthing": 1, "id": 8, "name": "truck"}, + {"color": [0, 0, 192], "isthing": 1, "id": 9, "name": "boat"}, + {"color": [250, 0, 30], "isthing": 1, "id": 15, "name": "bench"}, + {"color": [165, 42, 42], "isthing": 1, "id": 16, "name": "bird"}, + {"color": [255, 77, 255], "isthing": 1, "id": 17, "name": "cat"}, + {"color": [0, 226, 252], "isthing": 1, "id": 18, "name": "dog"}, + {"color": [182, 182, 255], "isthing": 1, "id": 19, "name": "horse"}, + {"color": [0, 82, 0], "isthing": 1, "id": 20, "name": "sheep"}, + {"color": [120, 166, 157], "isthing": 1, "id": 21, "name": "cow"}, + {"color": [110, 76, 0], "isthing": 1, "id": 22, "name": "elephant"}, + {"color": [174, 57, 255], "isthing": 1, "id": 23, "name": "bear"}, + {"color": [199, 100, 0], "isthing": 1, "id": 24, "name": "zebra"}, + {"color": [72, 0, 118], "isthing": 1, "id": 25, "name": "giraffe"}, + {"color": [255, 179, 240], "isthing": 1, "id": 27, "name": "backpack"}, + {"color": [0, 125, 92], "isthing": 1, "id": 28, "name": "umbrella"}, + {"color": [209, 0, 151], "isthing": 1, "id": 31, "name": "handbag"}, + {"color": [188, 208, 182], "isthing": 1, "id": 32, "name": "tie"}, + {"color": [0, 220, 176], "isthing": 1, "id": 33, "name": "suitcase"}, + {"color": [255, 99, 164], "isthing": 1, "id": 34, "name": "frisbee"}, + {"color": [92, 0, 73], "isthing": 1, "id": 35, "name": "skis"}, + {"color": [133, 129, 255], "isthing": 1, "id": 36, "name": "snowboard"}, + {"color": [0, 228, 0], "isthing": 1, "id": 38, "name": "kite"}, + {"color": [134, 134, 103], "isthing": 1, "id": 41, "name": "skateboard"}, + {"color": [145, 148, 174], "isthing": 1, "id": 42, "name": "surfboard"}, + {"color": [197, 226, 255], "isthing": 1, "id": 44, "name": "bottle"}, + {"color": [109, 63, 54], "isthing": 1, "id": 47, "name": "cup"}, + {"color": [207, 138, 255], "isthing": 1, "id": 48, "name": "fork"}, + {"color": [151, 0, 95], "isthing": 1, "id": 49, "name": "knife"}, + {"color": [9, 80, 61], "isthing": 1, "id": 50, "name": "spoon"}, + {"color": [84, 105, 51], "isthing": 1, "id": 51, "name": "bowl"}, + {"color": [74, 65, 105], "isthing": 1, "id": 52, "name": "banana"}, + {"color": [166, 196, 102], "isthing": 1, "id": 53, "name": "apple"}, + {"color": [208, 195, 210], "isthing": 1, "id": 54, "name": "sandwich"}, + {"color": [255, 109, 65], "isthing": 1, "id": 55, "name": "orange"}, + {"color": [0, 143, 149], "isthing": 1, "id": 56, "name": "broccoli"}, + {"color": [179, 0, 194], "isthing": 1, "id": 57, "name": "carrot"}, + {"color": [5, 121, 0], "isthing": 1, "id": 59, "name": "pizza"}, + {"color": [227, 255, 205], "isthing": 1, "id": 60, "name": "donut"}, + {"color": [147, 186, 208], "isthing": 1, "id": 61, "name": "cake"}, + {"color": [153, 69, 1], "isthing": 1, "id": 62, "name": "chair"}, + {"color": [3, 95, 161], "isthing": 1, "id": 63, "name": "couch"}, + {"color": [119, 0, 170], "isthing": 1, "id": 65, "name": "bed"}, + {"color": [0, 165, 120], "isthing": 1, "id": 70, "name": "toilet"}, + {"color": [183, 130, 88], "isthing": 1, "id": 72, "name": "tv"}, + {"color": [95, 32, 0], "isthing": 1, "id": 73, "name": "laptop"}, + {"color": [130, 114, 135], "isthing": 1, "id": 74, "name": "mouse"}, + {"color": [110, 129, 133], "isthing": 1, "id": 75, "name": "remote"}, + {"color": [166, 74, 118], "isthing": 1, "id": 76, "name": "keyboard"}, + {"color": [79, 210, 114], "isthing": 1, "id": 78, "name": "microwave"}, + {"color": [178, 90, 62], "isthing": 1, "id": 79, "name": "oven"}, + {"color": [65, 70, 15], "isthing": 1, "id": 80, "name": "toaster"}, + {"color": [127, 167, 115], "isthing": 1, "id": 81, "name": "sink"}, + {"color": [59, 105, 106], "isthing": 1, "id": 82, "name": "refrigerator"}, + {"color": [142, 108, 45], "isthing": 1, "id": 84, "name": "book"}, + {"color": [196, 172, 0], "isthing": 1, "id": 85, "name": "clock"}, + {"color": [95, 54, 80], "isthing": 1, "id": 86, "name": "vase"}, + {"color": [128, 76, 255], "isthing": 1, "id": 87, "name": "scissors"}, + {"color": [191, 162, 208], "isthing": 1, "id": 90, "name": "toothbrush"}, + ], + } + +# Classes not used in COCO_OVD_CATEGORIES +NOT_USED = [ + {"color": [250, 170, 30], "isthing": 1, "id": 10, "name": "traffic light"}, + {"color": [100, 170, 30], "isthing": 1, "id": 11, "name": "fire hydrant"}, + {"color": [220, 220, 0], "isthing": 1, "id": 13, "name": "stop sign"}, + {"color": [175, 116, 175], "isthing": 1, "id": 14, "name": "parking meter"}, + {"color": [78, 180, 255], "isthing": 1, "id": 37, "name": "sports ball"}, + {"color": [174, 255, 243], "isthing": 1, "id": 39, "name": "baseball bat"}, + {"color": [45, 89, 255], "isthing": 1, "id": 40, "name": "baseball glove"}, + {"color": [255, 208, 186], "isthing": 1, "id": 43, "name": "tennis racket"}, + {"color": [171, 134, 1], "isthing": 1, "id": 46, "name": "wine glass"}, + {"color": [209, 99, 106], "isthing": 1, "id": 58, "name": "hot dog"}, + {"color": [163, 255, 0], "isthing": 1, "id": 64, "name": "potted plant"}, + {"color": [0, 182, 199], "isthing": 1, "id": 67, "name": "dining table"}, + {"color": [219, 142, 185], "isthing": 1, "id": 77, "name": "cell phone"}, + {"color": [201, 57, 1], "isthing": 1, "id": 88, "name": "teddy bear"}, + {"color": [246, 0, 122], "isthing": 1, "id": 89, "name": "hair drier"}, + {"color": [255, 255, 128], "isthing": 0, "id": 92, "name": "banner"}, + {"color": [147, 211, 203], "isthing": 0, "id": 93, "name": "blanket"}, + {"color": [150, 100, 100], "isthing": 0, "id": 95, "name": "bridge"}, + {"color": [168, 171, 172], "isthing": 0, "id": 100, "name": "cardboard"}, + {"color": [146, 112, 198], "isthing": 0, "id": 107, "name": "counter"}, + {"color": [210, 170, 100], "isthing": 0, "id": 109, "name": "curtain"}, + {"color": [92, 136, 89], "isthing": 0, "id": 112, "name": "door-stuff"}, + {"color": [218, 88, 184], "isthing": 0, "id": 118, "name": "floor-wood"}, + {"color": [241, 129, 0], "isthing": 0, "id": 119, "name": "flower"}, + {"color": [217, 17, 255], "isthing": 0, "id": 122, "name": "fruit"}, + {"color": [124, 74, 181], "isthing": 0, "id": 125, "name": "gravel"}, + {"color": [70, 70, 70], "isthing": 0, "id": 128, "name": "house"}, + {"color": [255, 228, 255], "isthing": 0, "id": 130, "name": "light"}, + {"color": [154, 208, 0], "isthing": 0, "id": 133, "name": "mirror-stuff"}, + {"color": [193, 0, 92], "isthing": 0, "id": 138, "name": "net"}, + {"color": [76, 91, 113], "isthing": 0, "id": 141, "name": "pillow"}, + {"color": [255, 180, 195], "isthing": 0, "id": 144, "name": "platform"}, + {"color": [106, 154, 176], "isthing": 0, "id": 145, "name": "playingfield"}, + {"color": [230, 150, 140], "isthing": 0, "id": 147, "name": "railroad"}, + {"color": [60, 143, 255], "isthing": 0, "id": 148, "name": "river"}, + {"color": [128, 64, 128], "isthing": 0, "id": 149, "name": "road"}, + {"color": [92, 82, 55], "isthing": 0, "id": 151, "name": "roof"}, + {"color": [254, 212, 124], "isthing": 0, "id": 154, "name": "sand"}, + {"color": [73, 77, 174], "isthing": 0, "id": 155, "name": "sea"}, + {"color": [255, 160, 98], "isthing": 0, "id": 156, "name": "shelf"}, + {"color": [255, 255, 255], "isthing": 0, "id": 159, "name": "snow"}, + {"color": [104, 84, 109], "isthing": 0, "id": 161, "name": "stairs"}, + {"color": [169, 164, 131], "isthing": 0, "id": 166, "name": "tent"}, + {"color": [225, 199, 255], "isthing": 0, "id": 168, "name": "towel"}, + {"color": [135, 158, 223], "isthing": 0, "id": 175, "name": "wall-stone"}, + {"color": [137, 54, 74], "isthing": 0, "id": 171, "name": "wall-brick"}, + {"color": [7, 246, 231], "isthing": 0, "id": 176, "name": "wall-tile"}, + {"color": [107, 255, 200], "isthing": 0, "id": 177, "name": "wall-wood"}, + {"color": [58, 41, 149], "isthing": 0, "id": 178, "name": "water-other"}, + {"color": [183, 121, 142], "isthing": 0, "id": 180, "name": "window-blind"}, + {"color": [255, 73, 97], "isthing": 0, "id": 181, "name": "window-other"}, + {"color": [107, 142, 35], "isthing": 0, "id": 184, "name": "tree-merged"}, + {"color": [190, 153, 153], "isthing": 0, "id": 185, "name": "fence-merged"}, + {"color": [146, 139, 141], "isthing": 0, "id": 186, "name": "ceiling-merged"}, + {"color": [70, 130, 180], "isthing": 0, "id": 187, "name": "sky-other-merged"}, + {"color": [134, 199, 156], "isthing": 0, "id": 188, "name": "cabinet-merged"}, + {"color": [209, 226, 140], "isthing": 0, "id": 189, "name": "table-merged"}, + {"color": [96, 36, 108], "isthing": 0, "id": 190, "name": "floor-other-merged"}, + {"color": [96, 96, 96], "isthing": 0, "id": 191, "name": "pavement-merged"}, + {"color": [64, 170, 64], "isthing": 0, "id": 192, "name": "mountain-merged"}, + {"color": [152, 251, 152], "isthing": 0, "id": 193, "name": "grass-merged"}, + {"color": [208, 229, 228], "isthing": 0, "id": 194, "name": "dirt-merged"}, + {"color": [206, 186, 171], "isthing": 0, "id": 195, "name": "paper-merged"}, + {"color": [152, 161, 64], "isthing": 0, "id": 196, "name": "food-other-merged"}, + {"color": [116, 112, 0], "isthing": 0, "id": 197, "name": "building-other-merged"}, + {"color": [0, 114, 143], "isthing": 0, "id": 198, "name": "rock-merged"}, + {"color": [102, 102, 156], "isthing": 0, "id": 199, "name": "wall-other-merged"}, + {"color": [250, 141, 255], "isthing": 0, "id": 200, "name": "rug-merged"}, + ] + +# All coco categories, together with their nice-looking visualization colors +# It's from https://github.com/cocodataset/panopticapi/blob/master/panoptic_coco_categories.json +COCO_CATEGORIES = [ + {"color": [220, 20, 60], "isthing": 1, "id": 1, "name": "person"}, + {"color": [119, 11, 32], "isthing": 1, "id": 2, "name": "bicycle"}, + {"color": [0, 0, 142], "isthing": 1, "id": 3, "name": "car"}, + {"color": [0, 0, 230], "isthing": 1, "id": 4, "name": "motorcycle"}, + {"color": [106, 0, 228], "isthing": 1, "id": 5, "name": "airplane"}, + {"color": [0, 60, 100], "isthing": 1, "id": 6, "name": "bus"}, + {"color": [0, 80, 100], "isthing": 1, "id": 7, "name": "train"}, + {"color": [0, 0, 70], "isthing": 1, "id": 8, "name": "truck"}, + {"color": [0, 0, 192], "isthing": 1, "id": 9, "name": "boat"}, + {"color": [250, 170, 30], "isthing": 1, "id": 10, "name": "traffic light"}, + {"color": [100, 170, 30], "isthing": 1, "id": 11, "name": "fire hydrant"}, + {"color": [220, 220, 0], "isthing": 1, "id": 13, "name": "stop sign"}, + {"color": [175, 116, 175], "isthing": 1, "id": 14, "name": "parking meter"}, + {"color": [250, 0, 30], "isthing": 1, "id": 15, "name": "bench"}, + {"color": [165, 42, 42], "isthing": 1, "id": 16, "name": "bird"}, + {"color": [255, 77, 255], "isthing": 1, "id": 17, "name": "cat"}, + {"color": [0, 226, 252], "isthing": 1, "id": 18, "name": "dog"}, + {"color": [182, 182, 255], "isthing": 1, "id": 19, "name": "horse"}, + {"color": [0, 82, 0], "isthing": 1, "id": 20, "name": "sheep"}, + {"color": [120, 166, 157], "isthing": 1, "id": 21, "name": "cow"}, + {"color": [110, 76, 0], "isthing": 1, "id": 22, "name": "elephant"}, + {"color": [174, 57, 255], "isthing": 1, "id": 23, "name": "bear"}, + {"color": [199, 100, 0], "isthing": 1, "id": 24, "name": "zebra"}, + {"color": [72, 0, 118], "isthing": 1, "id": 25, "name": "giraffe"}, + {"color": [255, 179, 240], "isthing": 1, "id": 27, "name": "backpack"}, + {"color": [0, 125, 92], "isthing": 1, "id": 28, "name": "umbrella"}, + {"color": [209, 0, 151], "isthing": 1, "id": 31, "name": "handbag"}, + {"color": [188, 208, 182], "isthing": 1, "id": 32, "name": "tie"}, + {"color": [0, 220, 176], "isthing": 1, "id": 33, "name": "suitcase"}, + {"color": [255, 99, 164], "isthing": 1, "id": 34, "name": "frisbee"}, + {"color": [92, 0, 73], "isthing": 1, "id": 35, "name": "skis"}, + {"color": [133, 129, 255], "isthing": 1, "id": 36, "name": "snowboard"}, + {"color": [78, 180, 255], "isthing": 1, "id": 37, "name": "sports ball"}, + {"color": [0, 228, 0], "isthing": 1, "id": 38, "name": "kite"}, + {"color": [174, 255, 243], "isthing": 1, "id": 39, "name": "baseball bat"}, + {"color": [45, 89, 255], "isthing": 1, "id": 40, "name": "baseball glove"}, + {"color": [134, 134, 103], "isthing": 1, "id": 41, "name": "skateboard"}, + {"color": [145, 148, 174], "isthing": 1, "id": 42, "name": "surfboard"}, + {"color": [255, 208, 186], "isthing": 1, "id": 43, "name": "tennis racket"}, + {"color": [197, 226, 255], "isthing": 1, "id": 44, "name": "bottle"}, + {"color": [171, 134, 1], "isthing": 1, "id": 46, "name": "wine glass"}, + {"color": [109, 63, 54], "isthing": 1, "id": 47, "name": "cup"}, + {"color": [207, 138, 255], "isthing": 1, "id": 48, "name": "fork"}, + {"color": [151, 0, 95], "isthing": 1, "id": 49, "name": "knife"}, + {"color": [9, 80, 61], "isthing": 1, "id": 50, "name": "spoon"}, + {"color": [84, 105, 51], "isthing": 1, "id": 51, "name": "bowl"}, + {"color": [74, 65, 105], "isthing": 1, "id": 52, "name": "banana"}, + {"color": [166, 196, 102], "isthing": 1, "id": 53, "name": "apple"}, + {"color": [208, 195, 210], "isthing": 1, "id": 54, "name": "sandwich"}, + {"color": [255, 109, 65], "isthing": 1, "id": 55, "name": "orange"}, + {"color": [0, 143, 149], "isthing": 1, "id": 56, "name": "broccoli"}, + {"color": [179, 0, 194], "isthing": 1, "id": 57, "name": "carrot"}, + {"color": [209, 99, 106], "isthing": 1, "id": 58, "name": "hot dog"}, + {"color": [5, 121, 0], "isthing": 1, "id": 59, "name": "pizza"}, + {"color": [227, 255, 205], "isthing": 1, "id": 60, "name": "donut"}, + {"color": [147, 186, 208], "isthing": 1, "id": 61, "name": "cake"}, + {"color": [153, 69, 1], "isthing": 1, "id": 62, "name": "chair"}, + {"color": [3, 95, 161], "isthing": 1, "id": 63, "name": "couch"}, + {"color": [163, 255, 0], "isthing": 1, "id": 64, "name": "potted plant"}, + {"color": [119, 0, 170], "isthing": 1, "id": 65, "name": "bed"}, + {"color": [0, 182, 199], "isthing": 1, "id": 67, "name": "dining table"}, + {"color": [0, 165, 120], "isthing": 1, "id": 70, "name": "toilet"}, + {"color": [183, 130, 88], "isthing": 1, "id": 72, "name": "tv"}, + {"color": [95, 32, 0], "isthing": 1, "id": 73, "name": "laptop"}, + {"color": [130, 114, 135], "isthing": 1, "id": 74, "name": "mouse"}, + {"color": [110, 129, 133], "isthing": 1, "id": 75, "name": "remote"}, + {"color": [166, 74, 118], "isthing": 1, "id": 76, "name": "keyboard"}, + {"color": [219, 142, 185], "isthing": 1, "id": 77, "name": "cell phone"}, + {"color": [79, 210, 114], "isthing": 1, "id": 78, "name": "microwave"}, + {"color": [178, 90, 62], "isthing": 1, "id": 79, "name": "oven"}, + {"color": [65, 70, 15], "isthing": 1, "id": 80, "name": "toaster"}, + {"color": [127, 167, 115], "isthing": 1, "id": 81, "name": "sink"}, + {"color": [59, 105, 106], "isthing": 1, "id": 82, "name": "refrigerator"}, + {"color": [142, 108, 45], "isthing": 1, "id": 84, "name": "book"}, + {"color": [196, 172, 0], "isthing": 1, "id": 85, "name": "clock"}, + {"color": [95, 54, 80], "isthing": 1, "id": 86, "name": "vase"}, + {"color": [128, 76, 255], "isthing": 1, "id": 87, "name": "scissors"}, + {"color": [201, 57, 1], "isthing": 1, "id": 88, "name": "teddy bear"}, + {"color": [246, 0, 122], "isthing": 1, "id": 89, "name": "hair drier"}, + {"color": [191, 162, 208], "isthing": 1, "id": 90, "name": "toothbrush"}, + {"color": [255, 255, 128], "isthing": 0, "id": 92, "name": "banner"}, + {"color": [147, 211, 203], "isthing": 0, "id": 93, "name": "blanket"}, + {"color": [150, 100, 100], "isthing": 0, "id": 95, "name": "bridge"}, + {"color": [168, 171, 172], "isthing": 0, "id": 100, "name": "cardboard"}, + {"color": [146, 112, 198], "isthing": 0, "id": 107, "name": "counter"}, + {"color": [210, 170, 100], "isthing": 0, "id": 109, "name": "curtain"}, + {"color": [92, 136, 89], "isthing": 0, "id": 112, "name": "door-stuff"}, + {"color": [218, 88, 184], "isthing": 0, "id": 118, "name": "floor-wood"}, + {"color": [241, 129, 0], "isthing": 0, "id": 119, "name": "flower"}, + {"color": [217, 17, 255], "isthing": 0, "id": 122, "name": "fruit"}, + {"color": [124, 74, 181], "isthing": 0, "id": 125, "name": "gravel"}, + {"color": [70, 70, 70], "isthing": 0, "id": 128, "name": "house"}, + {"color": [255, 228, 255], "isthing": 0, "id": 130, "name": "light"}, + {"color": [154, 208, 0], "isthing": 0, "id": 133, "name": "mirror-stuff"}, + {"color": [193, 0, 92], "isthing": 0, "id": 138, "name": "net"}, + {"color": [76, 91, 113], "isthing": 0, "id": 141, "name": "pillow"}, + {"color": [255, 180, 195], "isthing": 0, "id": 144, "name": "platform"}, + {"color": [106, 154, 176], "isthing": 0, "id": 145, "name": "playingfield"}, + {"color": [230, 150, 140], "isthing": 0, "id": 147, "name": "railroad"}, + {"color": [60, 143, 255], "isthing": 0, "id": 148, "name": "river"}, + {"color": [128, 64, 128], "isthing": 0, "id": 149, "name": "road"}, + {"color": [92, 82, 55], "isthing": 0, "id": 151, "name": "roof"}, + {"color": [254, 212, 124], "isthing": 0, "id": 154, "name": "sand"}, + {"color": [73, 77, 174], "isthing": 0, "id": 155, "name": "sea"}, + {"color": [255, 160, 98], "isthing": 0, "id": 156, "name": "shelf"}, + {"color": [255, 255, 255], "isthing": 0, "id": 159, "name": "snow"}, + {"color": [104, 84, 109], "isthing": 0, "id": 161, "name": "stairs"}, + {"color": [169, 164, 131], "isthing": 0, "id": 166, "name": "tent"}, + {"color": [225, 199, 255], "isthing": 0, "id": 168, "name": "towel"}, + {"color": [137, 54, 74], "isthing": 0, "id": 171, "name": "wall-brick"}, + {"color": [135, 158, 223], "isthing": 0, "id": 175, "name": "wall-stone"}, + {"color": [7, 246, 231], "isthing": 0, "id": 176, "name": "wall-tile"}, + {"color": [107, 255, 200], "isthing": 0, "id": 177, "name": "wall-wood"}, + {"color": [58, 41, 149], "isthing": 0, "id": 178, "name": "water-other"}, + {"color": [183, 121, 142], "isthing": 0, "id": 180, "name": "window-blind"}, + {"color": [255, 73, 97], "isthing": 0, "id": 181, "name": "window-other"}, + {"color": [107, 142, 35], "isthing": 0, "id": 184, "name": "tree-merged"}, + {"color": [190, 153, 153], "isthing": 0, "id": 185, "name": "fence-merged"}, + {"color": [146, 139, 141], "isthing": 0, "id": 186, "name": "ceiling-merged"}, + {"color": [70, 130, 180], "isthing": 0, "id": 187, "name": "sky-other-merged"}, + {"color": [134, 199, 156], "isthing": 0, "id": 188, "name": "cabinet-merged"}, + {"color": [209, 226, 140], "isthing": 0, "id": 189, "name": "table-merged"}, + {"color": [96, 36, 108], "isthing": 0, "id": 190, "name": "floor-other-merged"}, + {"color": [96, 96, 96], "isthing": 0, "id": 191, "name": "pavement-merged"}, + {"color": [64, 170, 64], "isthing": 0, "id": 192, "name": "mountain-merged"}, + {"color": [152, 251, 152], "isthing": 0, "id": 193, "name": "grass-merged"}, + {"color": [208, 229, 228], "isthing": 0, "id": 194, "name": "dirt-merged"}, + {"color": [206, 186, 171], "isthing": 0, "id": 195, "name": "paper-merged"}, + {"color": [152, 161, 64], "isthing": 0, "id": 196, "name": "food-other-merged"}, + {"color": [116, 112, 0], "isthing": 0, "id": 197, "name": "building-other-merged"}, + {"color": [0, 114, 143], "isthing": 0, "id": 198, "name": "rock-merged"}, + {"color": [102, 102, 156], "isthing": 0, "id": 199, "name": "wall-other-merged"}, + {"color": [250, 141, 255], "isthing": 0, "id": 200, "name": "rug-merged"}, +] + +# fmt: off +COCO_PERSON_KEYPOINT_NAMES = ( + "nose", + "left_eye", "right_eye", + "left_ear", "right_ear", + "left_shoulder", "right_shoulder", + "left_elbow", "right_elbow", + "left_wrist", "right_wrist", + "left_hip", "right_hip", + "left_knee", "right_knee", + "left_ankle", "right_ankle", +) +# fmt: on + +# Pairs of keypoints that should be exchanged under horizontal flipping +COCO_PERSON_KEYPOINT_FLIP_MAP = ( + ("left_eye", "right_eye"), + ("left_ear", "right_ear"), + ("left_shoulder", "right_shoulder"), + ("left_elbow", "right_elbow"), + ("left_wrist", "right_wrist"), + ("left_hip", "right_hip"), + ("left_knee", "right_knee"), + ("left_ankle", "right_ankle"), +) + +# rules for pairs of keypoints to draw a line between, and the line color to use. +KEYPOINT_CONNECTION_RULES = [ + # face + ("left_ear", "left_eye", (102, 204, 255)), + ("right_ear", "right_eye", (51, 153, 255)), + ("left_eye", "nose", (102, 0, 204)), + ("nose", "right_eye", (51, 102, 255)), + # upper-body + ("left_shoulder", "right_shoulder", (255, 128, 0)), + ("left_shoulder", "left_elbow", (153, 255, 204)), + ("right_shoulder", "right_elbow", (128, 229, 255)), + ("left_elbow", "left_wrist", (153, 255, 153)), + ("right_elbow", "right_wrist", (102, 255, 224)), + # lower-body + ("left_hip", "right_hip", (255, 102, 0)), + ("left_hip", "left_knee", (255, 255, 77)), + ("right_hip", "right_knee", (153, 255, 204)), + ("left_knee", "left_ankle", (191, 255, 128)), + ("right_knee", "right_ankle", (255, 195, 77)), +] + +# All Cityscapes categories, together with their nice-looking visualization colors +# It's from https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/helpers/labels.py # noqa +CITYSCAPES_CATEGORIES = [ + {"color": (128, 64, 128), "isthing": 0, "id": 7, "trainId": 0, "name": "road"}, + {"color": (244, 35, 232), "isthing": 0, "id": 8, "trainId": 1, "name": "sidewalk"}, + {"color": (70, 70, 70), "isthing": 0, "id": 11, "trainId": 2, "name": "building"}, + {"color": (102, 102, 156), "isthing": 0, "id": 12, "trainId": 3, "name": "wall"}, + {"color": (190, 153, 153), "isthing": 0, "id": 13, "trainId": 4, "name": "fence"}, + {"color": (153, 153, 153), "isthing": 0, "id": 17, "trainId": 5, "name": "pole"}, + {"color": (250, 170, 30), "isthing": 0, "id": 19, "trainId": 6, "name": "traffic light"}, + {"color": (220, 220, 0), "isthing": 0, "id": 20, "trainId": 7, "name": "traffic sign"}, + {"color": (107, 142, 35), "isthing": 0, "id": 21, "trainId": 8, "name": "vegetation"}, + {"color": (152, 251, 152), "isthing": 0, "id": 22, "trainId": 9, "name": "terrain"}, + {"color": (70, 130, 180), "isthing": 0, "id": 23, "trainId": 10, "name": "sky"}, + {"color": (220, 20, 60), "isthing": 1, "id": 24, "trainId": 11, "name": "person"}, + {"color": (255, 0, 0), "isthing": 1, "id": 25, "trainId": 12, "name": "rider"}, + {"color": (0, 0, 142), "isthing": 1, "id": 26, "trainId": 13, "name": "car"}, + {"color": (0, 0, 70), "isthing": 1, "id": 27, "trainId": 14, "name": "truck"}, + {"color": (0, 60, 100), "isthing": 1, "id": 28, "trainId": 15, "name": "bus"}, + {"color": (0, 80, 100), "isthing": 1, "id": 31, "trainId": 16, "name": "train"}, + {"color": (0, 0, 230), "isthing": 1, "id": 32, "trainId": 17, "name": "motorcycle"}, + {"color": (119, 11, 32), "isthing": 1, "id": 33, "trainId": 18, "name": "bicycle"}, +] + +# fmt: off +ADE20K_SEM_SEG_CATEGORIES = [ + "wall", "building", "sky", "floor", "tree", "ceiling", "road, route", "bed", "window ", "grass", "cabinet", "sidewalk, pavement", "person", "earth, ground", "door", "table", "mountain, mount", "plant", "curtain", "chair", "car", "water", "painting, picture", "sofa", "shelf", "house", "sea", "mirror", "rug", "field", "armchair", "seat", "fence", "desk", "rock, stone", "wardrobe, closet, press", "lamp", "tub", "rail", "cushion", "base, pedestal, stand", "box", "column, pillar", "signboard, sign", "chest of drawers, chest, bureau, dresser", "counter", "sand", "sink", "skyscraper", "fireplace", "refrigerator, icebox", "grandstand, covered stand", "path", "stairs", "runway", "case, display case, showcase, vitrine", "pool table, billiard table, snooker table", "pillow", "screen door, screen", "stairway, staircase", "river", "bridge, span", "bookcase", "blind, screen", "coffee table", "toilet, can, commode, crapper, pot, potty, stool, throne", "flower", "book", "hill", "bench", "countertop", "stove", "palm, palm tree", "kitchen island", "computer", "swivel chair", "boat", "bar", "arcade machine", "hovel, hut, hutch, shack, shanty", "bus", "towel", "light", "truck", "tower", "chandelier", "awning, sunshade, sunblind", "street lamp", "booth", "tv", "plane", "dirt track", "clothes", "pole", "land, ground, soil", "bannister, banister, balustrade, balusters, handrail", "escalator, moving staircase, moving stairway", "ottoman, pouf, pouffe, puff, hassock", "bottle", "buffet, counter, sideboard", "poster, posting, placard, notice, bill, card", "stage", "van", "ship", "fountain", "conveyer belt, conveyor belt, conveyer, conveyor, transporter", "canopy", "washer, automatic washer, washing machine", "plaything, toy", "pool", "stool", "barrel, cask", "basket, handbasket", "falls", "tent", "bag", "minibike, motorbike", "cradle", "oven", "ball", "food, solid food", "step, stair", "tank, storage tank", "trade name", "microwave", "pot", "animal", "bicycle", "lake", "dishwasher", "screen", "blanket, cover", "sculpture", "hood, exhaust hood", "sconce", "vase", "traffic light", "tray", "trash can", "fan", "pier", "crt screen", "plate", "monitor", "bulletin board", "shower", "radiator", "glass, drinking glass", "clock", "flag", # noqa +] +# After processed by `prepare_ade20k_sem_seg.py`, id 255 means ignore +# fmt: on + + +def _get_coco_instances_meta(): + thing_ids = [k["id"] for k in COCO_CATEGORIES if k["isthing"] == 1] + thing_colors = [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 1] + assert len(thing_ids) == 80, len(thing_ids) + # Mapping from the incontiguous COCO category id to an id in [0, 79] + thing_dataset_id_to_contiguous_id = {k: i for i, k in enumerate(thing_ids)} + thing_classes = [k["name"] for k in COCO_CATEGORIES if k["isthing"] == 1] + ret = { + "thing_dataset_id_to_contiguous_id": thing_dataset_id_to_contiguous_id, + "thing_classes": thing_classes, + "thing_colors": thing_colors, + } + return ret + + +def _get_coco_panoptic_separated_meta(): + """ + Returns metadata for "separated" version of the panoptic segmentation dataset. + """ + stuff_ids = [k["id"] for k in COCO_CATEGORIES if k["isthing"] == 0] + assert len(stuff_ids) == 53, len(stuff_ids) + + # For semantic segmentation, this mapping maps from contiguous stuff id + # (in [0, 53], used in models) to ids in the dataset (used for processing results) + # The id 0 is mapped to an extra category "thing". + stuff_dataset_id_to_contiguous_id = {k: i + 1 for i, k in enumerate(stuff_ids)} + # When converting COCO panoptic annotations to semantic annotations + # We label the "thing" category to 0 + stuff_dataset_id_to_contiguous_id[0] = 0 + + # 54 names for COCO stuff categories (including "things") + stuff_classes = ["things"] + [ + k["name"].replace("-other", "").replace("-merged", "") + for k in COCO_CATEGORIES + if k["isthing"] == 0 + ] + + # NOTE: I randomly picked a color for things + stuff_colors = [[82, 18, 128]] + [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 0] + ret = { + "stuff_dataset_id_to_contiguous_id": stuff_dataset_id_to_contiguous_id, + "stuff_classes": stuff_classes, + "stuff_colors": stuff_colors, + } + ret.update(_get_coco_instances_meta()) + return ret + + +def _get_builtin_metadata(dataset_name): + if dataset_name == "coco": + return _get_coco_instances_meta() + if dataset_name == "coco_panoptic_separated": + return _get_coco_panoptic_separated_meta() + elif dataset_name == "coco_panoptic_standard": + meta = {} + # The following metadata maps contiguous id from [0, #thing categories + + # #stuff categories) to their names and colors. We have to replica of the + # same name and color under "thing_*" and "stuff_*" because the current + # visualization function in D2 handles thing and class classes differently + # due to some heuristic used in Panoptic FPN. We keep the same naming to + # enable reusing existing visualization functions. + thing_classes = [k["name"] for k in COCO_CATEGORIES] + thing_colors = [k["color"] for k in COCO_CATEGORIES] + stuff_classes = [k["name"] for k in COCO_CATEGORIES] + stuff_colors = [k["color"] for k in COCO_CATEGORIES] + + meta["thing_classes"] = thing_classes + meta["thing_colors"] = thing_colors + meta["stuff_classes"] = stuff_classes + meta["stuff_colors"] = stuff_colors + + # Convert category id for training: + # category id: like semantic segmentation, it is the class id for each + # pixel. Since there are some classes not used in evaluation, the category + # id is not always contiguous and thus we have two set of category ids: + # - original category id: category id in the original dataset, mainly + # used for evaluation. + # - contiguous category id: [0, #classes), in order to train the linear + # softmax classifier. + thing_dataset_id_to_contiguous_id = {} + stuff_dataset_id_to_contiguous_id = {} + + for i, cat in enumerate(COCO_CATEGORIES): + if cat["isthing"]: + thing_dataset_id_to_contiguous_id[cat["id"]] = i + else: + stuff_dataset_id_to_contiguous_id[cat["id"]] = i + + meta["thing_dataset_id_to_contiguous_id"] = thing_dataset_id_to_contiguous_id + meta["stuff_dataset_id_to_contiguous_id"] = stuff_dataset_id_to_contiguous_id + + return meta + elif dataset_name == "coco_person": + return { + "thing_classes": ["person"], + "keypoint_names": COCO_PERSON_KEYPOINT_NAMES, + "keypoint_flip_map": COCO_PERSON_KEYPOINT_FLIP_MAP, + "keypoint_connection_rules": KEYPOINT_CONNECTION_RULES, + } + elif dataset_name == "cityscapes": + # fmt: off + CITYSCAPES_THING_CLASSES = [ + "person", "rider", "car", "truck", + "bus", "train", "motorcycle", "bicycle", + ] + CITYSCAPES_STUFF_CLASSES = [ + "road", "sidewalk", "building", "wall", "fence", "pole", "traffic light", + "traffic sign", "vegetation", "terrain", "sky", "person", "rider", "car", + "truck", "bus", "train", "motorcycle", "bicycle", + ] + # fmt: on + return { + "thing_classes": CITYSCAPES_THING_CLASSES, + "stuff_classes": CITYSCAPES_STUFF_CLASSES, + } + raise KeyError("No built-in metadata for dataset {}".format(dataset_name)) diff --git a/detectron2/data/datasets/cityscapes.py b/detectron2/data/datasets/cityscapes.py new file mode 100644 index 0000000000000000000000000000000000000000..1e84a5bdb3d4e410d8eef4b80a5d4c099a180104 --- /dev/null +++ b/detectron2/data/datasets/cityscapes.py @@ -0,0 +1,329 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import functools +import json +import logging +import multiprocessing as mp +import numpy as np +import os +from itertools import chain +import pycocotools.mask as mask_util +from PIL import Image + +from detectron2.structures import BoxMode +from detectron2.utils.comm import get_world_size +from detectron2.utils.file_io import PathManager +from detectron2.utils.logger import setup_logger + +try: + import cv2 # noqa +except ImportError: + # OpenCV is an optional dependency at the moment + pass + + +logger = logging.getLogger(__name__) + + +def _get_cityscapes_files(image_dir, gt_dir): + files = [] + # scan through the directory + cities = PathManager.ls(image_dir) + logger.info(f"{len(cities)} cities found in '{image_dir}'.") + for city in cities: + city_img_dir = os.path.join(image_dir, city) + city_gt_dir = os.path.join(gt_dir, city) + for basename in PathManager.ls(city_img_dir): + image_file = os.path.join(city_img_dir, basename) + + suffix = "leftImg8bit.png" + assert basename.endswith(suffix), basename + basename = basename[: -len(suffix)] + + instance_file = os.path.join(city_gt_dir, basename + "gtFine_instanceIds.png") + label_file = os.path.join(city_gt_dir, basename + "gtFine_labelIds.png") + json_file = os.path.join(city_gt_dir, basename + "gtFine_polygons.json") + + files.append((image_file, instance_file, label_file, json_file)) + assert len(files), "No images found in {}".format(image_dir) + for f in files[0]: + assert PathManager.isfile(f), f + return files + + +def load_cityscapes_instances(image_dir, gt_dir, from_json=True, to_polygons=True): + """ + Args: + image_dir (str): path to the raw dataset. e.g., "~/cityscapes/leftImg8bit/train". + gt_dir (str): path to the raw annotations. e.g., "~/cityscapes/gtFine/train". + from_json (bool): whether to read annotations from the raw json file or the png files. + to_polygons (bool): whether to represent the segmentation as polygons + (COCO's format) instead of masks (cityscapes's format). + + Returns: + list[dict]: a list of dicts in Detectron2 standard format. (See + `Using Custom Datasets `_ ) + """ + if from_json: + assert to_polygons, ( + "Cityscapes's json annotations are in polygon format. " + "Converting to mask format is not supported now." + ) + files = _get_cityscapes_files(image_dir, gt_dir) + + logger.info("Preprocessing cityscapes annotations ...") + # This is still not fast: all workers will execute duplicate works and will + # take up to 10m on a 8GPU server. + pool = mp.Pool(processes=max(mp.cpu_count() // get_world_size() // 2, 4)) + + ret = pool.map( + functools.partial(_cityscapes_files_to_dict, from_json=from_json, to_polygons=to_polygons), + files, + ) + logger.info("Loaded {} images from {}".format(len(ret), image_dir)) + + # Map cityscape ids to contiguous ids + from cityscapesscripts.helpers.labels import labels + + labels = [l for l in labels if l.hasInstances and not l.ignoreInEval] + dataset_id_to_contiguous_id = {l.id: idx for idx, l in enumerate(labels)} + for dict_per_image in ret: + for anno in dict_per_image["annotations"]: + anno["category_id"] = dataset_id_to_contiguous_id[anno["category_id"]] + return ret + + +def load_cityscapes_semantic(image_dir, gt_dir): + """ + Args: + image_dir (str): path to the raw dataset. e.g., "~/cityscapes/leftImg8bit/train". + gt_dir (str): path to the raw annotations. e.g., "~/cityscapes/gtFine/train". + + Returns: + list[dict]: a list of dict, each has "file_name" and + "sem_seg_file_name". + """ + ret = [] + # gt_dir is small and contain many small files. make sense to fetch to local first + gt_dir = PathManager.get_local_path(gt_dir) + for image_file, _, label_file, json_file in _get_cityscapes_files(image_dir, gt_dir): + label_file = label_file.replace("labelIds", "labelTrainIds") + + with PathManager.open(json_file, "r") as f: + jsonobj = json.load(f) + ret.append( + { + "file_name": image_file, + "sem_seg_file_name": label_file, + "height": jsonobj["imgHeight"], + "width": jsonobj["imgWidth"], + } + ) + assert len(ret), f"No images found in {image_dir}!" + assert PathManager.isfile( + ret[0]["sem_seg_file_name"] + ), "Please generate labelTrainIds.png with cityscapesscripts/preparation/createTrainIdLabelImgs.py" # noqa + return ret + + +def _cityscapes_files_to_dict(files, from_json, to_polygons): + """ + Parse cityscapes annotation files to a instance segmentation dataset dict. + + Args: + files (tuple): consists of (image_file, instance_id_file, label_id_file, json_file) + from_json (bool): whether to read annotations from the raw json file or the png files. + to_polygons (bool): whether to represent the segmentation as polygons + (COCO's format) instead of masks (cityscapes's format). + + Returns: + A dict in Detectron2 Dataset format. + """ + from cityscapesscripts.helpers.labels import id2label, name2label + + image_file, instance_id_file, _, json_file = files + + annos = [] + + if from_json: + from shapely.geometry import MultiPolygon, Polygon + + with PathManager.open(json_file, "r") as f: + jsonobj = json.load(f) + ret = { + "file_name": image_file, + "image_id": os.path.basename(image_file), + "height": jsonobj["imgHeight"], + "width": jsonobj["imgWidth"], + } + + # `polygons_union` contains the union of all valid polygons. + polygons_union = Polygon() + + # CityscapesScripts draw the polygons in sequential order + # and each polygon *overwrites* existing ones. See + # (https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/preparation/json2instanceImg.py) # noqa + # We use reverse order, and each polygon *avoids* early ones. + # This will resolve the ploygon overlaps in the same way as CityscapesScripts. + for obj in jsonobj["objects"][::-1]: + if "deleted" in obj: # cityscapes data format specific + continue + label_name = obj["label"] + + try: + label = name2label[label_name] + except KeyError: + if label_name.endswith("group"): # crowd area + label = name2label[label_name[: -len("group")]] + else: + raise + if label.id < 0: # cityscapes data format + continue + + # Cityscapes's raw annotations uses integer coordinates + # Therefore +0.5 here + poly_coord = np.asarray(obj["polygon"], dtype="f4") + 0.5 + # CityscapesScript uses PIL.ImageDraw.polygon to rasterize + # polygons for evaluation. This function operates in integer space + # and draws each pixel whose center falls into the polygon. + # Therefore it draws a polygon which is 0.5 "fatter" in expectation. + # We therefore dilate the input polygon by 0.5 as our input. + poly = Polygon(poly_coord).buffer(0.5, resolution=4) + + if not label.hasInstances or label.ignoreInEval: + # even if we won't store the polygon it still contributes to overlaps resolution + polygons_union = polygons_union.union(poly) + continue + + # Take non-overlapping part of the polygon + poly_wo_overlaps = poly.difference(polygons_union) + if poly_wo_overlaps.is_empty: + continue + polygons_union = polygons_union.union(poly) + + anno = {} + anno["iscrowd"] = label_name.endswith("group") + anno["category_id"] = label.id + + if isinstance(poly_wo_overlaps, Polygon): + poly_list = [poly_wo_overlaps] + elif isinstance(poly_wo_overlaps, MultiPolygon): + poly_list = poly_wo_overlaps.geoms + else: + raise NotImplementedError("Unknown geometric structure {}".format(poly_wo_overlaps)) + + poly_coord = [] + for poly_el in poly_list: + # COCO API can work only with exterior boundaries now, hence we store only them. + # TODO: store both exterior and interior boundaries once other parts of the + # codebase support holes in polygons. + poly_coord.append(list(chain(*poly_el.exterior.coords))) + anno["segmentation"] = poly_coord + (xmin, ymin, xmax, ymax) = poly_wo_overlaps.bounds + + anno["bbox"] = (xmin, ymin, xmax, ymax) + anno["bbox_mode"] = BoxMode.XYXY_ABS + + annos.append(anno) + else: + # See also the official annotation parsing scripts at + # https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/instances2dict.py # noqa + with PathManager.open(instance_id_file, "rb") as f: + inst_image = np.asarray(Image.open(f), order="F") + # ids < 24 are stuff labels (filtering them first is about 5% faster) + flattened_ids = np.unique(inst_image[inst_image >= 24]) + + ret = { + "file_name": image_file, + "image_id": os.path.basename(image_file), + "height": inst_image.shape[0], + "width": inst_image.shape[1], + } + + for instance_id in flattened_ids: + # For non-crowd annotations, instance_id // 1000 is the label_id + # Crowd annotations have <1000 instance ids + label_id = instance_id // 1000 if instance_id >= 1000 else instance_id + label = id2label[label_id] + if not label.hasInstances or label.ignoreInEval: + continue + + anno = {} + anno["iscrowd"] = instance_id < 1000 + anno["category_id"] = label.id + + mask = np.asarray(inst_image == instance_id, dtype=np.uint8, order="F") + + inds = np.nonzero(mask) + ymin, ymax = inds[0].min(), inds[0].max() + xmin, xmax = inds[1].min(), inds[1].max() + anno["bbox"] = (xmin, ymin, xmax, ymax) + if xmax <= xmin or ymax <= ymin: + continue + anno["bbox_mode"] = BoxMode.XYXY_ABS + if to_polygons: + # This conversion comes from D4809743 and D5171122, + # when Mask-RCNN was first developed. + contours = cv2.findContours(mask.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)[ + -2 + ] + polygons = [c.reshape(-1).tolist() for c in contours if len(c) >= 3] + # opencv's can produce invalid polygons + if len(polygons) == 0: + continue + anno["segmentation"] = polygons + else: + anno["segmentation"] = mask_util.encode(mask[:, :, None])[0] + annos.append(anno) + ret["annotations"] = annos + return ret + + +if __name__ == "__main__": + """ + Test the cityscapes dataset loader. + + Usage: + python -m detectron2.data.datasets.cityscapes \ + cityscapes/leftImg8bit/train cityscapes/gtFine/train + """ + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("image_dir") + parser.add_argument("gt_dir") + parser.add_argument("--type", choices=["instance", "semantic"], default="instance") + args = parser.parse_args() + from detectron2.data.catalog import Metadata + from detectron2.utils.visualizer import Visualizer + from cityscapesscripts.helpers.labels import labels + + logger = setup_logger(name=__name__) + + dirname = "cityscapes-data-vis" + os.makedirs(dirname, exist_ok=True) + + if args.type == "instance": + dicts = load_cityscapes_instances( + args.image_dir, args.gt_dir, from_json=True, to_polygons=True + ) + logger.info("Done loading {} samples.".format(len(dicts))) + + thing_classes = [k.name for k in labels if k.hasInstances and not k.ignoreInEval] + meta = Metadata().set(thing_classes=thing_classes) + + else: + dicts = load_cityscapes_semantic(args.image_dir, args.gt_dir) + logger.info("Done loading {} samples.".format(len(dicts))) + + stuff_classes = [k.name for k in labels if k.trainId != 255] + stuff_colors = [k.color for k in labels if k.trainId != 255] + meta = Metadata().set(stuff_classes=stuff_classes, stuff_colors=stuff_colors) + + for d in dicts: + img = np.array(Image.open(PathManager.open(d["file_name"], "rb"))) + visualizer = Visualizer(img, metadata=meta) + vis = visualizer.draw_dataset_dict(d) + # cv2.imshow("a", vis.get_image()[:, :, ::-1]) + # cv2.waitKey() + fpath = os.path.join(dirname, os.path.basename(d["file_name"])) + vis.save(fpath) diff --git a/detectron2/data/datasets/cityscapes_panoptic.py b/detectron2/data/datasets/cityscapes_panoptic.py new file mode 100644 index 0000000000000000000000000000000000000000..48c136f1623261b079591065fec7c7fc38165076 --- /dev/null +++ b/detectron2/data/datasets/cityscapes_panoptic.py @@ -0,0 +1,187 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import json +import logging +import os + +from detectron2.data import DatasetCatalog, MetadataCatalog +from detectron2.data.datasets.builtin_meta import CITYSCAPES_CATEGORIES +from detectron2.utils.file_io import PathManager + +""" +This file contains functions to register the Cityscapes panoptic dataset to the DatasetCatalog. +""" + + +logger = logging.getLogger(__name__) + + +def get_cityscapes_panoptic_files(image_dir, gt_dir, json_info): + files = [] + # scan through the directory + cities = PathManager.ls(image_dir) + logger.info(f"{len(cities)} cities found in '{image_dir}'.") + image_dict = {} + for city in cities: + city_img_dir = os.path.join(image_dir, city) + for basename in PathManager.ls(city_img_dir): + image_file = os.path.join(city_img_dir, basename) + + suffix = "_leftImg8bit.png" + assert basename.endswith(suffix), basename + basename = os.path.basename(basename)[: -len(suffix)] + + image_dict[basename] = image_file + + for ann in json_info["annotations"]: + image_file = image_dict.get(ann["image_id"], None) + assert image_file is not None, "No image {} found for annotation {}".format( + ann["image_id"], ann["file_name"] + ) + label_file = os.path.join(gt_dir, ann["file_name"]) + segments_info = ann["segments_info"] + + files.append((image_file, label_file, segments_info)) + + assert len(files), "No images found in {}".format(image_dir) + assert PathManager.isfile(files[0][0]), files[0][0] + assert PathManager.isfile(files[0][1]), files[0][1] + return files + + +def load_cityscapes_panoptic(image_dir, gt_dir, gt_json, meta): + """ + Args: + image_dir (str): path to the raw dataset. e.g., "~/cityscapes/leftImg8bit/train". + gt_dir (str): path to the raw annotations. e.g., + "~/cityscapes/gtFine/cityscapes_panoptic_train". + gt_json (str): path to the json file. e.g., + "~/cityscapes/gtFine/cityscapes_panoptic_train.json". + meta (dict): dictionary containing "thing_dataset_id_to_contiguous_id" + and "stuff_dataset_id_to_contiguous_id" to map category ids to + contiguous ids for training. + + Returns: + list[dict]: a list of dicts in Detectron2 standard format. (See + `Using Custom Datasets `_ ) + """ + + def _convert_category_id(segment_info, meta): + if segment_info["category_id"] in meta["thing_dataset_id_to_contiguous_id"]: + segment_info["category_id"] = meta["thing_dataset_id_to_contiguous_id"][ + segment_info["category_id"] + ] + else: + segment_info["category_id"] = meta["stuff_dataset_id_to_contiguous_id"][ + segment_info["category_id"] + ] + return segment_info + + assert os.path.exists( + gt_json + ), "Please run `python cityscapesscripts/preparation/createPanopticImgs.py` to generate label files." # noqa + with open(gt_json) as f: + json_info = json.load(f) + files = get_cityscapes_panoptic_files(image_dir, gt_dir, json_info) + ret = [] + for image_file, label_file, segments_info in files: + sem_label_file = ( + image_file.replace("leftImg8bit", "gtFine").split(".")[0] + "_labelTrainIds.png" + ) + segments_info = [_convert_category_id(x, meta) for x in segments_info] + ret.append( + { + "file_name": image_file, + "image_id": "_".join( + os.path.splitext(os.path.basename(image_file))[0].split("_")[:3] + ), + "sem_seg_file_name": sem_label_file, + "pan_seg_file_name": label_file, + "segments_info": segments_info, + } + ) + assert len(ret), f"No images found in {image_dir}!" + assert PathManager.isfile( + ret[0]["sem_seg_file_name"] + ), "Please generate labelTrainIds.png with cityscapesscripts/preparation/createTrainIdLabelImgs.py" # noqa + assert PathManager.isfile( + ret[0]["pan_seg_file_name"] + ), "Please generate panoptic annotation with python cityscapesscripts/preparation/createPanopticImgs.py" # noqa + return ret + + +_RAW_CITYSCAPES_PANOPTIC_SPLITS = { + "cityscapes_fine_panoptic_train": ( + "cityscapes/leftImg8bit/train", + "cityscapes/gtFine/cityscapes_panoptic_train", + "cityscapes/gtFine/cityscapes_panoptic_train.json", + ), + "cityscapes_fine_panoptic_val": ( + "cityscapes/leftImg8bit/val", + "cityscapes/gtFine/cityscapes_panoptic_val", + "cityscapes/gtFine/cityscapes_panoptic_val.json", + ), + # "cityscapes_fine_panoptic_test": not supported yet +} + + +def register_all_cityscapes_panoptic(root): + meta = {} + # The following metadata maps contiguous id from [0, #thing categories + + # #stuff categories) to their names and colors. We have to replica of the + # same name and color under "thing_*" and "stuff_*" because the current + # visualization function in D2 handles thing and class classes differently + # due to some heuristic used in Panoptic FPN. We keep the same naming to + # enable reusing existing visualization functions. + thing_classes = [k["name"] for k in CITYSCAPES_CATEGORIES] + thing_colors = [k["color"] for k in CITYSCAPES_CATEGORIES] + stuff_classes = [k["name"] for k in CITYSCAPES_CATEGORIES] + stuff_colors = [k["color"] for k in CITYSCAPES_CATEGORIES] + + meta["thing_classes"] = thing_classes + meta["thing_colors"] = thing_colors + meta["stuff_classes"] = stuff_classes + meta["stuff_colors"] = stuff_colors + + # There are three types of ids in cityscapes panoptic segmentation: + # (1) category id: like semantic segmentation, it is the class id for each + # pixel. Since there are some classes not used in evaluation, the category + # id is not always contiguous and thus we have two set of category ids: + # - original category id: category id in the original dataset, mainly + # used for evaluation. + # - contiguous category id: [0, #classes), in order to train the classifier + # (2) instance id: this id is used to differentiate different instances from + # the same category. For "stuff" classes, the instance id is always 0; for + # "thing" classes, the instance id starts from 1 and 0 is reserved for + # ignored instances (e.g. crowd annotation). + # (3) panoptic id: this is the compact id that encode both category and + # instance id by: category_id * 1000 + instance_id. + thing_dataset_id_to_contiguous_id = {} + stuff_dataset_id_to_contiguous_id = {} + + for k in CITYSCAPES_CATEGORIES: + if k["isthing"] == 1: + thing_dataset_id_to_contiguous_id[k["id"]] = k["trainId"] + else: + stuff_dataset_id_to_contiguous_id[k["id"]] = k["trainId"] + + meta["thing_dataset_id_to_contiguous_id"] = thing_dataset_id_to_contiguous_id + meta["stuff_dataset_id_to_contiguous_id"] = stuff_dataset_id_to_contiguous_id + + for key, (image_dir, gt_dir, gt_json) in _RAW_CITYSCAPES_PANOPTIC_SPLITS.items(): + image_dir = os.path.join(root, image_dir) + gt_dir = os.path.join(root, gt_dir) + gt_json = os.path.join(root, gt_json) + + DatasetCatalog.register( + key, lambda x=image_dir, y=gt_dir, z=gt_json: load_cityscapes_panoptic(x, y, z, meta) + ) + MetadataCatalog.get(key).set( + panoptic_root=gt_dir, + image_root=image_dir, + panoptic_json=gt_json, + gt_dir=gt_dir.replace("cityscapes_panoptic_", ""), + evaluator_type="cityscapes_panoptic_seg", + ignore_label=255, + label_divisor=1000, + **meta, + ) diff --git a/detectron2/data/datasets/clip_prompt_utils.py b/detectron2/data/datasets/clip_prompt_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..665f4adaf7093bb50dfc90c686a20d7a646a53d2 --- /dev/null +++ b/detectron2/data/datasets/clip_prompt_utils.py @@ -0,0 +1,441 @@ +import gzip +import html +import os +from functools import lru_cache + +import ftfy +import regex as re +import torch +import numpy as np +from typing import Union, List + +from .lvis_v1_categories import LVIS_CATEGORIES as LVIS_V1_CATEGORIES +from .coco_zeroshot_categories import COCO_UNSEEN_CLS, COCO_SEEN_CLS, COCO_OVD_ALL_CLS, COCO_80_ALL_CLS + +# https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py +@lru_cache() +def default_bpe(): + return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz") + + +@lru_cache() +def bytes_to_unicode(): + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a signficant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1)) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8+n) + n += 1 + cs = [chr(n) for n in cs] + return dict(zip(bs, cs)) + + +def get_pairs(word): + """Return set of symbol pairs in a word. + Word is represented as tuple of symbols (symbols being variable-length strings). + """ + pairs = set() + prev_char = word[0] + for char in word[1:]: + pairs.add((prev_char, char)) + prev_char = char + return pairs + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r'\s+', ' ', text) + text = text.strip() + return text + + +class SimpleTokenizer(object): + def __init__(self, bpe_path: str = default_bpe()): + self.byte_encoder = bytes_to_unicode() + self.byte_decoder = {v: k for k, v in self.byte_encoder.items()} + merges = gzip.open(bpe_path).read().decode("utf-8").split('\n') + merges = merges[1:49152-256-2+1] + merges = [tuple(merge.split()) for merge in merges] + vocab = list(bytes_to_unicode().values()) + vocab = vocab + [v+'' for v in vocab] + self.vocab = vocab + for merge in merges: + vocab.append(''.join(merge)) + vocab.extend(['<|startoftext|>', '<|endoftext|>']) + self.encoder = dict(zip(vocab, range(len(vocab)))) + self.decoder = {v: k for k, v in self.encoder.items()} + self.bpe_ranks = dict(zip(merges, range(len(merges)))) + self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'} + self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE) + + def bpe(self, token): + if token in self.cache: + return self.cache[token] + word = tuple(token[:-1]) + ( token[-1] + '',) + pairs = get_pairs(word) + + if not pairs: + return token+'' + + while True: + bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf'))) + if bigram not in self.bpe_ranks: + break + first, second = bigram + new_word = [] + i = 0 + while i < len(word): + try: + j = word.index(first, i) + new_word.extend(word[i:j]) + i = j + except: + new_word.extend(word[i:]) + break + + if word[i] == first and i < len(word)-1 and word[i+1] == second: + new_word.append(first+second) + i += 2 + else: + new_word.append(word[i]) + i += 1 + new_word = tuple(new_word) + word = new_word + if len(word) == 1: + break + else: + pairs = get_pairs(word) + word = ' '.join(word) + self.cache[token] = word + return word + + def encode(self, text, return_link=False): + bpe_tokens = [] + text = whitespace_clean(basic_clean(text)).lower() + str2id_links = [] # link original sentence word to the tokenized ids of its subwords + for token in re.findall(self.pat, text): + this_link = [token] + token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8')) + ids = [self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' ')] + bpe_tokens.extend(ids) + this_link.append(ids) + str2id_links.append(this_link) + if return_link: + return bpe_tokens, str2id_links + return bpe_tokens + + def decode(self, tokens): + text = ''.join([self.decoder[token] for token in tokens]) + text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('', ' ') + return text + + +# https://github.com/openai/CLIP/blob/main/clip/clip.py +#_tokenizer = SimpleTokenizer() + +def tokenize(texts: Union[str, List[str]], context_length: int = 77): + if isinstance(texts, str): + texts = [texts] + + sot_token = _tokenizer.encoder["<|startoftext|>"] + eot_token = _tokenizer.encoder["<|endoftext|>"] + all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts] + result = torch.zeros(len(all_tokens), context_length, dtype=torch.long) + + for i, tokens in enumerate(all_tokens): + if len(tokens) > context_length: + raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}") + result[i, :len(tokens)] = torch.tensor(tokens) + + return result + + +# prompt_engineering.py +def get_prompt_templates(): + # prompt_templates = [ + # 'There is a {} in the scene.', + # 'There is the {} in the scene.', + # 'a photo of a {} in the scene.', + # 'a photo of the {} in the scene.', + # 'a photo of one {} in the scene.', + + # 'itap of a {}.', + # 'itap of my {}.', # itap: I took a picture of + # 'itap of the {}.', + # 'a photo of a {}.', + # 'a photo of my {}.', + # 'a photo of the {}.', + # 'a photo of one {}.', + # 'a photo of many {}.', + + # 'a good photo of a {}.', + # 'a good photo of the {}.', + # 'a bad photo of a {}.', + # 'a bad photo of the {}.', + # 'a photo of a nice {}.', + # 'a photo of the nice {}.', + # 'a photo of a cool {}.', + # 'a photo of the cool {}.', + # 'a photo of a weird {}.', + # 'a photo of the weird {}.', + + # 'a photo of a small {}.', + # 'a photo of the small {}.', + # 'a photo of a large {}.', + # 'a photo of the large {}.', + + # 'a photo of a clean {}.', + # 'a photo of the clean {}.', + # 'a photo of a dirty {}.', + # 'a photo of the dirty {}.', + + # 'a bright photo of a {}.', + # 'a bright photo of the {}.', + # 'a dark photo of a {}.', + # 'a dark photo of the {}.', + + # 'a photo of a hard to see {}.', + # 'a photo of the hard to see {}.', + # 'a low resolution photo of a {}.', + # 'a low resolution photo of the {}.', + # 'a cropped photo of a {}.', + # 'a cropped photo of the {}.', + # 'a close-up photo of a {}.', + # 'a close-up photo of the {}.', + # 'a jpeg corrupted photo of a {}.', + # 'a jpeg corrupted photo of the {}.', + # 'a blurry photo of a {}.', + # 'a blurry photo of the {}.', + # 'a pixelated photo of a {}.', + # 'a pixelated photo of the {}.', + + # 'a black and white photo of the {}.', + # 'a black and white photo of a {}.', + + # 'a plastic {}.', + # 'the plastic {}.', + + # 'a toy {}.', + # 'the toy {}.', + # 'a plushie {}.', + # 'the plushie {}.', + # 'a cartoon {}.', + # 'the cartoon {}.', + + # 'an embroidered {}.', + # 'the embroidered {}.', + + # 'a painting of the {}.', + # 'a painting of a {}.', + # ] + + prompt_templates = [ + '{}.', + 'a photo of a {}.', + 'a bad photo of a {}.', + 'a photo of many {}.', + 'a sculpture of a {}.', + 'a photo of the hard to see {}.', + 'a low resolution photo of the {}.', + 'a rendering of a {}.', + 'graffiti of a {}.', + 'a bad photo of the {}.', + 'a cropped photo of the {}.', + 'a tattoo of a {}.', + 'the embroidered {}.', + 'a photo of a hard to see {}.', + 'a bright photo of a {}.', + 'a photo of a clean {}.', + 'a photo of a dirty {}.', + 'a dark photo of the {}.', + 'a drawing of a {}.', + 'a photo of my {}.', + 'the plastic {}.', + 'a photo of the cool {}.', + 'a close-up photo of a {}.', + 'a black and white photo of the {}.', + 'a painting of the {}.', + 'a painting of a {}.', + 'a pixelated photo of the {}.', + 'a sculpture of the {}.', + 'a bright photo of the {}.', + 'a cropped photo of a {}.', + 'a plastic {}.', + 'a photo of the dirty {}.', + 'a jpeg corrupted photo of a {}.', + 'a blurry photo of the {}.', + 'a photo of the {}.', + 'a good photo of the {}.', + 'a rendering of the {}.', + 'a {} in a video game.', + 'a photo of one {}.', + 'a doodle of a {}.', + 'a close-up photo of the {}.', + 'the origami {}.', + 'the {} in a video game.', + 'a sketch of a {}.', + 'a doodle of the {}.', + 'a origami {}.', + 'a low resolution photo of a {}.', + 'the toy {}.', + 'a rendition of the {}.', + 'a photo of the clean {}.', + 'a photo of a large {}.', + 'a rendition of a {}.', + 'a photo of a nice {}.', + 'a photo of a weird {}.', + 'a blurry photo of a {}.', + 'a cartoon {}.', + 'art of a {}.', + 'a sketch of the {}.', + 'a embroidered {}.', + 'a pixelated photo of a {}.', + 'itap of the {}.', + 'a jpeg corrupted photo of the {}.', + 'a good photo of a {}.', + 'a plushie {}.', + 'a photo of the nice {}.', + 'a photo of the small {}.', + 'a photo of the weird {}.', + 'the cartoon {}.', + 'art of the {}.', + 'a drawing of the {}.', + 'a photo of the large {}.', + 'a black and white photo of a {}.', + 'the plushie {}.', + 'a dark photo of a {}.', + 'itap of a {}.', + 'graffiti of the {}.', + 'a toy {}.', + 'itap of my {}.', + 'a photo of a cool {}.', + 'a photo of a small {}.', + 'a tattoo of the {}.', + ] + return prompt_templates + +def prompt_engineering(classnames, template=""): + return template.replace('{}', classnames.replace(',', '').replace('+', ' ')) + +# clip_img_tsv.py +def convert_example_to_features_bpe(text, tokenizer, sot_token, eot_token, context_length=77): + """ + Convert a raw sample (pair of sentences as tokenized strings) into a proper training sample. + :param tokenizer: Tokenizer + :return: List, a list containing token id, padded by 0 + """ + assert isinstance(text, str) + input_ids = [sot_token] + tokenizer.encode(text) + [eot_token] + if len(input_ids) > context_length: + input_ids = input_ids[:context_length] + input_ids = np.array(input_ids) + + pad_input_ids = np.zeros(context_length) + pad_input_ids[:input_ids.shape[0]] = input_ids + + return pad_input_ids + +def get_cls_names(filter_novel=False, coco=None, from_file=False): + """ return a list of strings with each string as name of a class + """ + # the names are stored in a txt file + if from_file: + # coco_det_cls = {COCO_80_ALL_CLS[key]: key for key in COCO_80_ALL_CLS} + # # not found in nouns {'skis': 31, 'sports ball': 33, 'hot dog': 53, 'potted plant': 59, 'scissors': 77, 'hair drier': 79} + # coco_det_cls['ski'] = 81 + # coco_det_cls['scissor'] = 82 + # with open('/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/trained_models/concept_pool/COCO_Caption_nouns_4688.txt','w') as g: + # with open(from_file, 'r') as f: + # cnt = 0 + # for row in f: + # if row.split(",")[0] not in coco_det_cls: + # g.write(row) + # cnt += 1 + # else: + # coco_det_cls.pop(row.split(",")[0]) + names = [] + with open(from_file, 'r') as f: + for row in f: + names.append(row.split(",")[0]) + return names + # classes' names + if coco == 'target': + return COCO_UNSEEN_CLS + elif coco == 'base': + return COCO_SEEN_CLS + elif coco == 'all': + return COCO_OVD_ALL_CLS + elif coco == 'all_80': + return [COCO_80_ALL_CLS[i+1] for i in range(80)] + assert len(LVIS_V1_CATEGORIES) == 1203 + cat_ids = [k["id"] for k in LVIS_V1_CATEGORIES] + assert min(cat_ids) == 1 and max(cat_ids) == len( + cat_ids + ), "Category ids are not in [1, #categories], as expected" + # Ensure that the category list is sorted by id + lvis_categories = sorted(LVIS_V1_CATEGORIES, key=lambda x: x["id"]) + if filter_novel: + class_names = [cls_meta['name'] for cls_meta in lvis_categories if cls_meta['frequency'] != 'r'] + else: + class_names = [cls_meta['name'] for cls_meta in lvis_categories] + + # remove or replace special symbols + class_names = [cls_n.replace("_", " ") for cls_n in class_names] + class_names = [cls_n.replace("(", "") for cls_n in class_names] + class_names = [cls_n.replace(")", "") for cls_n in class_names] + return class_names + +def pre_tokenize(class_names): + """ + pre-tokenize class names + :param class_names: List, a list of class names + :param tokenizer: Tokenizer, SimpleTokenizer() + :return: Tensor, containing all prompts for all classes, [#cls, #prompts, context_length] + """ + # tokenizer + tokenizer = SimpleTokenizer() + sot_token = tokenizer.encoder["<|startoftext|>"] + eot_token = tokenizer.encoder["<|endoftext|>"] + + # prompt engineering + prompt_templates = get_prompt_templates() + input_ids_all = [] + for k in range(len(class_names)): + v = class_names[k] + if isinstance(v, str): + vs = [v] + elif isinstance(v, list): + vs = v + t1s = [] + for v in vs: + for pt in prompt_templates: + t1s.append(prompt_engineering(v, template=pt)) + input_ids = [] + for t1 in t1s: + this_input_ids = convert_example_to_features_bpe(t1, tokenizer, sot_token, eot_token) + input_ids.append(torch.tensor(this_input_ids, dtype=torch.long)) + + input_ids_all.append(torch.stack(input_ids, 0)) + + input_ids_all_classes = torch.stack(input_ids_all, 0) + return input_ids_all_classes + + +if __name__ == "__main__": + flatten_input_ids = pre_tokenize() diff --git a/detectron2/data/datasets/coco.py b/detectron2/data/datasets/coco.py new file mode 100644 index 0000000000000000000000000000000000000000..ed4f7ccb20efa3b54c719783e279c381ca5d8587 --- /dev/null +++ b/detectron2/data/datasets/coco.py @@ -0,0 +1,539 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import contextlib +import datetime +import io +import json +import logging +import numpy as np +import os +import shutil +import pycocotools.mask as mask_util +from fvcore.common.timer import Timer +from iopath.common.file_io import file_lock +from PIL import Image + +from detectron2.structures import Boxes, BoxMode, PolygonMasks, RotatedBoxes +from detectron2.utils.file_io import PathManager + +from .. import DatasetCatalog, MetadataCatalog + +""" +This file contains functions to parse COCO-format annotations into dicts in "Detectron2 format". +""" + + +logger = logging.getLogger(__name__) + +__all__ = ["load_coco_json", "load_sem_seg", "convert_to_coco_json", "register_coco_instances"] + + +def load_coco_json(json_file, image_root, dataset_name=None, extra_annotation_keys=None): + """ + Load a json file with COCO's instances annotation format. + Currently supports instance detection, instance segmentation, + and person keypoints annotations. + + Args: + json_file (str): full path to the json file in COCO instances annotation format. + image_root (str or path-like): the directory where the images in this json file exists. + dataset_name (str or None): the name of the dataset (e.g., coco_2017_train). + When provided, this function will also do the following: + + * Put "thing_classes" into the metadata associated with this dataset. + * Map the category ids into a contiguous range (needed by standard dataset format), + and add "thing_dataset_id_to_contiguous_id" to the metadata associated + with this dataset. + + This option should usually be provided, unless users need to load + the original json content and apply more processing manually. + extra_annotation_keys (list[str]): list of per-annotation keys that should also be + loaded into the dataset dict (besides "iscrowd", "bbox", "keypoints", + "category_id", "segmentation"). The values for these keys will be returned as-is. + For example, the densepose annotations are loaded in this way. + + Returns: + list[dict]: a list of dicts in Detectron2 standard dataset dicts format (See + `Using Custom Datasets `_ ) when `dataset_name` is not None. + If `dataset_name` is None, the returned `category_ids` may be + incontiguous and may not conform to the Detectron2 standard format. + + Notes: + 1. This function does not read the image files. + The results do not have the "image" field. + """ + from pycocotools.coco import COCO + + timer = Timer() + json_file = PathManager.get_local_path(json_file) + with contextlib.redirect_stdout(io.StringIO()): + coco_api = COCO(json_file) + if timer.seconds() > 1: + logger.info("Loading {} takes {:.2f} seconds.".format(json_file, timer.seconds())) + + id_map = None + if dataset_name is not None: + meta = MetadataCatalog.get(dataset_name) + cat_ids = sorted(coco_api.getCatIds()) + cats = coco_api.loadCats(cat_ids) + # The categories in a custom json file may not be sorted. + thing_classes = [c["name"] for c in sorted(cats, key=lambda x: x["id"])] + meta.thing_classes = thing_classes + + # In COCO, certain category ids are artificially removed, + # and by convention they are always ignored. + # We deal with COCO's id issue and translate + # the category ids to contiguous ids in [0, 80). + + # It works by looking at the "categories" field in the json, therefore + # if users' own json also have incontiguous ids, we'll + # apply this mapping as well but print a warning. + if not (min(cat_ids) == 1 and max(cat_ids) == len(cat_ids)): + if "coco" not in dataset_name: + logger.warning( + """ +Category ids in annotations are not in [1, #categories]! We'll apply a mapping for you. +""" + ) + id_map = {v: i for i, v in enumerate(cat_ids)} + meta.thing_dataset_id_to_contiguous_id = id_map + + # sort indices for reproducible results + img_ids = sorted(coco_api.imgs.keys()) + # imgs is a list of dicts, each looks something like: + # {'license': 4, + # 'url': 'http://farm6.staticflickr.com/5454/9413846304_881d5e5c3b_z.jpg', + # 'file_name': 'COCO_val2014_000000001268.jpg', + # 'height': 427, + # 'width': 640, + # 'date_captured': '2013-11-17 05:57:24', + # 'id': 1268} + imgs = coco_api.loadImgs(img_ids) + # anns is a list[list[dict]], where each dict is an annotation + # record for an object. The inner list enumerates the objects in an image + # and the outer list enumerates over images. Example of anns[0]: + # [{'segmentation': [[192.81, + # 247.09, + # ... + # 219.03, + # 249.06]], + # 'area': 1035.749, + # 'iscrowd': 0, + # 'image_id': 1268, + # 'bbox': [192.81, 224.8, 74.73, 33.43], + # 'category_id': 16, + # 'id': 42986}, + # ...] + anns = [coco_api.imgToAnns[img_id] for img_id in img_ids] + total_num_valid_anns = sum([len(x) for x in anns]) + total_num_anns = len(coco_api.anns) + if total_num_valid_anns < total_num_anns: + logger.warning( + f"{json_file} contains {total_num_anns} annotations, but only " + f"{total_num_valid_anns} of them match to images in the file." + ) + + if "minival" not in json_file: + # The popular valminusminival & minival annotations for COCO2014 contain this bug. + # However the ratio of buggy annotations there is tiny and does not affect accuracy. + # Therefore we explicitly white-list them. + ann_ids = [ann["id"] for anns_per_image in anns for ann in anns_per_image] + assert len(set(ann_ids)) == len(ann_ids), "Annotation ids in '{}' are not unique!".format( + json_file + ) + + imgs_anns = list(zip(imgs, anns)) + logger.info("Loaded {} images in COCO format from {}".format(len(imgs_anns), json_file)) + + dataset_dicts = [] + + ann_keys = ["iscrowd", "bbox", "keypoints", "category_id"] + (extra_annotation_keys or []) + + num_instances_without_valid_segmentation = 0 + + for (img_dict, anno_dict_list) in imgs_anns: + record = {} + record["file_name"] = os.path.join(image_root, img_dict["file_name"]) + record["height"] = img_dict["height"] + record["width"] = img_dict["width"] + image_id = record["image_id"] = img_dict["id"] + + objs = [] + for anno in anno_dict_list: + # Check that the image_id in this annotation is the same as + # the image_id we're looking at. + # This fails only when the data parsing logic or the annotation file is buggy. + + # The original COCO valminusminival2014 & minival2014 annotation files + # actually contains bugs that, together with certain ways of using COCO API, + # can trigger this assertion. + assert anno["image_id"] == image_id + + assert anno.get("ignore", 0) == 0, '"ignore" in COCO json file is not supported.' + + obj = {key: anno[key] for key in ann_keys if key in anno} + if "bbox" in obj and len(obj["bbox"]) == 0: + raise ValueError( + f"One annotation of image {image_id} contains empty 'bbox' value! " + "This json does not have valid COCO format." + ) + + segm = anno.get("segmentation", None) + if segm: # either list[list[float]] or dict(RLE) + if isinstance(segm, dict): + if isinstance(segm["counts"], list): + # convert to compressed RLE + segm = mask_util.frPyObjects(segm, *segm["size"]) + else: + # filter out invalid polygons (< 3 points) + segm = [poly for poly in segm if len(poly) % 2 == 0 and len(poly) >= 6] + if len(segm) == 0: + num_instances_without_valid_segmentation += 1 + continue # ignore this instance + obj["segmentation"] = segm + + keypts = anno.get("keypoints", None) + if keypts: # list[int] + for idx, v in enumerate(keypts): + if idx % 3 != 2: + # COCO's segmentation coordinates are floating points in [0, H or W], + # but keypoint coordinates are integers in [0, H-1 or W-1] + # Therefore we assume the coordinates are "pixel indices" and + # add 0.5 to convert to floating point coordinates. + keypts[idx] = v + 0.5 + obj["keypoints"] = keypts + + obj["bbox_mode"] = BoxMode.XYWH_ABS + if id_map: + annotation_category_id = obj["category_id"] + try: + obj["category_id"] = id_map[annotation_category_id] + except KeyError as e: + raise KeyError( + f"Encountered category_id={annotation_category_id} " + "but this id does not exist in 'categories' of the json file." + ) from e + objs.append(obj) + record["annotations"] = objs + dataset_dicts.append(record) + + if num_instances_without_valid_segmentation > 0: + logger.warning( + "Filtered out {} instances without valid segmentation. ".format( + num_instances_without_valid_segmentation + ) + + "There might be issues in your dataset generation process. Please " + "check https://detectron2.readthedocs.io/en/latest/tutorials/datasets.html carefully" + ) + return dataset_dicts + + +def load_sem_seg(gt_root, image_root, gt_ext="png", image_ext="jpg"): + """ + Load semantic segmentation datasets. All files under "gt_root" with "gt_ext" extension are + treated as ground truth annotations and all files under "image_root" with "image_ext" extension + as input images. Ground truth and input images are matched using file paths relative to + "gt_root" and "image_root" respectively without taking into account file extensions. + This works for COCO as well as some other datasets. + + Args: + gt_root (str): full path to ground truth semantic segmentation files. Semantic segmentation + annotations are stored as images with integer values in pixels that represent + corresponding semantic labels. + image_root (str): the directory where the input images are. + gt_ext (str): file extension for ground truth annotations. + image_ext (str): file extension for input images. + + Returns: + list[dict]: + a list of dicts in detectron2 standard format without instance-level + annotation. + + Notes: + 1. This function does not read the image and ground truth files. + The results do not have the "image" and "sem_seg" fields. + """ + + # We match input images with ground truth based on their relative filepaths (without file + # extensions) starting from 'image_root' and 'gt_root' respectively. + def file2id(folder_path, file_path): + # extract relative path starting from `folder_path` + image_id = os.path.normpath(os.path.relpath(file_path, start=folder_path)) + # remove file extension + image_id = os.path.splitext(image_id)[0] + return image_id + + input_files = sorted( + (os.path.join(image_root, f) for f in PathManager.ls(image_root) if f.endswith(image_ext)), + key=lambda file_path: file2id(image_root, file_path), + ) + gt_files = sorted( + (os.path.join(gt_root, f) for f in PathManager.ls(gt_root) if f.endswith(gt_ext)), + key=lambda file_path: file2id(gt_root, file_path), + ) + + assert len(gt_files) > 0, "No annotations found in {}.".format(gt_root) + + # Use the intersection, so that val2017_100 annotations can run smoothly with val2017 images + if len(input_files) != len(gt_files): + logger.warn( + "Directory {} and {} has {} and {} files, respectively.".format( + image_root, gt_root, len(input_files), len(gt_files) + ) + ) + input_basenames = [os.path.basename(f)[: -len(image_ext)] for f in input_files] + gt_basenames = [os.path.basename(f)[: -len(gt_ext)] for f in gt_files] + intersect = list(set(input_basenames) & set(gt_basenames)) + # sort, otherwise each worker may obtain a list[dict] in different order + intersect = sorted(intersect) + logger.warn("Will use their intersection of {} files.".format(len(intersect))) + input_files = [os.path.join(image_root, f + image_ext) for f in intersect] + gt_files = [os.path.join(gt_root, f + gt_ext) for f in intersect] + + logger.info( + "Loaded {} images with semantic segmentation from {}".format(len(input_files), image_root) + ) + + dataset_dicts = [] + for (img_path, gt_path) in zip(input_files, gt_files): + record = {} + record["file_name"] = img_path + record["sem_seg_file_name"] = gt_path + dataset_dicts.append(record) + + return dataset_dicts + + +def convert_to_coco_dict(dataset_name): + """ + Convert an instance detection/segmentation or keypoint detection dataset + in detectron2's standard format into COCO json format. + + Generic dataset description can be found here: + https://detectron2.readthedocs.io/tutorials/datasets.html#register-a-dataset + + COCO data format description can be found here: + http://cocodataset.org/#format-data + + Args: + dataset_name (str): + name of the source dataset + Must be registered in DatastCatalog and in detectron2's standard format. + Must have corresponding metadata "thing_classes" + Returns: + coco_dict: serializable dict in COCO json format + """ + + dataset_dicts = DatasetCatalog.get(dataset_name) + metadata = MetadataCatalog.get(dataset_name) + + # unmap the category mapping ids for COCO + if hasattr(metadata, "thing_dataset_id_to_contiguous_id"): + reverse_id_mapping = {v: k for k, v in metadata.thing_dataset_id_to_contiguous_id.items()} + reverse_id_mapper = lambda contiguous_id: reverse_id_mapping[contiguous_id] # noqa + else: + reverse_id_mapper = lambda contiguous_id: contiguous_id # noqa + + categories = [ + {"id": reverse_id_mapper(id), "name": name} + for id, name in enumerate(metadata.thing_classes) + ] + + logger.info("Converting dataset dicts into COCO format") + coco_images = [] + coco_annotations = [] + + for image_id, image_dict in enumerate(dataset_dicts): + coco_image = { + "id": image_dict.get("image_id", image_id), + "width": int(image_dict["width"]), + "height": int(image_dict["height"]), + "file_name": str(image_dict["file_name"]), + } + coco_images.append(coco_image) + + anns_per_image = image_dict.get("annotations", []) + for annotation in anns_per_image: + # create a new dict with only COCO fields + coco_annotation = {} + + # COCO requirement: XYWH box format for axis-align and XYWHA for rotated + bbox = annotation["bbox"] + if isinstance(bbox, np.ndarray): + if bbox.ndim != 1: + raise ValueError(f"bbox has to be 1-dimensional. Got shape={bbox.shape}.") + bbox = bbox.tolist() + if len(bbox) not in [4, 5]: + raise ValueError(f"bbox has to has length 4 or 5. Got {bbox}.") + from_bbox_mode = annotation["bbox_mode"] + to_bbox_mode = BoxMode.XYWH_ABS if len(bbox) == 4 else BoxMode.XYWHA_ABS + bbox = BoxMode.convert(bbox, from_bbox_mode, to_bbox_mode) + + # COCO requirement: instance area + if "segmentation" in annotation: + # Computing areas for instances by counting the pixels + segmentation = annotation["segmentation"] + # TODO: check segmentation type: RLE, BinaryMask or Polygon + if isinstance(segmentation, list): + polygons = PolygonMasks([segmentation]) + area = polygons.area()[0].item() + elif isinstance(segmentation, dict): # RLE + area = mask_util.area(segmentation).item() + else: + raise TypeError(f"Unknown segmentation type {type(segmentation)}!") + else: + # Computing areas using bounding boxes + if to_bbox_mode == BoxMode.XYWH_ABS: + bbox_xy = BoxMode.convert(bbox, to_bbox_mode, BoxMode.XYXY_ABS) + area = Boxes([bbox_xy]).area()[0].item() + else: + area = RotatedBoxes([bbox]).area()[0].item() + + if "keypoints" in annotation: + keypoints = annotation["keypoints"] # list[int] + for idx, v in enumerate(keypoints): + if idx % 3 != 2: + # COCO's segmentation coordinates are floating points in [0, H or W], + # but keypoint coordinates are integers in [0, H-1 or W-1] + # For COCO format consistency we substract 0.5 + # https://github.com/facebookresearch/detectron2/pull/175#issuecomment-551202163 + keypoints[idx] = v - 0.5 + if "num_keypoints" in annotation: + num_keypoints = annotation["num_keypoints"] + else: + num_keypoints = sum(kp > 0 for kp in keypoints[2::3]) + + # COCO requirement: + # linking annotations to images + # "id" field must start with 1 + coco_annotation["id"] = len(coco_annotations) + 1 + coco_annotation["image_id"] = coco_image["id"] + coco_annotation["bbox"] = [round(float(x), 3) for x in bbox] + coco_annotation["area"] = float(area) + coco_annotation["iscrowd"] = int(annotation.get("iscrowd", 0)) + coco_annotation["category_id"] = int(reverse_id_mapper(annotation["category_id"])) + + # Add optional fields + if "keypoints" in annotation: + coco_annotation["keypoints"] = keypoints + coco_annotation["num_keypoints"] = num_keypoints + + if "segmentation" in annotation: + seg = coco_annotation["segmentation"] = annotation["segmentation"] + if isinstance(seg, dict): # RLE + counts = seg["counts"] + if not isinstance(counts, str): + # make it json-serializable + seg["counts"] = counts.decode("ascii") + + coco_annotations.append(coco_annotation) + + logger.info( + "Conversion finished, " + f"#images: {len(coco_images)}, #annotations: {len(coco_annotations)}" + ) + + info = { + "date_created": str(datetime.datetime.now()), + "description": "Automatically generated COCO json file for Detectron2.", + } + coco_dict = {"info": info, "images": coco_images, "categories": categories, "licenses": None} + if len(coco_annotations) > 0: + coco_dict["annotations"] = coco_annotations + return coco_dict + + +def convert_to_coco_json(dataset_name, output_file, allow_cached=True): + """ + Converts dataset into COCO format and saves it to a json file. + dataset_name must be registered in DatasetCatalog and in detectron2's standard format. + + Args: + dataset_name: + reference from the config file to the catalogs + must be registered in DatasetCatalog and in detectron2's standard format + output_file: path of json file that will be saved to + allow_cached: if json file is already present then skip conversion + """ + + # TODO: The dataset or the conversion script *may* change, + # a checksum would be useful for validating the cached data + + PathManager.mkdirs(os.path.dirname(output_file)) + with file_lock(output_file): + if PathManager.exists(output_file) and allow_cached: + logger.warning( + f"Using previously cached COCO format annotations at '{output_file}'. " + "You need to clear the cache file if your dataset has been modified." + ) + else: + logger.info(f"Converting annotations of dataset '{dataset_name}' to COCO format ...)") + coco_dict = convert_to_coco_dict(dataset_name) + + logger.info(f"Caching COCO format annotations at '{output_file}' ...") + tmp_file = output_file + ".tmp" + with PathManager.open(tmp_file, "w") as f: + json.dump(coco_dict, f) + shutil.move(tmp_file, output_file) + + +def register_coco_instances(name, metadata, json_file, image_root): + """ + Register a dataset in COCO's json annotation format for + instance detection, instance segmentation and keypoint detection. + (i.e., Type 1 and 2 in http://cocodataset.org/#format-data. + `instances*.json` and `person_keypoints*.json` in the dataset). + + This is an example of how to register a new dataset. + You can do something similar to this function, to register new datasets. + + Args: + name (str): the name that identifies a dataset, e.g. "coco_2014_train". + metadata (dict): extra metadata associated with this dataset. You can + leave it as an empty dict. + json_file (str): path to the json instance annotation file. + image_root (str or path-like): directory which contains all the images. + """ + assert isinstance(name, str), name + assert isinstance(json_file, (str, os.PathLike)), json_file + assert isinstance(image_root, (str, os.PathLike)), image_root + # 1. register a function which returns dicts + DatasetCatalog.register(name, lambda: load_coco_json(json_file, image_root, name)) + + # 2. Optionally, add metadata about this dataset, + # since they might be useful in evaluation, visualization or logging + MetadataCatalog.get(name).set( + json_file=json_file, image_root=image_root, evaluator_type="coco", **metadata + ) + + +if __name__ == "__main__": + """ + Test the COCO json dataset loader. + + Usage: + python -m detectron2.data.datasets.coco \ + path/to/json path/to/image_root dataset_name + + "dataset_name" can be "coco_2014_minival_100", or other + pre-registered ones + """ + from detectron2.utils.logger import setup_logger + from detectron2.utils.visualizer import Visualizer + import detectron2.data.datasets # noqa # add pre-defined metadata + import sys + + logger = setup_logger(name=__name__) + assert sys.argv[3] in DatasetCatalog.list() + meta = MetadataCatalog.get(sys.argv[3]) + + dicts = load_coco_json(sys.argv[1], sys.argv[2], sys.argv[3]) + logger.info("Done loading {} samples.".format(len(dicts))) + + dirname = "coco-data-vis" + os.makedirs(dirname, exist_ok=True) + for d in dicts: + img = np.array(Image.open(d["file_name"])) + visualizer = Visualizer(img, metadata=meta) + vis = visualizer.draw_dataset_dict(d) + fpath = os.path.join(dirname, os.path.basename(d["file_name"])) + vis.save(fpath) diff --git a/detectron2/data/datasets/coco_panoptic.py b/detectron2/data/datasets/coco_panoptic.py new file mode 100644 index 0000000000000000000000000000000000000000..b8dae44317b556610d7fed39017e082d7e855956 --- /dev/null +++ b/detectron2/data/datasets/coco_panoptic.py @@ -0,0 +1,228 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import json +import os + +from detectron2.data import DatasetCatalog, MetadataCatalog +from detectron2.utils.file_io import PathManager + +from .coco import load_coco_json, load_sem_seg + +__all__ = ["register_coco_panoptic", "register_coco_panoptic_separated"] + + +def load_coco_panoptic_json(json_file, image_dir, gt_dir, meta): + """ + Args: + image_dir (str): path to the raw dataset. e.g., "~/coco/train2017". + gt_dir (str): path to the raw annotations. e.g., "~/coco/panoptic_train2017". + json_file (str): path to the json file. e.g., "~/coco/annotations/panoptic_train2017.json". + + Returns: + list[dict]: a list of dicts in Detectron2 standard format. (See + `Using Custom Datasets `_ ) + """ + + def _convert_category_id(segment_info, meta): + if segment_info["category_id"] in meta["thing_dataset_id_to_contiguous_id"]: + segment_info["category_id"] = meta["thing_dataset_id_to_contiguous_id"][ + segment_info["category_id"] + ] + segment_info["isthing"] = True + else: + segment_info["category_id"] = meta["stuff_dataset_id_to_contiguous_id"][ + segment_info["category_id"] + ] + segment_info["isthing"] = False + return segment_info + + with PathManager.open(json_file) as f: + json_info = json.load(f) + + ret = [] + for ann in json_info["annotations"]: + image_id = int(ann["image_id"]) + # TODO: currently we assume image and label has the same filename but + # different extension, and images have extension ".jpg" for COCO. Need + # to make image extension a user-provided argument if we extend this + # function to support other COCO-like datasets. + image_file = os.path.join(image_dir, os.path.splitext(ann["file_name"])[0] + ".jpg") + label_file = os.path.join(gt_dir, ann["file_name"]) + segments_info = [_convert_category_id(x, meta) for x in ann["segments_info"]] + ret.append( + { + "file_name": image_file, + "image_id": image_id, + "pan_seg_file_name": label_file, + "segments_info": segments_info, + } + ) + assert len(ret), f"No images found in {image_dir}!" + assert PathManager.isfile(ret[0]["file_name"]), ret[0]["file_name"] + assert PathManager.isfile(ret[0]["pan_seg_file_name"]), ret[0]["pan_seg_file_name"] + return ret + + +def register_coco_panoptic( + name, metadata, image_root, panoptic_root, panoptic_json, instances_json=None +): + """ + Register a "standard" version of COCO panoptic segmentation dataset named `name`. + The dictionaries in this registered dataset follows detectron2's standard format. + Hence it's called "standard". + + Args: + name (str): the name that identifies a dataset, + e.g. "coco_2017_train_panoptic" + metadata (dict): extra metadata associated with this dataset. + image_root (str): directory which contains all the images + panoptic_root (str): directory which contains panoptic annotation images in COCO format + panoptic_json (str): path to the json panoptic annotation file in COCO format + sem_seg_root (none): not used, to be consistent with + `register_coco_panoptic_separated`. + instances_json (str): path to the json instance annotation file + """ + panoptic_name = name + DatasetCatalog.register( + panoptic_name, + lambda: load_coco_panoptic_json(panoptic_json, image_root, panoptic_root, metadata), + ) + MetadataCatalog.get(panoptic_name).set( + panoptic_root=panoptic_root, + image_root=image_root, + panoptic_json=panoptic_json, + json_file=instances_json, + evaluator_type="coco_panoptic_seg", + ignore_label=255, + label_divisor=1000, + **metadata, + ) + + +def register_coco_panoptic_separated( + name, metadata, image_root, panoptic_root, panoptic_json, sem_seg_root, instances_json +): + """ + Register a "separated" version of COCO panoptic segmentation dataset named `name`. + The annotations in this registered dataset will contain both instance annotations and + semantic annotations, each with its own contiguous ids. Hence it's called "separated". + + It follows the setting used by the PanopticFPN paper: + + 1. The instance annotations directly come from polygons in the COCO + instances annotation task, rather than from the masks in the COCO panoptic annotations. + + The two format have small differences: + Polygons in the instance annotations may have overlaps. + The mask annotations are produced by labeling the overlapped polygons + with depth ordering. + + 2. The semantic annotations are converted from panoptic annotations, where + all "things" are assigned a semantic id of 0. + All semantic categories will therefore have ids in contiguous + range [1, #stuff_categories]. + + This function will also register a pure semantic segmentation dataset + named ``name + '_stuffonly'``. + + Args: + name (str): the name that identifies a dataset, + e.g. "coco_2017_train_panoptic" + metadata (dict): extra metadata associated with this dataset. + image_root (str): directory which contains all the images + panoptic_root (str): directory which contains panoptic annotation images + panoptic_json (str): path to the json panoptic annotation file + sem_seg_root (str): directory which contains all the ground truth segmentation annotations. + instances_json (str): path to the json instance annotation file + """ + panoptic_name = name + "_separated" + DatasetCatalog.register( + panoptic_name, + lambda: merge_to_panoptic( + load_coco_json(instances_json, image_root, panoptic_name), + load_sem_seg(sem_seg_root, image_root), + ), + ) + MetadataCatalog.get(panoptic_name).set( + panoptic_root=panoptic_root, + image_root=image_root, + panoptic_json=panoptic_json, + sem_seg_root=sem_seg_root, + json_file=instances_json, # TODO rename + evaluator_type="coco_panoptic_seg", + ignore_label=255, + **metadata, + ) + + semantic_name = name + "_stuffonly" + DatasetCatalog.register(semantic_name, lambda: load_sem_seg(sem_seg_root, image_root)) + MetadataCatalog.get(semantic_name).set( + sem_seg_root=sem_seg_root, + image_root=image_root, + evaluator_type="sem_seg", + ignore_label=255, + **metadata, + ) + + +def merge_to_panoptic(detection_dicts, sem_seg_dicts): + """ + Create dataset dicts for panoptic segmentation, by + merging two dicts using "file_name" field to match their entries. + + Args: + detection_dicts (list[dict]): lists of dicts for object detection or instance segmentation. + sem_seg_dicts (list[dict]): lists of dicts for semantic segmentation. + + Returns: + list[dict] (one per input image): Each dict contains all (key, value) pairs from dicts in + both detection_dicts and sem_seg_dicts that correspond to the same image. + The function assumes that the same key in different dicts has the same value. + """ + results = [] + sem_seg_file_to_entry = {x["file_name"]: x for x in sem_seg_dicts} + assert len(sem_seg_file_to_entry) > 0 + + for det_dict in detection_dicts: + dic = copy.copy(det_dict) + dic.update(sem_seg_file_to_entry[dic["file_name"]]) + results.append(dic) + return results + + +if __name__ == "__main__": + """ + Test the COCO panoptic dataset loader. + + Usage: + python -m detectron2.data.datasets.coco_panoptic \ + path/to/image_root path/to/panoptic_root path/to/panoptic_json dataset_name 10 + + "dataset_name" can be "coco_2017_train_panoptic", or other + pre-registered ones + """ + from detectron2.utils.logger import setup_logger + from detectron2.utils.visualizer import Visualizer + import detectron2.data.datasets # noqa # add pre-defined metadata + import sys + from PIL import Image + import numpy as np + + logger = setup_logger(name=__name__) + assert sys.argv[4] in DatasetCatalog.list() + meta = MetadataCatalog.get(sys.argv[4]) + + dicts = load_coco_panoptic_json(sys.argv[3], sys.argv[1], sys.argv[2], meta.as_dict()) + logger.info("Done loading {} samples.".format(len(dicts))) + + dirname = "coco-data-vis" + os.makedirs(dirname, exist_ok=True) + num_imgs_to_vis = int(sys.argv[5]) + for i, d in enumerate(dicts): + img = np.array(Image.open(d["file_name"])) + visualizer = Visualizer(img, metadata=meta) + vis = visualizer.draw_dataset_dict(d) + fpath = os.path.join(dirname, os.path.basename(d["file_name"])) + vis.save(fpath) + if i + 1 >= num_imgs_to_vis: + break diff --git a/detectron2/data/datasets/coco_zeroshot_categories.py b/detectron2/data/datasets/coco_zeroshot_categories.py new file mode 100644 index 0000000000000000000000000000000000000000..8ebaec6600c218a534d45c7401f90cc20bca8b09 --- /dev/null +++ b/detectron2/data/datasets/coco_zeroshot_categories.py @@ -0,0 +1,206 @@ +# COCO categories for zero-shot setting +# 65 categories in total, 48 base categories for training, 17 unseen categories are only used in testing +# from http://ankan.umiacs.io/files/mscoco_seen_classes.json, http://ankan.umiacs.io/files/mscoco_unseen_classes.json + +# 17 class names in order, obtained from load_coco_json() function +COCO_UNSEEN_CLS = ['airplane', 'bus', 'cat', 'dog', 'cow', 'elephant', 'umbrella', \ + 'tie', 'snowboard', 'skateboard', 'cup', 'knife', 'cake', 'couch', 'keyboard', \ + 'sink', 'scissors'] + +# 48 class names in order, obtained from load_coco_json() function +COCO_SEEN_CLS = ['person', 'bicycle', 'car', 'motorcycle', 'train', 'truck', \ + 'boat', 'bench', 'bird', 'horse', 'sheep', 'bear', 'zebra', 'giraffe', \ + 'backpack', 'handbag', 'suitcase', 'frisbee', 'skis', 'kite', 'surfboard', \ + 'bottle', 'fork', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', \ + 'broccoli', 'carrot', 'pizza', 'donut', 'chair', 'bed', 'toilet', 'tv', \ + 'laptop', 'mouse', 'remote', 'microwave', 'oven', 'toaster', \ + 'refrigerator', 'book', 'clock', 'vase', 'toothbrush'] + +# 65 class names in order, obtained from load_coco_json() function +COCO_OVD_ALL_CLS = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', \ + 'bus', 'train', 'truck', 'boat', 'bench', 'bird', 'cat', 'dog', 'horse', \ + 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', \ + 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'kite', 'skateboard', \ + 'surfboard', 'bottle', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', \ + 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'pizza', 'donut', 'cake', \ + 'chair', 'couch', 'bed', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', \ + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', \ + 'scissors', 'toothbrush'] + +# 80 class names +COCO_80_ALL_CLS = {1: 'person', + 2: 'bicycle', + 3: 'car', + 4: 'motorcycle', + 5: 'airplane', + 6: 'bus', + 7: 'train', + 8: 'truck', + 9: 'boat', + 10: 'traffic light', + 11: 'fire hydrant', + 12: 'stop sign', + 13: 'parking meter', + 14: 'bench', + 15: 'bird', + 16: 'cat', + 17: 'dog', + 18: 'horse', + 19: 'sheep', + 20: 'cow', + 21: 'elephant', + 22: 'bear', + 23: 'zebra', + 24: 'giraffe', + 25: 'backpack', + 26: 'umbrella', + 27: 'handbag', + 28: 'tie', + 29: 'suitcase', + 30: 'frisbee', + 31: 'skis', + 32: 'snowboard', + 33: 'sports ball', + 34: 'kite', + 35: 'baseball bat', + 36: 'baseball glove', + 37: 'skateboard', + 38: 'surfboard', + 39: 'tennis racket', + 40: 'bottle', + 41: 'wine glass', + 42: 'cup', + 43: 'fork', + 44: 'knife', + 45: 'spoon', + 46: 'bowl', + 47: 'banana', + 48: 'apple', + 49: 'sandwich', + 50: 'orange', + 51: 'broccoli', + 52: 'carrot', + 53: 'hot dog', + 54: 'pizza', + 55: 'donut', + 56: 'cake', + 57: 'chair', + 58: 'couch', + 59: 'potted plant', + 60: 'bed', + 61: 'dining table', + 62: 'toilet', + 63: 'tv', + 64: 'laptop', + 65: 'mouse', + 66: 'remote', + 67: 'keyboard', + 68: 'cell phone', + 69: 'microwave', + 70: 'oven', + 71: 'toaster', + 72: 'sink', + 73: 'refrigerator', + 74: 'book', + 75: 'clock', + 76: 'vase', + 77: 'scissors', + 78: 'teddy bear', + 79: 'hair drier', + 80: 'toothbrush'} + +if __name__ == "__main__": + # from https://github.com/alirezazareian/ovr-cnn/blob/master/ipynb/001.ipynb + # Create zero-shot setting data split in COCO + import json + import ipdb + + with open('./datasets/coco/annotations/instances_train2017.json', 'r') as fin: + coco_train_anno_all = json.load(fin) + + with open('./datasets/coco/annotations/instances_train2017.json', 'r') as fin: + coco_train_anno_seen = json.load(fin) + + with open('./datasets/coco/annotations/instances_train2017.json', 'r') as fin: + coco_train_anno_unseen = json.load(fin) + + with open('./datasets/coco/annotations/instances_val2017.json', 'r') as fin: + coco_val_anno_all = json.load(fin) + + with open('./datasets/coco/annotations/instances_val2017.json', 'r') as fin: + coco_val_anno_seen = json.load(fin) + + with open('./datasets/coco/annotations/instances_val2017.json', 'r') as fin: + coco_val_anno_unseen = json.load(fin) + + labels_seen = COCO_SEEN_CLS + labels_unseen = COCO_UNSEEN_CLS + labels_all = [item['name'] for item in coco_val_anno_all['categories']] # 80 class names + # len(labels_seen), len(labels_unseen) + # set(labels_seen) - set(labels_all) + # set(labels_unseen) - set(labels_all) + + class_id_to_split = {} # {1: 'seen', 2: 'seen', 3: 'seen', 4: 'seen', 5: 'unseen',...} + class_name_to_split = {} # {'person': 'seen', 'bicycle': 'seen', 'car': 'seen', 'motorcycle': 'seen', 'airplane': 'unseen',...} + for item in coco_val_anno_all['categories']: + if item['name'] in labels_seen: + class_id_to_split[item['id']] = 'seen' + class_name_to_split[item['name']] = 'seen' + elif item['name'] in labels_unseen: + class_id_to_split[item['id']] = 'unseen' + class_name_to_split[item['name']] = 'unseen' + + # class_name_to_emb = {} + # with open('../datasets/coco/zero-shot/glove.6B.300d.txt', 'r') as fin: + # for row in fin: + # row_tk = row.split() + # if row_tk[0] in class_name_to_split: + # class_name_to_emb[row_tk[0]] = [float(num) for num in row_tk[1:]] + # len(class_name_to_emb), len(class_name_to_split) + + def filter_annotation(anno_dict, split_name_list): + """ + COCO annotations have fields: dict_keys(['info', 'licenses', 'images', 'annotations', 'categories']) + This function (1) filters the category metadata (list) in 'categories'; + (2) filter instance annotation in 'annotations'; (3) filter image metadata (list) in 'images + """ + filtered_categories = [] + for item in anno_dict['categories']: + if class_id_to_split.get(item['id']) in split_name_list: + #item['embedding'] = class_name_to_emb[item['name']] + item['split'] = class_id_to_split.get(item['id']) + filtered_categories.append(item) + anno_dict['categories'] = filtered_categories + + filtered_images = [] + filtered_annotations = [] + useful_image_ids = set() + for item in anno_dict['annotations']: + if class_id_to_split.get(item['category_id']) in split_name_list: + filtered_annotations.append(item) + useful_image_ids.add(item['image_id']) + for item in anno_dict['images']: + if item['id'] in useful_image_ids: + filtered_images.append(item) + anno_dict['annotations'] = filtered_annotations + anno_dict['images'] = filtered_images + + filter_annotation(coco_train_anno_seen, ['seen']) + filter_annotation(coco_train_anno_unseen, ['unseen']) + filter_annotation(coco_train_anno_all, ['seen', 'unseen']) + filter_annotation(coco_val_anno_seen, ['seen']) + filter_annotation(coco_val_anno_unseen, ['unseen']) + filter_annotation(coco_val_anno_all, ['seen', 'unseen']) + + with open('./datasets/coco/annotations/ovd_ins_train2017_b.json', 'w') as fout: + json.dump(coco_train_anno_seen, fout) + with open('./datasets/coco/annotations/ovd_ins_train2017_t.json', 'w') as fout: + json.dump(coco_train_anno_unseen, fout) + with open('./datasets/coco/annotations/ovd_ins_train2017_all.json', 'w') as fout: + json.dump(coco_train_anno_all, fout) + with open('./datasets/coco/annotations/ovd_ins_val2017_b.json', 'w') as fout: + json.dump(coco_val_anno_seen, fout) + with open('./datasets/coco/annotations/ovd_ins_val2017_t.json', 'w') as fout: + json.dump(coco_val_anno_unseen, fout) + with open('./datasets/coco/annotations/ovd_ins_val2017_all.json', 'w') as fout: + json.dump(coco_val_anno_all, fout) \ No newline at end of file diff --git a/detectron2/data/datasets/lvis.py b/detectron2/data/datasets/lvis.py new file mode 100644 index 0000000000000000000000000000000000000000..2248129e798baefda037a8dddf7abe3c8f15dd40 --- /dev/null +++ b/detectron2/data/datasets/lvis.py @@ -0,0 +1,357 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import os +from fvcore.common.timer import Timer + +from detectron2.data import DatasetCatalog, MetadataCatalog +from detectron2.structures import BoxMode +from detectron2.utils.file_io import PathManager + +from .builtin_meta import _get_coco_instances_meta +from .lvis_v0_5_categories import LVIS_CATEGORIES as LVIS_V0_5_CATEGORIES +from .lvis_v1_categories import LVIS_CATEGORIES as LVIS_V1_CATEGORIES + +import torch +import numpy as np +""" +This file contains functions to parse LVIS-format annotations into dicts in the +"Detectron2 format". +""" + +logger = logging.getLogger(__name__) + +__all__ = ["load_lvis_json", "register_lvis_instances", "get_lvis_instances_meta"] + + +def register_lvis_instances(name, metadata, json_file, image_root): + """ + Register a dataset in LVIS's json annotation format for instance detection and segmentation. + + Args: + name (str): a name that identifies the dataset, e.g. "lvis_v0.5_train". + metadata (dict): extra metadata associated with this dataset. It can be an empty dict. + json_file (str): path to the json instance annotation file. + image_root (str or path-like): directory which contains all the images. + """ + DatasetCatalog.register(name, lambda: load_lvis_json(json_file, image_root, name)) + MetadataCatalog.get(name).set( + json_file=json_file, image_root=image_root, evaluator_type="lvis", **metadata + ) + + +def load_lvis_json_original(json_file, image_root, dataset_name=None, filter_open_cls=True, clip_gt_crop=True, max_gt_per_img=500): + """ + Load a json file in LVIS's annotation format. + + Args: + json_file (str): full path to the LVIS json annotation file. + image_root (str): the directory where the images in this json file exists. + dataset_name (str): the name of the dataset (e.g., "lvis_v0.5_train"). + If provided, this function will put "thing_classes" into the metadata + associated with this dataset. + filter_open_cls: open-set setting, filter the open-set categories during training + clip_gt_crop: must filter images with empty annotations or too many GT bbox, + even if in testing (eg, use CLIP on GT regions) + Returns: + list[dict]: a list of dicts in Detectron2 standard format. (See + `Using Custom Datasets `_ ) + + Notes: + 1. This function does not read the image files. + The results do not have the "image" field. + """ + from lvis import LVIS + + if 'train' in dataset_name: #'zeroshot' in dataset_name and 'train' in dataset_name: # openset setting, filter the novel classes during training + filter_open_cls = True + else: + filter_open_cls = False + + json_file = PathManager.get_local_path(json_file) + + timer = Timer() + lvis_api = LVIS(json_file) + if timer.seconds() > 1: + logger.info("Loading {} takes {:.2f} seconds.".format(json_file, timer.seconds())) + + if dataset_name is not None: + meta = get_lvis_instances_meta(dataset_name) + MetadataCatalog.get(dataset_name).set(**meta) + + # sort indices for reproducible results + img_ids = sorted(lvis_api.imgs.keys()) + # imgs is a list of dicts, each looks something like: + # {'license': 4, + # 'url': 'http://farm6.staticflickr.com/5454/9413846304_881d5e5c3b_z.jpg', + # 'file_name': 'COCO_val2014_000000001268.jpg', + # 'height': 427, + # 'width': 640, + # 'date_captured': '2013-11-17 05:57:24', + # 'id': 1268} + imgs = lvis_api.load_imgs(img_ids) + # anns is a list[list[dict]], where each dict is an annotation + # record for an object. The inner list enumerates the objects in an image + # and the outer list enumerates over images. Example of anns[0]: + # [{'segmentation': [[192.81, + # 247.09, + # ... + # 219.03, + # 249.06]], + # 'area': 1035.749, + # 'image_id': 1268, + # 'bbox': [192.81, 224.8, 74.73, 33.43], + # 'category_id': 16, + # 'id': 42986}, + # ...] + anns = [lvis_api.img_ann_map[img_id] for img_id in img_ids] + + # Sanity check that each annotation has a unique id + ann_ids = [ann["id"] for anns_per_image in anns for ann in anns_per_image] + assert len(set(ann_ids)) == len(ann_ids), "Annotation ids in '{}' are not unique".format( + json_file + ) + + imgs_anns = list(zip(imgs, anns)) + + logger.info("Loaded {} images in the LVIS format from {}".format(len(imgs_anns), json_file)) + + def get_file_name(img_root, img_dict): + # Determine the path including the split folder ("train2017", "val2017", "test2017") from + # the coco_url field. Example: + # 'coco_url': 'http://images.cocodataset.org/train2017/000000155379.jpg' + split_folder, file_name = img_dict["coco_url"].split("/")[-2:] + return os.path.join(img_root + split_folder, file_name) + + dataset_dicts = [] + cls_type_dict = {cls_meta['id']: cls_meta['frequency'] for cls_meta in lvis_api.dataset['categories']} # map cls id to cls type + area_dict = {'r': [], 'c': [], 'f': []} # calculate box area for each type of class + # import os + # from PIL import Image + # custom_img_path = 'datasets/epic_sample_frames' + # custom_img_list = [os.path.join(custom_img_path, item) for item in os.listdir(custom_img_path)] + # cnt = 0 + for (img_dict, anno_dict_list) in imgs_anns: + record = {} + record["file_name"] = get_file_name(image_root, img_dict) + # record["file_name"] = custom_img_list[cnt]; cnt += 1; + # if cnt == 46: + # break # get_file_name(image_root, img_dict) + # img_file = Image.open(record["file_name"]) + record["height"] = img_dict["height"] + record["width"] = img_dict["width"] + # record["height"] = img_file.size[1] # img_dict["height"] + # record["width"] = img_file.size[0] # img_dict["width"] + record["not_exhaustive_category_ids"] = img_dict.get("not_exhaustive_category_ids", []) + record["neg_category_ids"] = img_dict.get("neg_category_ids", []) + image_id = record["image_id"] = img_dict["id"] + + objs = [] + for anno in anno_dict_list: + # Check that the image_id in this annotation is the same as + # the image_id we're looking at. + # This fails only when the data parsing logic or the annotation file is buggy. + assert anno["image_id"] == image_id + obj = {"bbox": anno["bbox"], "bbox_mode": BoxMode.XYWH_ABS} + # LVIS data loader can be used to load COCO dataset categories. In this case `meta` + # variable will have a field with COCO-specific category mapping. + if dataset_name is not None and "thing_dataset_id_to_contiguous_id" in meta: + obj["category_id"] = meta["thing_dataset_id_to_contiguous_id"][anno["category_id"]] + else: + obj["category_id"] = anno["category_id"] - 1 # Convert 1-indexed to 0-indexed + obj['frequency'] = cls_type_dict[anno["category_id"]] # used for open-set filtering + if filter_open_cls: # filter categories for open-set training + if obj['frequency'] == 'r': + continue + area_dict[obj['frequency']].append(anno["bbox"][2] * anno["bbox"][3]) + + segm = anno["segmentation"] # list[list[float]] + # filter out invalid polygons (< 3 points) + valid_segm = [poly for poly in segm if len(poly) % 2 == 0 and len(poly) >= 6] + assert len(segm) == len( + valid_segm + ), "Annotation contains an invalid polygon with < 3 points" + assert len(segm) > 0 + obj["segmentation"] = segm + objs.append(obj) + if (filter_open_cls or clip_gt_crop) and len(objs) == 0: # no annotation for this image + continue + record["annotations"] = objs + dataset_dicts.append(record) + + # For the training in open-set setting, map original category id to new category id number (base categories) + if filter_open_cls: + # get new category id in order + old_to_new = {} + for i in range(len(cls_type_dict)): + if cls_type_dict[i+1] != 'r': # cls_type_dict is 1-indexed + old_to_new[i] = len(old_to_new) + # map annotation to new category id + for record in dataset_dicts: + record.pop('not_exhaustive_category_ids') # won't be used + record.pop('neg_category_ids') # won't be used + for obj in record['annotations']: + obj['category_id'] = old_to_new[obj['category_id']] # 0-indexed id + assert obj['frequency'] != 'r' + logger.info("\n\nModel will be trained in the open-set setting! {} / {} categories are kept.\n".format(len(old_to_new),len(cls_type_dict))) + # calculate box area for each type of class + area_lst = np.array([0, 400, 1600, 2500, 5000, 10000, 22500, 224 * 224, 90000, 160000, 1e8]) + # rare_cls = np.histogram(np.array(area_dict['r']), bins=area_lst)[0] + # common_cls = np.histogram(np.array(area_dict['c']), bins=area_lst)[0] + # freq_cls = np.histogram(np.array(area_dict['f']), bins=area_lst)[0] + # print("rare classes: {}; \ncommon classes: {}; \nfrequent classes: {}".format(rare_cls/rare_cls.sum()*100, common_cls/common_cls.sum()*100, freq_cls/freq_cls.sum()*100)) + # # apply CLIP on GT regions: some images has large number of GT bbox (eg, 759), remove them, otherwise, OOM + if clip_gt_crop: + # len_num = sorted([len(item["annotations"]) for item in dataset_dicts], reverse=True) + dataset_dicts = sorted(dataset_dicts, key=lambda x: len(x["annotations"]), reverse=True) + for record in dataset_dicts: + record["annotations"] = record["annotations"][:max_gt_per_img] # only <10 / 20k images in test have >300 GT boxes + #dataset_dicts = sorted(dataset_dicts, key=lambda x: len(x["annotations"]))[:12] #[12000:14000] # + #dataset_dicts = sorted(dataset_dicts, key=lambda x: len(x["annotations"]))[-1200:-1000] + #eval_cls_acc(dataset_dicts, area_lst) + return dataset_dicts + +def load_lvis_json(json_file, image_root, dataset_name=None, filter_open_cls=True, clip_gt_crop=True, max_gt_per_img=500, custom_img_path='datasets/custom_images'): + """ + This is a tentitive function for loading custom images. + Given a folder of images (eg, 'datasets/custom_images'), load their meta data into a dictionary + """ + import os + from PIL import Image + custom_img_list = [os.path.join(custom_img_path, item) for item in os.listdir(custom_img_path)] + + dataset_dicts = [] + for f_i, file_name in enumerate(custom_img_list): + record = {} + record["file_name"] = file_name + img_file = Image.open(record["file_name"]) + record["height"] = img_file.size[1] + record["width"] = img_file.size[0] + record["image_id"] = f_i + + dataset_dicts.append(record) + + return dataset_dicts + +def eval_cls_acc(dataset_dicts, area_lst): + #pred_file = '/home/v-yiwuzhong/projects/detectron2-open-set/output/rcnn_gt_crop/vit/instances_predictions.pth' + #pred_file = '/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/results/test_CLIP_rcnn_resnet50_crop_regions_perclassnms/inference/instances_predictions.pth' + #pred_file = '/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/results/test_CLIP_rcnn_vitb32_crop_regions_perclassnms/inference/instances_predictions.pth' + #pred_file = '/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/results/test_CLIP_fast_rcnn_resnet50_roifeatmap/inference/instances_predictions.pth' + #pred_file = '/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/results/test_CLIP_fast_rcnn_resnet50_supmrcnnbaselinefpn/inference/instances_predictions.pth' + #pred_file = '/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/results/test_CLIP_fast_rcnn_resnet50_supmrcnnbaselinec4/inference/instances_predictions.pth' + pred_file = '/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/results/test_CLIP_fast_rcnn_resnet50_e1-3-3gtbox/inference/instances_predictions.pth' + predictions = torch.load(pred_file) + correct = 0 + wrong = 0 + area_threshold = area_lst[1:-1] # np.array([400, 1600, 2500, 5000, 10000, 22500, 224 * 224, 90000, 160000]) + acc_list = [[0, 0] for i in range(area_threshold.shape[0] + 1)] + small_cnt = 0 + for preds, gts in zip(predictions, dataset_dicts): + assert preds['image_id'] == gts['image_id'] # same image + #assert len(preds['instances']) == len(gts['annotations']) + box_seen = {} # keep a set for the predicted boxes that have been checked + for pred, gt in zip(preds['instances'], gts['annotations']): + if pred['bbox'][0] in box_seen: # duplicate box due to perclass NMS + continue + else: + box_seen[pred['bbox'][0]] = 1 + if np.sum(np.array(pred['bbox']) - np.array(gt['bbox'])) < 1.0: # same box + pass + else: # has been NMS and shuffled + for gt in gts['annotations']: + if np.sum(np.array(pred['bbox']) - np.array(gt['bbox'])) < 1.0: # same box + break + assert np.sum(np.array(pred['bbox']) - np.array(gt['bbox'])) < 1.0 # same box + this_area = gt['bbox'][2] * gt['bbox'][3] + block = (area_threshold < this_area).nonzero()[0].shape[0] + if pred['category_id'] == gt['category_id']: # matched + correct += 1 + acc_list[block][0] += 1 + else: + wrong += 1 + acc_list[block][1] += 1 + + print("\n\nGot correct {} and wrong {}. Accuracy is {} / {} = {}\n\n".format(correct,wrong,correct,correct+wrong,correct/(correct+wrong))) + block_acc = [100 * acc_list[i][0] / (acc_list[i][0] + acc_list[i][1]) for i in range(len(acc_list))] + block_acc = [round(i, 1) for i in block_acc] + print("Block accuracy: {}".format(block_acc)) + block_num = [acc_list[i][0] + acc_list[i][1] for i in range(len(acc_list))] + block_num = list(block_num / np.sum(block_num) * 100) + block_num = [round(i, 1) for i in block_num] + print("Block #instances: {}".format(block_num)) + return + +def get_lvis_instances_meta(dataset_name): + """ + Load LVIS metadata. + + Args: + dataset_name (str): LVIS dataset name without the split name (e.g., "lvis_v0.5"). + + Returns: + dict: LVIS metadata with keys: thing_classes + """ + if "cocofied" in dataset_name: + return _get_coco_instances_meta() + if "v0.5" in dataset_name: + return _get_lvis_instances_meta_v0_5() + elif "v1" in dataset_name: + return _get_lvis_instances_meta_v1() + raise ValueError("No built-in metadata for dataset {}".format(dataset_name)) + + +def _get_lvis_instances_meta_v0_5(): + assert len(LVIS_V0_5_CATEGORIES) == 1230 + cat_ids = [k["id"] for k in LVIS_V0_5_CATEGORIES] + assert min(cat_ids) == 1 and max(cat_ids) == len( + cat_ids + ), "Category ids are not in [1, #categories], as expected" + # Ensure that the category list is sorted by id + lvis_categories = sorted(LVIS_V0_5_CATEGORIES, key=lambda x: x["id"]) + thing_classes = [k["synonyms"][0] for k in lvis_categories] + meta = {"thing_classes": thing_classes} + return meta + + +def _get_lvis_instances_meta_v1(): + assert len(LVIS_V1_CATEGORIES) == 1203 + cat_ids = [k["id"] for k in LVIS_V1_CATEGORIES] + assert min(cat_ids) == 1 and max(cat_ids) == len( + cat_ids + ), "Category ids are not in [1, #categories], as expected" + # Ensure that the category list is sorted by id + lvis_categories = sorted(LVIS_V1_CATEGORIES, key=lambda x: x["id"]) + thing_classes = [k["synonyms"][0] for k in lvis_categories] + meta = {"thing_classes": thing_classes} + return meta + + +if __name__ == "__main__": + """ + Test the LVIS json dataset loader. + + Usage: + python -m detectron2.data.datasets.lvis \ + path/to/json path/to/image_root dataset_name vis_limit + """ + import sys + import numpy as np + from detectron2.utils.logger import setup_logger + from PIL import Image + import detectron2.data.datasets # noqa # add pre-defined metadata + from detectron2.utils.visualizer import Visualizer + + logger = setup_logger(name=__name__) + meta = MetadataCatalog.get(sys.argv[3]) + + dicts = load_lvis_json(sys.argv[1], sys.argv[2], sys.argv[3]) + logger.info("Done loading {} samples.".format(len(dicts))) + + dirname = "lvis-data-vis" + os.makedirs(dirname, exist_ok=True) + for d in dicts[: int(sys.argv[4])]: + img = np.array(Image.open(d["file_name"])) + visualizer = Visualizer(img, metadata=meta) + vis = visualizer.draw_dataset_dict(d) + fpath = os.path.join(dirname, os.path.basename(d["file_name"])) + vis.save(fpath) diff --git a/detectron2/data/datasets/lvis_v0_5_categories.py b/detectron2/data/datasets/lvis_v0_5_categories.py new file mode 100644 index 0000000000000000000000000000000000000000..d3dab6198da614937b08682f4c9edf52bdf1d236 --- /dev/null +++ b/detectron2/data/datasets/lvis_v0_5_categories.py @@ -0,0 +1,13 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# Autogen with +# with open("lvis_v0.5_val.json", "r") as f: +# a = json.load(f) +# c = a["categories"] +# for x in c: +# del x["image_count"] +# del x["instance_count"] +# LVIS_CATEGORIES = repr(c) + " # noqa" + +# fmt: off +LVIS_CATEGORIES = [{'frequency': 'r', 'id': 1, 'synset': 'acorn.n.01', 'synonyms': ['acorn'], 'def': 'nut from an oak tree', 'name': 'acorn'}, {'frequency': 'c', 'id': 2, 'synset': 'aerosol.n.02', 'synonyms': ['aerosol_can', 'spray_can'], 'def': 'a dispenser that holds a substance under pressure', 'name': 'aerosol_can'}, {'frequency': 'f', 'id': 3, 'synset': 'air_conditioner.n.01', 'synonyms': ['air_conditioner'], 'def': 'a machine that keeps air cool and dry', 'name': 'air_conditioner'}, {'frequency': 'f', 'id': 4, 'synset': 'airplane.n.01', 'synonyms': ['airplane', 'aeroplane'], 'def': 'an aircraft that has a fixed wing and is powered by propellers or jets', 'name': 'airplane'}, {'frequency': 'c', 'id': 5, 'synset': 'alarm_clock.n.01', 'synonyms': ['alarm_clock'], 'def': 'a clock that wakes a sleeper at some preset time', 'name': 'alarm_clock'}, {'frequency': 'c', 'id': 6, 'synset': 'alcohol.n.01', 'synonyms': ['alcohol', 'alcoholic_beverage'], 'def': 'a liquor or brew containing alcohol as the active agent', 'name': 'alcohol'}, {'frequency': 'r', 'id': 7, 'synset': 'alligator.n.02', 'synonyms': ['alligator', 'gator'], 'def': 'amphibious reptiles related to crocodiles but with shorter broader snouts', 'name': 'alligator'}, {'frequency': 'c', 'id': 8, 'synset': 'almond.n.02', 'synonyms': ['almond'], 'def': 'oval-shaped edible seed of the almond tree', 'name': 'almond'}, {'frequency': 'c', 'id': 9, 'synset': 'ambulance.n.01', 'synonyms': ['ambulance'], 'def': 'a vehicle that takes people to and from hospitals', 'name': 'ambulance'}, {'frequency': 'r', 'id': 10, 'synset': 'amplifier.n.01', 'synonyms': ['amplifier'], 'def': 'electronic equipment that increases strength of signals', 'name': 'amplifier'}, {'frequency': 'c', 'id': 11, 'synset': 'anklet.n.03', 'synonyms': ['anklet', 'ankle_bracelet'], 'def': 'an ornament worn around the ankle', 'name': 'anklet'}, {'frequency': 'f', 'id': 12, 'synset': 'antenna.n.01', 'synonyms': ['antenna', 'aerial', 'transmitting_aerial'], 'def': 'an electrical device that sends or receives radio or television signals', 'name': 'antenna'}, {'frequency': 'f', 'id': 13, 'synset': 'apple.n.01', 'synonyms': ['apple'], 'def': 'fruit with red or yellow or green skin and sweet to tart crisp whitish flesh', 'name': 'apple'}, {'frequency': 'r', 'id': 14, 'synset': 'apple_juice.n.01', 'synonyms': ['apple_juice'], 'def': 'the juice of apples', 'name': 'apple_juice'}, {'frequency': 'r', 'id': 15, 'synset': 'applesauce.n.01', 'synonyms': ['applesauce'], 'def': 'puree of stewed apples usually sweetened and spiced', 'name': 'applesauce'}, {'frequency': 'r', 'id': 16, 'synset': 'apricot.n.02', 'synonyms': ['apricot'], 'def': 'downy yellow to rosy-colored fruit resembling a small peach', 'name': 'apricot'}, {'frequency': 'f', 'id': 17, 'synset': 'apron.n.01', 'synonyms': ['apron'], 'def': 'a garment of cloth that is tied about the waist and worn to protect clothing', 'name': 'apron'}, {'frequency': 'c', 'id': 18, 'synset': 'aquarium.n.01', 'synonyms': ['aquarium', 'fish_tank'], 'def': 'a tank/pool/bowl filled with water for keeping live fish and underwater animals', 'name': 'aquarium'}, {'frequency': 'c', 'id': 19, 'synset': 'armband.n.02', 'synonyms': ['armband'], 'def': 'a band worn around the upper arm', 'name': 'armband'}, {'frequency': 'f', 'id': 20, 'synset': 'armchair.n.01', 'synonyms': ['armchair'], 'def': 'chair with a support on each side for arms', 'name': 'armchair'}, {'frequency': 'r', 'id': 21, 'synset': 'armoire.n.01', 'synonyms': ['armoire'], 'def': 'a large wardrobe or cabinet', 'name': 'armoire'}, {'frequency': 'r', 'id': 22, 'synset': 'armor.n.01', 'synonyms': ['armor', 'armour'], 'def': 'protective covering made of metal and used in combat', 'name': 'armor'}, {'frequency': 'c', 'id': 23, 'synset': 'artichoke.n.02', 'synonyms': ['artichoke'], 'def': 'a thistlelike flower head with edible fleshy leaves and heart', 'name': 'artichoke'}, {'frequency': 'f', 'id': 24, 'synset': 'ashcan.n.01', 'synonyms': ['trash_can', 'garbage_can', 'wastebin', 'dustbin', 'trash_barrel', 'trash_bin'], 'def': 'a bin that holds rubbish until it is collected', 'name': 'trash_can'}, {'frequency': 'c', 'id': 25, 'synset': 'ashtray.n.01', 'synonyms': ['ashtray'], 'def': "a receptacle for the ash from smokers' cigars or cigarettes", 'name': 'ashtray'}, {'frequency': 'c', 'id': 26, 'synset': 'asparagus.n.02', 'synonyms': ['asparagus'], 'def': 'edible young shoots of the asparagus plant', 'name': 'asparagus'}, {'frequency': 'c', 'id': 27, 'synset': 'atomizer.n.01', 'synonyms': ['atomizer', 'atomiser', 'spray', 'sprayer', 'nebulizer', 'nebuliser'], 'def': 'a dispenser that turns a liquid (such as perfume) into a fine mist', 'name': 'atomizer'}, {'frequency': 'c', 'id': 28, 'synset': 'avocado.n.01', 'synonyms': ['avocado'], 'def': 'a pear-shaped fruit with green or blackish skin and rich yellowish pulp enclosing a single large seed', 'name': 'avocado'}, {'frequency': 'c', 'id': 29, 'synset': 'award.n.02', 'synonyms': ['award', 'accolade'], 'def': 'a tangible symbol signifying approval or distinction', 'name': 'award'}, {'frequency': 'f', 'id': 30, 'synset': 'awning.n.01', 'synonyms': ['awning'], 'def': 'a canopy made of canvas to shelter people or things from rain or sun', 'name': 'awning'}, {'frequency': 'r', 'id': 31, 'synset': 'ax.n.01', 'synonyms': ['ax', 'axe'], 'def': 'an edge tool with a heavy bladed head mounted across a handle', 'name': 'ax'}, {'frequency': 'f', 'id': 32, 'synset': 'baby_buggy.n.01', 'synonyms': ['baby_buggy', 'baby_carriage', 'perambulator', 'pram', 'stroller'], 'def': 'a small vehicle with four wheels in which a baby or child is pushed around', 'name': 'baby_buggy'}, {'frequency': 'c', 'id': 33, 'synset': 'backboard.n.01', 'synonyms': ['basketball_backboard'], 'def': 'a raised vertical board with basket attached; used to play basketball', 'name': 'basketball_backboard'}, {'frequency': 'f', 'id': 34, 'synset': 'backpack.n.01', 'synonyms': ['backpack', 'knapsack', 'packsack', 'rucksack', 'haversack'], 'def': 'a bag carried by a strap on your back or shoulder', 'name': 'backpack'}, {'frequency': 'f', 'id': 35, 'synset': 'bag.n.04', 'synonyms': ['handbag', 'purse', 'pocketbook'], 'def': 'a container used for carrying money and small personal items or accessories', 'name': 'handbag'}, {'frequency': 'f', 'id': 36, 'synset': 'bag.n.06', 'synonyms': ['suitcase', 'baggage', 'luggage'], 'def': 'cases used to carry belongings when traveling', 'name': 'suitcase'}, {'frequency': 'c', 'id': 37, 'synset': 'bagel.n.01', 'synonyms': ['bagel', 'beigel'], 'def': 'glazed yeast-raised doughnut-shaped roll with hard crust', 'name': 'bagel'}, {'frequency': 'r', 'id': 38, 'synset': 'bagpipe.n.01', 'synonyms': ['bagpipe'], 'def': 'a tubular wind instrument; the player blows air into a bag and squeezes it out', 'name': 'bagpipe'}, {'frequency': 'r', 'id': 39, 'synset': 'baguet.n.01', 'synonyms': ['baguet', 'baguette'], 'def': 'narrow French stick loaf', 'name': 'baguet'}, {'frequency': 'r', 'id': 40, 'synset': 'bait.n.02', 'synonyms': ['bait', 'lure'], 'def': 'something used to lure fish or other animals into danger so they can be trapped or killed', 'name': 'bait'}, {'frequency': 'f', 'id': 41, 'synset': 'ball.n.06', 'synonyms': ['ball'], 'def': 'a spherical object used as a plaything', 'name': 'ball'}, {'frequency': 'r', 'id': 42, 'synset': 'ballet_skirt.n.01', 'synonyms': ['ballet_skirt', 'tutu'], 'def': 'very short skirt worn by ballerinas', 'name': 'ballet_skirt'}, {'frequency': 'f', 'id': 43, 'synset': 'balloon.n.01', 'synonyms': ['balloon'], 'def': 'large tough nonrigid bag filled with gas or heated air', 'name': 'balloon'}, {'frequency': 'c', 'id': 44, 'synset': 'bamboo.n.02', 'synonyms': ['bamboo'], 'def': 'woody tropical grass having hollow woody stems', 'name': 'bamboo'}, {'frequency': 'f', 'id': 45, 'synset': 'banana.n.02', 'synonyms': ['banana'], 'def': 'elongated crescent-shaped yellow fruit with soft sweet flesh', 'name': 'banana'}, {'frequency': 'r', 'id': 46, 'synset': 'band_aid.n.01', 'synonyms': ['Band_Aid'], 'def': 'trade name for an adhesive bandage to cover small cuts or blisters', 'name': 'Band_Aid'}, {'frequency': 'c', 'id': 47, 'synset': 'bandage.n.01', 'synonyms': ['bandage'], 'def': 'a piece of soft material that covers and protects an injured part of the body', 'name': 'bandage'}, {'frequency': 'c', 'id': 48, 'synset': 'bandanna.n.01', 'synonyms': ['bandanna', 'bandana'], 'def': 'large and brightly colored handkerchief; often used as a neckerchief', 'name': 'bandanna'}, {'frequency': 'r', 'id': 49, 'synset': 'banjo.n.01', 'synonyms': ['banjo'], 'def': 'a stringed instrument of the guitar family with a long neck and circular body', 'name': 'banjo'}, {'frequency': 'f', 'id': 50, 'synset': 'banner.n.01', 'synonyms': ['banner', 'streamer'], 'def': 'long strip of cloth or paper used for decoration or advertising', 'name': 'banner'}, {'frequency': 'r', 'id': 51, 'synset': 'barbell.n.01', 'synonyms': ['barbell'], 'def': 'a bar to which heavy discs are attached at each end; used in weightlifting', 'name': 'barbell'}, {'frequency': 'r', 'id': 52, 'synset': 'barge.n.01', 'synonyms': ['barge'], 'def': 'a flatbottom boat for carrying heavy loads (especially on canals)', 'name': 'barge'}, {'frequency': 'f', 'id': 53, 'synset': 'barrel.n.02', 'synonyms': ['barrel', 'cask'], 'def': 'a cylindrical container that holds liquids', 'name': 'barrel'}, {'frequency': 'c', 'id': 54, 'synset': 'barrette.n.01', 'synonyms': ['barrette'], 'def': "a pin for holding women's hair in place", 'name': 'barrette'}, {'frequency': 'c', 'id': 55, 'synset': 'barrow.n.03', 'synonyms': ['barrow', 'garden_cart', 'lawn_cart', 'wheelbarrow'], 'def': 'a cart for carrying small loads; has handles and one or more wheels', 'name': 'barrow'}, {'frequency': 'f', 'id': 56, 'synset': 'base.n.03', 'synonyms': ['baseball_base'], 'def': 'a place that the runner must touch before scoring', 'name': 'baseball_base'}, {'frequency': 'f', 'id': 57, 'synset': 'baseball.n.02', 'synonyms': ['baseball'], 'def': 'a ball used in playing baseball', 'name': 'baseball'}, {'frequency': 'f', 'id': 58, 'synset': 'baseball_bat.n.01', 'synonyms': ['baseball_bat'], 'def': 'an implement used in baseball by the batter', 'name': 'baseball_bat'}, {'frequency': 'f', 'id': 59, 'synset': 'baseball_cap.n.01', 'synonyms': ['baseball_cap', 'jockey_cap', 'golf_cap'], 'def': 'a cap with a bill', 'name': 'baseball_cap'}, {'frequency': 'f', 'id': 60, 'synset': 'baseball_glove.n.01', 'synonyms': ['baseball_glove', 'baseball_mitt'], 'def': 'the handwear used by fielders in playing baseball', 'name': 'baseball_glove'}, {'frequency': 'f', 'id': 61, 'synset': 'basket.n.01', 'synonyms': ['basket', 'handbasket'], 'def': 'a container that is usually woven and has handles', 'name': 'basket'}, {'frequency': 'c', 'id': 62, 'synset': 'basket.n.03', 'synonyms': ['basketball_hoop'], 'def': 'metal hoop supporting a net through which players try to throw the basketball', 'name': 'basketball_hoop'}, {'frequency': 'c', 'id': 63, 'synset': 'basketball.n.02', 'synonyms': ['basketball'], 'def': 'an inflated ball used in playing basketball', 'name': 'basketball'}, {'frequency': 'r', 'id': 64, 'synset': 'bass_horn.n.01', 'synonyms': ['bass_horn', 'sousaphone', 'tuba'], 'def': 'the lowest brass wind instrument', 'name': 'bass_horn'}, {'frequency': 'r', 'id': 65, 'synset': 'bat.n.01', 'synonyms': ['bat_(animal)'], 'def': 'nocturnal mouselike mammal with forelimbs modified to form membranous wings', 'name': 'bat_(animal)'}, {'frequency': 'f', 'id': 66, 'synset': 'bath_mat.n.01', 'synonyms': ['bath_mat'], 'def': 'a heavy towel or mat to stand on while drying yourself after a bath', 'name': 'bath_mat'}, {'frequency': 'f', 'id': 67, 'synset': 'bath_towel.n.01', 'synonyms': ['bath_towel'], 'def': 'a large towel; to dry yourself after a bath', 'name': 'bath_towel'}, {'frequency': 'c', 'id': 68, 'synset': 'bathrobe.n.01', 'synonyms': ['bathrobe'], 'def': 'a loose-fitting robe of towelling; worn after a bath or swim', 'name': 'bathrobe'}, {'frequency': 'f', 'id': 69, 'synset': 'bathtub.n.01', 'synonyms': ['bathtub', 'bathing_tub'], 'def': 'a large open container that you fill with water and use to wash the body', 'name': 'bathtub'}, {'frequency': 'r', 'id': 70, 'synset': 'batter.n.02', 'synonyms': ['batter_(food)'], 'def': 'a liquid or semiliquid mixture, as of flour, eggs, and milk, used in cooking', 'name': 'batter_(food)'}, {'frequency': 'c', 'id': 71, 'synset': 'battery.n.02', 'synonyms': ['battery'], 'def': 'a portable device that produces electricity', 'name': 'battery'}, {'frequency': 'r', 'id': 72, 'synset': 'beach_ball.n.01', 'synonyms': ['beachball'], 'def': 'large and light ball; for play at the seaside', 'name': 'beachball'}, {'frequency': 'c', 'id': 73, 'synset': 'bead.n.01', 'synonyms': ['bead'], 'def': 'a small ball with a hole through the middle used for ornamentation, jewellery, etc.', 'name': 'bead'}, {'frequency': 'r', 'id': 74, 'synset': 'beaker.n.01', 'synonyms': ['beaker'], 'def': 'a flatbottomed jar made of glass or plastic; used for chemistry', 'name': 'beaker'}, {'frequency': 'c', 'id': 75, 'synset': 'bean_curd.n.01', 'synonyms': ['bean_curd', 'tofu'], 'def': 'cheeselike food made of curdled soybean milk', 'name': 'bean_curd'}, {'frequency': 'c', 'id': 76, 'synset': 'beanbag.n.01', 'synonyms': ['beanbag'], 'def': 'a bag filled with dried beans or similar items; used in games or to sit on', 'name': 'beanbag'}, {'frequency': 'f', 'id': 77, 'synset': 'beanie.n.01', 'synonyms': ['beanie', 'beany'], 'def': 'a small skullcap; formerly worn by schoolboys and college freshmen', 'name': 'beanie'}, {'frequency': 'f', 'id': 78, 'synset': 'bear.n.01', 'synonyms': ['bear'], 'def': 'large carnivorous or omnivorous mammals with shaggy coats and claws', 'name': 'bear'}, {'frequency': 'f', 'id': 79, 'synset': 'bed.n.01', 'synonyms': ['bed'], 'def': 'a piece of furniture that provides a place to sleep', 'name': 'bed'}, {'frequency': 'c', 'id': 80, 'synset': 'bedspread.n.01', 'synonyms': ['bedspread', 'bedcover', 'bed_covering', 'counterpane', 'spread'], 'def': 'decorative cover for a bed', 'name': 'bedspread'}, {'frequency': 'f', 'id': 81, 'synset': 'beef.n.01', 'synonyms': ['cow'], 'def': 'cattle that are reared for their meat', 'name': 'cow'}, {'frequency': 'c', 'id': 82, 'synset': 'beef.n.02', 'synonyms': ['beef_(food)', 'boeuf_(food)'], 'def': 'meat from an adult domestic bovine', 'name': 'beef_(food)'}, {'frequency': 'r', 'id': 83, 'synset': 'beeper.n.01', 'synonyms': ['beeper', 'pager'], 'def': 'an device that beeps when the person carrying it is being paged', 'name': 'beeper'}, {'frequency': 'f', 'id': 84, 'synset': 'beer_bottle.n.01', 'synonyms': ['beer_bottle'], 'def': 'a bottle that holds beer', 'name': 'beer_bottle'}, {'frequency': 'c', 'id': 85, 'synset': 'beer_can.n.01', 'synonyms': ['beer_can'], 'def': 'a can that holds beer', 'name': 'beer_can'}, {'frequency': 'r', 'id': 86, 'synset': 'beetle.n.01', 'synonyms': ['beetle'], 'def': 'insect with hard wing covers', 'name': 'beetle'}, {'frequency': 'f', 'id': 87, 'synset': 'bell.n.01', 'synonyms': ['bell'], 'def': 'a hollow device made of metal that makes a ringing sound when struck', 'name': 'bell'}, {'frequency': 'f', 'id': 88, 'synset': 'bell_pepper.n.02', 'synonyms': ['bell_pepper', 'capsicum'], 'def': 'large bell-shaped sweet pepper in green or red or yellow or orange or black varieties', 'name': 'bell_pepper'}, {'frequency': 'f', 'id': 89, 'synset': 'belt.n.02', 'synonyms': ['belt'], 'def': 'a band to tie or buckle around the body (usually at the waist)', 'name': 'belt'}, {'frequency': 'f', 'id': 90, 'synset': 'belt_buckle.n.01', 'synonyms': ['belt_buckle'], 'def': 'the buckle used to fasten a belt', 'name': 'belt_buckle'}, {'frequency': 'f', 'id': 91, 'synset': 'bench.n.01', 'synonyms': ['bench'], 'def': 'a long seat for more than one person', 'name': 'bench'}, {'frequency': 'c', 'id': 92, 'synset': 'beret.n.01', 'synonyms': ['beret'], 'def': 'a cap with no brim or bill; made of soft cloth', 'name': 'beret'}, {'frequency': 'c', 'id': 93, 'synset': 'bib.n.02', 'synonyms': ['bib'], 'def': 'a napkin tied under the chin of a child while eating', 'name': 'bib'}, {'frequency': 'r', 'id': 94, 'synset': 'bible.n.01', 'synonyms': ['Bible'], 'def': 'the sacred writings of the Christian religions', 'name': 'Bible'}, {'frequency': 'f', 'id': 95, 'synset': 'bicycle.n.01', 'synonyms': ['bicycle', 'bike_(bicycle)'], 'def': 'a wheeled vehicle that has two wheels and is moved by foot pedals', 'name': 'bicycle'}, {'frequency': 'f', 'id': 96, 'synset': 'bill.n.09', 'synonyms': ['visor', 'vizor'], 'def': 'a brim that projects to the front to shade the eyes', 'name': 'visor'}, {'frequency': 'c', 'id': 97, 'synset': 'binder.n.03', 'synonyms': ['binder', 'ring-binder'], 'def': 'holds loose papers or magazines', 'name': 'binder'}, {'frequency': 'c', 'id': 98, 'synset': 'binoculars.n.01', 'synonyms': ['binoculars', 'field_glasses', 'opera_glasses'], 'def': 'an optical instrument designed for simultaneous use by both eyes', 'name': 'binoculars'}, {'frequency': 'f', 'id': 99, 'synset': 'bird.n.01', 'synonyms': ['bird'], 'def': 'animal characterized by feathers and wings', 'name': 'bird'}, {'frequency': 'r', 'id': 100, 'synset': 'bird_feeder.n.01', 'synonyms': ['birdfeeder'], 'def': 'an outdoor device that supplies food for wild birds', 'name': 'birdfeeder'}, {'frequency': 'r', 'id': 101, 'synset': 'birdbath.n.01', 'synonyms': ['birdbath'], 'def': 'an ornamental basin (usually in a garden) for birds to bathe in', 'name': 'birdbath'}, {'frequency': 'c', 'id': 102, 'synset': 'birdcage.n.01', 'synonyms': ['birdcage'], 'def': 'a cage in which a bird can be kept', 'name': 'birdcage'}, {'frequency': 'c', 'id': 103, 'synset': 'birdhouse.n.01', 'synonyms': ['birdhouse'], 'def': 'a shelter for birds', 'name': 'birdhouse'}, {'frequency': 'f', 'id': 104, 'synset': 'birthday_cake.n.01', 'synonyms': ['birthday_cake'], 'def': 'decorated cake served at a birthday party', 'name': 'birthday_cake'}, {'frequency': 'r', 'id': 105, 'synset': 'birthday_card.n.01', 'synonyms': ['birthday_card'], 'def': 'a card expressing a birthday greeting', 'name': 'birthday_card'}, {'frequency': 'r', 'id': 106, 'synset': 'biscuit.n.01', 'synonyms': ['biscuit_(bread)'], 'def': 'small round bread leavened with baking-powder or soda', 'name': 'biscuit_(bread)'}, {'frequency': 'r', 'id': 107, 'synset': 'black_flag.n.01', 'synonyms': ['pirate_flag'], 'def': 'a flag usually bearing a white skull and crossbones on a black background', 'name': 'pirate_flag'}, {'frequency': 'c', 'id': 108, 'synset': 'black_sheep.n.02', 'synonyms': ['black_sheep'], 'def': 'sheep with a black coat', 'name': 'black_sheep'}, {'frequency': 'c', 'id': 109, 'synset': 'blackboard.n.01', 'synonyms': ['blackboard', 'chalkboard'], 'def': 'sheet of slate; for writing with chalk', 'name': 'blackboard'}, {'frequency': 'f', 'id': 110, 'synset': 'blanket.n.01', 'synonyms': ['blanket'], 'def': 'bedding that keeps a person warm in bed', 'name': 'blanket'}, {'frequency': 'c', 'id': 111, 'synset': 'blazer.n.01', 'synonyms': ['blazer', 'sport_jacket', 'sport_coat', 'sports_jacket', 'sports_coat'], 'def': 'lightweight jacket; often striped in the colors of a club or school', 'name': 'blazer'}, {'frequency': 'f', 'id': 112, 'synset': 'blender.n.01', 'synonyms': ['blender', 'liquidizer', 'liquidiser'], 'def': 'an electrically powered mixer that mix or chop or liquefy foods', 'name': 'blender'}, {'frequency': 'r', 'id': 113, 'synset': 'blimp.n.02', 'synonyms': ['blimp'], 'def': 'a small nonrigid airship used for observation or as a barrage balloon', 'name': 'blimp'}, {'frequency': 'c', 'id': 114, 'synset': 'blinker.n.01', 'synonyms': ['blinker', 'flasher'], 'def': 'a light that flashes on and off; used as a signal or to send messages', 'name': 'blinker'}, {'frequency': 'c', 'id': 115, 'synset': 'blueberry.n.02', 'synonyms': ['blueberry'], 'def': 'sweet edible dark-blue berries of blueberry plants', 'name': 'blueberry'}, {'frequency': 'r', 'id': 116, 'synset': 'boar.n.02', 'synonyms': ['boar'], 'def': 'an uncastrated male hog', 'name': 'boar'}, {'frequency': 'r', 'id': 117, 'synset': 'board.n.09', 'synonyms': ['gameboard'], 'def': 'a flat portable surface (usually rectangular) designed for board games', 'name': 'gameboard'}, {'frequency': 'f', 'id': 118, 'synset': 'boat.n.01', 'synonyms': ['boat', 'ship_(boat)'], 'def': 'a vessel for travel on water', 'name': 'boat'}, {'frequency': 'c', 'id': 119, 'synset': 'bobbin.n.01', 'synonyms': ['bobbin', 'spool', 'reel'], 'def': 'a thing around which thread/tape/film or other flexible materials can be wound', 'name': 'bobbin'}, {'frequency': 'r', 'id': 120, 'synset': 'bobby_pin.n.01', 'synonyms': ['bobby_pin', 'hairgrip'], 'def': 'a flat wire hairpin used to hold bobbed hair in place', 'name': 'bobby_pin'}, {'frequency': 'c', 'id': 121, 'synset': 'boiled_egg.n.01', 'synonyms': ['boiled_egg', 'coddled_egg'], 'def': 'egg cooked briefly in the shell in gently boiling water', 'name': 'boiled_egg'}, {'frequency': 'r', 'id': 122, 'synset': 'bolo_tie.n.01', 'synonyms': ['bolo_tie', 'bolo', 'bola_tie', 'bola'], 'def': 'a cord fastened around the neck with an ornamental clasp and worn as a necktie', 'name': 'bolo_tie'}, {'frequency': 'c', 'id': 123, 'synset': 'bolt.n.03', 'synonyms': ['deadbolt'], 'def': 'the part of a lock that is engaged or withdrawn with a key', 'name': 'deadbolt'}, {'frequency': 'f', 'id': 124, 'synset': 'bolt.n.06', 'synonyms': ['bolt'], 'def': 'a screw that screws into a nut to form a fastener', 'name': 'bolt'}, {'frequency': 'r', 'id': 125, 'synset': 'bonnet.n.01', 'synonyms': ['bonnet'], 'def': 'a hat tied under the chin', 'name': 'bonnet'}, {'frequency': 'f', 'id': 126, 'synset': 'book.n.01', 'synonyms': ['book'], 'def': 'a written work or composition that has been published', 'name': 'book'}, {'frequency': 'r', 'id': 127, 'synset': 'book_bag.n.01', 'synonyms': ['book_bag'], 'def': 'a bag in which students carry their books', 'name': 'book_bag'}, {'frequency': 'c', 'id': 128, 'synset': 'bookcase.n.01', 'synonyms': ['bookcase'], 'def': 'a piece of furniture with shelves for storing books', 'name': 'bookcase'}, {'frequency': 'c', 'id': 129, 'synset': 'booklet.n.01', 'synonyms': ['booklet', 'brochure', 'leaflet', 'pamphlet'], 'def': 'a small book usually having a paper cover', 'name': 'booklet'}, {'frequency': 'r', 'id': 130, 'synset': 'bookmark.n.01', 'synonyms': ['bookmark', 'bookmarker'], 'def': 'a marker (a piece of paper or ribbon) placed between the pages of a book', 'name': 'bookmark'}, {'frequency': 'r', 'id': 131, 'synset': 'boom.n.04', 'synonyms': ['boom_microphone', 'microphone_boom'], 'def': 'a pole carrying an overhead microphone projected over a film or tv set', 'name': 'boom_microphone'}, {'frequency': 'f', 'id': 132, 'synset': 'boot.n.01', 'synonyms': ['boot'], 'def': 'footwear that covers the whole foot and lower leg', 'name': 'boot'}, {'frequency': 'f', 'id': 133, 'synset': 'bottle.n.01', 'synonyms': ['bottle'], 'def': 'a glass or plastic vessel used for storing drinks or other liquids', 'name': 'bottle'}, {'frequency': 'c', 'id': 134, 'synset': 'bottle_opener.n.01', 'synonyms': ['bottle_opener'], 'def': 'an opener for removing caps or corks from bottles', 'name': 'bottle_opener'}, {'frequency': 'c', 'id': 135, 'synset': 'bouquet.n.01', 'synonyms': ['bouquet'], 'def': 'an arrangement of flowers that is usually given as a present', 'name': 'bouquet'}, {'frequency': 'r', 'id': 136, 'synset': 'bow.n.04', 'synonyms': ['bow_(weapon)'], 'def': 'a weapon for shooting arrows', 'name': 'bow_(weapon)'}, {'frequency': 'f', 'id': 137, 'synset': 'bow.n.08', 'synonyms': ['bow_(decorative_ribbons)'], 'def': 'a decorative interlacing of ribbons', 'name': 'bow_(decorative_ribbons)'}, {'frequency': 'f', 'id': 138, 'synset': 'bow_tie.n.01', 'synonyms': ['bow-tie', 'bowtie'], 'def': "a man's tie that ties in a bow", 'name': 'bow-tie'}, {'frequency': 'f', 'id': 139, 'synset': 'bowl.n.03', 'synonyms': ['bowl'], 'def': 'a dish that is round and open at the top for serving foods', 'name': 'bowl'}, {'frequency': 'r', 'id': 140, 'synset': 'bowl.n.08', 'synonyms': ['pipe_bowl'], 'def': 'a small round container that is open at the top for holding tobacco', 'name': 'pipe_bowl'}, {'frequency': 'c', 'id': 141, 'synset': 'bowler_hat.n.01', 'synonyms': ['bowler_hat', 'bowler', 'derby_hat', 'derby', 'plug_hat'], 'def': 'a felt hat that is round and hard with a narrow brim', 'name': 'bowler_hat'}, {'frequency': 'r', 'id': 142, 'synset': 'bowling_ball.n.01', 'synonyms': ['bowling_ball'], 'def': 'a large ball with finger holes used in the sport of bowling', 'name': 'bowling_ball'}, {'frequency': 'r', 'id': 143, 'synset': 'bowling_pin.n.01', 'synonyms': ['bowling_pin'], 'def': 'a club-shaped wooden object used in bowling', 'name': 'bowling_pin'}, {'frequency': 'r', 'id': 144, 'synset': 'boxing_glove.n.01', 'synonyms': ['boxing_glove'], 'def': 'large glove coverings the fists of a fighter worn for the sport of boxing', 'name': 'boxing_glove'}, {'frequency': 'c', 'id': 145, 'synset': 'brace.n.06', 'synonyms': ['suspenders'], 'def': 'elastic straps that hold trousers up (usually used in the plural)', 'name': 'suspenders'}, {'frequency': 'f', 'id': 146, 'synset': 'bracelet.n.02', 'synonyms': ['bracelet', 'bangle'], 'def': 'jewelry worn around the wrist for decoration', 'name': 'bracelet'}, {'frequency': 'r', 'id': 147, 'synset': 'brass.n.07', 'synonyms': ['brass_plaque'], 'def': 'a memorial made of brass', 'name': 'brass_plaque'}, {'frequency': 'c', 'id': 148, 'synset': 'brassiere.n.01', 'synonyms': ['brassiere', 'bra', 'bandeau'], 'def': 'an undergarment worn by women to support their breasts', 'name': 'brassiere'}, {'frequency': 'c', 'id': 149, 'synset': 'bread-bin.n.01', 'synonyms': ['bread-bin', 'breadbox'], 'def': 'a container used to keep bread or cake in', 'name': 'bread-bin'}, {'frequency': 'r', 'id': 150, 'synset': 'breechcloth.n.01', 'synonyms': ['breechcloth', 'breechclout', 'loincloth'], 'def': 'a garment that provides covering for the loins', 'name': 'breechcloth'}, {'frequency': 'c', 'id': 151, 'synset': 'bridal_gown.n.01', 'synonyms': ['bridal_gown', 'wedding_gown', 'wedding_dress'], 'def': 'a gown worn by the bride at a wedding', 'name': 'bridal_gown'}, {'frequency': 'c', 'id': 152, 'synset': 'briefcase.n.01', 'synonyms': ['briefcase'], 'def': 'a case with a handle; for carrying papers or files or books', 'name': 'briefcase'}, {'frequency': 'c', 'id': 153, 'synset': 'bristle_brush.n.01', 'synonyms': ['bristle_brush'], 'def': 'a brush that is made with the short stiff hairs of an animal or plant', 'name': 'bristle_brush'}, {'frequency': 'f', 'id': 154, 'synset': 'broccoli.n.01', 'synonyms': ['broccoli'], 'def': 'plant with dense clusters of tight green flower buds', 'name': 'broccoli'}, {'frequency': 'r', 'id': 155, 'synset': 'brooch.n.01', 'synonyms': ['broach'], 'def': 'a decorative pin worn by women', 'name': 'broach'}, {'frequency': 'c', 'id': 156, 'synset': 'broom.n.01', 'synonyms': ['broom'], 'def': 'bundle of straws or twigs attached to a long handle; used for cleaning', 'name': 'broom'}, {'frequency': 'c', 'id': 157, 'synset': 'brownie.n.03', 'synonyms': ['brownie'], 'def': 'square or bar of very rich chocolate cake usually with nuts', 'name': 'brownie'}, {'frequency': 'c', 'id': 158, 'synset': 'brussels_sprouts.n.01', 'synonyms': ['brussels_sprouts'], 'def': 'the small edible cabbage-like buds growing along a stalk', 'name': 'brussels_sprouts'}, {'frequency': 'r', 'id': 159, 'synset': 'bubble_gum.n.01', 'synonyms': ['bubble_gum'], 'def': 'a kind of chewing gum that can be blown into bubbles', 'name': 'bubble_gum'}, {'frequency': 'f', 'id': 160, 'synset': 'bucket.n.01', 'synonyms': ['bucket', 'pail'], 'def': 'a roughly cylindrical vessel that is open at the top', 'name': 'bucket'}, {'frequency': 'r', 'id': 161, 'synset': 'buggy.n.01', 'synonyms': ['horse_buggy'], 'def': 'a small lightweight carriage; drawn by a single horse', 'name': 'horse_buggy'}, {'frequency': 'c', 'id': 162, 'synset': 'bull.n.11', 'synonyms': ['bull'], 'def': 'mature male cow', 'name': 'bull'}, {'frequency': 'r', 'id': 163, 'synset': 'bulldog.n.01', 'synonyms': ['bulldog'], 'def': 'a thickset short-haired dog with a large head and strong undershot lower jaw', 'name': 'bulldog'}, {'frequency': 'r', 'id': 164, 'synset': 'bulldozer.n.01', 'synonyms': ['bulldozer', 'dozer'], 'def': 'large powerful tractor; a large blade in front flattens areas of ground', 'name': 'bulldozer'}, {'frequency': 'c', 'id': 165, 'synset': 'bullet_train.n.01', 'synonyms': ['bullet_train'], 'def': 'a high-speed passenger train', 'name': 'bullet_train'}, {'frequency': 'c', 'id': 166, 'synset': 'bulletin_board.n.02', 'synonyms': ['bulletin_board', 'notice_board'], 'def': 'a board that hangs on a wall; displays announcements', 'name': 'bulletin_board'}, {'frequency': 'r', 'id': 167, 'synset': 'bulletproof_vest.n.01', 'synonyms': ['bulletproof_vest'], 'def': 'a vest capable of resisting the impact of a bullet', 'name': 'bulletproof_vest'}, {'frequency': 'c', 'id': 168, 'synset': 'bullhorn.n.01', 'synonyms': ['bullhorn', 'megaphone'], 'def': 'a portable loudspeaker with built-in microphone and amplifier', 'name': 'bullhorn'}, {'frequency': 'r', 'id': 169, 'synset': 'bully_beef.n.01', 'synonyms': ['corned_beef', 'corn_beef'], 'def': 'beef cured or pickled in brine', 'name': 'corned_beef'}, {'frequency': 'f', 'id': 170, 'synset': 'bun.n.01', 'synonyms': ['bun', 'roll'], 'def': 'small rounded bread either plain or sweet', 'name': 'bun'}, {'frequency': 'c', 'id': 171, 'synset': 'bunk_bed.n.01', 'synonyms': ['bunk_bed'], 'def': 'beds built one above the other', 'name': 'bunk_bed'}, {'frequency': 'f', 'id': 172, 'synset': 'buoy.n.01', 'synonyms': ['buoy'], 'def': 'a float attached by rope to the seabed to mark channels in a harbor or underwater hazards', 'name': 'buoy'}, {'frequency': 'r', 'id': 173, 'synset': 'burrito.n.01', 'synonyms': ['burrito'], 'def': 'a flour tortilla folded around a filling', 'name': 'burrito'}, {'frequency': 'f', 'id': 174, 'synset': 'bus.n.01', 'synonyms': ['bus_(vehicle)', 'autobus', 'charabanc', 'double-decker', 'motorbus', 'motorcoach'], 'def': 'a vehicle carrying many passengers; used for public transport', 'name': 'bus_(vehicle)'}, {'frequency': 'c', 'id': 175, 'synset': 'business_card.n.01', 'synonyms': ['business_card'], 'def': "a card on which are printed the person's name and business affiliation", 'name': 'business_card'}, {'frequency': 'c', 'id': 176, 'synset': 'butcher_knife.n.01', 'synonyms': ['butcher_knife'], 'def': 'a large sharp knife for cutting or trimming meat', 'name': 'butcher_knife'}, {'frequency': 'c', 'id': 177, 'synset': 'butter.n.01', 'synonyms': ['butter'], 'def': 'an edible emulsion of fat globules made by churning milk or cream; for cooking and table use', 'name': 'butter'}, {'frequency': 'c', 'id': 178, 'synset': 'butterfly.n.01', 'synonyms': ['butterfly'], 'def': 'insect typically having a slender body with knobbed antennae and broad colorful wings', 'name': 'butterfly'}, {'frequency': 'f', 'id': 179, 'synset': 'button.n.01', 'synonyms': ['button'], 'def': 'a round fastener sewn to shirts and coats etc to fit through buttonholes', 'name': 'button'}, {'frequency': 'f', 'id': 180, 'synset': 'cab.n.03', 'synonyms': ['cab_(taxi)', 'taxi', 'taxicab'], 'def': 'a car that takes passengers where they want to go in exchange for money', 'name': 'cab_(taxi)'}, {'frequency': 'r', 'id': 181, 'synset': 'cabana.n.01', 'synonyms': ['cabana'], 'def': 'a small tent used as a dressing room beside the sea or a swimming pool', 'name': 'cabana'}, {'frequency': 'r', 'id': 182, 'synset': 'cabin_car.n.01', 'synonyms': ['cabin_car', 'caboose'], 'def': 'a car on a freight train for use of the train crew; usually the last car on the train', 'name': 'cabin_car'}, {'frequency': 'f', 'id': 183, 'synset': 'cabinet.n.01', 'synonyms': ['cabinet'], 'def': 'a piece of furniture resembling a cupboard with doors and shelves and drawers', 'name': 'cabinet'}, {'frequency': 'r', 'id': 184, 'synset': 'cabinet.n.03', 'synonyms': ['locker', 'storage_locker'], 'def': 'a storage compartment for clothes and valuables; usually it has a lock', 'name': 'locker'}, {'frequency': 'f', 'id': 185, 'synset': 'cake.n.03', 'synonyms': ['cake'], 'def': 'baked goods made from or based on a mixture of flour, sugar, eggs, and fat', 'name': 'cake'}, {'frequency': 'c', 'id': 186, 'synset': 'calculator.n.02', 'synonyms': ['calculator'], 'def': 'a small machine that is used for mathematical calculations', 'name': 'calculator'}, {'frequency': 'f', 'id': 187, 'synset': 'calendar.n.02', 'synonyms': ['calendar'], 'def': 'a list or register of events (appointments/social events/court cases, etc)', 'name': 'calendar'}, {'frequency': 'c', 'id': 188, 'synset': 'calf.n.01', 'synonyms': ['calf'], 'def': 'young of domestic cattle', 'name': 'calf'}, {'frequency': 'c', 'id': 189, 'synset': 'camcorder.n.01', 'synonyms': ['camcorder'], 'def': 'a portable television camera and videocassette recorder', 'name': 'camcorder'}, {'frequency': 'c', 'id': 190, 'synset': 'camel.n.01', 'synonyms': ['camel'], 'def': 'cud-chewing mammal used as a draft or saddle animal in desert regions', 'name': 'camel'}, {'frequency': 'f', 'id': 191, 'synset': 'camera.n.01', 'synonyms': ['camera'], 'def': 'equipment for taking photographs', 'name': 'camera'}, {'frequency': 'c', 'id': 192, 'synset': 'camera_lens.n.01', 'synonyms': ['camera_lens'], 'def': 'a lens that focuses the image in a camera', 'name': 'camera_lens'}, {'frequency': 'c', 'id': 193, 'synset': 'camper.n.02', 'synonyms': ['camper_(vehicle)', 'camping_bus', 'motor_home'], 'def': 'a recreational vehicle equipped for camping out while traveling', 'name': 'camper_(vehicle)'}, {'frequency': 'f', 'id': 194, 'synset': 'can.n.01', 'synonyms': ['can', 'tin_can'], 'def': 'airtight sealed metal container for food or drink or paint etc.', 'name': 'can'}, {'frequency': 'c', 'id': 195, 'synset': 'can_opener.n.01', 'synonyms': ['can_opener', 'tin_opener'], 'def': 'a device for cutting cans open', 'name': 'can_opener'}, {'frequency': 'r', 'id': 196, 'synset': 'candelabrum.n.01', 'synonyms': ['candelabrum', 'candelabra'], 'def': 'branched candlestick; ornamental; has several lights', 'name': 'candelabrum'}, {'frequency': 'f', 'id': 197, 'synset': 'candle.n.01', 'synonyms': ['candle', 'candlestick'], 'def': 'stick of wax with a wick in the middle', 'name': 'candle'}, {'frequency': 'f', 'id': 198, 'synset': 'candlestick.n.01', 'synonyms': ['candle_holder'], 'def': 'a holder with sockets for candles', 'name': 'candle_holder'}, {'frequency': 'r', 'id': 199, 'synset': 'candy_bar.n.01', 'synonyms': ['candy_bar'], 'def': 'a candy shaped as a bar', 'name': 'candy_bar'}, {'frequency': 'c', 'id': 200, 'synset': 'candy_cane.n.01', 'synonyms': ['candy_cane'], 'def': 'a hard candy in the shape of a rod (usually with stripes)', 'name': 'candy_cane'}, {'frequency': 'c', 'id': 201, 'synset': 'cane.n.01', 'synonyms': ['walking_cane'], 'def': 'a stick that people can lean on to help them walk', 'name': 'walking_cane'}, {'frequency': 'c', 'id': 202, 'synset': 'canister.n.02', 'synonyms': ['canister', 'cannister'], 'def': 'metal container for storing dry foods such as tea or flour', 'name': 'canister'}, {'frequency': 'r', 'id': 203, 'synset': 'cannon.n.02', 'synonyms': ['cannon'], 'def': 'heavy gun fired from a tank', 'name': 'cannon'}, {'frequency': 'c', 'id': 204, 'synset': 'canoe.n.01', 'synonyms': ['canoe'], 'def': 'small and light boat; pointed at both ends; propelled with a paddle', 'name': 'canoe'}, {'frequency': 'r', 'id': 205, 'synset': 'cantaloup.n.02', 'synonyms': ['cantaloup', 'cantaloupe'], 'def': 'the fruit of a cantaloup vine; small to medium-sized melon with yellowish flesh', 'name': 'cantaloup'}, {'frequency': 'r', 'id': 206, 'synset': 'canteen.n.01', 'synonyms': ['canteen'], 'def': 'a flask for carrying water; used by soldiers or travelers', 'name': 'canteen'}, {'frequency': 'c', 'id': 207, 'synset': 'cap.n.01', 'synonyms': ['cap_(headwear)'], 'def': 'a tight-fitting headwear', 'name': 'cap_(headwear)'}, {'frequency': 'f', 'id': 208, 'synset': 'cap.n.02', 'synonyms': ['bottle_cap', 'cap_(container_lid)'], 'def': 'a top (as for a bottle)', 'name': 'bottle_cap'}, {'frequency': 'r', 'id': 209, 'synset': 'cape.n.02', 'synonyms': ['cape'], 'def': 'a sleeveless garment like a cloak but shorter', 'name': 'cape'}, {'frequency': 'c', 'id': 210, 'synset': 'cappuccino.n.01', 'synonyms': ['cappuccino', 'coffee_cappuccino'], 'def': 'equal parts of espresso and steamed milk', 'name': 'cappuccino'}, {'frequency': 'f', 'id': 211, 'synset': 'car.n.01', 'synonyms': ['car_(automobile)', 'auto_(automobile)', 'automobile'], 'def': 'a motor vehicle with four wheels', 'name': 'car_(automobile)'}, {'frequency': 'f', 'id': 212, 'synset': 'car.n.02', 'synonyms': ['railcar_(part_of_a_train)', 'railway_car_(part_of_a_train)', 'railroad_car_(part_of_a_train)'], 'def': 'a wheeled vehicle adapted to the rails of railroad', 'name': 'railcar_(part_of_a_train)'}, {'frequency': 'r', 'id': 213, 'synset': 'car.n.04', 'synonyms': ['elevator_car'], 'def': 'where passengers ride up and down', 'name': 'elevator_car'}, {'frequency': 'r', 'id': 214, 'synset': 'car_battery.n.01', 'synonyms': ['car_battery', 'automobile_battery'], 'def': 'a battery in a motor vehicle', 'name': 'car_battery'}, {'frequency': 'c', 'id': 215, 'synset': 'card.n.02', 'synonyms': ['identity_card'], 'def': 'a card certifying the identity of the bearer', 'name': 'identity_card'}, {'frequency': 'c', 'id': 216, 'synset': 'card.n.03', 'synonyms': ['card'], 'def': 'a rectangular piece of paper used to send messages (e.g. greetings or pictures)', 'name': 'card'}, {'frequency': 'r', 'id': 217, 'synset': 'cardigan.n.01', 'synonyms': ['cardigan'], 'def': 'knitted jacket that is fastened up the front with buttons or a zipper', 'name': 'cardigan'}, {'frequency': 'r', 'id': 218, 'synset': 'cargo_ship.n.01', 'synonyms': ['cargo_ship', 'cargo_vessel'], 'def': 'a ship designed to carry cargo', 'name': 'cargo_ship'}, {'frequency': 'r', 'id': 219, 'synset': 'carnation.n.01', 'synonyms': ['carnation'], 'def': 'plant with pink to purple-red spice-scented usually double flowers', 'name': 'carnation'}, {'frequency': 'c', 'id': 220, 'synset': 'carriage.n.02', 'synonyms': ['horse_carriage'], 'def': 'a vehicle with wheels drawn by one or more horses', 'name': 'horse_carriage'}, {'frequency': 'f', 'id': 221, 'synset': 'carrot.n.01', 'synonyms': ['carrot'], 'def': 'deep orange edible root of the cultivated carrot plant', 'name': 'carrot'}, {'frequency': 'c', 'id': 222, 'synset': 'carryall.n.01', 'synonyms': ['tote_bag'], 'def': 'a capacious bag or basket', 'name': 'tote_bag'}, {'frequency': 'c', 'id': 223, 'synset': 'cart.n.01', 'synonyms': ['cart'], 'def': 'a heavy open wagon usually having two wheels and drawn by an animal', 'name': 'cart'}, {'frequency': 'c', 'id': 224, 'synset': 'carton.n.02', 'synonyms': ['carton'], 'def': 'a box made of cardboard; opens by flaps on top', 'name': 'carton'}, {'frequency': 'c', 'id': 225, 'synset': 'cash_register.n.01', 'synonyms': ['cash_register', 'register_(for_cash_transactions)'], 'def': 'a cashbox with an adding machine to register transactions', 'name': 'cash_register'}, {'frequency': 'r', 'id': 226, 'synset': 'casserole.n.01', 'synonyms': ['casserole'], 'def': 'food cooked and served in a casserole', 'name': 'casserole'}, {'frequency': 'r', 'id': 227, 'synset': 'cassette.n.01', 'synonyms': ['cassette'], 'def': 'a container that holds a magnetic tape used for recording or playing sound or video', 'name': 'cassette'}, {'frequency': 'c', 'id': 228, 'synset': 'cast.n.05', 'synonyms': ['cast', 'plaster_cast', 'plaster_bandage'], 'def': 'bandage consisting of a firm covering that immobilizes broken bones while they heal', 'name': 'cast'}, {'frequency': 'f', 'id': 229, 'synset': 'cat.n.01', 'synonyms': ['cat'], 'def': 'a domestic house cat', 'name': 'cat'}, {'frequency': 'c', 'id': 230, 'synset': 'cauliflower.n.02', 'synonyms': ['cauliflower'], 'def': 'edible compact head of white undeveloped flowers', 'name': 'cauliflower'}, {'frequency': 'r', 'id': 231, 'synset': 'caviar.n.01', 'synonyms': ['caviar', 'caviare'], 'def': "salted roe of sturgeon or other large fish; usually served as an hors d'oeuvre", 'name': 'caviar'}, {'frequency': 'c', 'id': 232, 'synset': 'cayenne.n.02', 'synonyms': ['cayenne_(spice)', 'cayenne_pepper_(spice)', 'red_pepper_(spice)'], 'def': 'ground pods and seeds of pungent red peppers of the genus Capsicum', 'name': 'cayenne_(spice)'}, {'frequency': 'c', 'id': 233, 'synset': 'cd_player.n.01', 'synonyms': ['CD_player'], 'def': 'electronic equipment for playing compact discs (CDs)', 'name': 'CD_player'}, {'frequency': 'c', 'id': 234, 'synset': 'celery.n.01', 'synonyms': ['celery'], 'def': 'widely cultivated herb with aromatic leaf stalks that are eaten raw or cooked', 'name': 'celery'}, {'frequency': 'f', 'id': 235, 'synset': 'cellular_telephone.n.01', 'synonyms': ['cellular_telephone', 'cellular_phone', 'cellphone', 'mobile_phone', 'smart_phone'], 'def': 'a hand-held mobile telephone', 'name': 'cellular_telephone'}, {'frequency': 'r', 'id': 236, 'synset': 'chain_mail.n.01', 'synonyms': ['chain_mail', 'ring_mail', 'chain_armor', 'chain_armour', 'ring_armor', 'ring_armour'], 'def': '(Middle Ages) flexible armor made of interlinked metal rings', 'name': 'chain_mail'}, {'frequency': 'f', 'id': 237, 'synset': 'chair.n.01', 'synonyms': ['chair'], 'def': 'a seat for one person, with a support for the back', 'name': 'chair'}, {'frequency': 'r', 'id': 238, 'synset': 'chaise_longue.n.01', 'synonyms': ['chaise_longue', 'chaise', 'daybed'], 'def': 'a long chair; for reclining', 'name': 'chaise_longue'}, {'frequency': 'r', 'id': 239, 'synset': 'champagne.n.01', 'synonyms': ['champagne'], 'def': 'a white sparkling wine produced in Champagne or resembling that produced there', 'name': 'champagne'}, {'frequency': 'f', 'id': 240, 'synset': 'chandelier.n.01', 'synonyms': ['chandelier'], 'def': 'branched lighting fixture; often ornate; hangs from the ceiling', 'name': 'chandelier'}, {'frequency': 'r', 'id': 241, 'synset': 'chap.n.04', 'synonyms': ['chap'], 'def': 'leather leggings without a seat; worn over trousers by cowboys to protect their legs', 'name': 'chap'}, {'frequency': 'r', 'id': 242, 'synset': 'checkbook.n.01', 'synonyms': ['checkbook', 'chequebook'], 'def': 'a book issued to holders of checking accounts', 'name': 'checkbook'}, {'frequency': 'r', 'id': 243, 'synset': 'checkerboard.n.01', 'synonyms': ['checkerboard'], 'def': 'a board having 64 squares of two alternating colors', 'name': 'checkerboard'}, {'frequency': 'c', 'id': 244, 'synset': 'cherry.n.03', 'synonyms': ['cherry'], 'def': 'a red fruit with a single hard stone', 'name': 'cherry'}, {'frequency': 'r', 'id': 245, 'synset': 'chessboard.n.01', 'synonyms': ['chessboard'], 'def': 'a checkerboard used to play chess', 'name': 'chessboard'}, {'frequency': 'r', 'id': 246, 'synset': 'chest_of_drawers.n.01', 'synonyms': ['chest_of_drawers_(furniture)', 'bureau_(furniture)', 'chest_(furniture)'], 'def': 'furniture with drawers for keeping clothes', 'name': 'chest_of_drawers_(furniture)'}, {'frequency': 'c', 'id': 247, 'synset': 'chicken.n.02', 'synonyms': ['chicken_(animal)'], 'def': 'a domestic fowl bred for flesh or eggs', 'name': 'chicken_(animal)'}, {'frequency': 'c', 'id': 248, 'synset': 'chicken_wire.n.01', 'synonyms': ['chicken_wire'], 'def': 'a galvanized wire network with a hexagonal mesh; used to build fences', 'name': 'chicken_wire'}, {'frequency': 'r', 'id': 249, 'synset': 'chickpea.n.01', 'synonyms': ['chickpea', 'garbanzo'], 'def': 'the seed of the chickpea plant; usually dried', 'name': 'chickpea'}, {'frequency': 'r', 'id': 250, 'synset': 'chihuahua.n.03', 'synonyms': ['Chihuahua'], 'def': 'an old breed of tiny short-haired dog with protruding eyes from Mexico', 'name': 'Chihuahua'}, {'frequency': 'r', 'id': 251, 'synset': 'chili.n.02', 'synonyms': ['chili_(vegetable)', 'chili_pepper_(vegetable)', 'chilli_(vegetable)', 'chilly_(vegetable)', 'chile_(vegetable)'], 'def': 'very hot and finely tapering pepper of special pungency', 'name': 'chili_(vegetable)'}, {'frequency': 'r', 'id': 252, 'synset': 'chime.n.01', 'synonyms': ['chime', 'gong'], 'def': 'an instrument consisting of a set of bells that are struck with a hammer', 'name': 'chime'}, {'frequency': 'r', 'id': 253, 'synset': 'chinaware.n.01', 'synonyms': ['chinaware'], 'def': 'dishware made of high quality porcelain', 'name': 'chinaware'}, {'frequency': 'c', 'id': 254, 'synset': 'chip.n.04', 'synonyms': ['crisp_(potato_chip)', 'potato_chip'], 'def': 'a thin crisp slice of potato fried in deep fat', 'name': 'crisp_(potato_chip)'}, {'frequency': 'r', 'id': 255, 'synset': 'chip.n.06', 'synonyms': ['poker_chip'], 'def': 'a small disk-shaped counter used to represent money when gambling', 'name': 'poker_chip'}, {'frequency': 'c', 'id': 256, 'synset': 'chocolate_bar.n.01', 'synonyms': ['chocolate_bar'], 'def': 'a bar of chocolate candy', 'name': 'chocolate_bar'}, {'frequency': 'c', 'id': 257, 'synset': 'chocolate_cake.n.01', 'synonyms': ['chocolate_cake'], 'def': 'cake containing chocolate', 'name': 'chocolate_cake'}, {'frequency': 'r', 'id': 258, 'synset': 'chocolate_milk.n.01', 'synonyms': ['chocolate_milk'], 'def': 'milk flavored with chocolate syrup', 'name': 'chocolate_milk'}, {'frequency': 'r', 'id': 259, 'synset': 'chocolate_mousse.n.01', 'synonyms': ['chocolate_mousse'], 'def': 'dessert mousse made with chocolate', 'name': 'chocolate_mousse'}, {'frequency': 'f', 'id': 260, 'synset': 'choker.n.03', 'synonyms': ['choker', 'collar', 'neckband'], 'def': 'necklace that fits tightly around the neck', 'name': 'choker'}, {'frequency': 'f', 'id': 261, 'synset': 'chopping_board.n.01', 'synonyms': ['chopping_board', 'cutting_board', 'chopping_block'], 'def': 'a wooden board where meats or vegetables can be cut', 'name': 'chopping_board'}, {'frequency': 'c', 'id': 262, 'synset': 'chopstick.n.01', 'synonyms': ['chopstick'], 'def': 'one of a pair of slender sticks used as oriental tableware to eat food with', 'name': 'chopstick'}, {'frequency': 'f', 'id': 263, 'synset': 'christmas_tree.n.05', 'synonyms': ['Christmas_tree'], 'def': 'an ornamented evergreen used as a Christmas decoration', 'name': 'Christmas_tree'}, {'frequency': 'c', 'id': 264, 'synset': 'chute.n.02', 'synonyms': ['slide'], 'def': 'sloping channel through which things can descend', 'name': 'slide'}, {'frequency': 'r', 'id': 265, 'synset': 'cider.n.01', 'synonyms': ['cider', 'cyder'], 'def': 'a beverage made from juice pressed from apples', 'name': 'cider'}, {'frequency': 'r', 'id': 266, 'synset': 'cigar_box.n.01', 'synonyms': ['cigar_box'], 'def': 'a box for holding cigars', 'name': 'cigar_box'}, {'frequency': 'c', 'id': 267, 'synset': 'cigarette.n.01', 'synonyms': ['cigarette'], 'def': 'finely ground tobacco wrapped in paper; for smoking', 'name': 'cigarette'}, {'frequency': 'c', 'id': 268, 'synset': 'cigarette_case.n.01', 'synonyms': ['cigarette_case', 'cigarette_pack'], 'def': 'a small flat case for holding cigarettes', 'name': 'cigarette_case'}, {'frequency': 'f', 'id': 269, 'synset': 'cistern.n.02', 'synonyms': ['cistern', 'water_tank'], 'def': 'a tank that holds the water used to flush a toilet', 'name': 'cistern'}, {'frequency': 'r', 'id': 270, 'synset': 'clarinet.n.01', 'synonyms': ['clarinet'], 'def': 'a single-reed instrument with a straight tube', 'name': 'clarinet'}, {'frequency': 'r', 'id': 271, 'synset': 'clasp.n.01', 'synonyms': ['clasp'], 'def': 'a fastener (as a buckle or hook) that is used to hold two things together', 'name': 'clasp'}, {'frequency': 'c', 'id': 272, 'synset': 'cleansing_agent.n.01', 'synonyms': ['cleansing_agent', 'cleanser', 'cleaner'], 'def': 'a preparation used in cleaning something', 'name': 'cleansing_agent'}, {'frequency': 'r', 'id': 273, 'synset': 'clementine.n.01', 'synonyms': ['clementine'], 'def': 'a variety of mandarin orange', 'name': 'clementine'}, {'frequency': 'c', 'id': 274, 'synset': 'clip.n.03', 'synonyms': ['clip'], 'def': 'any of various small fasteners used to hold loose articles together', 'name': 'clip'}, {'frequency': 'c', 'id': 275, 'synset': 'clipboard.n.01', 'synonyms': ['clipboard'], 'def': 'a small writing board with a clip at the top for holding papers', 'name': 'clipboard'}, {'frequency': 'f', 'id': 276, 'synset': 'clock.n.01', 'synonyms': ['clock', 'timepiece', 'timekeeper'], 'def': 'a timepiece that shows the time of day', 'name': 'clock'}, {'frequency': 'f', 'id': 277, 'synset': 'clock_tower.n.01', 'synonyms': ['clock_tower'], 'def': 'a tower with a large clock visible high up on an outside face', 'name': 'clock_tower'}, {'frequency': 'c', 'id': 278, 'synset': 'clothes_hamper.n.01', 'synonyms': ['clothes_hamper', 'laundry_basket', 'clothes_basket'], 'def': 'a hamper that holds dirty clothes to be washed or wet clothes to be dried', 'name': 'clothes_hamper'}, {'frequency': 'c', 'id': 279, 'synset': 'clothespin.n.01', 'synonyms': ['clothespin', 'clothes_peg'], 'def': 'wood or plastic fastener; for holding clothes on a clothesline', 'name': 'clothespin'}, {'frequency': 'r', 'id': 280, 'synset': 'clutch_bag.n.01', 'synonyms': ['clutch_bag'], 'def': "a woman's strapless purse that is carried in the hand", 'name': 'clutch_bag'}, {'frequency': 'f', 'id': 281, 'synset': 'coaster.n.03', 'synonyms': ['coaster'], 'def': 'a covering (plate or mat) that protects the surface of a table', 'name': 'coaster'}, {'frequency': 'f', 'id': 282, 'synset': 'coat.n.01', 'synonyms': ['coat'], 'def': 'an outer garment that has sleeves and covers the body from shoulder down', 'name': 'coat'}, {'frequency': 'c', 'id': 283, 'synset': 'coat_hanger.n.01', 'synonyms': ['coat_hanger', 'clothes_hanger', 'dress_hanger'], 'def': "a hanger that is shaped like a person's shoulders", 'name': 'coat_hanger'}, {'frequency': 'r', 'id': 284, 'synset': 'coatrack.n.01', 'synonyms': ['coatrack', 'hatrack'], 'def': 'a rack with hooks for temporarily holding coats and hats', 'name': 'coatrack'}, {'frequency': 'c', 'id': 285, 'synset': 'cock.n.04', 'synonyms': ['cock', 'rooster'], 'def': 'adult male chicken', 'name': 'cock'}, {'frequency': 'c', 'id': 286, 'synset': 'coconut.n.02', 'synonyms': ['coconut', 'cocoanut'], 'def': 'large hard-shelled brown oval nut with a fibrous husk', 'name': 'coconut'}, {'frequency': 'r', 'id': 287, 'synset': 'coffee_filter.n.01', 'synonyms': ['coffee_filter'], 'def': 'filter (usually of paper) that passes the coffee and retains the coffee grounds', 'name': 'coffee_filter'}, {'frequency': 'f', 'id': 288, 'synset': 'coffee_maker.n.01', 'synonyms': ['coffee_maker', 'coffee_machine'], 'def': 'a kitchen appliance for brewing coffee automatically', 'name': 'coffee_maker'}, {'frequency': 'f', 'id': 289, 'synset': 'coffee_table.n.01', 'synonyms': ['coffee_table', 'cocktail_table'], 'def': 'low table where magazines can be placed and coffee or cocktails are served', 'name': 'coffee_table'}, {'frequency': 'c', 'id': 290, 'synset': 'coffeepot.n.01', 'synonyms': ['coffeepot'], 'def': 'tall pot in which coffee is brewed', 'name': 'coffeepot'}, {'frequency': 'r', 'id': 291, 'synset': 'coil.n.05', 'synonyms': ['coil'], 'def': 'tubing that is wound in a spiral', 'name': 'coil'}, {'frequency': 'c', 'id': 292, 'synset': 'coin.n.01', 'synonyms': ['coin'], 'def': 'a flat metal piece (usually a disc) used as money', 'name': 'coin'}, {'frequency': 'r', 'id': 293, 'synset': 'colander.n.01', 'synonyms': ['colander', 'cullender'], 'def': 'bowl-shaped strainer; used to wash or drain foods', 'name': 'colander'}, {'frequency': 'c', 'id': 294, 'synset': 'coleslaw.n.01', 'synonyms': ['coleslaw', 'slaw'], 'def': 'basically shredded cabbage', 'name': 'coleslaw'}, {'frequency': 'r', 'id': 295, 'synset': 'coloring_material.n.01', 'synonyms': ['coloring_material', 'colouring_material'], 'def': 'any material used for its color', 'name': 'coloring_material'}, {'frequency': 'r', 'id': 296, 'synset': 'combination_lock.n.01', 'synonyms': ['combination_lock'], 'def': 'lock that can be opened only by turning dials in a special sequence', 'name': 'combination_lock'}, {'frequency': 'c', 'id': 297, 'synset': 'comforter.n.04', 'synonyms': ['pacifier', 'teething_ring'], 'def': 'device used for an infant to suck or bite on', 'name': 'pacifier'}, {'frequency': 'r', 'id': 298, 'synset': 'comic_book.n.01', 'synonyms': ['comic_book'], 'def': 'a magazine devoted to comic strips', 'name': 'comic_book'}, {'frequency': 'f', 'id': 299, 'synset': 'computer_keyboard.n.01', 'synonyms': ['computer_keyboard', 'keyboard_(computer)'], 'def': 'a keyboard that is a data input device for computers', 'name': 'computer_keyboard'}, {'frequency': 'r', 'id': 300, 'synset': 'concrete_mixer.n.01', 'synonyms': ['concrete_mixer', 'cement_mixer'], 'def': 'a machine with a large revolving drum in which cement/concrete is mixed', 'name': 'concrete_mixer'}, {'frequency': 'f', 'id': 301, 'synset': 'cone.n.01', 'synonyms': ['cone', 'traffic_cone'], 'def': 'a cone-shaped object used to direct traffic', 'name': 'cone'}, {'frequency': 'f', 'id': 302, 'synset': 'control.n.09', 'synonyms': ['control', 'controller'], 'def': 'a mechanism that controls the operation of a machine', 'name': 'control'}, {'frequency': 'r', 'id': 303, 'synset': 'convertible.n.01', 'synonyms': ['convertible_(automobile)'], 'def': 'a car that has top that can be folded or removed', 'name': 'convertible_(automobile)'}, {'frequency': 'r', 'id': 304, 'synset': 'convertible.n.03', 'synonyms': ['sofa_bed'], 'def': 'a sofa that can be converted into a bed', 'name': 'sofa_bed'}, {'frequency': 'c', 'id': 305, 'synset': 'cookie.n.01', 'synonyms': ['cookie', 'cooky', 'biscuit_(cookie)'], 'def': "any of various small flat sweet cakes (`biscuit' is the British term)", 'name': 'cookie'}, {'frequency': 'r', 'id': 306, 'synset': 'cookie_jar.n.01', 'synonyms': ['cookie_jar', 'cooky_jar'], 'def': 'a jar in which cookies are kept (and sometimes money is hidden)', 'name': 'cookie_jar'}, {'frequency': 'r', 'id': 307, 'synset': 'cooking_utensil.n.01', 'synonyms': ['cooking_utensil'], 'def': 'a kitchen utensil made of material that does not melt easily; used for cooking', 'name': 'cooking_utensil'}, {'frequency': 'f', 'id': 308, 'synset': 'cooler.n.01', 'synonyms': ['cooler_(for_food)', 'ice_chest'], 'def': 'an insulated box for storing food often with ice', 'name': 'cooler_(for_food)'}, {'frequency': 'c', 'id': 309, 'synset': 'cork.n.04', 'synonyms': ['cork_(bottle_plug)', 'bottle_cork'], 'def': 'the plug in the mouth of a bottle (especially a wine bottle)', 'name': 'cork_(bottle_plug)'}, {'frequency': 'r', 'id': 310, 'synset': 'corkboard.n.01', 'synonyms': ['corkboard'], 'def': 'a sheet consisting of cork granules', 'name': 'corkboard'}, {'frequency': 'r', 'id': 311, 'synset': 'corkscrew.n.01', 'synonyms': ['corkscrew', 'bottle_screw'], 'def': 'a bottle opener that pulls corks', 'name': 'corkscrew'}, {'frequency': 'c', 'id': 312, 'synset': 'corn.n.03', 'synonyms': ['edible_corn', 'corn', 'maize'], 'def': 'ears of corn that can be prepared and served for human food', 'name': 'edible_corn'}, {'frequency': 'r', 'id': 313, 'synset': 'cornbread.n.01', 'synonyms': ['cornbread'], 'def': 'bread made primarily of cornmeal', 'name': 'cornbread'}, {'frequency': 'c', 'id': 314, 'synset': 'cornet.n.01', 'synonyms': ['cornet', 'horn', 'trumpet'], 'def': 'a brass musical instrument with a narrow tube and a flared bell and many valves', 'name': 'cornet'}, {'frequency': 'c', 'id': 315, 'synset': 'cornice.n.01', 'synonyms': ['cornice', 'valance', 'valance_board', 'pelmet'], 'def': 'a decorative framework to conceal curtain fixtures at the top of a window casing', 'name': 'cornice'}, {'frequency': 'r', 'id': 316, 'synset': 'cornmeal.n.01', 'synonyms': ['cornmeal'], 'def': 'coarsely ground corn', 'name': 'cornmeal'}, {'frequency': 'r', 'id': 317, 'synset': 'corset.n.01', 'synonyms': ['corset', 'girdle'], 'def': "a woman's close-fitting foundation garment", 'name': 'corset'}, {'frequency': 'r', 'id': 318, 'synset': 'cos.n.02', 'synonyms': ['romaine_lettuce'], 'def': 'lettuce with long dark-green leaves in a loosely packed elongated head', 'name': 'romaine_lettuce'}, {'frequency': 'c', 'id': 319, 'synset': 'costume.n.04', 'synonyms': ['costume'], 'def': 'the attire characteristic of a country or a time or a social class', 'name': 'costume'}, {'frequency': 'r', 'id': 320, 'synset': 'cougar.n.01', 'synonyms': ['cougar', 'puma', 'catamount', 'mountain_lion', 'panther'], 'def': 'large American feline resembling a lion', 'name': 'cougar'}, {'frequency': 'r', 'id': 321, 'synset': 'coverall.n.01', 'synonyms': ['coverall'], 'def': 'a loose-fitting protective garment that is worn over other clothing', 'name': 'coverall'}, {'frequency': 'r', 'id': 322, 'synset': 'cowbell.n.01', 'synonyms': ['cowbell'], 'def': 'a bell hung around the neck of cow so that the cow can be easily located', 'name': 'cowbell'}, {'frequency': 'f', 'id': 323, 'synset': 'cowboy_hat.n.01', 'synonyms': ['cowboy_hat', 'ten-gallon_hat'], 'def': 'a hat with a wide brim and a soft crown; worn by American ranch hands', 'name': 'cowboy_hat'}, {'frequency': 'r', 'id': 324, 'synset': 'crab.n.01', 'synonyms': ['crab_(animal)'], 'def': 'decapod having eyes on short stalks and a broad flattened shell and pincers', 'name': 'crab_(animal)'}, {'frequency': 'c', 'id': 325, 'synset': 'cracker.n.01', 'synonyms': ['cracker'], 'def': 'a thin crisp wafer', 'name': 'cracker'}, {'frequency': 'r', 'id': 326, 'synset': 'crape.n.01', 'synonyms': ['crape', 'crepe', 'French_pancake'], 'def': 'small very thin pancake', 'name': 'crape'}, {'frequency': 'f', 'id': 327, 'synset': 'crate.n.01', 'synonyms': ['crate'], 'def': 'a rugged box (usually made of wood); used for shipping', 'name': 'crate'}, {'frequency': 'r', 'id': 328, 'synset': 'crayon.n.01', 'synonyms': ['crayon', 'wax_crayon'], 'def': 'writing or drawing implement made of a colored stick of composition wax', 'name': 'crayon'}, {'frequency': 'r', 'id': 329, 'synset': 'cream_pitcher.n.01', 'synonyms': ['cream_pitcher'], 'def': 'a small pitcher for serving cream', 'name': 'cream_pitcher'}, {'frequency': 'r', 'id': 330, 'synset': 'credit_card.n.01', 'synonyms': ['credit_card', 'charge_card', 'debit_card'], 'def': 'a card, usually plastic, used to pay for goods and services', 'name': 'credit_card'}, {'frequency': 'c', 'id': 331, 'synset': 'crescent_roll.n.01', 'synonyms': ['crescent_roll', 'croissant'], 'def': 'very rich flaky crescent-shaped roll', 'name': 'crescent_roll'}, {'frequency': 'c', 'id': 332, 'synset': 'crib.n.01', 'synonyms': ['crib', 'cot'], 'def': 'baby bed with high sides made of slats', 'name': 'crib'}, {'frequency': 'c', 'id': 333, 'synset': 'crock.n.03', 'synonyms': ['crock_pot', 'earthenware_jar'], 'def': 'an earthen jar (made of baked clay)', 'name': 'crock_pot'}, {'frequency': 'f', 'id': 334, 'synset': 'crossbar.n.01', 'synonyms': ['crossbar'], 'def': 'a horizontal bar that goes across something', 'name': 'crossbar'}, {'frequency': 'r', 'id': 335, 'synset': 'crouton.n.01', 'synonyms': ['crouton'], 'def': 'a small piece of toasted or fried bread; served in soup or salads', 'name': 'crouton'}, {'frequency': 'r', 'id': 336, 'synset': 'crow.n.01', 'synonyms': ['crow'], 'def': 'black birds having a raucous call', 'name': 'crow'}, {'frequency': 'c', 'id': 337, 'synset': 'crown.n.04', 'synonyms': ['crown'], 'def': 'an ornamental jeweled headdress signifying sovereignty', 'name': 'crown'}, {'frequency': 'c', 'id': 338, 'synset': 'crucifix.n.01', 'synonyms': ['crucifix'], 'def': 'representation of the cross on which Jesus died', 'name': 'crucifix'}, {'frequency': 'c', 'id': 339, 'synset': 'cruise_ship.n.01', 'synonyms': ['cruise_ship', 'cruise_liner'], 'def': 'a passenger ship used commercially for pleasure cruises', 'name': 'cruise_ship'}, {'frequency': 'c', 'id': 340, 'synset': 'cruiser.n.01', 'synonyms': ['police_cruiser', 'patrol_car', 'police_car', 'squad_car'], 'def': 'a car in which policemen cruise the streets', 'name': 'police_cruiser'}, {'frequency': 'c', 'id': 341, 'synset': 'crumb.n.03', 'synonyms': ['crumb'], 'def': 'small piece of e.g. bread or cake', 'name': 'crumb'}, {'frequency': 'r', 'id': 342, 'synset': 'crutch.n.01', 'synonyms': ['crutch'], 'def': 'a wooden or metal staff that fits under the armpit and reaches to the ground', 'name': 'crutch'}, {'frequency': 'c', 'id': 343, 'synset': 'cub.n.03', 'synonyms': ['cub_(animal)'], 'def': 'the young of certain carnivorous mammals such as the bear or wolf or lion', 'name': 'cub_(animal)'}, {'frequency': 'r', 'id': 344, 'synset': 'cube.n.05', 'synonyms': ['cube', 'square_block'], 'def': 'a block in the (approximate) shape of a cube', 'name': 'cube'}, {'frequency': 'f', 'id': 345, 'synset': 'cucumber.n.02', 'synonyms': ['cucumber', 'cuke'], 'def': 'cylindrical green fruit with thin green rind and white flesh eaten as a vegetable', 'name': 'cucumber'}, {'frequency': 'c', 'id': 346, 'synset': 'cufflink.n.01', 'synonyms': ['cufflink'], 'def': 'jewelry consisting of linked buttons used to fasten the cuffs of a shirt', 'name': 'cufflink'}, {'frequency': 'f', 'id': 347, 'synset': 'cup.n.01', 'synonyms': ['cup'], 'def': 'a small open container usually used for drinking; usually has a handle', 'name': 'cup'}, {'frequency': 'c', 'id': 348, 'synset': 'cup.n.08', 'synonyms': ['trophy_cup'], 'def': 'a metal vessel with handles that is awarded as a trophy to a competition winner', 'name': 'trophy_cup'}, {'frequency': 'c', 'id': 349, 'synset': 'cupcake.n.01', 'synonyms': ['cupcake'], 'def': 'small cake baked in a muffin tin', 'name': 'cupcake'}, {'frequency': 'r', 'id': 350, 'synset': 'curler.n.01', 'synonyms': ['hair_curler', 'hair_roller', 'hair_crimper'], 'def': 'a cylindrical tube around which the hair is wound to curl it', 'name': 'hair_curler'}, {'frequency': 'r', 'id': 351, 'synset': 'curling_iron.n.01', 'synonyms': ['curling_iron'], 'def': 'a cylindrical home appliance that heats hair that has been curled around it', 'name': 'curling_iron'}, {'frequency': 'f', 'id': 352, 'synset': 'curtain.n.01', 'synonyms': ['curtain', 'drapery'], 'def': 'hanging cloth used as a blind (especially for a window)', 'name': 'curtain'}, {'frequency': 'f', 'id': 353, 'synset': 'cushion.n.03', 'synonyms': ['cushion'], 'def': 'a soft bag filled with air or padding such as feathers or foam rubber', 'name': 'cushion'}, {'frequency': 'r', 'id': 354, 'synset': 'custard.n.01', 'synonyms': ['custard'], 'def': 'sweetened mixture of milk and eggs baked or boiled or frozen', 'name': 'custard'}, {'frequency': 'c', 'id': 355, 'synset': 'cutter.n.06', 'synonyms': ['cutting_tool'], 'def': 'a cutting implement; a tool for cutting', 'name': 'cutting_tool'}, {'frequency': 'r', 'id': 356, 'synset': 'cylinder.n.04', 'synonyms': ['cylinder'], 'def': 'a cylindrical container', 'name': 'cylinder'}, {'frequency': 'r', 'id': 357, 'synset': 'cymbal.n.01', 'synonyms': ['cymbal'], 'def': 'a percussion instrument consisting of a concave brass disk', 'name': 'cymbal'}, {'frequency': 'r', 'id': 358, 'synset': 'dachshund.n.01', 'synonyms': ['dachshund', 'dachsie', 'badger_dog'], 'def': 'small long-bodied short-legged breed of dog having a short sleek coat and long drooping ears', 'name': 'dachshund'}, {'frequency': 'r', 'id': 359, 'synset': 'dagger.n.01', 'synonyms': ['dagger'], 'def': 'a short knife with a pointed blade used for piercing or stabbing', 'name': 'dagger'}, {'frequency': 'r', 'id': 360, 'synset': 'dartboard.n.01', 'synonyms': ['dartboard'], 'def': 'a circular board of wood or cork used as the target in the game of darts', 'name': 'dartboard'}, {'frequency': 'r', 'id': 361, 'synset': 'date.n.08', 'synonyms': ['date_(fruit)'], 'def': 'sweet edible fruit of the date palm with a single long woody seed', 'name': 'date_(fruit)'}, {'frequency': 'f', 'id': 362, 'synset': 'deck_chair.n.01', 'synonyms': ['deck_chair', 'beach_chair'], 'def': 'a folding chair for use outdoors; a wooden frame supports a length of canvas', 'name': 'deck_chair'}, {'frequency': 'c', 'id': 363, 'synset': 'deer.n.01', 'synonyms': ['deer', 'cervid'], 'def': "distinguished from Bovidae by the male's having solid deciduous antlers", 'name': 'deer'}, {'frequency': 'c', 'id': 364, 'synset': 'dental_floss.n.01', 'synonyms': ['dental_floss', 'floss'], 'def': 'a soft thread for cleaning the spaces between the teeth', 'name': 'dental_floss'}, {'frequency': 'f', 'id': 365, 'synset': 'desk.n.01', 'synonyms': ['desk'], 'def': 'a piece of furniture with a writing surface and usually drawers or other compartments', 'name': 'desk'}, {'frequency': 'r', 'id': 366, 'synset': 'detergent.n.01', 'synonyms': ['detergent'], 'def': 'a surface-active chemical widely used in industry and laundering', 'name': 'detergent'}, {'frequency': 'c', 'id': 367, 'synset': 'diaper.n.01', 'synonyms': ['diaper'], 'def': 'garment consisting of a folded cloth drawn up between the legs and fastened at the waist', 'name': 'diaper'}, {'frequency': 'r', 'id': 368, 'synset': 'diary.n.01', 'synonyms': ['diary', 'journal'], 'def': 'a daily written record of (usually personal) experiences and observations', 'name': 'diary'}, {'frequency': 'r', 'id': 369, 'synset': 'die.n.01', 'synonyms': ['die', 'dice'], 'def': 'a small cube with 1 to 6 spots on the six faces; used in gambling', 'name': 'die'}, {'frequency': 'r', 'id': 370, 'synset': 'dinghy.n.01', 'synonyms': ['dinghy', 'dory', 'rowboat'], 'def': 'a small boat of shallow draft with seats and oars with which it is propelled', 'name': 'dinghy'}, {'frequency': 'f', 'id': 371, 'synset': 'dining_table.n.01', 'synonyms': ['dining_table'], 'def': 'a table at which meals are served', 'name': 'dining_table'}, {'frequency': 'r', 'id': 372, 'synset': 'dinner_jacket.n.01', 'synonyms': ['tux', 'tuxedo'], 'def': 'semiformal evening dress for men', 'name': 'tux'}, {'frequency': 'c', 'id': 373, 'synset': 'dish.n.01', 'synonyms': ['dish'], 'def': 'a piece of dishware normally used as a container for holding or serving food', 'name': 'dish'}, {'frequency': 'c', 'id': 374, 'synset': 'dish.n.05', 'synonyms': ['dish_antenna'], 'def': 'directional antenna consisting of a parabolic reflector', 'name': 'dish_antenna'}, {'frequency': 'c', 'id': 375, 'synset': 'dishrag.n.01', 'synonyms': ['dishrag', 'dishcloth'], 'def': 'a cloth for washing dishes', 'name': 'dishrag'}, {'frequency': 'c', 'id': 376, 'synset': 'dishtowel.n.01', 'synonyms': ['dishtowel', 'tea_towel'], 'def': 'a towel for drying dishes', 'name': 'dishtowel'}, {'frequency': 'f', 'id': 377, 'synset': 'dishwasher.n.01', 'synonyms': ['dishwasher', 'dishwashing_machine'], 'def': 'a machine for washing dishes', 'name': 'dishwasher'}, {'frequency': 'r', 'id': 378, 'synset': 'dishwasher_detergent.n.01', 'synonyms': ['dishwasher_detergent', 'dishwashing_detergent', 'dishwashing_liquid'], 'def': 'a low-sudsing detergent designed for use in dishwashers', 'name': 'dishwasher_detergent'}, {'frequency': 'r', 'id': 379, 'synset': 'diskette.n.01', 'synonyms': ['diskette', 'floppy', 'floppy_disk'], 'def': 'a small plastic magnetic disk enclosed in a stiff envelope used to store data', 'name': 'diskette'}, {'frequency': 'c', 'id': 380, 'synset': 'dispenser.n.01', 'synonyms': ['dispenser'], 'def': 'a container so designed that the contents can be used in prescribed amounts', 'name': 'dispenser'}, {'frequency': 'c', 'id': 381, 'synset': 'dixie_cup.n.01', 'synonyms': ['Dixie_cup', 'paper_cup'], 'def': 'a disposable cup made of paper; for holding drinks', 'name': 'Dixie_cup'}, {'frequency': 'f', 'id': 382, 'synset': 'dog.n.01', 'synonyms': ['dog'], 'def': 'a common domesticated dog', 'name': 'dog'}, {'frequency': 'f', 'id': 383, 'synset': 'dog_collar.n.01', 'synonyms': ['dog_collar'], 'def': 'a collar for a dog', 'name': 'dog_collar'}, {'frequency': 'c', 'id': 384, 'synset': 'doll.n.01', 'synonyms': ['doll'], 'def': 'a toy replica of a HUMAN (NOT AN ANIMAL)', 'name': 'doll'}, {'frequency': 'r', 'id': 385, 'synset': 'dollar.n.02', 'synonyms': ['dollar', 'dollar_bill', 'one_dollar_bill'], 'def': 'a piece of paper money worth one dollar', 'name': 'dollar'}, {'frequency': 'r', 'id': 386, 'synset': 'dolphin.n.02', 'synonyms': ['dolphin'], 'def': 'any of various small toothed whales with a beaklike snout; larger than porpoises', 'name': 'dolphin'}, {'frequency': 'c', 'id': 387, 'synset': 'domestic_ass.n.01', 'synonyms': ['domestic_ass', 'donkey'], 'def': 'domestic beast of burden descended from the African wild ass; patient but stubborn', 'name': 'domestic_ass'}, {'frequency': 'r', 'id': 388, 'synset': 'domino.n.03', 'synonyms': ['eye_mask'], 'def': 'a mask covering the upper part of the face but with holes for the eyes', 'name': 'eye_mask'}, {'frequency': 'r', 'id': 389, 'synset': 'doorbell.n.01', 'synonyms': ['doorbell', 'buzzer'], 'def': 'a button at an outer door that gives a ringing or buzzing signal when pushed', 'name': 'doorbell'}, {'frequency': 'f', 'id': 390, 'synset': 'doorknob.n.01', 'synonyms': ['doorknob', 'doorhandle'], 'def': "a knob used to open a door (often called `doorhandle' in Great Britain)", 'name': 'doorknob'}, {'frequency': 'c', 'id': 391, 'synset': 'doormat.n.02', 'synonyms': ['doormat', 'welcome_mat'], 'def': 'a mat placed outside an exterior door for wiping the shoes before entering', 'name': 'doormat'}, {'frequency': 'f', 'id': 392, 'synset': 'doughnut.n.02', 'synonyms': ['doughnut', 'donut'], 'def': 'a small ring-shaped friedcake', 'name': 'doughnut'}, {'frequency': 'r', 'id': 393, 'synset': 'dove.n.01', 'synonyms': ['dove'], 'def': 'any of numerous small pigeons', 'name': 'dove'}, {'frequency': 'r', 'id': 394, 'synset': 'dragonfly.n.01', 'synonyms': ['dragonfly'], 'def': 'slender-bodied non-stinging insect having iridescent wings that are outspread at rest', 'name': 'dragonfly'}, {'frequency': 'f', 'id': 395, 'synset': 'drawer.n.01', 'synonyms': ['drawer'], 'def': 'a boxlike container in a piece of furniture; made so as to slide in and out', 'name': 'drawer'}, {'frequency': 'c', 'id': 396, 'synset': 'drawers.n.01', 'synonyms': ['underdrawers', 'boxers', 'boxershorts'], 'def': 'underpants worn by men', 'name': 'underdrawers'}, {'frequency': 'f', 'id': 397, 'synset': 'dress.n.01', 'synonyms': ['dress', 'frock'], 'def': 'a one-piece garment for a woman; has skirt and bodice', 'name': 'dress'}, {'frequency': 'c', 'id': 398, 'synset': 'dress_hat.n.01', 'synonyms': ['dress_hat', 'high_hat', 'opera_hat', 'silk_hat', 'top_hat'], 'def': "a man's hat with a tall crown; usually covered with silk or with beaver fur", 'name': 'dress_hat'}, {'frequency': 'c', 'id': 399, 'synset': 'dress_suit.n.01', 'synonyms': ['dress_suit'], 'def': 'formalwear consisting of full evening dress for men', 'name': 'dress_suit'}, {'frequency': 'c', 'id': 400, 'synset': 'dresser.n.05', 'synonyms': ['dresser'], 'def': 'a cabinet with shelves', 'name': 'dresser'}, {'frequency': 'c', 'id': 401, 'synset': 'drill.n.01', 'synonyms': ['drill'], 'def': 'a tool with a sharp rotating point for making holes in hard materials', 'name': 'drill'}, {'frequency': 'r', 'id': 402, 'synset': 'drinking_fountain.n.01', 'synonyms': ['drinking_fountain'], 'def': 'a public fountain to provide a jet of drinking water', 'name': 'drinking_fountain'}, {'frequency': 'r', 'id': 403, 'synset': 'drone.n.04', 'synonyms': ['drone'], 'def': 'an aircraft without a pilot that is operated by remote control', 'name': 'drone'}, {'frequency': 'r', 'id': 404, 'synset': 'dropper.n.01', 'synonyms': ['dropper', 'eye_dropper'], 'def': 'pipet consisting of a small tube with a vacuum bulb at one end for drawing liquid in and releasing it a drop at a time', 'name': 'dropper'}, {'frequency': 'c', 'id': 405, 'synset': 'drum.n.01', 'synonyms': ['drum_(musical_instrument)'], 'def': 'a musical percussion instrument; usually consists of a hollow cylinder with a membrane stretched across each end', 'name': 'drum_(musical_instrument)'}, {'frequency': 'r', 'id': 406, 'synset': 'drumstick.n.02', 'synonyms': ['drumstick'], 'def': 'a stick used for playing a drum', 'name': 'drumstick'}, {'frequency': 'f', 'id': 407, 'synset': 'duck.n.01', 'synonyms': ['duck'], 'def': 'small web-footed broad-billed swimming bird', 'name': 'duck'}, {'frequency': 'r', 'id': 408, 'synset': 'duckling.n.02', 'synonyms': ['duckling'], 'def': 'young duck', 'name': 'duckling'}, {'frequency': 'c', 'id': 409, 'synset': 'duct_tape.n.01', 'synonyms': ['duct_tape'], 'def': 'a wide silvery adhesive tape', 'name': 'duct_tape'}, {'frequency': 'f', 'id': 410, 'synset': 'duffel_bag.n.01', 'synonyms': ['duffel_bag', 'duffle_bag', 'duffel', 'duffle'], 'def': 'a large cylindrical bag of heavy cloth', 'name': 'duffel_bag'}, {'frequency': 'r', 'id': 411, 'synset': 'dumbbell.n.01', 'synonyms': ['dumbbell'], 'def': 'an exercising weight with two ball-like ends connected by a short handle', 'name': 'dumbbell'}, {'frequency': 'c', 'id': 412, 'synset': 'dumpster.n.01', 'synonyms': ['dumpster'], 'def': 'a container designed to receive and transport and dump waste', 'name': 'dumpster'}, {'frequency': 'r', 'id': 413, 'synset': 'dustpan.n.02', 'synonyms': ['dustpan'], 'def': 'a short-handled receptacle into which dust can be swept', 'name': 'dustpan'}, {'frequency': 'r', 'id': 414, 'synset': 'dutch_oven.n.02', 'synonyms': ['Dutch_oven'], 'def': 'iron or earthenware cooking pot; used for stews', 'name': 'Dutch_oven'}, {'frequency': 'c', 'id': 415, 'synset': 'eagle.n.01', 'synonyms': ['eagle'], 'def': 'large birds of prey noted for their broad wings and strong soaring flight', 'name': 'eagle'}, {'frequency': 'f', 'id': 416, 'synset': 'earphone.n.01', 'synonyms': ['earphone', 'earpiece', 'headphone'], 'def': 'device for listening to audio that is held over or inserted into the ear', 'name': 'earphone'}, {'frequency': 'r', 'id': 417, 'synset': 'earplug.n.01', 'synonyms': ['earplug'], 'def': 'a soft plug that is inserted into the ear canal to block sound', 'name': 'earplug'}, {'frequency': 'f', 'id': 418, 'synset': 'earring.n.01', 'synonyms': ['earring'], 'def': 'jewelry to ornament the ear', 'name': 'earring'}, {'frequency': 'c', 'id': 419, 'synset': 'easel.n.01', 'synonyms': ['easel'], 'def': "an upright tripod for displaying something (usually an artist's canvas)", 'name': 'easel'}, {'frequency': 'r', 'id': 420, 'synset': 'eclair.n.01', 'synonyms': ['eclair'], 'def': 'oblong cream puff', 'name': 'eclair'}, {'frequency': 'r', 'id': 421, 'synset': 'eel.n.01', 'synonyms': ['eel'], 'def': 'an elongate fish with fatty flesh', 'name': 'eel'}, {'frequency': 'f', 'id': 422, 'synset': 'egg.n.02', 'synonyms': ['egg', 'eggs'], 'def': 'oval reproductive body of a fowl (especially a hen) used as food', 'name': 'egg'}, {'frequency': 'r', 'id': 423, 'synset': 'egg_roll.n.01', 'synonyms': ['egg_roll', 'spring_roll'], 'def': 'minced vegetables and meat wrapped in a pancake and fried', 'name': 'egg_roll'}, {'frequency': 'c', 'id': 424, 'synset': 'egg_yolk.n.01', 'synonyms': ['egg_yolk', 'yolk_(egg)'], 'def': 'the yellow spherical part of an egg', 'name': 'egg_yolk'}, {'frequency': 'c', 'id': 425, 'synset': 'eggbeater.n.02', 'synonyms': ['eggbeater', 'eggwhisk'], 'def': 'a mixer for beating eggs or whipping cream', 'name': 'eggbeater'}, {'frequency': 'c', 'id': 426, 'synset': 'eggplant.n.01', 'synonyms': ['eggplant', 'aubergine'], 'def': 'egg-shaped vegetable having a shiny skin typically dark purple', 'name': 'eggplant'}, {'frequency': 'r', 'id': 427, 'synset': 'electric_chair.n.01', 'synonyms': ['electric_chair'], 'def': 'a chair-shaped instrument of execution by electrocution', 'name': 'electric_chair'}, {'frequency': 'f', 'id': 428, 'synset': 'electric_refrigerator.n.01', 'synonyms': ['refrigerator'], 'def': 'a refrigerator in which the coolant is pumped around by an electric motor', 'name': 'refrigerator'}, {'frequency': 'f', 'id': 429, 'synset': 'elephant.n.01', 'synonyms': ['elephant'], 'def': 'a common elephant', 'name': 'elephant'}, {'frequency': 'r', 'id': 430, 'synset': 'elk.n.01', 'synonyms': ['elk', 'moose'], 'def': 'large northern deer with enormous flattened antlers in the male', 'name': 'elk'}, {'frequency': 'c', 'id': 431, 'synset': 'envelope.n.01', 'synonyms': ['envelope'], 'def': 'a flat (usually rectangular) container for a letter, thin package, etc.', 'name': 'envelope'}, {'frequency': 'c', 'id': 432, 'synset': 'eraser.n.01', 'synonyms': ['eraser'], 'def': 'an implement used to erase something', 'name': 'eraser'}, {'frequency': 'r', 'id': 433, 'synset': 'escargot.n.01', 'synonyms': ['escargot'], 'def': 'edible snail usually served in the shell with a sauce of melted butter and garlic', 'name': 'escargot'}, {'frequency': 'r', 'id': 434, 'synset': 'eyepatch.n.01', 'synonyms': ['eyepatch'], 'def': 'a protective cloth covering for an injured eye', 'name': 'eyepatch'}, {'frequency': 'r', 'id': 435, 'synset': 'falcon.n.01', 'synonyms': ['falcon'], 'def': 'birds of prey having long pointed powerful wings adapted for swift flight', 'name': 'falcon'}, {'frequency': 'f', 'id': 436, 'synset': 'fan.n.01', 'synonyms': ['fan'], 'def': 'a device for creating a current of air by movement of a surface or surfaces', 'name': 'fan'}, {'frequency': 'f', 'id': 437, 'synset': 'faucet.n.01', 'synonyms': ['faucet', 'spigot', 'tap'], 'def': 'a regulator for controlling the flow of a liquid from a reservoir', 'name': 'faucet'}, {'frequency': 'r', 'id': 438, 'synset': 'fedora.n.01', 'synonyms': ['fedora'], 'def': 'a hat made of felt with a creased crown', 'name': 'fedora'}, {'frequency': 'r', 'id': 439, 'synset': 'ferret.n.02', 'synonyms': ['ferret'], 'def': 'domesticated albino variety of the European polecat bred for hunting rats and rabbits', 'name': 'ferret'}, {'frequency': 'c', 'id': 440, 'synset': 'ferris_wheel.n.01', 'synonyms': ['Ferris_wheel'], 'def': 'a large wheel with suspended seats that remain upright as the wheel rotates', 'name': 'Ferris_wheel'}, {'frequency': 'r', 'id': 441, 'synset': 'ferry.n.01', 'synonyms': ['ferry', 'ferryboat'], 'def': 'a boat that transports people or vehicles across a body of water and operates on a regular schedule', 'name': 'ferry'}, {'frequency': 'r', 'id': 442, 'synset': 'fig.n.04', 'synonyms': ['fig_(fruit)'], 'def': 'fleshy sweet pear-shaped yellowish or purple fruit eaten fresh or preserved or dried', 'name': 'fig_(fruit)'}, {'frequency': 'c', 'id': 443, 'synset': 'fighter.n.02', 'synonyms': ['fighter_jet', 'fighter_aircraft', 'attack_aircraft'], 'def': 'a high-speed military or naval airplane designed to destroy enemy targets', 'name': 'fighter_jet'}, {'frequency': 'f', 'id': 444, 'synset': 'figurine.n.01', 'synonyms': ['figurine'], 'def': 'a small carved or molded figure', 'name': 'figurine'}, {'frequency': 'c', 'id': 445, 'synset': 'file.n.03', 'synonyms': ['file_cabinet', 'filing_cabinet'], 'def': 'office furniture consisting of a container for keeping papers in order', 'name': 'file_cabinet'}, {'frequency': 'r', 'id': 446, 'synset': 'file.n.04', 'synonyms': ['file_(tool)'], 'def': 'a steel hand tool with small sharp teeth on some or all of its surfaces; used for smoothing wood or metal', 'name': 'file_(tool)'}, {'frequency': 'f', 'id': 447, 'synset': 'fire_alarm.n.02', 'synonyms': ['fire_alarm', 'smoke_alarm'], 'def': 'an alarm that is tripped off by fire or smoke', 'name': 'fire_alarm'}, {'frequency': 'c', 'id': 448, 'synset': 'fire_engine.n.01', 'synonyms': ['fire_engine', 'fire_truck'], 'def': 'large trucks that carry firefighters and equipment to the site of a fire', 'name': 'fire_engine'}, {'frequency': 'c', 'id': 449, 'synset': 'fire_extinguisher.n.01', 'synonyms': ['fire_extinguisher', 'extinguisher'], 'def': 'a manually operated device for extinguishing small fires', 'name': 'fire_extinguisher'}, {'frequency': 'c', 'id': 450, 'synset': 'fire_hose.n.01', 'synonyms': ['fire_hose'], 'def': 'a large hose that carries water from a fire hydrant to the site of the fire', 'name': 'fire_hose'}, {'frequency': 'f', 'id': 451, 'synset': 'fireplace.n.01', 'synonyms': ['fireplace'], 'def': 'an open recess in a wall at the base of a chimney where a fire can be built', 'name': 'fireplace'}, {'frequency': 'f', 'id': 452, 'synset': 'fireplug.n.01', 'synonyms': ['fireplug', 'fire_hydrant', 'hydrant'], 'def': 'an upright hydrant for drawing water to use in fighting a fire', 'name': 'fireplug'}, {'frequency': 'c', 'id': 453, 'synset': 'fish.n.01', 'synonyms': ['fish'], 'def': 'any of various mostly cold-blooded aquatic vertebrates usually having scales and breathing through gills', 'name': 'fish'}, {'frequency': 'r', 'id': 454, 'synset': 'fish.n.02', 'synonyms': ['fish_(food)'], 'def': 'the flesh of fish used as food', 'name': 'fish_(food)'}, {'frequency': 'r', 'id': 455, 'synset': 'fishbowl.n.02', 'synonyms': ['fishbowl', 'goldfish_bowl'], 'def': 'a transparent bowl in which small fish are kept', 'name': 'fishbowl'}, {'frequency': 'r', 'id': 456, 'synset': 'fishing_boat.n.01', 'synonyms': ['fishing_boat', 'fishing_vessel'], 'def': 'a vessel for fishing', 'name': 'fishing_boat'}, {'frequency': 'c', 'id': 457, 'synset': 'fishing_rod.n.01', 'synonyms': ['fishing_rod', 'fishing_pole'], 'def': 'a rod that is used in fishing to extend the fishing line', 'name': 'fishing_rod'}, {'frequency': 'f', 'id': 458, 'synset': 'flag.n.01', 'synonyms': ['flag'], 'def': 'emblem usually consisting of a rectangular piece of cloth of distinctive design (do not include pole)', 'name': 'flag'}, {'frequency': 'f', 'id': 459, 'synset': 'flagpole.n.02', 'synonyms': ['flagpole', 'flagstaff'], 'def': 'a tall staff or pole on which a flag is raised', 'name': 'flagpole'}, {'frequency': 'c', 'id': 460, 'synset': 'flamingo.n.01', 'synonyms': ['flamingo'], 'def': 'large pink web-footed bird with down-bent bill', 'name': 'flamingo'}, {'frequency': 'c', 'id': 461, 'synset': 'flannel.n.01', 'synonyms': ['flannel'], 'def': 'a soft light woolen fabric; used for clothing', 'name': 'flannel'}, {'frequency': 'r', 'id': 462, 'synset': 'flash.n.10', 'synonyms': ['flash', 'flashbulb'], 'def': 'a lamp for providing momentary light to take a photograph', 'name': 'flash'}, {'frequency': 'c', 'id': 463, 'synset': 'flashlight.n.01', 'synonyms': ['flashlight', 'torch'], 'def': 'a small portable battery-powered electric lamp', 'name': 'flashlight'}, {'frequency': 'r', 'id': 464, 'synset': 'fleece.n.03', 'synonyms': ['fleece'], 'def': 'a soft bulky fabric with deep pile; used chiefly for clothing', 'name': 'fleece'}, {'frequency': 'f', 'id': 465, 'synset': 'flip-flop.n.02', 'synonyms': ['flip-flop_(sandal)'], 'def': 'a backless sandal held to the foot by a thong between two toes', 'name': 'flip-flop_(sandal)'}, {'frequency': 'c', 'id': 466, 'synset': 'flipper.n.01', 'synonyms': ['flipper_(footwear)', 'fin_(footwear)'], 'def': 'a shoe to aid a person in swimming', 'name': 'flipper_(footwear)'}, {'frequency': 'f', 'id': 467, 'synset': 'flower_arrangement.n.01', 'synonyms': ['flower_arrangement', 'floral_arrangement'], 'def': 'a decorative arrangement of flowers', 'name': 'flower_arrangement'}, {'frequency': 'c', 'id': 468, 'synset': 'flute.n.02', 'synonyms': ['flute_glass', 'champagne_flute'], 'def': 'a tall narrow wineglass', 'name': 'flute_glass'}, {'frequency': 'r', 'id': 469, 'synset': 'foal.n.01', 'synonyms': ['foal'], 'def': 'a young horse', 'name': 'foal'}, {'frequency': 'c', 'id': 470, 'synset': 'folding_chair.n.01', 'synonyms': ['folding_chair'], 'def': 'a chair that can be folded flat for storage', 'name': 'folding_chair'}, {'frequency': 'c', 'id': 471, 'synset': 'food_processor.n.01', 'synonyms': ['food_processor'], 'def': 'a kitchen appliance for shredding, blending, chopping, or slicing food', 'name': 'food_processor'}, {'frequency': 'c', 'id': 472, 'synset': 'football.n.02', 'synonyms': ['football_(American)'], 'def': 'the inflated oblong ball used in playing American football', 'name': 'football_(American)'}, {'frequency': 'r', 'id': 473, 'synset': 'football_helmet.n.01', 'synonyms': ['football_helmet'], 'def': 'a padded helmet with a face mask to protect the head of football players', 'name': 'football_helmet'}, {'frequency': 'c', 'id': 474, 'synset': 'footstool.n.01', 'synonyms': ['footstool', 'footrest'], 'def': 'a low seat or a stool to rest the feet of a seated person', 'name': 'footstool'}, {'frequency': 'f', 'id': 475, 'synset': 'fork.n.01', 'synonyms': ['fork'], 'def': 'cutlery used for serving and eating food', 'name': 'fork'}, {'frequency': 'r', 'id': 476, 'synset': 'forklift.n.01', 'synonyms': ['forklift'], 'def': 'an industrial vehicle with a power operated fork in front that can be inserted under loads to lift and move them', 'name': 'forklift'}, {'frequency': 'r', 'id': 477, 'synset': 'freight_car.n.01', 'synonyms': ['freight_car'], 'def': 'a railway car that carries freight', 'name': 'freight_car'}, {'frequency': 'r', 'id': 478, 'synset': 'french_toast.n.01', 'synonyms': ['French_toast'], 'def': 'bread slice dipped in egg and milk and fried', 'name': 'French_toast'}, {'frequency': 'c', 'id': 479, 'synset': 'freshener.n.01', 'synonyms': ['freshener', 'air_freshener'], 'def': 'anything that freshens', 'name': 'freshener'}, {'frequency': 'f', 'id': 480, 'synset': 'frisbee.n.01', 'synonyms': ['frisbee'], 'def': 'a light, plastic disk propelled with a flip of the wrist for recreation or competition', 'name': 'frisbee'}, {'frequency': 'c', 'id': 481, 'synset': 'frog.n.01', 'synonyms': ['frog', 'toad', 'toad_frog'], 'def': 'a tailless stout-bodied amphibians with long hind limbs for leaping', 'name': 'frog'}, {'frequency': 'c', 'id': 482, 'synset': 'fruit_juice.n.01', 'synonyms': ['fruit_juice'], 'def': 'drink produced by squeezing or crushing fruit', 'name': 'fruit_juice'}, {'frequency': 'r', 'id': 483, 'synset': 'fruit_salad.n.01', 'synonyms': ['fruit_salad'], 'def': 'salad composed of fruits', 'name': 'fruit_salad'}, {'frequency': 'c', 'id': 484, 'synset': 'frying_pan.n.01', 'synonyms': ['frying_pan', 'frypan', 'skillet'], 'def': 'a pan used for frying foods', 'name': 'frying_pan'}, {'frequency': 'r', 'id': 485, 'synset': 'fudge.n.01', 'synonyms': ['fudge'], 'def': 'soft creamy candy', 'name': 'fudge'}, {'frequency': 'r', 'id': 486, 'synset': 'funnel.n.02', 'synonyms': ['funnel'], 'def': 'a cone-shaped utensil used to channel a substance into a container with a small mouth', 'name': 'funnel'}, {'frequency': 'c', 'id': 487, 'synset': 'futon.n.01', 'synonyms': ['futon'], 'def': 'a pad that is used for sleeping on the floor or on a raised frame', 'name': 'futon'}, {'frequency': 'r', 'id': 488, 'synset': 'gag.n.02', 'synonyms': ['gag', 'muzzle'], 'def': "restraint put into a person's mouth to prevent speaking or shouting", 'name': 'gag'}, {'frequency': 'r', 'id': 489, 'synset': 'garbage.n.03', 'synonyms': ['garbage'], 'def': 'a receptacle where waste can be discarded', 'name': 'garbage'}, {'frequency': 'c', 'id': 490, 'synset': 'garbage_truck.n.01', 'synonyms': ['garbage_truck'], 'def': 'a truck for collecting domestic refuse', 'name': 'garbage_truck'}, {'frequency': 'c', 'id': 491, 'synset': 'garden_hose.n.01', 'synonyms': ['garden_hose'], 'def': 'a hose used for watering a lawn or garden', 'name': 'garden_hose'}, {'frequency': 'c', 'id': 492, 'synset': 'gargle.n.01', 'synonyms': ['gargle', 'mouthwash'], 'def': 'a medicated solution used for gargling and rinsing the mouth', 'name': 'gargle'}, {'frequency': 'r', 'id': 493, 'synset': 'gargoyle.n.02', 'synonyms': ['gargoyle'], 'def': 'an ornament consisting of a grotesquely carved figure of a person or animal', 'name': 'gargoyle'}, {'frequency': 'c', 'id': 494, 'synset': 'garlic.n.02', 'synonyms': ['garlic', 'ail'], 'def': 'aromatic bulb used as seasoning', 'name': 'garlic'}, {'frequency': 'r', 'id': 495, 'synset': 'gasmask.n.01', 'synonyms': ['gasmask', 'respirator', 'gas_helmet'], 'def': 'a protective face mask with a filter', 'name': 'gasmask'}, {'frequency': 'r', 'id': 496, 'synset': 'gazelle.n.01', 'synonyms': ['gazelle'], 'def': 'small swift graceful antelope of Africa and Asia having lustrous eyes', 'name': 'gazelle'}, {'frequency': 'c', 'id': 497, 'synset': 'gelatin.n.02', 'synonyms': ['gelatin', 'jelly'], 'def': 'an edible jelly made with gelatin and used as a dessert or salad base or a coating for foods', 'name': 'gelatin'}, {'frequency': 'r', 'id': 498, 'synset': 'gem.n.02', 'synonyms': ['gemstone'], 'def': 'a crystalline rock that can be cut and polished for jewelry', 'name': 'gemstone'}, {'frequency': 'c', 'id': 499, 'synset': 'giant_panda.n.01', 'synonyms': ['giant_panda', 'panda', 'panda_bear'], 'def': 'large black-and-white herbivorous mammal of bamboo forests of China and Tibet', 'name': 'giant_panda'}, {'frequency': 'c', 'id': 500, 'synset': 'gift_wrap.n.01', 'synonyms': ['gift_wrap'], 'def': 'attractive wrapping paper suitable for wrapping gifts', 'name': 'gift_wrap'}, {'frequency': 'c', 'id': 501, 'synset': 'ginger.n.03', 'synonyms': ['ginger', 'gingerroot'], 'def': 'the root of the common ginger plant; used fresh as a seasoning', 'name': 'ginger'}, {'frequency': 'f', 'id': 502, 'synset': 'giraffe.n.01', 'synonyms': ['giraffe'], 'def': 'tall animal having a spotted coat and small horns and very long neck and legs', 'name': 'giraffe'}, {'frequency': 'c', 'id': 503, 'synset': 'girdle.n.02', 'synonyms': ['cincture', 'sash', 'waistband', 'waistcloth'], 'def': 'a band of material around the waist that strengthens a skirt or trousers', 'name': 'cincture'}, {'frequency': 'f', 'id': 504, 'synset': 'glass.n.02', 'synonyms': ['glass_(drink_container)', 'drinking_glass'], 'def': 'a container for holding liquids while drinking', 'name': 'glass_(drink_container)'}, {'frequency': 'c', 'id': 505, 'synset': 'globe.n.03', 'synonyms': ['globe'], 'def': 'a sphere on which a map (especially of the earth) is represented', 'name': 'globe'}, {'frequency': 'f', 'id': 506, 'synset': 'glove.n.02', 'synonyms': ['glove'], 'def': 'handwear covering the hand', 'name': 'glove'}, {'frequency': 'c', 'id': 507, 'synset': 'goat.n.01', 'synonyms': ['goat'], 'def': 'a common goat', 'name': 'goat'}, {'frequency': 'f', 'id': 508, 'synset': 'goggles.n.01', 'synonyms': ['goggles'], 'def': 'tight-fitting spectacles worn to protect the eyes', 'name': 'goggles'}, {'frequency': 'r', 'id': 509, 'synset': 'goldfish.n.01', 'synonyms': ['goldfish'], 'def': 'small golden or orange-red freshwater fishes used as pond or aquarium pets', 'name': 'goldfish'}, {'frequency': 'r', 'id': 510, 'synset': 'golf_club.n.02', 'synonyms': ['golf_club', 'golf-club'], 'def': 'golf equipment used by a golfer to hit a golf ball', 'name': 'golf_club'}, {'frequency': 'c', 'id': 511, 'synset': 'golfcart.n.01', 'synonyms': ['golfcart'], 'def': 'a small motor vehicle in which golfers can ride between shots', 'name': 'golfcart'}, {'frequency': 'r', 'id': 512, 'synset': 'gondola.n.02', 'synonyms': ['gondola_(boat)'], 'def': 'long narrow flat-bottomed boat propelled by sculling; traditionally used on canals of Venice', 'name': 'gondola_(boat)'}, {'frequency': 'c', 'id': 513, 'synset': 'goose.n.01', 'synonyms': ['goose'], 'def': 'loud, web-footed long-necked aquatic birds usually larger than ducks', 'name': 'goose'}, {'frequency': 'r', 'id': 514, 'synset': 'gorilla.n.01', 'synonyms': ['gorilla'], 'def': 'largest ape', 'name': 'gorilla'}, {'frequency': 'r', 'id': 515, 'synset': 'gourd.n.02', 'synonyms': ['gourd'], 'def': 'any of numerous inedible fruits with hard rinds', 'name': 'gourd'}, {'frequency': 'r', 'id': 516, 'synset': 'gown.n.04', 'synonyms': ['surgical_gown', 'scrubs_(surgical_clothing)'], 'def': 'protective garment worn by surgeons during operations', 'name': 'surgical_gown'}, {'frequency': 'f', 'id': 517, 'synset': 'grape.n.01', 'synonyms': ['grape'], 'def': 'any of various juicy fruit with green or purple skins; grow in clusters', 'name': 'grape'}, {'frequency': 'r', 'id': 518, 'synset': 'grasshopper.n.01', 'synonyms': ['grasshopper'], 'def': 'plant-eating insect with hind legs adapted for leaping', 'name': 'grasshopper'}, {'frequency': 'c', 'id': 519, 'synset': 'grater.n.01', 'synonyms': ['grater'], 'def': 'utensil with sharp perforations for shredding foods (as vegetables or cheese)', 'name': 'grater'}, {'frequency': 'c', 'id': 520, 'synset': 'gravestone.n.01', 'synonyms': ['gravestone', 'headstone', 'tombstone'], 'def': 'a stone that is used to mark a grave', 'name': 'gravestone'}, {'frequency': 'r', 'id': 521, 'synset': 'gravy_boat.n.01', 'synonyms': ['gravy_boat', 'gravy_holder'], 'def': 'a dish (often boat-shaped) for serving gravy or sauce', 'name': 'gravy_boat'}, {'frequency': 'c', 'id': 522, 'synset': 'green_bean.n.02', 'synonyms': ['green_bean'], 'def': 'a common bean plant cultivated for its slender green edible pods', 'name': 'green_bean'}, {'frequency': 'c', 'id': 523, 'synset': 'green_onion.n.01', 'synonyms': ['green_onion', 'spring_onion', 'scallion'], 'def': 'a young onion before the bulb has enlarged', 'name': 'green_onion'}, {'frequency': 'r', 'id': 524, 'synset': 'griddle.n.01', 'synonyms': ['griddle'], 'def': 'cooking utensil consisting of a flat heated surface on which food is cooked', 'name': 'griddle'}, {'frequency': 'r', 'id': 525, 'synset': 'grillroom.n.01', 'synonyms': ['grillroom', 'grill_(restaurant)'], 'def': 'a restaurant where food is cooked on a grill', 'name': 'grillroom'}, {'frequency': 'r', 'id': 526, 'synset': 'grinder.n.04', 'synonyms': ['grinder_(tool)'], 'def': 'a machine tool that polishes metal', 'name': 'grinder_(tool)'}, {'frequency': 'r', 'id': 527, 'synset': 'grits.n.01', 'synonyms': ['grits', 'hominy_grits'], 'def': 'coarsely ground corn boiled as a breakfast dish', 'name': 'grits'}, {'frequency': 'c', 'id': 528, 'synset': 'grizzly.n.01', 'synonyms': ['grizzly', 'grizzly_bear'], 'def': 'powerful brownish-yellow bear of the uplands of western North America', 'name': 'grizzly'}, {'frequency': 'c', 'id': 529, 'synset': 'grocery_bag.n.01', 'synonyms': ['grocery_bag'], 'def': "a sack for holding customer's groceries", 'name': 'grocery_bag'}, {'frequency': 'r', 'id': 530, 'synset': 'guacamole.n.01', 'synonyms': ['guacamole'], 'def': 'a dip made of mashed avocado mixed with chopped onions and other seasonings', 'name': 'guacamole'}, {'frequency': 'f', 'id': 531, 'synset': 'guitar.n.01', 'synonyms': ['guitar'], 'def': 'a stringed instrument usually having six strings; played by strumming or plucking', 'name': 'guitar'}, {'frequency': 'c', 'id': 532, 'synset': 'gull.n.02', 'synonyms': ['gull', 'seagull'], 'def': 'mostly white aquatic bird having long pointed wings and short legs', 'name': 'gull'}, {'frequency': 'c', 'id': 533, 'synset': 'gun.n.01', 'synonyms': ['gun'], 'def': 'a weapon that discharges a bullet at high velocity from a metal tube', 'name': 'gun'}, {'frequency': 'r', 'id': 534, 'synset': 'hair_spray.n.01', 'synonyms': ['hair_spray'], 'def': 'substance sprayed on the hair to hold it in place', 'name': 'hair_spray'}, {'frequency': 'c', 'id': 535, 'synset': 'hairbrush.n.01', 'synonyms': ['hairbrush'], 'def': "a brush used to groom a person's hair", 'name': 'hairbrush'}, {'frequency': 'c', 'id': 536, 'synset': 'hairnet.n.01', 'synonyms': ['hairnet'], 'def': 'a small net that someone wears over their hair to keep it in place', 'name': 'hairnet'}, {'frequency': 'c', 'id': 537, 'synset': 'hairpin.n.01', 'synonyms': ['hairpin'], 'def': "a double pronged pin used to hold women's hair in place", 'name': 'hairpin'}, {'frequency': 'f', 'id': 538, 'synset': 'ham.n.01', 'synonyms': ['ham', 'jambon', 'gammon'], 'def': 'meat cut from the thigh of a hog (usually smoked)', 'name': 'ham'}, {'frequency': 'c', 'id': 539, 'synset': 'hamburger.n.01', 'synonyms': ['hamburger', 'beefburger', 'burger'], 'def': 'a sandwich consisting of a patty of minced beef served on a bun', 'name': 'hamburger'}, {'frequency': 'c', 'id': 540, 'synset': 'hammer.n.02', 'synonyms': ['hammer'], 'def': 'a hand tool with a heavy head and a handle; used to deliver an impulsive force by striking', 'name': 'hammer'}, {'frequency': 'r', 'id': 541, 'synset': 'hammock.n.02', 'synonyms': ['hammock'], 'def': 'a hanging bed of canvas or rope netting (usually suspended between two trees)', 'name': 'hammock'}, {'frequency': 'r', 'id': 542, 'synset': 'hamper.n.02', 'synonyms': ['hamper'], 'def': 'a basket usually with a cover', 'name': 'hamper'}, {'frequency': 'r', 'id': 543, 'synset': 'hamster.n.01', 'synonyms': ['hamster'], 'def': 'short-tailed burrowing rodent with large cheek pouches', 'name': 'hamster'}, {'frequency': 'c', 'id': 544, 'synset': 'hand_blower.n.01', 'synonyms': ['hair_dryer'], 'def': 'a hand-held electric blower that can blow warm air onto the hair', 'name': 'hair_dryer'}, {'frequency': 'r', 'id': 545, 'synset': 'hand_glass.n.01', 'synonyms': ['hand_glass', 'hand_mirror'], 'def': 'a mirror intended to be held in the hand', 'name': 'hand_glass'}, {'frequency': 'f', 'id': 546, 'synset': 'hand_towel.n.01', 'synonyms': ['hand_towel', 'face_towel'], 'def': 'a small towel used to dry the hands or face', 'name': 'hand_towel'}, {'frequency': 'c', 'id': 547, 'synset': 'handcart.n.01', 'synonyms': ['handcart', 'pushcart', 'hand_truck'], 'def': 'wheeled vehicle that can be pushed by a person', 'name': 'handcart'}, {'frequency': 'r', 'id': 548, 'synset': 'handcuff.n.01', 'synonyms': ['handcuff'], 'def': 'shackle that consists of a metal loop that can be locked around the wrist', 'name': 'handcuff'}, {'frequency': 'c', 'id': 549, 'synset': 'handkerchief.n.01', 'synonyms': ['handkerchief'], 'def': 'a square piece of cloth used for wiping the eyes or nose or as a costume accessory', 'name': 'handkerchief'}, {'frequency': 'f', 'id': 550, 'synset': 'handle.n.01', 'synonyms': ['handle', 'grip', 'handgrip'], 'def': 'the appendage to an object that is designed to be held in order to use or move it', 'name': 'handle'}, {'frequency': 'r', 'id': 551, 'synset': 'handsaw.n.01', 'synonyms': ['handsaw', "carpenter's_saw"], 'def': 'a saw used with one hand for cutting wood', 'name': 'handsaw'}, {'frequency': 'r', 'id': 552, 'synset': 'hardback.n.01', 'synonyms': ['hardback_book', 'hardcover_book'], 'def': 'a book with cardboard or cloth or leather covers', 'name': 'hardback_book'}, {'frequency': 'r', 'id': 553, 'synset': 'harmonium.n.01', 'synonyms': ['harmonium', 'organ_(musical_instrument)', 'reed_organ_(musical_instrument)'], 'def': 'a free-reed instrument in which air is forced through the reeds by bellows', 'name': 'harmonium'}, {'frequency': 'f', 'id': 554, 'synset': 'hat.n.01', 'synonyms': ['hat'], 'def': 'headwear that protects the head from bad weather, sun, or worn for fashion', 'name': 'hat'}, {'frequency': 'r', 'id': 555, 'synset': 'hatbox.n.01', 'synonyms': ['hatbox'], 'def': 'a round piece of luggage for carrying hats', 'name': 'hatbox'}, {'frequency': 'r', 'id': 556, 'synset': 'hatch.n.03', 'synonyms': ['hatch'], 'def': 'a movable barrier covering a hatchway', 'name': 'hatch'}, {'frequency': 'c', 'id': 557, 'synset': 'head_covering.n.01', 'synonyms': ['veil'], 'def': 'a garment that covers the head and face', 'name': 'veil'}, {'frequency': 'f', 'id': 558, 'synset': 'headband.n.01', 'synonyms': ['headband'], 'def': 'a band worn around or over the head', 'name': 'headband'}, {'frequency': 'f', 'id': 559, 'synset': 'headboard.n.01', 'synonyms': ['headboard'], 'def': 'a vertical board or panel forming the head of a bedstead', 'name': 'headboard'}, {'frequency': 'f', 'id': 560, 'synset': 'headlight.n.01', 'synonyms': ['headlight', 'headlamp'], 'def': 'a powerful light with reflector; attached to the front of an automobile or locomotive', 'name': 'headlight'}, {'frequency': 'c', 'id': 561, 'synset': 'headscarf.n.01', 'synonyms': ['headscarf'], 'def': 'a kerchief worn over the head and tied under the chin', 'name': 'headscarf'}, {'frequency': 'r', 'id': 562, 'synset': 'headset.n.01', 'synonyms': ['headset'], 'def': 'receiver consisting of a pair of headphones', 'name': 'headset'}, {'frequency': 'c', 'id': 563, 'synset': 'headstall.n.01', 'synonyms': ['headstall_(for_horses)', 'headpiece_(for_horses)'], 'def': "the band that is the part of a bridle that fits around a horse's head", 'name': 'headstall_(for_horses)'}, {'frequency': 'r', 'id': 564, 'synset': 'hearing_aid.n.02', 'synonyms': ['hearing_aid'], 'def': 'an acoustic device used to direct sound to the ear of a hearing-impaired person', 'name': 'hearing_aid'}, {'frequency': 'c', 'id': 565, 'synset': 'heart.n.02', 'synonyms': ['heart'], 'def': 'a muscular organ; its contractions move the blood through the body', 'name': 'heart'}, {'frequency': 'c', 'id': 566, 'synset': 'heater.n.01', 'synonyms': ['heater', 'warmer'], 'def': 'device that heats water or supplies warmth to a room', 'name': 'heater'}, {'frequency': 'c', 'id': 567, 'synset': 'helicopter.n.01', 'synonyms': ['helicopter'], 'def': 'an aircraft without wings that obtains its lift from the rotation of overhead blades', 'name': 'helicopter'}, {'frequency': 'f', 'id': 568, 'synset': 'helmet.n.02', 'synonyms': ['helmet'], 'def': 'a protective headgear made of hard material to resist blows', 'name': 'helmet'}, {'frequency': 'r', 'id': 569, 'synset': 'heron.n.02', 'synonyms': ['heron'], 'def': 'grey or white wading bird with long neck and long legs and (usually) long bill', 'name': 'heron'}, {'frequency': 'c', 'id': 570, 'synset': 'highchair.n.01', 'synonyms': ['highchair', 'feeding_chair'], 'def': 'a chair for feeding a very young child', 'name': 'highchair'}, {'frequency': 'f', 'id': 571, 'synset': 'hinge.n.01', 'synonyms': ['hinge'], 'def': 'a joint that holds two parts together so that one can swing relative to the other', 'name': 'hinge'}, {'frequency': 'r', 'id': 572, 'synset': 'hippopotamus.n.01', 'synonyms': ['hippopotamus'], 'def': 'massive thick-skinned animal living in or around rivers of tropical Africa', 'name': 'hippopotamus'}, {'frequency': 'r', 'id': 573, 'synset': 'hockey_stick.n.01', 'synonyms': ['hockey_stick'], 'def': 'sports implement consisting of a stick used by hockey players to move the puck', 'name': 'hockey_stick'}, {'frequency': 'c', 'id': 574, 'synset': 'hog.n.03', 'synonyms': ['hog', 'pig'], 'def': 'domestic swine', 'name': 'hog'}, {'frequency': 'f', 'id': 575, 'synset': 'home_plate.n.01', 'synonyms': ['home_plate_(baseball)', 'home_base_(baseball)'], 'def': '(baseball) a rubber slab where the batter stands; it must be touched by a base runner in order to score', 'name': 'home_plate_(baseball)'}, {'frequency': 'c', 'id': 576, 'synset': 'honey.n.01', 'synonyms': ['honey'], 'def': 'a sweet yellow liquid produced by bees', 'name': 'honey'}, {'frequency': 'f', 'id': 577, 'synset': 'hood.n.06', 'synonyms': ['fume_hood', 'exhaust_hood'], 'def': 'metal covering leading to a vent that exhausts smoke or fumes', 'name': 'fume_hood'}, {'frequency': 'f', 'id': 578, 'synset': 'hook.n.05', 'synonyms': ['hook'], 'def': 'a curved or bent implement for suspending or pulling something', 'name': 'hook'}, {'frequency': 'f', 'id': 579, 'synset': 'horse.n.01', 'synonyms': ['horse'], 'def': 'a common horse', 'name': 'horse'}, {'frequency': 'f', 'id': 580, 'synset': 'hose.n.03', 'synonyms': ['hose', 'hosepipe'], 'def': 'a flexible pipe for conveying a liquid or gas', 'name': 'hose'}, {'frequency': 'r', 'id': 581, 'synset': 'hot-air_balloon.n.01', 'synonyms': ['hot-air_balloon'], 'def': 'balloon for travel through the air in a basket suspended below a large bag of heated air', 'name': 'hot-air_balloon'}, {'frequency': 'r', 'id': 582, 'synset': 'hot_plate.n.01', 'synonyms': ['hotplate'], 'def': 'a portable electric appliance for heating or cooking or keeping food warm', 'name': 'hotplate'}, {'frequency': 'c', 'id': 583, 'synset': 'hot_sauce.n.01', 'synonyms': ['hot_sauce'], 'def': 'a pungent peppery sauce', 'name': 'hot_sauce'}, {'frequency': 'r', 'id': 584, 'synset': 'hourglass.n.01', 'synonyms': ['hourglass'], 'def': 'a sandglass timer that runs for sixty minutes', 'name': 'hourglass'}, {'frequency': 'r', 'id': 585, 'synset': 'houseboat.n.01', 'synonyms': ['houseboat'], 'def': 'a barge that is designed and equipped for use as a dwelling', 'name': 'houseboat'}, {'frequency': 'r', 'id': 586, 'synset': 'hummingbird.n.01', 'synonyms': ['hummingbird'], 'def': 'tiny American bird having brilliant iridescent plumage and long slender bills', 'name': 'hummingbird'}, {'frequency': 'r', 'id': 587, 'synset': 'hummus.n.01', 'synonyms': ['hummus', 'humus', 'hommos', 'hoummos', 'humous'], 'def': 'a thick spread made from mashed chickpeas', 'name': 'hummus'}, {'frequency': 'c', 'id': 588, 'synset': 'ice_bear.n.01', 'synonyms': ['polar_bear'], 'def': 'white bear of Arctic regions', 'name': 'polar_bear'}, {'frequency': 'c', 'id': 589, 'synset': 'ice_cream.n.01', 'synonyms': ['icecream'], 'def': 'frozen dessert containing cream and sugar and flavoring', 'name': 'icecream'}, {'frequency': 'r', 'id': 590, 'synset': 'ice_lolly.n.01', 'synonyms': ['popsicle'], 'def': 'ice cream or water ice on a small wooden stick', 'name': 'popsicle'}, {'frequency': 'c', 'id': 591, 'synset': 'ice_maker.n.01', 'synonyms': ['ice_maker'], 'def': 'an appliance included in some electric refrigerators for making ice cubes', 'name': 'ice_maker'}, {'frequency': 'r', 'id': 592, 'synset': 'ice_pack.n.01', 'synonyms': ['ice_pack', 'ice_bag'], 'def': 'a waterproof bag filled with ice: applied to the body (especially the head) to cool or reduce swelling', 'name': 'ice_pack'}, {'frequency': 'r', 'id': 593, 'synset': 'ice_skate.n.01', 'synonyms': ['ice_skate'], 'def': 'skate consisting of a boot with a steel blade fitted to the sole', 'name': 'ice_skate'}, {'frequency': 'r', 'id': 594, 'synset': 'ice_tea.n.01', 'synonyms': ['ice_tea', 'iced_tea'], 'def': 'strong tea served over ice', 'name': 'ice_tea'}, {'frequency': 'c', 'id': 595, 'synset': 'igniter.n.01', 'synonyms': ['igniter', 'ignitor', 'lighter'], 'def': 'a substance or device used to start a fire', 'name': 'igniter'}, {'frequency': 'r', 'id': 596, 'synset': 'incense.n.01', 'synonyms': ['incense'], 'def': 'a substance that produces a fragrant odor when burned', 'name': 'incense'}, {'frequency': 'r', 'id': 597, 'synset': 'inhaler.n.01', 'synonyms': ['inhaler', 'inhalator'], 'def': 'a dispenser that produces a chemical vapor to be inhaled through mouth or nose', 'name': 'inhaler'}, {'frequency': 'c', 'id': 598, 'synset': 'ipod.n.01', 'synonyms': ['iPod'], 'def': 'a pocket-sized device used to play music files', 'name': 'iPod'}, {'frequency': 'c', 'id': 599, 'synset': 'iron.n.04', 'synonyms': ['iron_(for_clothing)', 'smoothing_iron_(for_clothing)'], 'def': 'home appliance consisting of a flat metal base that is heated and used to smooth cloth', 'name': 'iron_(for_clothing)'}, {'frequency': 'r', 'id': 600, 'synset': 'ironing_board.n.01', 'synonyms': ['ironing_board'], 'def': 'narrow padded board on collapsible supports; used for ironing clothes', 'name': 'ironing_board'}, {'frequency': 'f', 'id': 601, 'synset': 'jacket.n.01', 'synonyms': ['jacket'], 'def': 'a waist-length coat', 'name': 'jacket'}, {'frequency': 'r', 'id': 602, 'synset': 'jam.n.01', 'synonyms': ['jam'], 'def': 'preserve of crushed fruit', 'name': 'jam'}, {'frequency': 'f', 'id': 603, 'synset': 'jean.n.01', 'synonyms': ['jean', 'blue_jean', 'denim'], 'def': '(usually plural) close-fitting trousers of heavy denim for manual work or casual wear', 'name': 'jean'}, {'frequency': 'c', 'id': 604, 'synset': 'jeep.n.01', 'synonyms': ['jeep', 'landrover'], 'def': 'a car suitable for traveling over rough terrain', 'name': 'jeep'}, {'frequency': 'r', 'id': 605, 'synset': 'jelly_bean.n.01', 'synonyms': ['jelly_bean', 'jelly_egg'], 'def': 'sugar-glazed jellied candy', 'name': 'jelly_bean'}, {'frequency': 'f', 'id': 606, 'synset': 'jersey.n.03', 'synonyms': ['jersey', 'T-shirt', 'tee_shirt'], 'def': 'a close-fitting pullover shirt', 'name': 'jersey'}, {'frequency': 'c', 'id': 607, 'synset': 'jet.n.01', 'synonyms': ['jet_plane', 'jet-propelled_plane'], 'def': 'an airplane powered by one or more jet engines', 'name': 'jet_plane'}, {'frequency': 'c', 'id': 608, 'synset': 'jewelry.n.01', 'synonyms': ['jewelry', 'jewellery'], 'def': 'an adornment (as a bracelet or ring or necklace) made of precious metals and set with gems (or imitation gems)', 'name': 'jewelry'}, {'frequency': 'r', 'id': 609, 'synset': 'joystick.n.02', 'synonyms': ['joystick'], 'def': 'a control device for computers consisting of a vertical handle that can move freely in two directions', 'name': 'joystick'}, {'frequency': 'r', 'id': 610, 'synset': 'jump_suit.n.01', 'synonyms': ['jumpsuit'], 'def': "one-piece garment fashioned after a parachutist's uniform", 'name': 'jumpsuit'}, {'frequency': 'c', 'id': 611, 'synset': 'kayak.n.01', 'synonyms': ['kayak'], 'def': 'a small canoe consisting of a light frame made watertight with animal skins', 'name': 'kayak'}, {'frequency': 'r', 'id': 612, 'synset': 'keg.n.02', 'synonyms': ['keg'], 'def': 'small cask or barrel', 'name': 'keg'}, {'frequency': 'r', 'id': 613, 'synset': 'kennel.n.01', 'synonyms': ['kennel', 'doghouse'], 'def': 'outbuilding that serves as a shelter for a dog', 'name': 'kennel'}, {'frequency': 'c', 'id': 614, 'synset': 'kettle.n.01', 'synonyms': ['kettle', 'boiler'], 'def': 'a metal pot for stewing or boiling; usually has a lid', 'name': 'kettle'}, {'frequency': 'f', 'id': 615, 'synset': 'key.n.01', 'synonyms': ['key'], 'def': 'metal instrument used to unlock a lock', 'name': 'key'}, {'frequency': 'r', 'id': 616, 'synset': 'keycard.n.01', 'synonyms': ['keycard'], 'def': 'a plastic card used to gain access typically to a door', 'name': 'keycard'}, {'frequency': 'r', 'id': 617, 'synset': 'kilt.n.01', 'synonyms': ['kilt'], 'def': 'a knee-length pleated tartan skirt worn by men as part of the traditional dress in the Highlands of northern Scotland', 'name': 'kilt'}, {'frequency': 'c', 'id': 618, 'synset': 'kimono.n.01', 'synonyms': ['kimono'], 'def': 'a loose robe; imitated from robes originally worn by Japanese', 'name': 'kimono'}, {'frequency': 'f', 'id': 619, 'synset': 'kitchen_sink.n.01', 'synonyms': ['kitchen_sink'], 'def': 'a sink in a kitchen', 'name': 'kitchen_sink'}, {'frequency': 'c', 'id': 620, 'synset': 'kitchen_table.n.01', 'synonyms': ['kitchen_table'], 'def': 'a table in the kitchen', 'name': 'kitchen_table'}, {'frequency': 'f', 'id': 621, 'synset': 'kite.n.03', 'synonyms': ['kite'], 'def': 'plaything consisting of a light frame covered with tissue paper; flown in wind at end of a string', 'name': 'kite'}, {'frequency': 'c', 'id': 622, 'synset': 'kitten.n.01', 'synonyms': ['kitten', 'kitty'], 'def': 'young domestic cat', 'name': 'kitten'}, {'frequency': 'c', 'id': 623, 'synset': 'kiwi.n.03', 'synonyms': ['kiwi_fruit'], 'def': 'fuzzy brown egg-shaped fruit with slightly tart green flesh', 'name': 'kiwi_fruit'}, {'frequency': 'f', 'id': 624, 'synset': 'knee_pad.n.01', 'synonyms': ['knee_pad'], 'def': 'protective garment consisting of a pad worn by football or baseball or hockey players', 'name': 'knee_pad'}, {'frequency': 'f', 'id': 625, 'synset': 'knife.n.01', 'synonyms': ['knife'], 'def': 'tool with a blade and point used as a cutting instrument', 'name': 'knife'}, {'frequency': 'r', 'id': 626, 'synset': 'knight.n.02', 'synonyms': ['knight_(chess_piece)', 'horse_(chess_piece)'], 'def': 'a chess game piece shaped to resemble the head of a horse', 'name': 'knight_(chess_piece)'}, {'frequency': 'r', 'id': 627, 'synset': 'knitting_needle.n.01', 'synonyms': ['knitting_needle'], 'def': 'needle consisting of a slender rod with pointed ends; usually used in pairs', 'name': 'knitting_needle'}, {'frequency': 'f', 'id': 628, 'synset': 'knob.n.02', 'synonyms': ['knob'], 'def': 'a round handle often found on a door', 'name': 'knob'}, {'frequency': 'r', 'id': 629, 'synset': 'knocker.n.05', 'synonyms': ['knocker_(on_a_door)', 'doorknocker'], 'def': 'a device (usually metal and ornamental) attached by a hinge to a door', 'name': 'knocker_(on_a_door)'}, {'frequency': 'r', 'id': 630, 'synset': 'koala.n.01', 'synonyms': ['koala', 'koala_bear'], 'def': 'sluggish tailless Australian marsupial with grey furry ears and coat', 'name': 'koala'}, {'frequency': 'r', 'id': 631, 'synset': 'lab_coat.n.01', 'synonyms': ['lab_coat', 'laboratory_coat'], 'def': 'a light coat worn to protect clothing from substances used while working in a laboratory', 'name': 'lab_coat'}, {'frequency': 'f', 'id': 632, 'synset': 'ladder.n.01', 'synonyms': ['ladder'], 'def': 'steps consisting of two parallel members connected by rungs', 'name': 'ladder'}, {'frequency': 'c', 'id': 633, 'synset': 'ladle.n.01', 'synonyms': ['ladle'], 'def': 'a spoon-shaped vessel with a long handle frequently used to transfer liquids', 'name': 'ladle'}, {'frequency': 'r', 'id': 634, 'synset': 'ladybug.n.01', 'synonyms': ['ladybug', 'ladybeetle', 'ladybird_beetle'], 'def': 'small round bright-colored and spotted beetle, typically red and black', 'name': 'ladybug'}, {'frequency': 'c', 'id': 635, 'synset': 'lamb.n.01', 'synonyms': ['lamb_(animal)'], 'def': 'young sheep', 'name': 'lamb_(animal)'}, {'frequency': 'r', 'id': 636, 'synset': 'lamb_chop.n.01', 'synonyms': ['lamb-chop', 'lambchop'], 'def': 'chop cut from a lamb', 'name': 'lamb-chop'}, {'frequency': 'f', 'id': 637, 'synset': 'lamp.n.02', 'synonyms': ['lamp'], 'def': 'a piece of furniture holding one or more electric light bulbs', 'name': 'lamp'}, {'frequency': 'f', 'id': 638, 'synset': 'lamppost.n.01', 'synonyms': ['lamppost'], 'def': 'a metal post supporting an outdoor lamp (such as a streetlight)', 'name': 'lamppost'}, {'frequency': 'f', 'id': 639, 'synset': 'lampshade.n.01', 'synonyms': ['lampshade'], 'def': 'a protective ornamental shade used to screen a light bulb from direct view', 'name': 'lampshade'}, {'frequency': 'c', 'id': 640, 'synset': 'lantern.n.01', 'synonyms': ['lantern'], 'def': 'light in a transparent protective case', 'name': 'lantern'}, {'frequency': 'f', 'id': 641, 'synset': 'lanyard.n.02', 'synonyms': ['lanyard', 'laniard'], 'def': 'a cord worn around the neck to hold a knife or whistle, etc.', 'name': 'lanyard'}, {'frequency': 'f', 'id': 642, 'synset': 'laptop.n.01', 'synonyms': ['laptop_computer', 'notebook_computer'], 'def': 'a portable computer small enough to use in your lap', 'name': 'laptop_computer'}, {'frequency': 'r', 'id': 643, 'synset': 'lasagna.n.01', 'synonyms': ['lasagna', 'lasagne'], 'def': 'baked dish of layers of lasagna pasta with sauce and cheese and meat or vegetables', 'name': 'lasagna'}, {'frequency': 'c', 'id': 644, 'synset': 'latch.n.02', 'synonyms': ['latch'], 'def': 'a bar that can be lowered or slid into a groove to fasten a door or gate', 'name': 'latch'}, {'frequency': 'r', 'id': 645, 'synset': 'lawn_mower.n.01', 'synonyms': ['lawn_mower'], 'def': 'garden tool for mowing grass on lawns', 'name': 'lawn_mower'}, {'frequency': 'r', 'id': 646, 'synset': 'leather.n.01', 'synonyms': ['leather'], 'def': 'an animal skin made smooth and flexible by removing the hair and then tanning', 'name': 'leather'}, {'frequency': 'c', 'id': 647, 'synset': 'legging.n.01', 'synonyms': ['legging_(clothing)', 'leging_(clothing)', 'leg_covering'], 'def': 'a garment covering the leg (usually extending from the knee to the ankle)', 'name': 'legging_(clothing)'}, {'frequency': 'c', 'id': 648, 'synset': 'lego.n.01', 'synonyms': ['Lego', 'Lego_set'], 'def': "a child's plastic construction set for making models from blocks", 'name': 'Lego'}, {'frequency': 'f', 'id': 649, 'synset': 'lemon.n.01', 'synonyms': ['lemon'], 'def': 'yellow oval fruit with juicy acidic flesh', 'name': 'lemon'}, {'frequency': 'r', 'id': 650, 'synset': 'lemonade.n.01', 'synonyms': ['lemonade'], 'def': 'sweetened beverage of diluted lemon juice', 'name': 'lemonade'}, {'frequency': 'f', 'id': 651, 'synset': 'lettuce.n.02', 'synonyms': ['lettuce'], 'def': 'leafy plant commonly eaten in salad or on sandwiches', 'name': 'lettuce'}, {'frequency': 'f', 'id': 652, 'synset': 'license_plate.n.01', 'synonyms': ['license_plate', 'numberplate'], 'def': "a plate mounted on the front and back of car and bearing the car's registration number", 'name': 'license_plate'}, {'frequency': 'f', 'id': 653, 'synset': 'life_buoy.n.01', 'synonyms': ['life_buoy', 'lifesaver', 'life_belt', 'life_ring'], 'def': 'a ring-shaped life preserver used to prevent drowning (NOT a life-jacket or vest)', 'name': 'life_buoy'}, {'frequency': 'f', 'id': 654, 'synset': 'life_jacket.n.01', 'synonyms': ['life_jacket', 'life_vest'], 'def': 'life preserver consisting of a sleeveless jacket of buoyant or inflatable design', 'name': 'life_jacket'}, {'frequency': 'f', 'id': 655, 'synset': 'light_bulb.n.01', 'synonyms': ['lightbulb'], 'def': 'glass bulb or tube shaped electric device that emits light (DO NOT MARK LAMPS AS A WHOLE)', 'name': 'lightbulb'}, {'frequency': 'r', 'id': 656, 'synset': 'lightning_rod.n.02', 'synonyms': ['lightning_rod', 'lightning_conductor'], 'def': 'a metallic conductor that is attached to a high point and leads to the ground', 'name': 'lightning_rod'}, {'frequency': 'c', 'id': 657, 'synset': 'lime.n.06', 'synonyms': ['lime'], 'def': 'the green acidic fruit of any of various lime trees', 'name': 'lime'}, {'frequency': 'r', 'id': 658, 'synset': 'limousine.n.01', 'synonyms': ['limousine'], 'def': 'long luxurious car; usually driven by a chauffeur', 'name': 'limousine'}, {'frequency': 'r', 'id': 659, 'synset': 'linen.n.02', 'synonyms': ['linen_paper'], 'def': 'a high-quality paper made of linen fibers or with a linen finish', 'name': 'linen_paper'}, {'frequency': 'c', 'id': 660, 'synset': 'lion.n.01', 'synonyms': ['lion'], 'def': 'large gregarious predatory cat of Africa and India', 'name': 'lion'}, {'frequency': 'c', 'id': 661, 'synset': 'lip_balm.n.01', 'synonyms': ['lip_balm'], 'def': 'a balm applied to the lips', 'name': 'lip_balm'}, {'frequency': 'c', 'id': 662, 'synset': 'lipstick.n.01', 'synonyms': ['lipstick', 'lip_rouge'], 'def': 'makeup that is used to color the lips', 'name': 'lipstick'}, {'frequency': 'r', 'id': 663, 'synset': 'liquor.n.01', 'synonyms': ['liquor', 'spirits', 'hard_liquor', 'liqueur', 'cordial'], 'def': 'an alcoholic beverage that is distilled rather than fermented', 'name': 'liquor'}, {'frequency': 'r', 'id': 664, 'synset': 'lizard.n.01', 'synonyms': ['lizard'], 'def': 'a reptile with usually two pairs of legs and a tapering tail', 'name': 'lizard'}, {'frequency': 'r', 'id': 665, 'synset': 'loafer.n.02', 'synonyms': ['Loafer_(type_of_shoe)'], 'def': 'a low leather step-in shoe', 'name': 'Loafer_(type_of_shoe)'}, {'frequency': 'f', 'id': 666, 'synset': 'log.n.01', 'synonyms': ['log'], 'def': 'a segment of the trunk of a tree when stripped of branches', 'name': 'log'}, {'frequency': 'c', 'id': 667, 'synset': 'lollipop.n.02', 'synonyms': ['lollipop'], 'def': 'hard candy on a stick', 'name': 'lollipop'}, {'frequency': 'c', 'id': 668, 'synset': 'lotion.n.01', 'synonyms': ['lotion'], 'def': 'any of various cosmetic preparations that are applied to the skin', 'name': 'lotion'}, {'frequency': 'f', 'id': 669, 'synset': 'loudspeaker.n.01', 'synonyms': ['speaker_(stero_equipment)'], 'def': 'electronic device that produces sound often as part of a stereo system', 'name': 'speaker_(stero_equipment)'}, {'frequency': 'c', 'id': 670, 'synset': 'love_seat.n.01', 'synonyms': ['loveseat'], 'def': 'small sofa that seats two people', 'name': 'loveseat'}, {'frequency': 'r', 'id': 671, 'synset': 'machine_gun.n.01', 'synonyms': ['machine_gun'], 'def': 'a rapidly firing automatic gun', 'name': 'machine_gun'}, {'frequency': 'f', 'id': 672, 'synset': 'magazine.n.02', 'synonyms': ['magazine'], 'def': 'a paperback periodic publication', 'name': 'magazine'}, {'frequency': 'f', 'id': 673, 'synset': 'magnet.n.01', 'synonyms': ['magnet'], 'def': 'a device that attracts iron and produces a magnetic field', 'name': 'magnet'}, {'frequency': 'r', 'id': 674, 'synset': 'mail_slot.n.01', 'synonyms': ['mail_slot'], 'def': 'a slot (usually in a door) through which mail can be delivered', 'name': 'mail_slot'}, {'frequency': 'c', 'id': 675, 'synset': 'mailbox.n.01', 'synonyms': ['mailbox_(at_home)', 'letter_box_(at_home)'], 'def': 'a private box for delivery of mail', 'name': 'mailbox_(at_home)'}, {'frequency': 'r', 'id': 676, 'synset': 'mallet.n.01', 'synonyms': ['mallet'], 'def': 'a sports implement with a long handle and a hammer-like head used to hit a ball', 'name': 'mallet'}, {'frequency': 'r', 'id': 677, 'synset': 'mammoth.n.01', 'synonyms': ['mammoth'], 'def': 'any of numerous extinct elephants widely distributed in the Pleistocene', 'name': 'mammoth'}, {'frequency': 'c', 'id': 678, 'synset': 'mandarin.n.05', 'synonyms': ['mandarin_orange'], 'def': 'a somewhat flat reddish-orange loose skinned citrus of China', 'name': 'mandarin_orange'}, {'frequency': 'c', 'id': 679, 'synset': 'manger.n.01', 'synonyms': ['manger', 'trough'], 'def': 'a container (usually in a barn or stable) from which cattle or horses feed', 'name': 'manger'}, {'frequency': 'f', 'id': 680, 'synset': 'manhole.n.01', 'synonyms': ['manhole'], 'def': 'a hole (usually with a flush cover) through which a person can gain access to an underground structure', 'name': 'manhole'}, {'frequency': 'c', 'id': 681, 'synset': 'map.n.01', 'synonyms': ['map'], 'def': "a diagrammatic representation of the earth's surface (or part of it)", 'name': 'map'}, {'frequency': 'c', 'id': 682, 'synset': 'marker.n.03', 'synonyms': ['marker'], 'def': 'a writing implement for making a mark', 'name': 'marker'}, {'frequency': 'r', 'id': 683, 'synset': 'martini.n.01', 'synonyms': ['martini'], 'def': 'a cocktail made of gin (or vodka) with dry vermouth', 'name': 'martini'}, {'frequency': 'r', 'id': 684, 'synset': 'mascot.n.01', 'synonyms': ['mascot'], 'def': 'a person or animal that is adopted by a team or other group as a symbolic figure', 'name': 'mascot'}, {'frequency': 'c', 'id': 685, 'synset': 'mashed_potato.n.01', 'synonyms': ['mashed_potato'], 'def': 'potato that has been peeled and boiled and then mashed', 'name': 'mashed_potato'}, {'frequency': 'r', 'id': 686, 'synset': 'masher.n.02', 'synonyms': ['masher'], 'def': 'a kitchen utensil used for mashing (e.g. potatoes)', 'name': 'masher'}, {'frequency': 'f', 'id': 687, 'synset': 'mask.n.04', 'synonyms': ['mask', 'facemask'], 'def': 'a protective covering worn over the face', 'name': 'mask'}, {'frequency': 'f', 'id': 688, 'synset': 'mast.n.01', 'synonyms': ['mast'], 'def': 'a vertical spar for supporting sails', 'name': 'mast'}, {'frequency': 'c', 'id': 689, 'synset': 'mat.n.03', 'synonyms': ['mat_(gym_equipment)', 'gym_mat'], 'def': 'sports equipment consisting of a piece of thick padding on the floor for gymnastics', 'name': 'mat_(gym_equipment)'}, {'frequency': 'r', 'id': 690, 'synset': 'matchbox.n.01', 'synonyms': ['matchbox'], 'def': 'a box for holding matches', 'name': 'matchbox'}, {'frequency': 'f', 'id': 691, 'synset': 'mattress.n.01', 'synonyms': ['mattress'], 'def': 'a thick pad filled with resilient material used as a bed or part of a bed', 'name': 'mattress'}, {'frequency': 'c', 'id': 692, 'synset': 'measuring_cup.n.01', 'synonyms': ['measuring_cup'], 'def': 'graduated cup used to measure liquid or granular ingredients', 'name': 'measuring_cup'}, {'frequency': 'c', 'id': 693, 'synset': 'measuring_stick.n.01', 'synonyms': ['measuring_stick', 'ruler_(measuring_stick)', 'measuring_rod'], 'def': 'measuring instrument having a sequence of marks at regular intervals', 'name': 'measuring_stick'}, {'frequency': 'c', 'id': 694, 'synset': 'meatball.n.01', 'synonyms': ['meatball'], 'def': 'ground meat formed into a ball and fried or simmered in broth', 'name': 'meatball'}, {'frequency': 'c', 'id': 695, 'synset': 'medicine.n.02', 'synonyms': ['medicine'], 'def': 'something that treats or prevents or alleviates the symptoms of disease', 'name': 'medicine'}, {'frequency': 'r', 'id': 696, 'synset': 'melon.n.01', 'synonyms': ['melon'], 'def': 'fruit of the gourd family having a hard rind and sweet juicy flesh', 'name': 'melon'}, {'frequency': 'f', 'id': 697, 'synset': 'microphone.n.01', 'synonyms': ['microphone'], 'def': 'device for converting sound waves into electrical energy', 'name': 'microphone'}, {'frequency': 'r', 'id': 698, 'synset': 'microscope.n.01', 'synonyms': ['microscope'], 'def': 'magnifier of the image of small objects', 'name': 'microscope'}, {'frequency': 'f', 'id': 699, 'synset': 'microwave.n.02', 'synonyms': ['microwave_oven'], 'def': 'kitchen appliance that cooks food by passing an electromagnetic wave through it', 'name': 'microwave_oven'}, {'frequency': 'r', 'id': 700, 'synset': 'milestone.n.01', 'synonyms': ['milestone', 'milepost'], 'def': 'stone post at side of a road to show distances', 'name': 'milestone'}, {'frequency': 'c', 'id': 701, 'synset': 'milk.n.01', 'synonyms': ['milk'], 'def': 'a white nutritious liquid secreted by mammals and used as food by human beings', 'name': 'milk'}, {'frequency': 'f', 'id': 702, 'synset': 'minivan.n.01', 'synonyms': ['minivan'], 'def': 'a small box-shaped passenger van', 'name': 'minivan'}, {'frequency': 'r', 'id': 703, 'synset': 'mint.n.05', 'synonyms': ['mint_candy'], 'def': 'a candy that is flavored with a mint oil', 'name': 'mint_candy'}, {'frequency': 'f', 'id': 704, 'synset': 'mirror.n.01', 'synonyms': ['mirror'], 'def': 'polished surface that forms images by reflecting light', 'name': 'mirror'}, {'frequency': 'c', 'id': 705, 'synset': 'mitten.n.01', 'synonyms': ['mitten'], 'def': 'glove that encases the thumb separately and the other four fingers together', 'name': 'mitten'}, {'frequency': 'c', 'id': 706, 'synset': 'mixer.n.04', 'synonyms': ['mixer_(kitchen_tool)', 'stand_mixer'], 'def': 'a kitchen utensil that is used for mixing foods', 'name': 'mixer_(kitchen_tool)'}, {'frequency': 'c', 'id': 707, 'synset': 'money.n.03', 'synonyms': ['money'], 'def': 'the official currency issued by a government or national bank', 'name': 'money'}, {'frequency': 'f', 'id': 708, 'synset': 'monitor.n.04', 'synonyms': ['monitor_(computer_equipment) computer_monitor'], 'def': 'a computer monitor', 'name': 'monitor_(computer_equipment) computer_monitor'}, {'frequency': 'c', 'id': 709, 'synset': 'monkey.n.01', 'synonyms': ['monkey'], 'def': 'any of various long-tailed primates', 'name': 'monkey'}, {'frequency': 'f', 'id': 710, 'synset': 'motor.n.01', 'synonyms': ['motor'], 'def': 'machine that converts other forms of energy into mechanical energy and so imparts motion', 'name': 'motor'}, {'frequency': 'f', 'id': 711, 'synset': 'motor_scooter.n.01', 'synonyms': ['motor_scooter', 'scooter'], 'def': 'a wheeled vehicle with small wheels and a low-powered engine', 'name': 'motor_scooter'}, {'frequency': 'r', 'id': 712, 'synset': 'motor_vehicle.n.01', 'synonyms': ['motor_vehicle', 'automotive_vehicle'], 'def': 'a self-propelled wheeled vehicle that does not run on rails', 'name': 'motor_vehicle'}, {'frequency': 'r', 'id': 713, 'synset': 'motorboat.n.01', 'synonyms': ['motorboat', 'powerboat'], 'def': 'a boat propelled by an internal-combustion engine', 'name': 'motorboat'}, {'frequency': 'f', 'id': 714, 'synset': 'motorcycle.n.01', 'synonyms': ['motorcycle'], 'def': 'a motor vehicle with two wheels and a strong frame', 'name': 'motorcycle'}, {'frequency': 'f', 'id': 715, 'synset': 'mound.n.01', 'synonyms': ['mound_(baseball)', "pitcher's_mound"], 'def': '(baseball) the slight elevation on which the pitcher stands', 'name': 'mound_(baseball)'}, {'frequency': 'r', 'id': 716, 'synset': 'mouse.n.01', 'synonyms': ['mouse_(animal_rodent)'], 'def': 'a small rodent with pointed snouts and small ears on elongated bodies with slender usually hairless tails', 'name': 'mouse_(animal_rodent)'}, {'frequency': 'f', 'id': 717, 'synset': 'mouse.n.04', 'synonyms': ['mouse_(computer_equipment)', 'computer_mouse'], 'def': 'a computer input device that controls an on-screen pointer', 'name': 'mouse_(computer_equipment)'}, {'frequency': 'f', 'id': 718, 'synset': 'mousepad.n.01', 'synonyms': ['mousepad'], 'def': 'a small portable pad that provides an operating surface for a computer mouse', 'name': 'mousepad'}, {'frequency': 'c', 'id': 719, 'synset': 'muffin.n.01', 'synonyms': ['muffin'], 'def': 'a sweet quick bread baked in a cup-shaped pan', 'name': 'muffin'}, {'frequency': 'f', 'id': 720, 'synset': 'mug.n.04', 'synonyms': ['mug'], 'def': 'with handle and usually cylindrical', 'name': 'mug'}, {'frequency': 'f', 'id': 721, 'synset': 'mushroom.n.02', 'synonyms': ['mushroom'], 'def': 'a common mushroom', 'name': 'mushroom'}, {'frequency': 'r', 'id': 722, 'synset': 'music_stool.n.01', 'synonyms': ['music_stool', 'piano_stool'], 'def': 'a stool for piano players; usually adjustable in height', 'name': 'music_stool'}, {'frequency': 'r', 'id': 723, 'synset': 'musical_instrument.n.01', 'synonyms': ['musical_instrument', 'instrument_(musical)'], 'def': 'any of various devices or contrivances that can be used to produce musical tones or sounds', 'name': 'musical_instrument'}, {'frequency': 'r', 'id': 724, 'synset': 'nailfile.n.01', 'synonyms': ['nailfile'], 'def': 'a small flat file for shaping the nails', 'name': 'nailfile'}, {'frequency': 'r', 'id': 725, 'synset': 'nameplate.n.01', 'synonyms': ['nameplate'], 'def': 'a plate bearing a name', 'name': 'nameplate'}, {'frequency': 'f', 'id': 726, 'synset': 'napkin.n.01', 'synonyms': ['napkin', 'table_napkin', 'serviette'], 'def': 'a small piece of table linen or paper that is used to wipe the mouth and to cover the lap in order to protect clothing', 'name': 'napkin'}, {'frequency': 'r', 'id': 727, 'synset': 'neckerchief.n.01', 'synonyms': ['neckerchief'], 'def': 'a kerchief worn around the neck', 'name': 'neckerchief'}, {'frequency': 'f', 'id': 728, 'synset': 'necklace.n.01', 'synonyms': ['necklace'], 'def': 'jewelry consisting of a cord or chain (often bearing gems) worn about the neck as an ornament', 'name': 'necklace'}, {'frequency': 'f', 'id': 729, 'synset': 'necktie.n.01', 'synonyms': ['necktie', 'tie_(necktie)'], 'def': 'neckwear consisting of a long narrow piece of material worn under a collar and tied in knot at the front', 'name': 'necktie'}, {'frequency': 'r', 'id': 730, 'synset': 'needle.n.03', 'synonyms': ['needle'], 'def': 'a sharp pointed implement (usually metal)', 'name': 'needle'}, {'frequency': 'c', 'id': 731, 'synset': 'nest.n.01', 'synonyms': ['nest'], 'def': 'a structure in which animals lay eggs or give birth to their young', 'name': 'nest'}, {'frequency': 'r', 'id': 732, 'synset': 'newsstand.n.01', 'synonyms': ['newsstand'], 'def': 'a stall where newspapers and other periodicals are sold', 'name': 'newsstand'}, {'frequency': 'c', 'id': 733, 'synset': 'nightwear.n.01', 'synonyms': ['nightshirt', 'nightwear', 'sleepwear', 'nightclothes'], 'def': 'garments designed to be worn in bed', 'name': 'nightshirt'}, {'frequency': 'r', 'id': 734, 'synset': 'nosebag.n.01', 'synonyms': ['nosebag_(for_animals)', 'feedbag'], 'def': 'a canvas bag that is used to feed an animal (such as a horse); covers the muzzle and fastens at the top of the head', 'name': 'nosebag_(for_animals)'}, {'frequency': 'r', 'id': 735, 'synset': 'noseband.n.01', 'synonyms': ['noseband_(for_animals)', 'nosepiece_(for_animals)'], 'def': "a strap that is the part of a bridle that goes over the animal's nose", 'name': 'noseband_(for_animals)'}, {'frequency': 'f', 'id': 736, 'synset': 'notebook.n.01', 'synonyms': ['notebook'], 'def': 'a book with blank pages for recording notes or memoranda', 'name': 'notebook'}, {'frequency': 'c', 'id': 737, 'synset': 'notepad.n.01', 'synonyms': ['notepad'], 'def': 'a pad of paper for keeping notes', 'name': 'notepad'}, {'frequency': 'c', 'id': 738, 'synset': 'nut.n.03', 'synonyms': ['nut'], 'def': 'a small metal block (usually square or hexagonal) with internal screw thread to be fitted onto a bolt', 'name': 'nut'}, {'frequency': 'r', 'id': 739, 'synset': 'nutcracker.n.01', 'synonyms': ['nutcracker'], 'def': 'a hand tool used to crack nuts open', 'name': 'nutcracker'}, {'frequency': 'c', 'id': 740, 'synset': 'oar.n.01', 'synonyms': ['oar'], 'def': 'an implement used to propel or steer a boat', 'name': 'oar'}, {'frequency': 'r', 'id': 741, 'synset': 'octopus.n.01', 'synonyms': ['octopus_(food)'], 'def': 'tentacles of octopus prepared as food', 'name': 'octopus_(food)'}, {'frequency': 'r', 'id': 742, 'synset': 'octopus.n.02', 'synonyms': ['octopus_(animal)'], 'def': 'bottom-living cephalopod having a soft oval body with eight long tentacles', 'name': 'octopus_(animal)'}, {'frequency': 'c', 'id': 743, 'synset': 'oil_lamp.n.01', 'synonyms': ['oil_lamp', 'kerosene_lamp', 'kerosine_lamp'], 'def': 'a lamp that burns oil (as kerosine) for light', 'name': 'oil_lamp'}, {'frequency': 'c', 'id': 744, 'synset': 'olive_oil.n.01', 'synonyms': ['olive_oil'], 'def': 'oil from olives', 'name': 'olive_oil'}, {'frequency': 'r', 'id': 745, 'synset': 'omelet.n.01', 'synonyms': ['omelet', 'omelette'], 'def': 'beaten eggs cooked until just set; may be folded around e.g. ham or cheese or jelly', 'name': 'omelet'}, {'frequency': 'f', 'id': 746, 'synset': 'onion.n.01', 'synonyms': ['onion'], 'def': 'the bulb of an onion plant', 'name': 'onion'}, {'frequency': 'f', 'id': 747, 'synset': 'orange.n.01', 'synonyms': ['orange_(fruit)'], 'def': 'orange (FRUIT of an orange tree)', 'name': 'orange_(fruit)'}, {'frequency': 'c', 'id': 748, 'synset': 'orange_juice.n.01', 'synonyms': ['orange_juice'], 'def': 'bottled or freshly squeezed juice of oranges', 'name': 'orange_juice'}, {'frequency': 'r', 'id': 749, 'synset': 'oregano.n.01', 'synonyms': ['oregano', 'marjoram'], 'def': 'aromatic Eurasian perennial herb used in cooking and baking', 'name': 'oregano'}, {'frequency': 'c', 'id': 750, 'synset': 'ostrich.n.02', 'synonyms': ['ostrich'], 'def': 'fast-running African flightless bird with two-toed feet; largest living bird', 'name': 'ostrich'}, {'frequency': 'c', 'id': 751, 'synset': 'ottoman.n.03', 'synonyms': ['ottoman', 'pouf', 'pouffe', 'hassock'], 'def': 'thick cushion used as a seat', 'name': 'ottoman'}, {'frequency': 'c', 'id': 752, 'synset': 'overall.n.01', 'synonyms': ['overalls_(clothing)'], 'def': 'work clothing consisting of denim trousers usually with a bib and shoulder straps', 'name': 'overalls_(clothing)'}, {'frequency': 'c', 'id': 753, 'synset': 'owl.n.01', 'synonyms': ['owl'], 'def': 'nocturnal bird of prey with hawk-like beak and claws and large head with front-facing eyes', 'name': 'owl'}, {'frequency': 'c', 'id': 754, 'synset': 'packet.n.03', 'synonyms': ['packet'], 'def': 'a small package or bundle', 'name': 'packet'}, {'frequency': 'r', 'id': 755, 'synset': 'pad.n.03', 'synonyms': ['inkpad', 'inking_pad', 'stamp_pad'], 'def': 'absorbent material saturated with ink used to transfer ink evenly to a rubber stamp', 'name': 'inkpad'}, {'frequency': 'c', 'id': 756, 'synset': 'pad.n.04', 'synonyms': ['pad'], 'def': 'a flat mass of soft material used for protection, stuffing, or comfort', 'name': 'pad'}, {'frequency': 'c', 'id': 757, 'synset': 'paddle.n.04', 'synonyms': ['paddle', 'boat_paddle'], 'def': 'a short light oar used without an oarlock to propel a canoe or small boat', 'name': 'paddle'}, {'frequency': 'c', 'id': 758, 'synset': 'padlock.n.01', 'synonyms': ['padlock'], 'def': 'a detachable, portable lock', 'name': 'padlock'}, {'frequency': 'r', 'id': 759, 'synset': 'paintbox.n.01', 'synonyms': ['paintbox'], 'def': "a box containing a collection of cubes or tubes of artists' paint", 'name': 'paintbox'}, {'frequency': 'c', 'id': 760, 'synset': 'paintbrush.n.01', 'synonyms': ['paintbrush'], 'def': 'a brush used as an applicator to apply paint', 'name': 'paintbrush'}, {'frequency': 'f', 'id': 761, 'synset': 'painting.n.01', 'synonyms': ['painting'], 'def': 'graphic art consisting of an artistic composition made by applying paints to a surface', 'name': 'painting'}, {'frequency': 'c', 'id': 762, 'synset': 'pajama.n.02', 'synonyms': ['pajamas', 'pyjamas'], 'def': 'loose-fitting nightclothes worn for sleeping or lounging', 'name': 'pajamas'}, {'frequency': 'c', 'id': 763, 'synset': 'palette.n.02', 'synonyms': ['palette', 'pallet'], 'def': 'board that provides a flat surface on which artists mix paints and the range of colors used', 'name': 'palette'}, {'frequency': 'f', 'id': 764, 'synset': 'pan.n.01', 'synonyms': ['pan_(for_cooking)', 'cooking_pan'], 'def': 'cooking utensil consisting of a wide metal vessel', 'name': 'pan_(for_cooking)'}, {'frequency': 'r', 'id': 765, 'synset': 'pan.n.03', 'synonyms': ['pan_(metal_container)'], 'def': 'shallow container made of metal', 'name': 'pan_(metal_container)'}, {'frequency': 'c', 'id': 766, 'synset': 'pancake.n.01', 'synonyms': ['pancake'], 'def': 'a flat cake of thin batter fried on both sides on a griddle', 'name': 'pancake'}, {'frequency': 'r', 'id': 767, 'synset': 'pantyhose.n.01', 'synonyms': ['pantyhose'], 'def': "a woman's tights consisting of underpants and stockings", 'name': 'pantyhose'}, {'frequency': 'r', 'id': 768, 'synset': 'papaya.n.02', 'synonyms': ['papaya'], 'def': 'large oval melon-like tropical fruit with yellowish flesh', 'name': 'papaya'}, {'frequency': 'r', 'id': 769, 'synset': 'paper_clip.n.01', 'synonyms': ['paperclip'], 'def': 'a wire or plastic clip for holding sheets of paper together', 'name': 'paperclip'}, {'frequency': 'f', 'id': 770, 'synset': 'paper_plate.n.01', 'synonyms': ['paper_plate'], 'def': 'a disposable plate made of cardboard', 'name': 'paper_plate'}, {'frequency': 'f', 'id': 771, 'synset': 'paper_towel.n.01', 'synonyms': ['paper_towel'], 'def': 'a disposable towel made of absorbent paper', 'name': 'paper_towel'}, {'frequency': 'r', 'id': 772, 'synset': 'paperback_book.n.01', 'synonyms': ['paperback_book', 'paper-back_book', 'softback_book', 'soft-cover_book'], 'def': 'a book with paper covers', 'name': 'paperback_book'}, {'frequency': 'r', 'id': 773, 'synset': 'paperweight.n.01', 'synonyms': ['paperweight'], 'def': 'a weight used to hold down a stack of papers', 'name': 'paperweight'}, {'frequency': 'c', 'id': 774, 'synset': 'parachute.n.01', 'synonyms': ['parachute'], 'def': 'rescue equipment consisting of a device that fills with air and retards your fall', 'name': 'parachute'}, {'frequency': 'r', 'id': 775, 'synset': 'parakeet.n.01', 'synonyms': ['parakeet', 'parrakeet', 'parroket', 'paraquet', 'paroquet', 'parroquet'], 'def': 'any of numerous small slender long-tailed parrots', 'name': 'parakeet'}, {'frequency': 'c', 'id': 776, 'synset': 'parasail.n.01', 'synonyms': ['parasail_(sports)'], 'def': 'parachute that will lift a person up into the air when it is towed by a motorboat or a car', 'name': 'parasail_(sports)'}, {'frequency': 'r', 'id': 777, 'synset': 'parchment.n.01', 'synonyms': ['parchment'], 'def': 'a superior paper resembling sheepskin', 'name': 'parchment'}, {'frequency': 'r', 'id': 778, 'synset': 'parka.n.01', 'synonyms': ['parka', 'anorak'], 'def': "a kind of heavy jacket (`windcheater' is a British term)", 'name': 'parka'}, {'frequency': 'f', 'id': 779, 'synset': 'parking_meter.n.01', 'synonyms': ['parking_meter'], 'def': 'a coin-operated timer located next to a parking space', 'name': 'parking_meter'}, {'frequency': 'c', 'id': 780, 'synset': 'parrot.n.01', 'synonyms': ['parrot'], 'def': 'usually brightly colored tropical birds with short hooked beaks and the ability to mimic sounds', 'name': 'parrot'}, {'frequency': 'c', 'id': 781, 'synset': 'passenger_car.n.01', 'synonyms': ['passenger_car_(part_of_a_train)', 'coach_(part_of_a_train)'], 'def': 'a railcar where passengers ride', 'name': 'passenger_car_(part_of_a_train)'}, {'frequency': 'r', 'id': 782, 'synset': 'passenger_ship.n.01', 'synonyms': ['passenger_ship'], 'def': 'a ship built to carry passengers', 'name': 'passenger_ship'}, {'frequency': 'r', 'id': 783, 'synset': 'passport.n.02', 'synonyms': ['passport'], 'def': 'a document issued by a country to a citizen allowing that person to travel abroad and re-enter the home country', 'name': 'passport'}, {'frequency': 'f', 'id': 784, 'synset': 'pastry.n.02', 'synonyms': ['pastry'], 'def': 'any of various baked foods made of dough or batter', 'name': 'pastry'}, {'frequency': 'r', 'id': 785, 'synset': 'patty.n.01', 'synonyms': ['patty_(food)'], 'def': 'small flat mass of chopped food', 'name': 'patty_(food)'}, {'frequency': 'c', 'id': 786, 'synset': 'pea.n.01', 'synonyms': ['pea_(food)'], 'def': 'seed of a pea plant used for food', 'name': 'pea_(food)'}, {'frequency': 'c', 'id': 787, 'synset': 'peach.n.03', 'synonyms': ['peach'], 'def': 'downy juicy fruit with sweet yellowish or whitish flesh', 'name': 'peach'}, {'frequency': 'c', 'id': 788, 'synset': 'peanut_butter.n.01', 'synonyms': ['peanut_butter'], 'def': 'a spread made from ground peanuts', 'name': 'peanut_butter'}, {'frequency': 'c', 'id': 789, 'synset': 'pear.n.01', 'synonyms': ['pear'], 'def': 'sweet juicy gritty-textured fruit available in many varieties', 'name': 'pear'}, {'frequency': 'r', 'id': 790, 'synset': 'peeler.n.03', 'synonyms': ['peeler_(tool_for_fruit_and_vegetables)'], 'def': 'a device for peeling vegetables or fruits', 'name': 'peeler_(tool_for_fruit_and_vegetables)'}, {'frequency': 'r', 'id': 791, 'synset': 'pegboard.n.01', 'synonyms': ['pegboard'], 'def': 'a board perforated with regularly spaced holes into which pegs can be fitted', 'name': 'pegboard'}, {'frequency': 'c', 'id': 792, 'synset': 'pelican.n.01', 'synonyms': ['pelican'], 'def': 'large long-winged warm-water seabird having a large bill with a distensible pouch for fish', 'name': 'pelican'}, {'frequency': 'f', 'id': 793, 'synset': 'pen.n.01', 'synonyms': ['pen'], 'def': 'a writing implement with a point from which ink flows', 'name': 'pen'}, {'frequency': 'c', 'id': 794, 'synset': 'pencil.n.01', 'synonyms': ['pencil'], 'def': 'a thin cylindrical pointed writing implement made of wood and graphite', 'name': 'pencil'}, {'frequency': 'r', 'id': 795, 'synset': 'pencil_box.n.01', 'synonyms': ['pencil_box', 'pencil_case'], 'def': 'a box for holding pencils', 'name': 'pencil_box'}, {'frequency': 'r', 'id': 796, 'synset': 'pencil_sharpener.n.01', 'synonyms': ['pencil_sharpener'], 'def': 'a rotary implement for sharpening the point on pencils', 'name': 'pencil_sharpener'}, {'frequency': 'r', 'id': 797, 'synset': 'pendulum.n.01', 'synonyms': ['pendulum'], 'def': 'an apparatus consisting of an object mounted so that it swings freely under the influence of gravity', 'name': 'pendulum'}, {'frequency': 'c', 'id': 798, 'synset': 'penguin.n.01', 'synonyms': ['penguin'], 'def': 'short-legged flightless birds of cold southern regions having webbed feet and wings modified as flippers', 'name': 'penguin'}, {'frequency': 'r', 'id': 799, 'synset': 'pennant.n.02', 'synonyms': ['pennant'], 'def': 'a flag longer than it is wide (and often tapering)', 'name': 'pennant'}, {'frequency': 'r', 'id': 800, 'synset': 'penny.n.02', 'synonyms': ['penny_(coin)'], 'def': 'a coin worth one-hundredth of the value of the basic unit', 'name': 'penny_(coin)'}, {'frequency': 'c', 'id': 801, 'synset': 'pepper.n.03', 'synonyms': ['pepper', 'peppercorn'], 'def': 'pungent seasoning from the berry of the common pepper plant; whole or ground', 'name': 'pepper'}, {'frequency': 'c', 'id': 802, 'synset': 'pepper_mill.n.01', 'synonyms': ['pepper_mill', 'pepper_grinder'], 'def': 'a mill for grinding pepper', 'name': 'pepper_mill'}, {'frequency': 'c', 'id': 803, 'synset': 'perfume.n.02', 'synonyms': ['perfume'], 'def': 'a toiletry that emits and diffuses a fragrant odor', 'name': 'perfume'}, {'frequency': 'r', 'id': 804, 'synset': 'persimmon.n.02', 'synonyms': ['persimmon'], 'def': 'orange fruit resembling a plum; edible when fully ripe', 'name': 'persimmon'}, {'frequency': 'f', 'id': 805, 'synset': 'person.n.01', 'synonyms': ['baby', 'child', 'boy', 'girl', 'man', 'woman', 'person', 'human'], 'def': 'a human being', 'name': 'baby'}, {'frequency': 'r', 'id': 806, 'synset': 'pet.n.01', 'synonyms': ['pet'], 'def': 'a domesticated animal kept for companionship or amusement', 'name': 'pet'}, {'frequency': 'r', 'id': 807, 'synset': 'petfood.n.01', 'synonyms': ['petfood', 'pet-food'], 'def': 'food prepared for animal pets', 'name': 'petfood'}, {'frequency': 'r', 'id': 808, 'synset': 'pew.n.01', 'synonyms': ['pew_(church_bench)', 'church_bench'], 'def': 'long bench with backs; used in church by the congregation', 'name': 'pew_(church_bench)'}, {'frequency': 'r', 'id': 809, 'synset': 'phonebook.n.01', 'synonyms': ['phonebook', 'telephone_book', 'telephone_directory'], 'def': 'a directory containing an alphabetical list of telephone subscribers and their telephone numbers', 'name': 'phonebook'}, {'frequency': 'c', 'id': 810, 'synset': 'phonograph_record.n.01', 'synonyms': ['phonograph_record', 'phonograph_recording', 'record_(phonograph_recording)'], 'def': 'sound recording consisting of a typically black disk with a continuous groove', 'name': 'phonograph_record'}, {'frequency': 'c', 'id': 811, 'synset': 'piano.n.01', 'synonyms': ['piano'], 'def': 'a keyboard instrument that is played by depressing keys that cause hammers to strike tuned strings and produce sounds', 'name': 'piano'}, {'frequency': 'f', 'id': 812, 'synset': 'pickle.n.01', 'synonyms': ['pickle'], 'def': 'vegetables (especially cucumbers) preserved in brine or vinegar', 'name': 'pickle'}, {'frequency': 'f', 'id': 813, 'synset': 'pickup.n.01', 'synonyms': ['pickup_truck'], 'def': 'a light truck with an open body and low sides and a tailboard', 'name': 'pickup_truck'}, {'frequency': 'c', 'id': 814, 'synset': 'pie.n.01', 'synonyms': ['pie'], 'def': 'dish baked in pastry-lined pan often with a pastry top', 'name': 'pie'}, {'frequency': 'c', 'id': 815, 'synset': 'pigeon.n.01', 'synonyms': ['pigeon'], 'def': 'wild and domesticated birds having a heavy body and short legs', 'name': 'pigeon'}, {'frequency': 'r', 'id': 816, 'synset': 'piggy_bank.n.01', 'synonyms': ['piggy_bank', 'penny_bank'], 'def': "a child's coin bank (often shaped like a pig)", 'name': 'piggy_bank'}, {'frequency': 'f', 'id': 817, 'synset': 'pillow.n.01', 'synonyms': ['pillow'], 'def': 'a cushion to support the head of a sleeping person', 'name': 'pillow'}, {'frequency': 'r', 'id': 818, 'synset': 'pin.n.09', 'synonyms': ['pin_(non_jewelry)'], 'def': 'a small slender (often pointed) piece of wood or metal used to support or fasten or attach things', 'name': 'pin_(non_jewelry)'}, {'frequency': 'f', 'id': 819, 'synset': 'pineapple.n.02', 'synonyms': ['pineapple'], 'def': 'large sweet fleshy tropical fruit with a tuft of stiff leaves', 'name': 'pineapple'}, {'frequency': 'c', 'id': 820, 'synset': 'pinecone.n.01', 'synonyms': ['pinecone'], 'def': 'the seed-producing cone of a pine tree', 'name': 'pinecone'}, {'frequency': 'r', 'id': 821, 'synset': 'ping-pong_ball.n.01', 'synonyms': ['ping-pong_ball'], 'def': 'light hollow ball used in playing table tennis', 'name': 'ping-pong_ball'}, {'frequency': 'r', 'id': 822, 'synset': 'pinwheel.n.03', 'synonyms': ['pinwheel'], 'def': 'a toy consisting of vanes of colored paper or plastic that is pinned to a stick and spins when it is pointed into the wind', 'name': 'pinwheel'}, {'frequency': 'r', 'id': 823, 'synset': 'pipe.n.01', 'synonyms': ['tobacco_pipe'], 'def': 'a tube with a small bowl at one end; used for smoking tobacco', 'name': 'tobacco_pipe'}, {'frequency': 'f', 'id': 824, 'synset': 'pipe.n.02', 'synonyms': ['pipe', 'piping'], 'def': 'a long tube made of metal or plastic that is used to carry water or oil or gas etc.', 'name': 'pipe'}, {'frequency': 'r', 'id': 825, 'synset': 'pistol.n.01', 'synonyms': ['pistol', 'handgun'], 'def': 'a firearm that is held and fired with one hand', 'name': 'pistol'}, {'frequency': 'r', 'id': 826, 'synset': 'pita.n.01', 'synonyms': ['pita_(bread)', 'pocket_bread'], 'def': 'usually small round bread that can open into a pocket for filling', 'name': 'pita_(bread)'}, {'frequency': 'f', 'id': 827, 'synset': 'pitcher.n.02', 'synonyms': ['pitcher_(vessel_for_liquid)', 'ewer'], 'def': 'an open vessel with a handle and a spout for pouring', 'name': 'pitcher_(vessel_for_liquid)'}, {'frequency': 'r', 'id': 828, 'synset': 'pitchfork.n.01', 'synonyms': ['pitchfork'], 'def': 'a long-handled hand tool with sharp widely spaced prongs for lifting and pitching hay', 'name': 'pitchfork'}, {'frequency': 'f', 'id': 829, 'synset': 'pizza.n.01', 'synonyms': ['pizza'], 'def': 'Italian open pie made of thin bread dough spread with a spiced mixture of e.g. tomato sauce and cheese', 'name': 'pizza'}, {'frequency': 'f', 'id': 830, 'synset': 'place_mat.n.01', 'synonyms': ['place_mat'], 'def': 'a mat placed on a table for an individual place setting', 'name': 'place_mat'}, {'frequency': 'f', 'id': 831, 'synset': 'plate.n.04', 'synonyms': ['plate'], 'def': 'dish on which food is served or from which food is eaten', 'name': 'plate'}, {'frequency': 'c', 'id': 832, 'synset': 'platter.n.01', 'synonyms': ['platter'], 'def': 'a large shallow dish used for serving food', 'name': 'platter'}, {'frequency': 'r', 'id': 833, 'synset': 'playing_card.n.01', 'synonyms': ['playing_card'], 'def': 'one of a pack of cards that are used to play card games', 'name': 'playing_card'}, {'frequency': 'r', 'id': 834, 'synset': 'playpen.n.01', 'synonyms': ['playpen'], 'def': 'a portable enclosure in which babies may be left to play', 'name': 'playpen'}, {'frequency': 'c', 'id': 835, 'synset': 'pliers.n.01', 'synonyms': ['pliers', 'plyers'], 'def': 'a gripping hand tool with two hinged arms and (usually) serrated jaws', 'name': 'pliers'}, {'frequency': 'r', 'id': 836, 'synset': 'plow.n.01', 'synonyms': ['plow_(farm_equipment)', 'plough_(farm_equipment)'], 'def': 'a farm tool having one or more heavy blades to break the soil and cut a furrow prior to sowing', 'name': 'plow_(farm_equipment)'}, {'frequency': 'r', 'id': 837, 'synset': 'pocket_watch.n.01', 'synonyms': ['pocket_watch'], 'def': 'a watch that is carried in a small watch pocket', 'name': 'pocket_watch'}, {'frequency': 'c', 'id': 838, 'synset': 'pocketknife.n.01', 'synonyms': ['pocketknife'], 'def': 'a knife with a blade that folds into the handle; suitable for carrying in the pocket', 'name': 'pocketknife'}, {'frequency': 'c', 'id': 839, 'synset': 'poker.n.01', 'synonyms': ['poker_(fire_stirring_tool)', 'stove_poker', 'fire_hook'], 'def': 'fire iron consisting of a metal rod with a handle; used to stir a fire', 'name': 'poker_(fire_stirring_tool)'}, {'frequency': 'f', 'id': 840, 'synset': 'pole.n.01', 'synonyms': ['pole', 'post'], 'def': 'a long (usually round) rod of wood or metal or plastic', 'name': 'pole'}, {'frequency': 'r', 'id': 841, 'synset': 'police_van.n.01', 'synonyms': ['police_van', 'police_wagon', 'paddy_wagon', 'patrol_wagon'], 'def': 'van used by police to transport prisoners', 'name': 'police_van'}, {'frequency': 'f', 'id': 842, 'synset': 'polo_shirt.n.01', 'synonyms': ['polo_shirt', 'sport_shirt'], 'def': 'a shirt with short sleeves designed for comfort and casual wear', 'name': 'polo_shirt'}, {'frequency': 'r', 'id': 843, 'synset': 'poncho.n.01', 'synonyms': ['poncho'], 'def': 'a blanket-like cloak with a hole in the center for the head', 'name': 'poncho'}, {'frequency': 'c', 'id': 844, 'synset': 'pony.n.05', 'synonyms': ['pony'], 'def': 'any of various breeds of small gentle horses usually less than five feet high at the shoulder', 'name': 'pony'}, {'frequency': 'r', 'id': 845, 'synset': 'pool_table.n.01', 'synonyms': ['pool_table', 'billiard_table', 'snooker_table'], 'def': 'game equipment consisting of a heavy table on which pool is played', 'name': 'pool_table'}, {'frequency': 'f', 'id': 846, 'synset': 'pop.n.02', 'synonyms': ['pop_(soda)', 'soda_(pop)', 'tonic', 'soft_drink'], 'def': 'a sweet drink containing carbonated water and flavoring', 'name': 'pop_(soda)'}, {'frequency': 'r', 'id': 847, 'synset': 'portrait.n.02', 'synonyms': ['portrait', 'portrayal'], 'def': 'any likeness of a person, in any medium', 'name': 'portrait'}, {'frequency': 'c', 'id': 848, 'synset': 'postbox.n.01', 'synonyms': ['postbox_(public)', 'mailbox_(public)'], 'def': 'public box for deposit of mail', 'name': 'postbox_(public)'}, {'frequency': 'c', 'id': 849, 'synset': 'postcard.n.01', 'synonyms': ['postcard', 'postal_card', 'mailing-card'], 'def': 'a card for sending messages by post without an envelope', 'name': 'postcard'}, {'frequency': 'f', 'id': 850, 'synset': 'poster.n.01', 'synonyms': ['poster', 'placard'], 'def': 'a sign posted in a public place as an advertisement', 'name': 'poster'}, {'frequency': 'f', 'id': 851, 'synset': 'pot.n.01', 'synonyms': ['pot'], 'def': 'metal or earthenware cooking vessel that is usually round and deep; often has a handle and lid', 'name': 'pot'}, {'frequency': 'f', 'id': 852, 'synset': 'pot.n.04', 'synonyms': ['flowerpot'], 'def': 'a container in which plants are cultivated', 'name': 'flowerpot'}, {'frequency': 'f', 'id': 853, 'synset': 'potato.n.01', 'synonyms': ['potato'], 'def': 'an edible tuber native to South America', 'name': 'potato'}, {'frequency': 'c', 'id': 854, 'synset': 'potholder.n.01', 'synonyms': ['potholder'], 'def': 'an insulated pad for holding hot pots', 'name': 'potholder'}, {'frequency': 'c', 'id': 855, 'synset': 'pottery.n.01', 'synonyms': ['pottery', 'clayware'], 'def': 'ceramic ware made from clay and baked in a kiln', 'name': 'pottery'}, {'frequency': 'c', 'id': 856, 'synset': 'pouch.n.01', 'synonyms': ['pouch'], 'def': 'a small or medium size container for holding or carrying things', 'name': 'pouch'}, {'frequency': 'r', 'id': 857, 'synset': 'power_shovel.n.01', 'synonyms': ['power_shovel', 'excavator', 'digger'], 'def': 'a machine for excavating', 'name': 'power_shovel'}, {'frequency': 'c', 'id': 858, 'synset': 'prawn.n.01', 'synonyms': ['prawn', 'shrimp'], 'def': 'any of various edible decapod crustaceans', 'name': 'prawn'}, {'frequency': 'f', 'id': 859, 'synset': 'printer.n.03', 'synonyms': ['printer', 'printing_machine'], 'def': 'a machine that prints', 'name': 'printer'}, {'frequency': 'c', 'id': 860, 'synset': 'projectile.n.01', 'synonyms': ['projectile_(weapon)', 'missile'], 'def': 'a weapon that is forcibly thrown or projected at a targets', 'name': 'projectile_(weapon)'}, {'frequency': 'c', 'id': 861, 'synset': 'projector.n.02', 'synonyms': ['projector'], 'def': 'an optical instrument that projects an enlarged image onto a screen', 'name': 'projector'}, {'frequency': 'f', 'id': 862, 'synset': 'propeller.n.01', 'synonyms': ['propeller', 'propellor'], 'def': 'a mechanical device that rotates to push against air or water', 'name': 'propeller'}, {'frequency': 'r', 'id': 863, 'synset': 'prune.n.01', 'synonyms': ['prune'], 'def': 'dried plum', 'name': 'prune'}, {'frequency': 'r', 'id': 864, 'synset': 'pudding.n.01', 'synonyms': ['pudding'], 'def': 'any of various soft thick unsweetened baked dishes', 'name': 'pudding'}, {'frequency': 'r', 'id': 865, 'synset': 'puffer.n.02', 'synonyms': ['puffer_(fish)', 'pufferfish', 'blowfish', 'globefish'], 'def': 'fishes whose elongated spiny body can inflate itself with water or air to form a globe', 'name': 'puffer_(fish)'}, {'frequency': 'r', 'id': 866, 'synset': 'puffin.n.01', 'synonyms': ['puffin'], 'def': 'seabirds having short necks and brightly colored compressed bills', 'name': 'puffin'}, {'frequency': 'r', 'id': 867, 'synset': 'pug.n.01', 'synonyms': ['pug-dog'], 'def': 'small compact smooth-coated breed of Asiatic origin having a tightly curled tail and broad flat wrinkled muzzle', 'name': 'pug-dog'}, {'frequency': 'c', 'id': 868, 'synset': 'pumpkin.n.02', 'synonyms': ['pumpkin'], 'def': 'usually large pulpy deep-yellow round fruit of the squash family maturing in late summer or early autumn', 'name': 'pumpkin'}, {'frequency': 'r', 'id': 869, 'synset': 'punch.n.03', 'synonyms': ['puncher'], 'def': 'a tool for making holes or indentations', 'name': 'puncher'}, {'frequency': 'r', 'id': 870, 'synset': 'puppet.n.01', 'synonyms': ['puppet', 'marionette'], 'def': 'a small figure of a person operated from above with strings by a puppeteer', 'name': 'puppet'}, {'frequency': 'r', 'id': 871, 'synset': 'puppy.n.01', 'synonyms': ['puppy'], 'def': 'a young dog', 'name': 'puppy'}, {'frequency': 'r', 'id': 872, 'synset': 'quesadilla.n.01', 'synonyms': ['quesadilla'], 'def': 'a tortilla that is filled with cheese and heated', 'name': 'quesadilla'}, {'frequency': 'r', 'id': 873, 'synset': 'quiche.n.02', 'synonyms': ['quiche'], 'def': 'a tart filled with rich unsweetened custard; often contains other ingredients (as cheese or ham or seafood or vegetables)', 'name': 'quiche'}, {'frequency': 'f', 'id': 874, 'synset': 'quilt.n.01', 'synonyms': ['quilt', 'comforter'], 'def': 'bedding made of two layers of cloth filled with stuffing and stitched together', 'name': 'quilt'}, {'frequency': 'c', 'id': 875, 'synset': 'rabbit.n.01', 'synonyms': ['rabbit'], 'def': 'any of various burrowing animals of the family Leporidae having long ears and short tails', 'name': 'rabbit'}, {'frequency': 'r', 'id': 876, 'synset': 'racer.n.02', 'synonyms': ['race_car', 'racing_car'], 'def': 'a fast car that competes in races', 'name': 'race_car'}, {'frequency': 'c', 'id': 877, 'synset': 'racket.n.04', 'synonyms': ['racket', 'racquet'], 'def': 'a sports implement used to strike a ball in various games', 'name': 'racket'}, {'frequency': 'r', 'id': 878, 'synset': 'radar.n.01', 'synonyms': ['radar'], 'def': 'measuring instrument in which the echo of a pulse of microwave radiation is used to detect and locate distant objects', 'name': 'radar'}, {'frequency': 'c', 'id': 879, 'synset': 'radiator.n.03', 'synonyms': ['radiator'], 'def': 'a mechanism consisting of a metal honeycomb through which hot fluids circulate', 'name': 'radiator'}, {'frequency': 'c', 'id': 880, 'synset': 'radio_receiver.n.01', 'synonyms': ['radio_receiver', 'radio_set', 'radio', 'tuner_(radio)'], 'def': 'an electronic receiver that detects and demodulates and amplifies transmitted radio signals', 'name': 'radio_receiver'}, {'frequency': 'c', 'id': 881, 'synset': 'radish.n.03', 'synonyms': ['radish', 'daikon'], 'def': 'pungent edible root of any of various cultivated radish plants', 'name': 'radish'}, {'frequency': 'c', 'id': 882, 'synset': 'raft.n.01', 'synonyms': ['raft'], 'def': 'a flat float (usually made of logs or planks) that can be used for transport or as a platform for swimmers', 'name': 'raft'}, {'frequency': 'r', 'id': 883, 'synset': 'rag_doll.n.01', 'synonyms': ['rag_doll'], 'def': 'a cloth doll that is stuffed and (usually) painted', 'name': 'rag_doll'}, {'frequency': 'c', 'id': 884, 'synset': 'raincoat.n.01', 'synonyms': ['raincoat', 'waterproof_jacket'], 'def': 'a water-resistant coat', 'name': 'raincoat'}, {'frequency': 'c', 'id': 885, 'synset': 'ram.n.05', 'synonyms': ['ram_(animal)'], 'def': 'uncastrated adult male sheep', 'name': 'ram_(animal)'}, {'frequency': 'c', 'id': 886, 'synset': 'raspberry.n.02', 'synonyms': ['raspberry'], 'def': 'red or black edible aggregate berries usually smaller than the related blackberries', 'name': 'raspberry'}, {'frequency': 'r', 'id': 887, 'synset': 'rat.n.01', 'synonyms': ['rat'], 'def': 'any of various long-tailed rodents similar to but larger than a mouse', 'name': 'rat'}, {'frequency': 'c', 'id': 888, 'synset': 'razorblade.n.01', 'synonyms': ['razorblade'], 'def': 'a blade that has very sharp edge', 'name': 'razorblade'}, {'frequency': 'c', 'id': 889, 'synset': 'reamer.n.01', 'synonyms': ['reamer_(juicer)', 'juicer', 'juice_reamer'], 'def': 'a squeezer with a conical ridged center that is used for squeezing juice from citrus fruit', 'name': 'reamer_(juicer)'}, {'frequency': 'f', 'id': 890, 'synset': 'rearview_mirror.n.01', 'synonyms': ['rearview_mirror'], 'def': 'car mirror that reflects the view out of the rear window', 'name': 'rearview_mirror'}, {'frequency': 'c', 'id': 891, 'synset': 'receipt.n.02', 'synonyms': ['receipt'], 'def': 'an acknowledgment (usually tangible) that payment has been made', 'name': 'receipt'}, {'frequency': 'c', 'id': 892, 'synset': 'recliner.n.01', 'synonyms': ['recliner', 'reclining_chair', 'lounger_(chair)'], 'def': 'an armchair whose back can be lowered and foot can be raised to allow the sitter to recline in it', 'name': 'recliner'}, {'frequency': 'r', 'id': 893, 'synset': 'record_player.n.01', 'synonyms': ['record_player', 'phonograph_(record_player)', 'turntable'], 'def': 'machine in which rotating records cause a stylus to vibrate and the vibrations are amplified acoustically or electronically', 'name': 'record_player'}, {'frequency': 'r', 'id': 894, 'synset': 'red_cabbage.n.02', 'synonyms': ['red_cabbage'], 'def': 'compact head of purplish-red leaves', 'name': 'red_cabbage'}, {'frequency': 'f', 'id': 895, 'synset': 'reflector.n.01', 'synonyms': ['reflector'], 'def': 'device that reflects light, radiation, etc.', 'name': 'reflector'}, {'frequency': 'f', 'id': 896, 'synset': 'remote_control.n.01', 'synonyms': ['remote_control'], 'def': 'a device that can be used to control a machine or apparatus from a distance', 'name': 'remote_control'}, {'frequency': 'c', 'id': 897, 'synset': 'rhinoceros.n.01', 'synonyms': ['rhinoceros'], 'def': 'massive powerful herbivorous odd-toed ungulate of southeast Asia and Africa having very thick skin and one or two horns on the snout', 'name': 'rhinoceros'}, {'frequency': 'r', 'id': 898, 'synset': 'rib.n.03', 'synonyms': ['rib_(food)'], 'def': 'cut of meat including one or more ribs', 'name': 'rib_(food)'}, {'frequency': 'r', 'id': 899, 'synset': 'rifle.n.01', 'synonyms': ['rifle'], 'def': 'a shoulder firearm with a long barrel', 'name': 'rifle'}, {'frequency': 'f', 'id': 900, 'synset': 'ring.n.08', 'synonyms': ['ring'], 'def': 'jewelry consisting of a circlet of precious metal (often set with jewels) worn on the finger', 'name': 'ring'}, {'frequency': 'r', 'id': 901, 'synset': 'river_boat.n.01', 'synonyms': ['river_boat'], 'def': 'a boat used on rivers or to ply a river', 'name': 'river_boat'}, {'frequency': 'r', 'id': 902, 'synset': 'road_map.n.02', 'synonyms': ['road_map'], 'def': '(NOT A ROAD) a MAP showing roads (for automobile travel)', 'name': 'road_map'}, {'frequency': 'c', 'id': 903, 'synset': 'robe.n.01', 'synonyms': ['robe'], 'def': 'any loose flowing garment', 'name': 'robe'}, {'frequency': 'c', 'id': 904, 'synset': 'rocking_chair.n.01', 'synonyms': ['rocking_chair'], 'def': 'a chair mounted on rockers', 'name': 'rocking_chair'}, {'frequency': 'r', 'id': 905, 'synset': 'roller_skate.n.01', 'synonyms': ['roller_skate'], 'def': 'a shoe with pairs of rollers (small hard wheels) fixed to the sole', 'name': 'roller_skate'}, {'frequency': 'r', 'id': 906, 'synset': 'rollerblade.n.01', 'synonyms': ['Rollerblade'], 'def': 'an in-line variant of a roller skate', 'name': 'Rollerblade'}, {'frequency': 'c', 'id': 907, 'synset': 'rolling_pin.n.01', 'synonyms': ['rolling_pin'], 'def': 'utensil consisting of a cylinder (usually of wood) with a handle at each end; used to roll out dough', 'name': 'rolling_pin'}, {'frequency': 'r', 'id': 908, 'synset': 'root_beer.n.01', 'synonyms': ['root_beer'], 'def': 'carbonated drink containing extracts of roots and herbs', 'name': 'root_beer'}, {'frequency': 'c', 'id': 909, 'synset': 'router.n.02', 'synonyms': ['router_(computer_equipment)'], 'def': 'a device that forwards data packets between computer networks', 'name': 'router_(computer_equipment)'}, {'frequency': 'f', 'id': 910, 'synset': 'rubber_band.n.01', 'synonyms': ['rubber_band', 'elastic_band'], 'def': 'a narrow band of elastic rubber used to hold things (such as papers) together', 'name': 'rubber_band'}, {'frequency': 'c', 'id': 911, 'synset': 'runner.n.08', 'synonyms': ['runner_(carpet)'], 'def': 'a long narrow carpet', 'name': 'runner_(carpet)'}, {'frequency': 'f', 'id': 912, 'synset': 'sack.n.01', 'synonyms': ['plastic_bag', 'paper_bag'], 'def': "a bag made of paper or plastic for holding customer's purchases", 'name': 'plastic_bag'}, {'frequency': 'f', 'id': 913, 'synset': 'saddle.n.01', 'synonyms': ['saddle_(on_an_animal)'], 'def': 'a seat for the rider of a horse or camel', 'name': 'saddle_(on_an_animal)'}, {'frequency': 'f', 'id': 914, 'synset': 'saddle_blanket.n.01', 'synonyms': ['saddle_blanket', 'saddlecloth', 'horse_blanket'], 'def': 'stable gear consisting of a blanket placed under the saddle', 'name': 'saddle_blanket'}, {'frequency': 'c', 'id': 915, 'synset': 'saddlebag.n.01', 'synonyms': ['saddlebag'], 'def': 'a large bag (or pair of bags) hung over a saddle', 'name': 'saddlebag'}, {'frequency': 'r', 'id': 916, 'synset': 'safety_pin.n.01', 'synonyms': ['safety_pin'], 'def': 'a pin in the form of a clasp; has a guard so the point of the pin will not stick the user', 'name': 'safety_pin'}, {'frequency': 'c', 'id': 917, 'synset': 'sail.n.01', 'synonyms': ['sail'], 'def': 'a large piece of fabric by means of which wind is used to propel a sailing vessel', 'name': 'sail'}, {'frequency': 'c', 'id': 918, 'synset': 'salad.n.01', 'synonyms': ['salad'], 'def': 'food mixtures either arranged on a plate or tossed and served with a moist dressing; usually consisting of or including greens', 'name': 'salad'}, {'frequency': 'r', 'id': 919, 'synset': 'salad_plate.n.01', 'synonyms': ['salad_plate', 'salad_bowl'], 'def': 'a plate or bowl for individual servings of salad', 'name': 'salad_plate'}, {'frequency': 'r', 'id': 920, 'synset': 'salami.n.01', 'synonyms': ['salami'], 'def': 'highly seasoned fatty sausage of pork and beef usually dried', 'name': 'salami'}, {'frequency': 'r', 'id': 921, 'synset': 'salmon.n.01', 'synonyms': ['salmon_(fish)'], 'def': 'any of various large food and game fishes of northern waters', 'name': 'salmon_(fish)'}, {'frequency': 'r', 'id': 922, 'synset': 'salmon.n.03', 'synonyms': ['salmon_(food)'], 'def': 'flesh of any of various marine or freshwater fish of the family Salmonidae', 'name': 'salmon_(food)'}, {'frequency': 'r', 'id': 923, 'synset': 'salsa.n.01', 'synonyms': ['salsa'], 'def': 'spicy sauce of tomatoes and onions and chili peppers to accompany Mexican foods', 'name': 'salsa'}, {'frequency': 'f', 'id': 924, 'synset': 'saltshaker.n.01', 'synonyms': ['saltshaker'], 'def': 'a shaker with a perforated top for sprinkling salt', 'name': 'saltshaker'}, {'frequency': 'f', 'id': 925, 'synset': 'sandal.n.01', 'synonyms': ['sandal_(type_of_shoe)'], 'def': 'a shoe consisting of a sole fastened by straps to the foot', 'name': 'sandal_(type_of_shoe)'}, {'frequency': 'f', 'id': 926, 'synset': 'sandwich.n.01', 'synonyms': ['sandwich'], 'def': 'two (or more) slices of bread with a filling between them', 'name': 'sandwich'}, {'frequency': 'r', 'id': 927, 'synset': 'satchel.n.01', 'synonyms': ['satchel'], 'def': 'luggage consisting of a small case with a flat bottom and (usually) a shoulder strap', 'name': 'satchel'}, {'frequency': 'r', 'id': 928, 'synset': 'saucepan.n.01', 'synonyms': ['saucepan'], 'def': 'a deep pan with a handle; used for stewing or boiling', 'name': 'saucepan'}, {'frequency': 'f', 'id': 929, 'synset': 'saucer.n.02', 'synonyms': ['saucer'], 'def': 'a small shallow dish for holding a cup at the table', 'name': 'saucer'}, {'frequency': 'f', 'id': 930, 'synset': 'sausage.n.01', 'synonyms': ['sausage'], 'def': 'highly seasoned minced meat stuffed in casings', 'name': 'sausage'}, {'frequency': 'r', 'id': 931, 'synset': 'sawhorse.n.01', 'synonyms': ['sawhorse', 'sawbuck'], 'def': 'a framework for holding wood that is being sawed', 'name': 'sawhorse'}, {'frequency': 'r', 'id': 932, 'synset': 'sax.n.02', 'synonyms': ['saxophone'], 'def': "a wind instrument with a `J'-shaped form typically made of brass", 'name': 'saxophone'}, {'frequency': 'f', 'id': 933, 'synset': 'scale.n.07', 'synonyms': ['scale_(measuring_instrument)'], 'def': 'a measuring instrument for weighing; shows amount of mass', 'name': 'scale_(measuring_instrument)'}, {'frequency': 'r', 'id': 934, 'synset': 'scarecrow.n.01', 'synonyms': ['scarecrow', 'strawman'], 'def': 'an effigy in the shape of a man to frighten birds away from seeds', 'name': 'scarecrow'}, {'frequency': 'f', 'id': 935, 'synset': 'scarf.n.01', 'synonyms': ['scarf'], 'def': 'a garment worn around the head or neck or shoulders for warmth or decoration', 'name': 'scarf'}, {'frequency': 'c', 'id': 936, 'synset': 'school_bus.n.01', 'synonyms': ['school_bus'], 'def': 'a bus used to transport children to or from school', 'name': 'school_bus'}, {'frequency': 'f', 'id': 937, 'synset': 'scissors.n.01', 'synonyms': ['scissors'], 'def': 'a tool having two crossed pivoting blades with looped handles', 'name': 'scissors'}, {'frequency': 'c', 'id': 938, 'synset': 'scoreboard.n.01', 'synonyms': ['scoreboard'], 'def': 'a large board for displaying the score of a contest (and some other information)', 'name': 'scoreboard'}, {'frequency': 'c', 'id': 939, 'synset': 'scrambled_eggs.n.01', 'synonyms': ['scrambled_eggs'], 'def': 'eggs beaten and cooked to a soft firm consistency while stirring', 'name': 'scrambled_eggs'}, {'frequency': 'r', 'id': 940, 'synset': 'scraper.n.01', 'synonyms': ['scraper'], 'def': 'any of various hand tools for scraping', 'name': 'scraper'}, {'frequency': 'r', 'id': 941, 'synset': 'scratcher.n.03', 'synonyms': ['scratcher'], 'def': 'a device used for scratching', 'name': 'scratcher'}, {'frequency': 'c', 'id': 942, 'synset': 'screwdriver.n.01', 'synonyms': ['screwdriver'], 'def': 'a hand tool for driving screws; has a tip that fits into the head of a screw', 'name': 'screwdriver'}, {'frequency': 'c', 'id': 943, 'synset': 'scrub_brush.n.01', 'synonyms': ['scrubbing_brush'], 'def': 'a brush with short stiff bristles for heavy cleaning', 'name': 'scrubbing_brush'}, {'frequency': 'c', 'id': 944, 'synset': 'sculpture.n.01', 'synonyms': ['sculpture'], 'def': 'a three-dimensional work of art', 'name': 'sculpture'}, {'frequency': 'r', 'id': 945, 'synset': 'seabird.n.01', 'synonyms': ['seabird', 'seafowl'], 'def': 'a bird that frequents coastal waters and the open ocean: gulls; pelicans; gannets; cormorants; albatrosses; petrels; etc.', 'name': 'seabird'}, {'frequency': 'r', 'id': 946, 'synset': 'seahorse.n.02', 'synonyms': ['seahorse'], 'def': 'small fish with horse-like heads bent sharply downward and curled tails', 'name': 'seahorse'}, {'frequency': 'r', 'id': 947, 'synset': 'seaplane.n.01', 'synonyms': ['seaplane', 'hydroplane'], 'def': 'an airplane that can land on or take off from water', 'name': 'seaplane'}, {'frequency': 'c', 'id': 948, 'synset': 'seashell.n.01', 'synonyms': ['seashell'], 'def': 'the shell of a marine organism', 'name': 'seashell'}, {'frequency': 'r', 'id': 949, 'synset': 'seedling.n.01', 'synonyms': ['seedling'], 'def': 'young plant or tree grown from a seed', 'name': 'seedling'}, {'frequency': 'c', 'id': 950, 'synset': 'serving_dish.n.01', 'synonyms': ['serving_dish'], 'def': 'a dish used for serving food', 'name': 'serving_dish'}, {'frequency': 'r', 'id': 951, 'synset': 'sewing_machine.n.01', 'synonyms': ['sewing_machine'], 'def': 'a textile machine used as a home appliance for sewing', 'name': 'sewing_machine'}, {'frequency': 'r', 'id': 952, 'synset': 'shaker.n.03', 'synonyms': ['shaker'], 'def': 'a container in which something can be shaken', 'name': 'shaker'}, {'frequency': 'c', 'id': 953, 'synset': 'shampoo.n.01', 'synonyms': ['shampoo'], 'def': 'cleansing agent consisting of soaps or detergents used for washing the hair', 'name': 'shampoo'}, {'frequency': 'r', 'id': 954, 'synset': 'shark.n.01', 'synonyms': ['shark'], 'def': 'typically large carnivorous fishes with sharpe teeth', 'name': 'shark'}, {'frequency': 'r', 'id': 955, 'synset': 'sharpener.n.01', 'synonyms': ['sharpener'], 'def': 'any implement that is used to make something (an edge or a point) sharper', 'name': 'sharpener'}, {'frequency': 'r', 'id': 956, 'synset': 'sharpie.n.03', 'synonyms': ['Sharpie'], 'def': 'a pen with indelible ink that will write on any surface', 'name': 'Sharpie'}, {'frequency': 'r', 'id': 957, 'synset': 'shaver.n.03', 'synonyms': ['shaver_(electric)', 'electric_shaver', 'electric_razor'], 'def': 'a razor powered by an electric motor', 'name': 'shaver_(electric)'}, {'frequency': 'c', 'id': 958, 'synset': 'shaving_cream.n.01', 'synonyms': ['shaving_cream', 'shaving_soap'], 'def': 'toiletry consisting that forms a rich lather for softening the beard before shaving', 'name': 'shaving_cream'}, {'frequency': 'r', 'id': 959, 'synset': 'shawl.n.01', 'synonyms': ['shawl'], 'def': 'cloak consisting of an oblong piece of cloth used to cover the head and shoulders', 'name': 'shawl'}, {'frequency': 'r', 'id': 960, 'synset': 'shears.n.01', 'synonyms': ['shears'], 'def': 'large scissors with strong blades', 'name': 'shears'}, {'frequency': 'f', 'id': 961, 'synset': 'sheep.n.01', 'synonyms': ['sheep'], 'def': 'woolly usually horned ruminant mammal related to the goat', 'name': 'sheep'}, {'frequency': 'r', 'id': 962, 'synset': 'shepherd_dog.n.01', 'synonyms': ['shepherd_dog', 'sheepdog'], 'def': 'any of various usually long-haired breeds of dog reared to herd and guard sheep', 'name': 'shepherd_dog'}, {'frequency': 'r', 'id': 963, 'synset': 'sherbert.n.01', 'synonyms': ['sherbert', 'sherbet'], 'def': 'a frozen dessert made primarily of fruit juice and sugar', 'name': 'sherbert'}, {'frequency': 'r', 'id': 964, 'synset': 'shield.n.02', 'synonyms': ['shield'], 'def': 'armor carried on the arm to intercept blows', 'name': 'shield'}, {'frequency': 'f', 'id': 965, 'synset': 'shirt.n.01', 'synonyms': ['shirt'], 'def': 'a garment worn on the upper half of the body', 'name': 'shirt'}, {'frequency': 'f', 'id': 966, 'synset': 'shoe.n.01', 'synonyms': ['shoe', 'sneaker_(type_of_shoe)', 'tennis_shoe'], 'def': 'common footwear covering the foot', 'name': 'shoe'}, {'frequency': 'c', 'id': 967, 'synset': 'shopping_bag.n.01', 'synonyms': ['shopping_bag'], 'def': 'a bag made of plastic or strong paper (often with handles); used to transport goods after shopping', 'name': 'shopping_bag'}, {'frequency': 'c', 'id': 968, 'synset': 'shopping_cart.n.01', 'synonyms': ['shopping_cart'], 'def': 'a handcart that holds groceries or other goods while shopping', 'name': 'shopping_cart'}, {'frequency': 'f', 'id': 969, 'synset': 'short_pants.n.01', 'synonyms': ['short_pants', 'shorts_(clothing)', 'trunks_(clothing)'], 'def': 'trousers that end at or above the knee', 'name': 'short_pants'}, {'frequency': 'r', 'id': 970, 'synset': 'shot_glass.n.01', 'synonyms': ['shot_glass'], 'def': 'a small glass adequate to hold a single swallow of whiskey', 'name': 'shot_glass'}, {'frequency': 'c', 'id': 971, 'synset': 'shoulder_bag.n.01', 'synonyms': ['shoulder_bag'], 'def': 'a large handbag that can be carried by a strap looped over the shoulder', 'name': 'shoulder_bag'}, {'frequency': 'c', 'id': 972, 'synset': 'shovel.n.01', 'synonyms': ['shovel'], 'def': 'a hand tool for lifting loose material such as snow, dirt, etc.', 'name': 'shovel'}, {'frequency': 'f', 'id': 973, 'synset': 'shower.n.01', 'synonyms': ['shower_head'], 'def': 'a plumbing fixture that sprays water over you', 'name': 'shower_head'}, {'frequency': 'f', 'id': 974, 'synset': 'shower_curtain.n.01', 'synonyms': ['shower_curtain'], 'def': 'a curtain that keeps water from splashing out of the shower area', 'name': 'shower_curtain'}, {'frequency': 'r', 'id': 975, 'synset': 'shredder.n.01', 'synonyms': ['shredder_(for_paper)'], 'def': 'a device that shreds documents', 'name': 'shredder_(for_paper)'}, {'frequency': 'r', 'id': 976, 'synset': 'sieve.n.01', 'synonyms': ['sieve', 'screen_(sieve)'], 'def': 'a strainer for separating lumps from powdered material or grading particles', 'name': 'sieve'}, {'frequency': 'f', 'id': 977, 'synset': 'signboard.n.01', 'synonyms': ['signboard'], 'def': 'structure displaying a board on which advertisements can be posted', 'name': 'signboard'}, {'frequency': 'c', 'id': 978, 'synset': 'silo.n.01', 'synonyms': ['silo'], 'def': 'a cylindrical tower used for storing goods', 'name': 'silo'}, {'frequency': 'f', 'id': 979, 'synset': 'sink.n.01', 'synonyms': ['sink'], 'def': 'plumbing fixture consisting of a water basin fixed to a wall or floor and having a drainpipe', 'name': 'sink'}, {'frequency': 'f', 'id': 980, 'synset': 'skateboard.n.01', 'synonyms': ['skateboard'], 'def': 'a board with wheels that is ridden in a standing or crouching position and propelled by foot', 'name': 'skateboard'}, {'frequency': 'c', 'id': 981, 'synset': 'skewer.n.01', 'synonyms': ['skewer'], 'def': 'a long pin for holding meat in position while it is being roasted', 'name': 'skewer'}, {'frequency': 'f', 'id': 982, 'synset': 'ski.n.01', 'synonyms': ['ski'], 'def': 'sports equipment for skiing on snow', 'name': 'ski'}, {'frequency': 'f', 'id': 983, 'synset': 'ski_boot.n.01', 'synonyms': ['ski_boot'], 'def': 'a stiff boot that is fastened to a ski with a ski binding', 'name': 'ski_boot'}, {'frequency': 'f', 'id': 984, 'synset': 'ski_parka.n.01', 'synonyms': ['ski_parka', 'ski_jacket'], 'def': 'a parka to be worn while skiing', 'name': 'ski_parka'}, {'frequency': 'f', 'id': 985, 'synset': 'ski_pole.n.01', 'synonyms': ['ski_pole'], 'def': 'a pole with metal points used as an aid in skiing', 'name': 'ski_pole'}, {'frequency': 'f', 'id': 986, 'synset': 'skirt.n.02', 'synonyms': ['skirt'], 'def': 'a garment hanging from the waist; worn mainly by girls and women', 'name': 'skirt'}, {'frequency': 'c', 'id': 987, 'synset': 'sled.n.01', 'synonyms': ['sled', 'sledge', 'sleigh'], 'def': 'a vehicle or flat object for transportation over snow by sliding or pulled by dogs, etc.', 'name': 'sled'}, {'frequency': 'c', 'id': 988, 'synset': 'sleeping_bag.n.01', 'synonyms': ['sleeping_bag'], 'def': 'large padded bag designed to be slept in outdoors', 'name': 'sleeping_bag'}, {'frequency': 'r', 'id': 989, 'synset': 'sling.n.05', 'synonyms': ['sling_(bandage)', 'triangular_bandage'], 'def': 'bandage to support an injured forearm; slung over the shoulder or neck', 'name': 'sling_(bandage)'}, {'frequency': 'c', 'id': 990, 'synset': 'slipper.n.01', 'synonyms': ['slipper_(footwear)', 'carpet_slipper_(footwear)'], 'def': 'low footwear that can be slipped on and off easily; usually worn indoors', 'name': 'slipper_(footwear)'}, {'frequency': 'r', 'id': 991, 'synset': 'smoothie.n.02', 'synonyms': ['smoothie'], 'def': 'a thick smooth drink consisting of fresh fruit pureed with ice cream or yoghurt or milk', 'name': 'smoothie'}, {'frequency': 'r', 'id': 992, 'synset': 'snake.n.01', 'synonyms': ['snake', 'serpent'], 'def': 'limbless scaly elongate reptile; some are venomous', 'name': 'snake'}, {'frequency': 'f', 'id': 993, 'synset': 'snowboard.n.01', 'synonyms': ['snowboard'], 'def': 'a board that resembles a broad ski or a small surfboard; used in a standing position to slide down snow-covered slopes', 'name': 'snowboard'}, {'frequency': 'c', 'id': 994, 'synset': 'snowman.n.01', 'synonyms': ['snowman'], 'def': 'a figure of a person made of packed snow', 'name': 'snowman'}, {'frequency': 'c', 'id': 995, 'synset': 'snowmobile.n.01', 'synonyms': ['snowmobile'], 'def': 'tracked vehicle for travel on snow having skis in front', 'name': 'snowmobile'}, {'frequency': 'f', 'id': 996, 'synset': 'soap.n.01', 'synonyms': ['soap'], 'def': 'a cleansing agent made from the salts of vegetable or animal fats', 'name': 'soap'}, {'frequency': 'f', 'id': 997, 'synset': 'soccer_ball.n.01', 'synonyms': ['soccer_ball'], 'def': "an inflated ball used in playing soccer (called `football' outside of the United States)", 'name': 'soccer_ball'}, {'frequency': 'f', 'id': 998, 'synset': 'sock.n.01', 'synonyms': ['sock'], 'def': 'cloth covering for the foot; worn inside the shoe; reaches to between the ankle and the knee', 'name': 'sock'}, {'frequency': 'r', 'id': 999, 'synset': 'soda_fountain.n.02', 'synonyms': ['soda_fountain'], 'def': 'an apparatus for dispensing soda water', 'name': 'soda_fountain'}, {'frequency': 'r', 'id': 1000, 'synset': 'soda_water.n.01', 'synonyms': ['carbonated_water', 'club_soda', 'seltzer', 'sparkling_water'], 'def': 'effervescent beverage artificially charged with carbon dioxide', 'name': 'carbonated_water'}, {'frequency': 'f', 'id': 1001, 'synset': 'sofa.n.01', 'synonyms': ['sofa', 'couch', 'lounge'], 'def': 'an upholstered seat for more than one person', 'name': 'sofa'}, {'frequency': 'r', 'id': 1002, 'synset': 'softball.n.01', 'synonyms': ['softball'], 'def': 'ball used in playing softball', 'name': 'softball'}, {'frequency': 'c', 'id': 1003, 'synset': 'solar_array.n.01', 'synonyms': ['solar_array', 'solar_battery', 'solar_panel'], 'def': 'electrical device consisting of a large array of connected solar cells', 'name': 'solar_array'}, {'frequency': 'r', 'id': 1004, 'synset': 'sombrero.n.02', 'synonyms': ['sombrero'], 'def': 'a straw hat with a tall crown and broad brim; worn in American southwest and in Mexico', 'name': 'sombrero'}, {'frequency': 'c', 'id': 1005, 'synset': 'soup.n.01', 'synonyms': ['soup'], 'def': 'liquid food especially of meat or fish or vegetable stock often containing pieces of solid food', 'name': 'soup'}, {'frequency': 'r', 'id': 1006, 'synset': 'soup_bowl.n.01', 'synonyms': ['soup_bowl'], 'def': 'a bowl for serving soup', 'name': 'soup_bowl'}, {'frequency': 'c', 'id': 1007, 'synset': 'soupspoon.n.01', 'synonyms': ['soupspoon'], 'def': 'a spoon with a rounded bowl for eating soup', 'name': 'soupspoon'}, {'frequency': 'c', 'id': 1008, 'synset': 'sour_cream.n.01', 'synonyms': ['sour_cream', 'soured_cream'], 'def': 'soured light cream', 'name': 'sour_cream'}, {'frequency': 'r', 'id': 1009, 'synset': 'soya_milk.n.01', 'synonyms': ['soya_milk', 'soybean_milk', 'soymilk'], 'def': 'a milk substitute containing soybean flour and water; used in some infant formulas and in making tofu', 'name': 'soya_milk'}, {'frequency': 'r', 'id': 1010, 'synset': 'space_shuttle.n.01', 'synonyms': ['space_shuttle'], 'def': "a reusable spacecraft with wings for a controlled descent through the Earth's atmosphere", 'name': 'space_shuttle'}, {'frequency': 'r', 'id': 1011, 'synset': 'sparkler.n.02', 'synonyms': ['sparkler_(fireworks)'], 'def': 'a firework that burns slowly and throws out a shower of sparks', 'name': 'sparkler_(fireworks)'}, {'frequency': 'f', 'id': 1012, 'synset': 'spatula.n.02', 'synonyms': ['spatula'], 'def': 'a hand tool with a thin flexible blade used to mix or spread soft substances', 'name': 'spatula'}, {'frequency': 'r', 'id': 1013, 'synset': 'spear.n.01', 'synonyms': ['spear', 'lance'], 'def': 'a long pointed rod used as a tool or weapon', 'name': 'spear'}, {'frequency': 'f', 'id': 1014, 'synset': 'spectacles.n.01', 'synonyms': ['spectacles', 'specs', 'eyeglasses', 'glasses'], 'def': 'optical instrument consisting of a frame that holds a pair of lenses for correcting defective vision', 'name': 'spectacles'}, {'frequency': 'c', 'id': 1015, 'synset': 'spice_rack.n.01', 'synonyms': ['spice_rack'], 'def': 'a rack for displaying containers filled with spices', 'name': 'spice_rack'}, {'frequency': 'r', 'id': 1016, 'synset': 'spider.n.01', 'synonyms': ['spider'], 'def': 'predatory arachnid with eight legs, two poison fangs, two feelers, and usually two silk-spinning organs at the back end of the body', 'name': 'spider'}, {'frequency': 'c', 'id': 1017, 'synset': 'sponge.n.01', 'synonyms': ['sponge'], 'def': 'a porous mass usable to absorb water typically used for cleaning', 'name': 'sponge'}, {'frequency': 'f', 'id': 1018, 'synset': 'spoon.n.01', 'synonyms': ['spoon'], 'def': 'a piece of cutlery with a shallow bowl-shaped container and a handle', 'name': 'spoon'}, {'frequency': 'c', 'id': 1019, 'synset': 'sportswear.n.01', 'synonyms': ['sportswear', 'athletic_wear', 'activewear'], 'def': 'attire worn for sport or for casual wear', 'name': 'sportswear'}, {'frequency': 'c', 'id': 1020, 'synset': 'spotlight.n.02', 'synonyms': ['spotlight'], 'def': 'a lamp that produces a strong beam of light to illuminate a restricted area; used to focus attention of a stage performer', 'name': 'spotlight'}, {'frequency': 'r', 'id': 1021, 'synset': 'squirrel.n.01', 'synonyms': ['squirrel'], 'def': 'a kind of arboreal rodent having a long bushy tail', 'name': 'squirrel'}, {'frequency': 'c', 'id': 1022, 'synset': 'stapler.n.01', 'synonyms': ['stapler_(stapling_machine)'], 'def': 'a machine that inserts staples into sheets of paper in order to fasten them together', 'name': 'stapler_(stapling_machine)'}, {'frequency': 'r', 'id': 1023, 'synset': 'starfish.n.01', 'synonyms': ['starfish', 'sea_star'], 'def': 'echinoderms characterized by five arms extending from a central disk', 'name': 'starfish'}, {'frequency': 'f', 'id': 1024, 'synset': 'statue.n.01', 'synonyms': ['statue_(sculpture)'], 'def': 'a sculpture representing a human or animal', 'name': 'statue_(sculpture)'}, {'frequency': 'c', 'id': 1025, 'synset': 'steak.n.01', 'synonyms': ['steak_(food)'], 'def': 'a slice of meat cut from the fleshy part of an animal or large fish', 'name': 'steak_(food)'}, {'frequency': 'r', 'id': 1026, 'synset': 'steak_knife.n.01', 'synonyms': ['steak_knife'], 'def': 'a sharp table knife used in eating steak', 'name': 'steak_knife'}, {'frequency': 'r', 'id': 1027, 'synset': 'steamer.n.02', 'synonyms': ['steamer_(kitchen_appliance)'], 'def': 'a cooking utensil that can be used to cook food by steaming it', 'name': 'steamer_(kitchen_appliance)'}, {'frequency': 'f', 'id': 1028, 'synset': 'steering_wheel.n.01', 'synonyms': ['steering_wheel'], 'def': 'a handwheel that is used for steering', 'name': 'steering_wheel'}, {'frequency': 'r', 'id': 1029, 'synset': 'stencil.n.01', 'synonyms': ['stencil'], 'def': 'a sheet of material (metal, plastic, etc.) that has been perforated with a pattern; ink or paint can pass through the perforations to create the printed pattern on the surface below', 'name': 'stencil'}, {'frequency': 'r', 'id': 1030, 'synset': 'step_ladder.n.01', 'synonyms': ['stepladder'], 'def': 'a folding portable ladder hinged at the top', 'name': 'stepladder'}, {'frequency': 'c', 'id': 1031, 'synset': 'step_stool.n.01', 'synonyms': ['step_stool'], 'def': 'a stool that has one or two steps that fold under the seat', 'name': 'step_stool'}, {'frequency': 'c', 'id': 1032, 'synset': 'stereo.n.01', 'synonyms': ['stereo_(sound_system)'], 'def': 'electronic device for playing audio', 'name': 'stereo_(sound_system)'}, {'frequency': 'r', 'id': 1033, 'synset': 'stew.n.02', 'synonyms': ['stew'], 'def': 'food prepared by stewing especially meat or fish with vegetables', 'name': 'stew'}, {'frequency': 'r', 'id': 1034, 'synset': 'stirrer.n.02', 'synonyms': ['stirrer'], 'def': 'an implement used for stirring', 'name': 'stirrer'}, {'frequency': 'f', 'id': 1035, 'synset': 'stirrup.n.01', 'synonyms': ['stirrup'], 'def': "support consisting of metal loops into which rider's feet go", 'name': 'stirrup'}, {'frequency': 'c', 'id': 1036, 'synset': 'stocking.n.01', 'synonyms': ['stockings_(leg_wear)'], 'def': 'close-fitting hosiery to cover the foot and leg; come in matched pairs', 'name': 'stockings_(leg_wear)'}, {'frequency': 'f', 'id': 1037, 'synset': 'stool.n.01', 'synonyms': ['stool'], 'def': 'a simple seat without a back or arms', 'name': 'stool'}, {'frequency': 'f', 'id': 1038, 'synset': 'stop_sign.n.01', 'synonyms': ['stop_sign'], 'def': 'a traffic sign to notify drivers that they must come to a complete stop', 'name': 'stop_sign'}, {'frequency': 'f', 'id': 1039, 'synset': 'stoplight.n.01', 'synonyms': ['brake_light'], 'def': 'a red light on the rear of a motor vehicle that signals when the brakes are applied', 'name': 'brake_light'}, {'frequency': 'f', 'id': 1040, 'synset': 'stove.n.01', 'synonyms': ['stove', 'kitchen_stove', 'range_(kitchen_appliance)', 'kitchen_range', 'cooking_stove'], 'def': 'a kitchen appliance used for cooking food', 'name': 'stove'}, {'frequency': 'c', 'id': 1041, 'synset': 'strainer.n.01', 'synonyms': ['strainer'], 'def': 'a filter to retain larger pieces while smaller pieces and liquids pass through', 'name': 'strainer'}, {'frequency': 'f', 'id': 1042, 'synset': 'strap.n.01', 'synonyms': ['strap'], 'def': 'an elongated strip of material for binding things together or holding', 'name': 'strap'}, {'frequency': 'f', 'id': 1043, 'synset': 'straw.n.04', 'synonyms': ['straw_(for_drinking)', 'drinking_straw'], 'def': 'a thin paper or plastic tube used to suck liquids into the mouth', 'name': 'straw_(for_drinking)'}, {'frequency': 'f', 'id': 1044, 'synset': 'strawberry.n.01', 'synonyms': ['strawberry'], 'def': 'sweet fleshy red fruit', 'name': 'strawberry'}, {'frequency': 'f', 'id': 1045, 'synset': 'street_sign.n.01', 'synonyms': ['street_sign'], 'def': 'a sign visible from the street', 'name': 'street_sign'}, {'frequency': 'f', 'id': 1046, 'synset': 'streetlight.n.01', 'synonyms': ['streetlight', 'street_lamp'], 'def': 'a lamp supported on a lamppost; for illuminating a street', 'name': 'streetlight'}, {'frequency': 'r', 'id': 1047, 'synset': 'string_cheese.n.01', 'synonyms': ['string_cheese'], 'def': 'cheese formed in long strings twisted together', 'name': 'string_cheese'}, {'frequency': 'r', 'id': 1048, 'synset': 'stylus.n.02', 'synonyms': ['stylus'], 'def': 'a pointed tool for writing or drawing or engraving', 'name': 'stylus'}, {'frequency': 'r', 'id': 1049, 'synset': 'subwoofer.n.01', 'synonyms': ['subwoofer'], 'def': 'a loudspeaker that is designed to reproduce very low bass frequencies', 'name': 'subwoofer'}, {'frequency': 'r', 'id': 1050, 'synset': 'sugar_bowl.n.01', 'synonyms': ['sugar_bowl'], 'def': 'a dish in which sugar is served', 'name': 'sugar_bowl'}, {'frequency': 'r', 'id': 1051, 'synset': 'sugarcane.n.01', 'synonyms': ['sugarcane_(plant)'], 'def': 'juicy canes whose sap is a source of molasses and commercial sugar; fresh canes are sometimes chewed for the juice', 'name': 'sugarcane_(plant)'}, {'frequency': 'c', 'id': 1052, 'synset': 'suit.n.01', 'synonyms': ['suit_(clothing)'], 'def': 'a set of garments (usually including a jacket and trousers or skirt) for outerwear all of the same fabric and color', 'name': 'suit_(clothing)'}, {'frequency': 'c', 'id': 1053, 'synset': 'sunflower.n.01', 'synonyms': ['sunflower'], 'def': 'any plant of the genus Helianthus having large flower heads with dark disk florets and showy yellow rays', 'name': 'sunflower'}, {'frequency': 'f', 'id': 1054, 'synset': 'sunglasses.n.01', 'synonyms': ['sunglasses'], 'def': 'spectacles that are darkened or polarized to protect the eyes from the glare of the sun', 'name': 'sunglasses'}, {'frequency': 'c', 'id': 1055, 'synset': 'sunhat.n.01', 'synonyms': ['sunhat'], 'def': 'a hat with a broad brim that protects the face from direct exposure to the sun', 'name': 'sunhat'}, {'frequency': 'r', 'id': 1056, 'synset': 'sunscreen.n.01', 'synonyms': ['sunscreen', 'sunblock'], 'def': 'a cream spread on the skin; contains a chemical to filter out ultraviolet light and so protect from sunburn', 'name': 'sunscreen'}, {'frequency': 'f', 'id': 1057, 'synset': 'surfboard.n.01', 'synonyms': ['surfboard'], 'def': 'a narrow buoyant board for riding surf', 'name': 'surfboard'}, {'frequency': 'c', 'id': 1058, 'synset': 'sushi.n.01', 'synonyms': ['sushi'], 'def': 'rice (with raw fish) wrapped in seaweed', 'name': 'sushi'}, {'frequency': 'c', 'id': 1059, 'synset': 'swab.n.02', 'synonyms': ['mop'], 'def': 'cleaning implement consisting of absorbent material fastened to a handle; for cleaning floors', 'name': 'mop'}, {'frequency': 'c', 'id': 1060, 'synset': 'sweat_pants.n.01', 'synonyms': ['sweat_pants'], 'def': 'loose-fitting trousers with elastic cuffs; worn by athletes', 'name': 'sweat_pants'}, {'frequency': 'c', 'id': 1061, 'synset': 'sweatband.n.02', 'synonyms': ['sweatband'], 'def': 'a band of material tied around the forehead or wrist to absorb sweat', 'name': 'sweatband'}, {'frequency': 'f', 'id': 1062, 'synset': 'sweater.n.01', 'synonyms': ['sweater'], 'def': 'a crocheted or knitted garment covering the upper part of the body', 'name': 'sweater'}, {'frequency': 'f', 'id': 1063, 'synset': 'sweatshirt.n.01', 'synonyms': ['sweatshirt'], 'def': 'cotton knit pullover with long sleeves worn during athletic activity', 'name': 'sweatshirt'}, {'frequency': 'c', 'id': 1064, 'synset': 'sweet_potato.n.02', 'synonyms': ['sweet_potato'], 'def': 'the edible tuberous root of the sweet potato vine', 'name': 'sweet_potato'}, {'frequency': 'f', 'id': 1065, 'synset': 'swimsuit.n.01', 'synonyms': ['swimsuit', 'swimwear', 'bathing_suit', 'swimming_costume', 'bathing_costume', 'swimming_trunks', 'bathing_trunks'], 'def': 'garment worn for swimming', 'name': 'swimsuit'}, {'frequency': 'c', 'id': 1066, 'synset': 'sword.n.01', 'synonyms': ['sword'], 'def': 'a cutting or thrusting weapon that has a long metal blade', 'name': 'sword'}, {'frequency': 'r', 'id': 1067, 'synset': 'syringe.n.01', 'synonyms': ['syringe'], 'def': 'a medical instrument used to inject or withdraw fluids', 'name': 'syringe'}, {'frequency': 'r', 'id': 1068, 'synset': 'tabasco.n.02', 'synonyms': ['Tabasco_sauce'], 'def': 'very spicy sauce (trade name Tabasco) made from fully-aged red peppers', 'name': 'Tabasco_sauce'}, {'frequency': 'r', 'id': 1069, 'synset': 'table-tennis_table.n.01', 'synonyms': ['table-tennis_table', 'ping-pong_table'], 'def': 'a table used for playing table tennis', 'name': 'table-tennis_table'}, {'frequency': 'f', 'id': 1070, 'synset': 'table.n.02', 'synonyms': ['table'], 'def': 'a piece of furniture having a smooth flat top that is usually supported by one or more vertical legs', 'name': 'table'}, {'frequency': 'c', 'id': 1071, 'synset': 'table_lamp.n.01', 'synonyms': ['table_lamp'], 'def': 'a lamp that sits on a table', 'name': 'table_lamp'}, {'frequency': 'f', 'id': 1072, 'synset': 'tablecloth.n.01', 'synonyms': ['tablecloth'], 'def': 'a covering spread over a dining table', 'name': 'tablecloth'}, {'frequency': 'r', 'id': 1073, 'synset': 'tachometer.n.01', 'synonyms': ['tachometer'], 'def': 'measuring instrument for indicating speed of rotation', 'name': 'tachometer'}, {'frequency': 'r', 'id': 1074, 'synset': 'taco.n.02', 'synonyms': ['taco'], 'def': 'a small tortilla cupped around a filling', 'name': 'taco'}, {'frequency': 'f', 'id': 1075, 'synset': 'tag.n.02', 'synonyms': ['tag'], 'def': 'a label associated with something for the purpose of identification or information', 'name': 'tag'}, {'frequency': 'f', 'id': 1076, 'synset': 'taillight.n.01', 'synonyms': ['taillight', 'rear_light'], 'def': 'lamp (usually red) mounted at the rear of a motor vehicle', 'name': 'taillight'}, {'frequency': 'r', 'id': 1077, 'synset': 'tambourine.n.01', 'synonyms': ['tambourine'], 'def': 'a shallow drum with a single drumhead and with metallic disks in the sides', 'name': 'tambourine'}, {'frequency': 'r', 'id': 1078, 'synset': 'tank.n.01', 'synonyms': ['army_tank', 'armored_combat_vehicle', 'armoured_combat_vehicle'], 'def': 'an enclosed armored military vehicle; has a cannon and moves on caterpillar treads', 'name': 'army_tank'}, {'frequency': 'c', 'id': 1079, 'synset': 'tank.n.02', 'synonyms': ['tank_(storage_vessel)', 'storage_tank'], 'def': 'a large (usually metallic) vessel for holding gases or liquids', 'name': 'tank_(storage_vessel)'}, {'frequency': 'f', 'id': 1080, 'synset': 'tank_top.n.01', 'synonyms': ['tank_top_(clothing)'], 'def': 'a tight-fitting sleeveless shirt with wide shoulder straps and low neck and no front opening', 'name': 'tank_top_(clothing)'}, {'frequency': 'c', 'id': 1081, 'synset': 'tape.n.01', 'synonyms': ['tape_(sticky_cloth_or_paper)'], 'def': 'a long thin piece of cloth or paper as used for binding or fastening', 'name': 'tape_(sticky_cloth_or_paper)'}, {'frequency': 'c', 'id': 1082, 'synset': 'tape.n.04', 'synonyms': ['tape_measure', 'measuring_tape'], 'def': 'measuring instrument consisting of a narrow strip (cloth or metal) marked in inches or centimeters and used for measuring lengths', 'name': 'tape_measure'}, {'frequency': 'c', 'id': 1083, 'synset': 'tapestry.n.02', 'synonyms': ['tapestry'], 'def': 'a heavy textile with a woven design; used for curtains and upholstery', 'name': 'tapestry'}, {'frequency': 'f', 'id': 1084, 'synset': 'tarpaulin.n.01', 'synonyms': ['tarp'], 'def': 'waterproofed canvas', 'name': 'tarp'}, {'frequency': 'c', 'id': 1085, 'synset': 'tartan.n.01', 'synonyms': ['tartan', 'plaid'], 'def': 'a cloth having a crisscross design', 'name': 'tartan'}, {'frequency': 'c', 'id': 1086, 'synset': 'tassel.n.01', 'synonyms': ['tassel'], 'def': 'adornment consisting of a bunch of cords fastened at one end', 'name': 'tassel'}, {'frequency': 'r', 'id': 1087, 'synset': 'tea_bag.n.01', 'synonyms': ['tea_bag'], 'def': 'a measured amount of tea in a bag for an individual serving of tea', 'name': 'tea_bag'}, {'frequency': 'c', 'id': 1088, 'synset': 'teacup.n.02', 'synonyms': ['teacup'], 'def': 'a cup from which tea is drunk', 'name': 'teacup'}, {'frequency': 'c', 'id': 1089, 'synset': 'teakettle.n.01', 'synonyms': ['teakettle'], 'def': 'kettle for boiling water to make tea', 'name': 'teakettle'}, {'frequency': 'c', 'id': 1090, 'synset': 'teapot.n.01', 'synonyms': ['teapot'], 'def': 'pot for brewing tea; usually has a spout and handle', 'name': 'teapot'}, {'frequency': 'f', 'id': 1091, 'synset': 'teddy.n.01', 'synonyms': ['teddy_bear'], 'def': "plaything consisting of a child's toy bear (usually plush and stuffed with soft materials)", 'name': 'teddy_bear'}, {'frequency': 'f', 'id': 1092, 'synset': 'telephone.n.01', 'synonyms': ['telephone', 'phone', 'telephone_set'], 'def': 'electronic device for communicating by voice over long distances', 'name': 'telephone'}, {'frequency': 'c', 'id': 1093, 'synset': 'telephone_booth.n.01', 'synonyms': ['telephone_booth', 'phone_booth', 'call_box', 'telephone_box', 'telephone_kiosk'], 'def': 'booth for using a telephone', 'name': 'telephone_booth'}, {'frequency': 'f', 'id': 1094, 'synset': 'telephone_pole.n.01', 'synonyms': ['telephone_pole', 'telegraph_pole', 'telegraph_post'], 'def': 'tall pole supporting telephone wires', 'name': 'telephone_pole'}, {'frequency': 'r', 'id': 1095, 'synset': 'telephoto_lens.n.01', 'synonyms': ['telephoto_lens', 'zoom_lens'], 'def': 'a camera lens that magnifies the image', 'name': 'telephoto_lens'}, {'frequency': 'c', 'id': 1096, 'synset': 'television_camera.n.01', 'synonyms': ['television_camera', 'tv_camera'], 'def': 'television equipment for capturing and recording video', 'name': 'television_camera'}, {'frequency': 'f', 'id': 1097, 'synset': 'television_receiver.n.01', 'synonyms': ['television_set', 'tv', 'tv_set'], 'def': 'an electronic device that receives television signals and displays them on a screen', 'name': 'television_set'}, {'frequency': 'f', 'id': 1098, 'synset': 'tennis_ball.n.01', 'synonyms': ['tennis_ball'], 'def': 'ball about the size of a fist used in playing tennis', 'name': 'tennis_ball'}, {'frequency': 'f', 'id': 1099, 'synset': 'tennis_racket.n.01', 'synonyms': ['tennis_racket'], 'def': 'a racket used to play tennis', 'name': 'tennis_racket'}, {'frequency': 'r', 'id': 1100, 'synset': 'tequila.n.01', 'synonyms': ['tequila'], 'def': 'Mexican liquor made from fermented juices of an agave plant', 'name': 'tequila'}, {'frequency': 'c', 'id': 1101, 'synset': 'thermometer.n.01', 'synonyms': ['thermometer'], 'def': 'measuring instrument for measuring temperature', 'name': 'thermometer'}, {'frequency': 'c', 'id': 1102, 'synset': 'thermos.n.01', 'synonyms': ['thermos_bottle'], 'def': 'vacuum flask that preserves temperature of hot or cold drinks', 'name': 'thermos_bottle'}, {'frequency': 'c', 'id': 1103, 'synset': 'thermostat.n.01', 'synonyms': ['thermostat'], 'def': 'a regulator for automatically regulating temperature by starting or stopping the supply of heat', 'name': 'thermostat'}, {'frequency': 'r', 'id': 1104, 'synset': 'thimble.n.02', 'synonyms': ['thimble'], 'def': 'a small metal cap to protect the finger while sewing; can be used as a small container', 'name': 'thimble'}, {'frequency': 'c', 'id': 1105, 'synset': 'thread.n.01', 'synonyms': ['thread', 'yarn'], 'def': 'a fine cord of twisted fibers (of cotton or silk or wool or nylon etc.) used in sewing and weaving', 'name': 'thread'}, {'frequency': 'c', 'id': 1106, 'synset': 'thumbtack.n.01', 'synonyms': ['thumbtack', 'drawing_pin', 'pushpin'], 'def': 'a tack for attaching papers to a bulletin board or drawing board', 'name': 'thumbtack'}, {'frequency': 'c', 'id': 1107, 'synset': 'tiara.n.01', 'synonyms': ['tiara'], 'def': 'a jeweled headdress worn by women on formal occasions', 'name': 'tiara'}, {'frequency': 'c', 'id': 1108, 'synset': 'tiger.n.02', 'synonyms': ['tiger'], 'def': 'large feline of forests in most of Asia having a tawny coat with black stripes', 'name': 'tiger'}, {'frequency': 'c', 'id': 1109, 'synset': 'tights.n.01', 'synonyms': ['tights_(clothing)', 'leotards'], 'def': 'skintight knit hose covering the body from the waist to the feet worn by acrobats and dancers and as stockings by women and girls', 'name': 'tights_(clothing)'}, {'frequency': 'c', 'id': 1110, 'synset': 'timer.n.01', 'synonyms': ['timer', 'stopwatch'], 'def': 'a timepiece that measures a time interval and signals its end', 'name': 'timer'}, {'frequency': 'f', 'id': 1111, 'synset': 'tinfoil.n.01', 'synonyms': ['tinfoil'], 'def': 'foil made of tin or an alloy of tin and lead', 'name': 'tinfoil'}, {'frequency': 'r', 'id': 1112, 'synset': 'tinsel.n.01', 'synonyms': ['tinsel'], 'def': 'a showy decoration that is basically valueless', 'name': 'tinsel'}, {'frequency': 'f', 'id': 1113, 'synset': 'tissue.n.02', 'synonyms': ['tissue_paper'], 'def': 'a soft thin (usually translucent) paper', 'name': 'tissue_paper'}, {'frequency': 'c', 'id': 1114, 'synset': 'toast.n.01', 'synonyms': ['toast_(food)'], 'def': 'slice of bread that has been toasted', 'name': 'toast_(food)'}, {'frequency': 'f', 'id': 1115, 'synset': 'toaster.n.02', 'synonyms': ['toaster'], 'def': 'a kitchen appliance (usually electric) for toasting bread', 'name': 'toaster'}, {'frequency': 'c', 'id': 1116, 'synset': 'toaster_oven.n.01', 'synonyms': ['toaster_oven'], 'def': 'kitchen appliance consisting of a small electric oven for toasting or warming food', 'name': 'toaster_oven'}, {'frequency': 'f', 'id': 1117, 'synset': 'toilet.n.02', 'synonyms': ['toilet'], 'def': 'a plumbing fixture for defecation and urination', 'name': 'toilet'}, {'frequency': 'f', 'id': 1118, 'synset': 'toilet_tissue.n.01', 'synonyms': ['toilet_tissue', 'toilet_paper', 'bathroom_tissue'], 'def': 'a soft thin absorbent paper for use in toilets', 'name': 'toilet_tissue'}, {'frequency': 'f', 'id': 1119, 'synset': 'tomato.n.01', 'synonyms': ['tomato'], 'def': 'mildly acid red or yellow pulpy fruit eaten as a vegetable', 'name': 'tomato'}, {'frequency': 'c', 'id': 1120, 'synset': 'tongs.n.01', 'synonyms': ['tongs'], 'def': 'any of various devices for taking hold of objects; usually have two hinged legs with handles above and pointed hooks below', 'name': 'tongs'}, {'frequency': 'c', 'id': 1121, 'synset': 'toolbox.n.01', 'synonyms': ['toolbox'], 'def': 'a box or chest or cabinet for holding hand tools', 'name': 'toolbox'}, {'frequency': 'f', 'id': 1122, 'synset': 'toothbrush.n.01', 'synonyms': ['toothbrush'], 'def': 'small brush; has long handle; used to clean teeth', 'name': 'toothbrush'}, {'frequency': 'f', 'id': 1123, 'synset': 'toothpaste.n.01', 'synonyms': ['toothpaste'], 'def': 'a dentifrice in the form of a paste', 'name': 'toothpaste'}, {'frequency': 'c', 'id': 1124, 'synset': 'toothpick.n.01', 'synonyms': ['toothpick'], 'def': 'pick consisting of a small strip of wood or plastic; used to pick food from between the teeth', 'name': 'toothpick'}, {'frequency': 'c', 'id': 1125, 'synset': 'top.n.09', 'synonyms': ['cover'], 'def': 'covering for a hole (especially a hole in the top of a container)', 'name': 'cover'}, {'frequency': 'c', 'id': 1126, 'synset': 'tortilla.n.01', 'synonyms': ['tortilla'], 'def': 'thin unleavened pancake made from cornmeal or wheat flour', 'name': 'tortilla'}, {'frequency': 'c', 'id': 1127, 'synset': 'tow_truck.n.01', 'synonyms': ['tow_truck'], 'def': 'a truck equipped to hoist and pull wrecked cars (or to remove cars from no-parking zones)', 'name': 'tow_truck'}, {'frequency': 'f', 'id': 1128, 'synset': 'towel.n.01', 'synonyms': ['towel'], 'def': 'a rectangular piece of absorbent cloth (or paper) for drying or wiping', 'name': 'towel'}, {'frequency': 'f', 'id': 1129, 'synset': 'towel_rack.n.01', 'synonyms': ['towel_rack', 'towel_rail', 'towel_bar'], 'def': 'a rack consisting of one or more bars on which towels can be hung', 'name': 'towel_rack'}, {'frequency': 'f', 'id': 1130, 'synset': 'toy.n.03', 'synonyms': ['toy'], 'def': 'a device regarded as providing amusement', 'name': 'toy'}, {'frequency': 'c', 'id': 1131, 'synset': 'tractor.n.01', 'synonyms': ['tractor_(farm_equipment)'], 'def': 'a wheeled vehicle with large wheels; used in farming and other applications', 'name': 'tractor_(farm_equipment)'}, {'frequency': 'f', 'id': 1132, 'synset': 'traffic_light.n.01', 'synonyms': ['traffic_light'], 'def': 'a device to control vehicle traffic often consisting of three or more lights', 'name': 'traffic_light'}, {'frequency': 'r', 'id': 1133, 'synset': 'trail_bike.n.01', 'synonyms': ['dirt_bike'], 'def': 'a lightweight motorcycle equipped with rugged tires and suspension for off-road use', 'name': 'dirt_bike'}, {'frequency': 'c', 'id': 1134, 'synset': 'trailer_truck.n.01', 'synonyms': ['trailer_truck', 'tractor_trailer', 'trucking_rig', 'articulated_lorry', 'semi_truck'], 'def': 'a truck consisting of a tractor and trailer together', 'name': 'trailer_truck'}, {'frequency': 'f', 'id': 1135, 'synset': 'train.n.01', 'synonyms': ['train_(railroad_vehicle)', 'railroad_train'], 'def': 'public or private transport provided by a line of railway cars coupled together and drawn by a locomotive', 'name': 'train_(railroad_vehicle)'}, {'frequency': 'r', 'id': 1136, 'synset': 'trampoline.n.01', 'synonyms': ['trampoline'], 'def': 'gymnastic apparatus consisting of a strong canvas sheet attached with springs to a metal frame', 'name': 'trampoline'}, {'frequency': 'f', 'id': 1137, 'synset': 'tray.n.01', 'synonyms': ['tray'], 'def': 'an open receptacle for holding or displaying or serving articles or food', 'name': 'tray'}, {'frequency': 'r', 'id': 1138, 'synset': 'tree_house.n.01', 'synonyms': ['tree_house'], 'def': '(NOT A TREE) a PLAYHOUSE built in the branches of a tree', 'name': 'tree_house'}, {'frequency': 'r', 'id': 1139, 'synset': 'trench_coat.n.01', 'synonyms': ['trench_coat'], 'def': 'a military style raincoat; belted with deep pockets', 'name': 'trench_coat'}, {'frequency': 'r', 'id': 1140, 'synset': 'triangle.n.05', 'synonyms': ['triangle_(musical_instrument)'], 'def': 'a percussion instrument consisting of a metal bar bent in the shape of an open triangle', 'name': 'triangle_(musical_instrument)'}, {'frequency': 'r', 'id': 1141, 'synset': 'tricycle.n.01', 'synonyms': ['tricycle'], 'def': 'a vehicle with three wheels that is moved by foot pedals', 'name': 'tricycle'}, {'frequency': 'c', 'id': 1142, 'synset': 'tripod.n.01', 'synonyms': ['tripod'], 'def': 'a three-legged rack used for support', 'name': 'tripod'}, {'frequency': 'f', 'id': 1143, 'synset': 'trouser.n.01', 'synonyms': ['trousers', 'pants_(clothing)'], 'def': 'a garment extending from the waist to the knee or ankle, covering each leg separately', 'name': 'trousers'}, {'frequency': 'f', 'id': 1144, 'synset': 'truck.n.01', 'synonyms': ['truck'], 'def': 'an automotive vehicle suitable for hauling', 'name': 'truck'}, {'frequency': 'r', 'id': 1145, 'synset': 'truffle.n.03', 'synonyms': ['truffle_(chocolate)', 'chocolate_truffle'], 'def': 'creamy chocolate candy', 'name': 'truffle_(chocolate)'}, {'frequency': 'c', 'id': 1146, 'synset': 'trunk.n.02', 'synonyms': ['trunk'], 'def': 'luggage consisting of a large strong case used when traveling or for storage', 'name': 'trunk'}, {'frequency': 'r', 'id': 1147, 'synset': 'tub.n.02', 'synonyms': ['vat'], 'def': 'a large open vessel for holding or storing liquids', 'name': 'vat'}, {'frequency': 'c', 'id': 1148, 'synset': 'turban.n.01', 'synonyms': ['turban'], 'def': 'a traditional headdress consisting of a long scarf wrapped around the head', 'name': 'turban'}, {'frequency': 'r', 'id': 1149, 'synset': 'turkey.n.01', 'synonyms': ['turkey_(bird)'], 'def': 'large gallinaceous bird with fan-shaped tail; widely domesticated for food', 'name': 'turkey_(bird)'}, {'frequency': 'c', 'id': 1150, 'synset': 'turkey.n.04', 'synonyms': ['turkey_(food)'], 'def': 'flesh of large domesticated fowl usually roasted', 'name': 'turkey_(food)'}, {'frequency': 'r', 'id': 1151, 'synset': 'turnip.n.01', 'synonyms': ['turnip'], 'def': 'widely cultivated plant having a large fleshy edible white or yellow root', 'name': 'turnip'}, {'frequency': 'c', 'id': 1152, 'synset': 'turtle.n.02', 'synonyms': ['turtle'], 'def': 'any of various aquatic and land reptiles having a bony shell and flipper-like limbs for swimming', 'name': 'turtle'}, {'frequency': 'r', 'id': 1153, 'synset': 'turtleneck.n.01', 'synonyms': ['turtleneck_(clothing)', 'polo-neck'], 'def': 'a sweater or jersey with a high close-fitting collar', 'name': 'turtleneck_(clothing)'}, {'frequency': 'r', 'id': 1154, 'synset': 'typewriter.n.01', 'synonyms': ['typewriter'], 'def': 'hand-operated character printer for printing written messages one character at a time', 'name': 'typewriter'}, {'frequency': 'f', 'id': 1155, 'synset': 'umbrella.n.01', 'synonyms': ['umbrella'], 'def': 'a lightweight handheld collapsible canopy', 'name': 'umbrella'}, {'frequency': 'c', 'id': 1156, 'synset': 'underwear.n.01', 'synonyms': ['underwear', 'underclothes', 'underclothing', 'underpants'], 'def': 'undergarment worn next to the skin and under the outer garments', 'name': 'underwear'}, {'frequency': 'r', 'id': 1157, 'synset': 'unicycle.n.01', 'synonyms': ['unicycle'], 'def': 'a vehicle with a single wheel that is driven by pedals', 'name': 'unicycle'}, {'frequency': 'c', 'id': 1158, 'synset': 'urinal.n.01', 'synonyms': ['urinal'], 'def': 'a plumbing fixture (usually attached to the wall) used by men to urinate', 'name': 'urinal'}, {'frequency': 'r', 'id': 1159, 'synset': 'urn.n.01', 'synonyms': ['urn'], 'def': 'a large vase that usually has a pedestal or feet', 'name': 'urn'}, {'frequency': 'c', 'id': 1160, 'synset': 'vacuum.n.04', 'synonyms': ['vacuum_cleaner'], 'def': 'an electrical home appliance that cleans by suction', 'name': 'vacuum_cleaner'}, {'frequency': 'c', 'id': 1161, 'synset': 'valve.n.03', 'synonyms': ['valve'], 'def': 'control consisting of a mechanical device for controlling the flow of a fluid', 'name': 'valve'}, {'frequency': 'f', 'id': 1162, 'synset': 'vase.n.01', 'synonyms': ['vase'], 'def': 'an open jar of glass or porcelain used as an ornament or to hold flowers', 'name': 'vase'}, {'frequency': 'c', 'id': 1163, 'synset': 'vending_machine.n.01', 'synonyms': ['vending_machine'], 'def': 'a slot machine for selling goods', 'name': 'vending_machine'}, {'frequency': 'f', 'id': 1164, 'synset': 'vent.n.01', 'synonyms': ['vent', 'blowhole', 'air_vent'], 'def': 'a hole for the escape of gas or air', 'name': 'vent'}, {'frequency': 'c', 'id': 1165, 'synset': 'videotape.n.01', 'synonyms': ['videotape'], 'def': 'a video recording made on magnetic tape', 'name': 'videotape'}, {'frequency': 'r', 'id': 1166, 'synset': 'vinegar.n.01', 'synonyms': ['vinegar'], 'def': 'sour-tasting liquid produced usually by oxidation of the alcohol in wine or cider and used as a condiment or food preservative', 'name': 'vinegar'}, {'frequency': 'r', 'id': 1167, 'synset': 'violin.n.01', 'synonyms': ['violin', 'fiddle'], 'def': 'bowed stringed instrument that is the highest member of the violin family', 'name': 'violin'}, {'frequency': 'r', 'id': 1168, 'synset': 'vodka.n.01', 'synonyms': ['vodka'], 'def': 'unaged colorless liquor originating in Russia', 'name': 'vodka'}, {'frequency': 'r', 'id': 1169, 'synset': 'volleyball.n.02', 'synonyms': ['volleyball'], 'def': 'an inflated ball used in playing volleyball', 'name': 'volleyball'}, {'frequency': 'r', 'id': 1170, 'synset': 'vulture.n.01', 'synonyms': ['vulture'], 'def': 'any of various large birds of prey having naked heads and weak claws and feeding chiefly on carrion', 'name': 'vulture'}, {'frequency': 'c', 'id': 1171, 'synset': 'waffle.n.01', 'synonyms': ['waffle'], 'def': 'pancake batter baked in a waffle iron', 'name': 'waffle'}, {'frequency': 'r', 'id': 1172, 'synset': 'waffle_iron.n.01', 'synonyms': ['waffle_iron'], 'def': 'a kitchen appliance for baking waffles', 'name': 'waffle_iron'}, {'frequency': 'c', 'id': 1173, 'synset': 'wagon.n.01', 'synonyms': ['wagon'], 'def': 'any of various kinds of wheeled vehicles drawn by an animal or a tractor', 'name': 'wagon'}, {'frequency': 'c', 'id': 1174, 'synset': 'wagon_wheel.n.01', 'synonyms': ['wagon_wheel'], 'def': 'a wheel of a wagon', 'name': 'wagon_wheel'}, {'frequency': 'c', 'id': 1175, 'synset': 'walking_stick.n.01', 'synonyms': ['walking_stick'], 'def': 'a stick carried in the hand for support in walking', 'name': 'walking_stick'}, {'frequency': 'c', 'id': 1176, 'synset': 'wall_clock.n.01', 'synonyms': ['wall_clock'], 'def': 'a clock mounted on a wall', 'name': 'wall_clock'}, {'frequency': 'f', 'id': 1177, 'synset': 'wall_socket.n.01', 'synonyms': ['wall_socket', 'wall_plug', 'electric_outlet', 'electrical_outlet', 'outlet', 'electric_receptacle'], 'def': 'receptacle providing a place in a wiring system where current can be taken to run electrical devices', 'name': 'wall_socket'}, {'frequency': 'c', 'id': 1178, 'synset': 'wallet.n.01', 'synonyms': ['wallet', 'billfold'], 'def': 'a pocket-size case for holding papers and paper money', 'name': 'wallet'}, {'frequency': 'r', 'id': 1179, 'synset': 'walrus.n.01', 'synonyms': ['walrus'], 'def': 'either of two large northern marine mammals having ivory tusks and tough hide over thick blubber', 'name': 'walrus'}, {'frequency': 'r', 'id': 1180, 'synset': 'wardrobe.n.01', 'synonyms': ['wardrobe'], 'def': 'a tall piece of furniture that provides storage space for clothes; has a door and rails or hooks for hanging clothes', 'name': 'wardrobe'}, {'frequency': 'r', 'id': 1181, 'synset': 'wasabi.n.02', 'synonyms': ['wasabi'], 'def': 'the thick green root of the wasabi plant that the Japanese use in cooking and that tastes like strong horseradish', 'name': 'wasabi'}, {'frequency': 'c', 'id': 1182, 'synset': 'washer.n.03', 'synonyms': ['automatic_washer', 'washing_machine'], 'def': 'a home appliance for washing clothes and linens automatically', 'name': 'automatic_washer'}, {'frequency': 'f', 'id': 1183, 'synset': 'watch.n.01', 'synonyms': ['watch', 'wristwatch'], 'def': 'a small, portable timepiece', 'name': 'watch'}, {'frequency': 'f', 'id': 1184, 'synset': 'water_bottle.n.01', 'synonyms': ['water_bottle'], 'def': 'a bottle for holding water', 'name': 'water_bottle'}, {'frequency': 'c', 'id': 1185, 'synset': 'water_cooler.n.01', 'synonyms': ['water_cooler'], 'def': 'a device for cooling and dispensing drinking water', 'name': 'water_cooler'}, {'frequency': 'c', 'id': 1186, 'synset': 'water_faucet.n.01', 'synonyms': ['water_faucet', 'water_tap', 'tap_(water_faucet)'], 'def': 'a faucet for drawing water from a pipe or cask', 'name': 'water_faucet'}, {'frequency': 'r', 'id': 1187, 'synset': 'water_filter.n.01', 'synonyms': ['water_filter'], 'def': 'a filter to remove impurities from the water supply', 'name': 'water_filter'}, {'frequency': 'r', 'id': 1188, 'synset': 'water_heater.n.01', 'synonyms': ['water_heater', 'hot-water_heater'], 'def': 'a heater and storage tank to supply heated water', 'name': 'water_heater'}, {'frequency': 'r', 'id': 1189, 'synset': 'water_jug.n.01', 'synonyms': ['water_jug'], 'def': 'a jug that holds water', 'name': 'water_jug'}, {'frequency': 'r', 'id': 1190, 'synset': 'water_pistol.n.01', 'synonyms': ['water_gun', 'squirt_gun'], 'def': 'plaything consisting of a toy pistol that squirts water', 'name': 'water_gun'}, {'frequency': 'c', 'id': 1191, 'synset': 'water_scooter.n.01', 'synonyms': ['water_scooter', 'sea_scooter', 'jet_ski'], 'def': 'a motorboat resembling a motor scooter (NOT A SURFBOARD OR WATER SKI)', 'name': 'water_scooter'}, {'frequency': 'c', 'id': 1192, 'synset': 'water_ski.n.01', 'synonyms': ['water_ski'], 'def': 'broad ski for skimming over water towed by a speedboat (DO NOT MARK WATER)', 'name': 'water_ski'}, {'frequency': 'c', 'id': 1193, 'synset': 'water_tower.n.01', 'synonyms': ['water_tower'], 'def': 'a large reservoir for water', 'name': 'water_tower'}, {'frequency': 'c', 'id': 1194, 'synset': 'watering_can.n.01', 'synonyms': ['watering_can'], 'def': 'a container with a handle and a spout with a perforated nozzle; used to sprinkle water over plants', 'name': 'watering_can'}, {'frequency': 'c', 'id': 1195, 'synset': 'watermelon.n.02', 'synonyms': ['watermelon'], 'def': 'large oblong or roundish melon with a hard green rind and sweet watery red or occasionally yellowish pulp', 'name': 'watermelon'}, {'frequency': 'f', 'id': 1196, 'synset': 'weathervane.n.01', 'synonyms': ['weathervane', 'vane_(weathervane)', 'wind_vane'], 'def': 'mechanical device attached to an elevated structure; rotates freely to show the direction of the wind', 'name': 'weathervane'}, {'frequency': 'c', 'id': 1197, 'synset': 'webcam.n.01', 'synonyms': ['webcam'], 'def': 'a digital camera designed to take digital photographs and transmit them over the internet', 'name': 'webcam'}, {'frequency': 'c', 'id': 1198, 'synset': 'wedding_cake.n.01', 'synonyms': ['wedding_cake', 'bridecake'], 'def': 'a rich cake with two or more tiers and covered with frosting and decorations; served at a wedding reception', 'name': 'wedding_cake'}, {'frequency': 'c', 'id': 1199, 'synset': 'wedding_ring.n.01', 'synonyms': ['wedding_ring', 'wedding_band'], 'def': 'a ring given to the bride and/or groom at the wedding', 'name': 'wedding_ring'}, {'frequency': 'f', 'id': 1200, 'synset': 'wet_suit.n.01', 'synonyms': ['wet_suit'], 'def': 'a close-fitting garment made of a permeable material; worn in cold water to retain body heat', 'name': 'wet_suit'}, {'frequency': 'f', 'id': 1201, 'synset': 'wheel.n.01', 'synonyms': ['wheel'], 'def': 'a circular frame with spokes (or a solid disc) that can rotate on a shaft or axle', 'name': 'wheel'}, {'frequency': 'c', 'id': 1202, 'synset': 'wheelchair.n.01', 'synonyms': ['wheelchair'], 'def': 'a movable chair mounted on large wheels', 'name': 'wheelchair'}, {'frequency': 'c', 'id': 1203, 'synset': 'whipped_cream.n.01', 'synonyms': ['whipped_cream'], 'def': 'cream that has been beaten until light and fluffy', 'name': 'whipped_cream'}, {'frequency': 'r', 'id': 1204, 'synset': 'whiskey.n.01', 'synonyms': ['whiskey'], 'def': 'a liquor made from fermented mash of grain', 'name': 'whiskey'}, {'frequency': 'r', 'id': 1205, 'synset': 'whistle.n.03', 'synonyms': ['whistle'], 'def': 'a small wind instrument that produces a whistling sound by blowing into it', 'name': 'whistle'}, {'frequency': 'r', 'id': 1206, 'synset': 'wick.n.02', 'synonyms': ['wick'], 'def': 'a loosely woven cord in a candle or oil lamp that is lit on fire', 'name': 'wick'}, {'frequency': 'c', 'id': 1207, 'synset': 'wig.n.01', 'synonyms': ['wig'], 'def': 'hairpiece covering the head and made of real or synthetic hair', 'name': 'wig'}, {'frequency': 'c', 'id': 1208, 'synset': 'wind_chime.n.01', 'synonyms': ['wind_chime'], 'def': 'a decorative arrangement of pieces of metal or glass or pottery that hang together loosely so the wind can cause them to tinkle', 'name': 'wind_chime'}, {'frequency': 'c', 'id': 1209, 'synset': 'windmill.n.01', 'synonyms': ['windmill'], 'def': 'a mill that is powered by the wind', 'name': 'windmill'}, {'frequency': 'c', 'id': 1210, 'synset': 'window_box.n.01', 'synonyms': ['window_box_(for_plants)'], 'def': 'a container for growing plants on a windowsill', 'name': 'window_box_(for_plants)'}, {'frequency': 'f', 'id': 1211, 'synset': 'windshield_wiper.n.01', 'synonyms': ['windshield_wiper', 'windscreen_wiper', 'wiper_(for_windshield/screen)'], 'def': 'a mechanical device that cleans the windshield', 'name': 'windshield_wiper'}, {'frequency': 'c', 'id': 1212, 'synset': 'windsock.n.01', 'synonyms': ['windsock', 'air_sock', 'air-sleeve', 'wind_sleeve', 'wind_cone'], 'def': 'a truncated cloth cone mounted on a mast/pole; shows wind direction', 'name': 'windsock'}, {'frequency': 'f', 'id': 1213, 'synset': 'wine_bottle.n.01', 'synonyms': ['wine_bottle'], 'def': 'a bottle for holding wine', 'name': 'wine_bottle'}, {'frequency': 'r', 'id': 1214, 'synset': 'wine_bucket.n.01', 'synonyms': ['wine_bucket', 'wine_cooler'], 'def': 'a bucket of ice used to chill a bottle of wine', 'name': 'wine_bucket'}, {'frequency': 'f', 'id': 1215, 'synset': 'wineglass.n.01', 'synonyms': ['wineglass'], 'def': 'a glass that has a stem and in which wine is served', 'name': 'wineglass'}, {'frequency': 'r', 'id': 1216, 'synset': 'wing_chair.n.01', 'synonyms': ['wing_chair'], 'def': 'easy chair having wings on each side of a high back', 'name': 'wing_chair'}, {'frequency': 'c', 'id': 1217, 'synset': 'winker.n.02', 'synonyms': ['blinder_(for_horses)'], 'def': 'blinds that prevent a horse from seeing something on either side', 'name': 'blinder_(for_horses)'}, {'frequency': 'c', 'id': 1218, 'synset': 'wok.n.01', 'synonyms': ['wok'], 'def': 'pan with a convex bottom; used for frying in Chinese cooking', 'name': 'wok'}, {'frequency': 'r', 'id': 1219, 'synset': 'wolf.n.01', 'synonyms': ['wolf'], 'def': 'a wild carnivorous mammal of the dog family, living and hunting in packs', 'name': 'wolf'}, {'frequency': 'c', 'id': 1220, 'synset': 'wooden_spoon.n.02', 'synonyms': ['wooden_spoon'], 'def': 'a spoon made of wood', 'name': 'wooden_spoon'}, {'frequency': 'c', 'id': 1221, 'synset': 'wreath.n.01', 'synonyms': ['wreath'], 'def': 'an arrangement of flowers, leaves, or stems fastened in a ring', 'name': 'wreath'}, {'frequency': 'c', 'id': 1222, 'synset': 'wrench.n.03', 'synonyms': ['wrench', 'spanner'], 'def': 'a hand tool that is used to hold or twist a nut or bolt', 'name': 'wrench'}, {'frequency': 'c', 'id': 1223, 'synset': 'wristband.n.01', 'synonyms': ['wristband'], 'def': 'band consisting of a part of a sleeve that covers the wrist', 'name': 'wristband'}, {'frequency': 'f', 'id': 1224, 'synset': 'wristlet.n.01', 'synonyms': ['wristlet', 'wrist_band'], 'def': 'a band or bracelet worn around the wrist', 'name': 'wristlet'}, {'frequency': 'r', 'id': 1225, 'synset': 'yacht.n.01', 'synonyms': ['yacht'], 'def': 'an expensive vessel propelled by sail or power and used for cruising or racing', 'name': 'yacht'}, {'frequency': 'r', 'id': 1226, 'synset': 'yak.n.02', 'synonyms': ['yak'], 'def': 'large long-haired wild ox of Tibet often domesticated', 'name': 'yak'}, {'frequency': 'c', 'id': 1227, 'synset': 'yogurt.n.01', 'synonyms': ['yogurt', 'yoghurt', 'yoghourt'], 'def': 'a custard-like food made from curdled milk', 'name': 'yogurt'}, {'frequency': 'r', 'id': 1228, 'synset': 'yoke.n.07', 'synonyms': ['yoke_(animal_equipment)'], 'def': 'gear joining two animals at the neck; NOT egg yolk', 'name': 'yoke_(animal_equipment)'}, {'frequency': 'f', 'id': 1229, 'synset': 'zebra.n.01', 'synonyms': ['zebra'], 'def': 'any of several fleet black-and-white striped African equines', 'name': 'zebra'}, {'frequency': 'c', 'id': 1230, 'synset': 'zucchini.n.02', 'synonyms': ['zucchini', 'courgette'], 'def': 'small cucumber-shaped vegetable marrow; typically dark green', 'name': 'zucchini'}] # noqa +# fmt: on diff --git a/detectron2/data/datasets/lvis_v1_categories.py b/detectron2/data/datasets/lvis_v1_categories.py new file mode 100644 index 0000000000000000000000000000000000000000..7374e6968bb006f5d8c49e75d9d3b31ea3d77d05 --- /dev/null +++ b/detectron2/data/datasets/lvis_v1_categories.py @@ -0,0 +1,16 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# Autogen with +# with open("lvis_v1_val.json", "r") as f: +# a = json.load(f) +# c = a["categories"] +# for x in c: +# del x["image_count"] +# del x["instance_count"] +# LVIS_CATEGORIES = repr(c) + " # noqa" +# with open("/tmp/lvis_categories.py", "wt") as f: +# f.write(f"LVIS_CATEGORIES = {LVIS_CATEGORIES}") +# Then paste the contents of that file below + +# fmt: off +LVIS_CATEGORIES = [{'frequency': 'c', 'synset': 'aerosol.n.02', 'synonyms': ['aerosol_can', 'spray_can'], 'id': 1, 'def': 'a dispenser that holds a substance under pressure', 'name': 'aerosol_can'}, {'frequency': 'f', 'synset': 'air_conditioner.n.01', 'synonyms': ['air_conditioner'], 'id': 2, 'def': 'a machine that keeps air cool and dry', 'name': 'air_conditioner'}, {'frequency': 'f', 'synset': 'airplane.n.01', 'synonyms': ['airplane', 'aeroplane'], 'id': 3, 'def': 'an aircraft that has a fixed wing and is powered by propellers or jets', 'name': 'airplane'}, {'frequency': 'f', 'synset': 'alarm_clock.n.01', 'synonyms': ['alarm_clock'], 'id': 4, 'def': 'a clock that wakes a sleeper at some preset time', 'name': 'alarm_clock'}, {'frequency': 'c', 'synset': 'alcohol.n.01', 'synonyms': ['alcohol', 'alcoholic_beverage'], 'id': 5, 'def': 'a liquor or brew containing alcohol as the active agent', 'name': 'alcohol'}, {'frequency': 'c', 'synset': 'alligator.n.02', 'synonyms': ['alligator', 'gator'], 'id': 6, 'def': 'amphibious reptiles related to crocodiles but with shorter broader snouts', 'name': 'alligator'}, {'frequency': 'c', 'synset': 'almond.n.02', 'synonyms': ['almond'], 'id': 7, 'def': 'oval-shaped edible seed of the almond tree', 'name': 'almond'}, {'frequency': 'c', 'synset': 'ambulance.n.01', 'synonyms': ['ambulance'], 'id': 8, 'def': 'a vehicle that takes people to and from hospitals', 'name': 'ambulance'}, {'frequency': 'c', 'synset': 'amplifier.n.01', 'synonyms': ['amplifier'], 'id': 9, 'def': 'electronic equipment that increases strength of signals', 'name': 'amplifier'}, {'frequency': 'c', 'synset': 'anklet.n.03', 'synonyms': ['anklet', 'ankle_bracelet'], 'id': 10, 'def': 'an ornament worn around the ankle', 'name': 'anklet'}, {'frequency': 'f', 'synset': 'antenna.n.01', 'synonyms': ['antenna', 'aerial', 'transmitting_aerial'], 'id': 11, 'def': 'an electrical device that sends or receives radio or television signals', 'name': 'antenna'}, {'frequency': 'f', 'synset': 'apple.n.01', 'synonyms': ['apple'], 'id': 12, 'def': 'fruit with red or yellow or green skin and sweet to tart crisp whitish flesh', 'name': 'apple'}, {'frequency': 'r', 'synset': 'applesauce.n.01', 'synonyms': ['applesauce'], 'id': 13, 'def': 'puree of stewed apples usually sweetened and spiced', 'name': 'applesauce'}, {'frequency': 'r', 'synset': 'apricot.n.02', 'synonyms': ['apricot'], 'id': 14, 'def': 'downy yellow to rosy-colored fruit resembling a small peach', 'name': 'apricot'}, {'frequency': 'f', 'synset': 'apron.n.01', 'synonyms': ['apron'], 'id': 15, 'def': 'a garment of cloth that is tied about the waist and worn to protect clothing', 'name': 'apron'}, {'frequency': 'c', 'synset': 'aquarium.n.01', 'synonyms': ['aquarium', 'fish_tank'], 'id': 16, 'def': 'a tank/pool/bowl filled with water for keeping live fish and underwater animals', 'name': 'aquarium'}, {'frequency': 'r', 'synset': 'arctic.n.02', 'synonyms': ['arctic_(type_of_shoe)', 'galosh', 'golosh', 'rubber_(type_of_shoe)', 'gumshoe'], 'id': 17, 'def': 'a waterproof overshoe that protects shoes from water or snow', 'name': 'arctic_(type_of_shoe)'}, {'frequency': 'c', 'synset': 'armband.n.02', 'synonyms': ['armband'], 'id': 18, 'def': 'a band worn around the upper arm', 'name': 'armband'}, {'frequency': 'f', 'synset': 'armchair.n.01', 'synonyms': ['armchair'], 'id': 19, 'def': 'chair with a support on each side for arms', 'name': 'armchair'}, {'frequency': 'r', 'synset': 'armoire.n.01', 'synonyms': ['armoire'], 'id': 20, 'def': 'a large wardrobe or cabinet', 'name': 'armoire'}, {'frequency': 'r', 'synset': 'armor.n.01', 'synonyms': ['armor', 'armour'], 'id': 21, 'def': 'protective covering made of metal and used in combat', 'name': 'armor'}, {'frequency': 'c', 'synset': 'artichoke.n.02', 'synonyms': ['artichoke'], 'id': 22, 'def': 'a thistlelike flower head with edible fleshy leaves and heart', 'name': 'artichoke'}, {'frequency': 'f', 'synset': 'ashcan.n.01', 'synonyms': ['trash_can', 'garbage_can', 'wastebin', 'dustbin', 'trash_barrel', 'trash_bin'], 'id': 23, 'def': 'a bin that holds rubbish until it is collected', 'name': 'trash_can'}, {'frequency': 'c', 'synset': 'ashtray.n.01', 'synonyms': ['ashtray'], 'id': 24, 'def': "a receptacle for the ash from smokers' cigars or cigarettes", 'name': 'ashtray'}, {'frequency': 'c', 'synset': 'asparagus.n.02', 'synonyms': ['asparagus'], 'id': 25, 'def': 'edible young shoots of the asparagus plant', 'name': 'asparagus'}, {'frequency': 'c', 'synset': 'atomizer.n.01', 'synonyms': ['atomizer', 'atomiser', 'spray', 'sprayer', 'nebulizer', 'nebuliser'], 'id': 26, 'def': 'a dispenser that turns a liquid (such as perfume) into a fine mist', 'name': 'atomizer'}, {'frequency': 'f', 'synset': 'avocado.n.01', 'synonyms': ['avocado'], 'id': 27, 'def': 'a pear-shaped fruit with green or blackish skin and rich yellowish pulp enclosing a single large seed', 'name': 'avocado'}, {'frequency': 'c', 'synset': 'award.n.02', 'synonyms': ['award', 'accolade'], 'id': 28, 'def': 'a tangible symbol signifying approval or distinction', 'name': 'award'}, {'frequency': 'f', 'synset': 'awning.n.01', 'synonyms': ['awning'], 'id': 29, 'def': 'a canopy made of canvas to shelter people or things from rain or sun', 'name': 'awning'}, {'frequency': 'r', 'synset': 'ax.n.01', 'synonyms': ['ax', 'axe'], 'id': 30, 'def': 'an edge tool with a heavy bladed head mounted across a handle', 'name': 'ax'}, {'frequency': 'r', 'synset': 'baboon.n.01', 'synonyms': ['baboon'], 'id': 31, 'def': 'large terrestrial monkeys having doglike muzzles', 'name': 'baboon'}, {'frequency': 'f', 'synset': 'baby_buggy.n.01', 'synonyms': ['baby_buggy', 'baby_carriage', 'perambulator', 'pram', 'stroller'], 'id': 32, 'def': 'a small vehicle with four wheels in which a baby or child is pushed around', 'name': 'baby_buggy'}, {'frequency': 'c', 'synset': 'backboard.n.01', 'synonyms': ['basketball_backboard'], 'id': 33, 'def': 'a raised vertical board with basket attached; used to play basketball', 'name': 'basketball_backboard'}, {'frequency': 'f', 'synset': 'backpack.n.01', 'synonyms': ['backpack', 'knapsack', 'packsack', 'rucksack', 'haversack'], 'id': 34, 'def': 'a bag carried by a strap on your back or shoulder', 'name': 'backpack'}, {'frequency': 'f', 'synset': 'bag.n.04', 'synonyms': ['handbag', 'purse', 'pocketbook'], 'id': 35, 'def': 'a container used for carrying money and small personal items or accessories', 'name': 'handbag'}, {'frequency': 'f', 'synset': 'bag.n.06', 'synonyms': ['suitcase', 'baggage', 'luggage'], 'id': 36, 'def': 'cases used to carry belongings when traveling', 'name': 'suitcase'}, {'frequency': 'c', 'synset': 'bagel.n.01', 'synonyms': ['bagel', 'beigel'], 'id': 37, 'def': 'glazed yeast-raised doughnut-shaped roll with hard crust', 'name': 'bagel'}, {'frequency': 'r', 'synset': 'bagpipe.n.01', 'synonyms': ['bagpipe'], 'id': 38, 'def': 'a tubular wind instrument; the player blows air into a bag and squeezes it out', 'name': 'bagpipe'}, {'frequency': 'r', 'synset': 'baguet.n.01', 'synonyms': ['baguet', 'baguette'], 'id': 39, 'def': 'narrow French stick loaf', 'name': 'baguet'}, {'frequency': 'r', 'synset': 'bait.n.02', 'synonyms': ['bait', 'lure'], 'id': 40, 'def': 'something used to lure fish or other animals into danger so they can be trapped or killed', 'name': 'bait'}, {'frequency': 'f', 'synset': 'ball.n.06', 'synonyms': ['ball'], 'id': 41, 'def': 'a spherical object used as a plaything', 'name': 'ball'}, {'frequency': 'r', 'synset': 'ballet_skirt.n.01', 'synonyms': ['ballet_skirt', 'tutu'], 'id': 42, 'def': 'very short skirt worn by ballerinas', 'name': 'ballet_skirt'}, {'frequency': 'f', 'synset': 'balloon.n.01', 'synonyms': ['balloon'], 'id': 43, 'def': 'large tough nonrigid bag filled with gas or heated air', 'name': 'balloon'}, {'frequency': 'c', 'synset': 'bamboo.n.02', 'synonyms': ['bamboo'], 'id': 44, 'def': 'woody tropical grass having hollow woody stems', 'name': 'bamboo'}, {'frequency': 'f', 'synset': 'banana.n.02', 'synonyms': ['banana'], 'id': 45, 'def': 'elongated crescent-shaped yellow fruit with soft sweet flesh', 'name': 'banana'}, {'frequency': 'c', 'synset': 'band_aid.n.01', 'synonyms': ['Band_Aid'], 'id': 46, 'def': 'trade name for an adhesive bandage to cover small cuts or blisters', 'name': 'Band_Aid'}, {'frequency': 'c', 'synset': 'bandage.n.01', 'synonyms': ['bandage'], 'id': 47, 'def': 'a piece of soft material that covers and protects an injured part of the body', 'name': 'bandage'}, {'frequency': 'f', 'synset': 'bandanna.n.01', 'synonyms': ['bandanna', 'bandana'], 'id': 48, 'def': 'large and brightly colored handkerchief; often used as a neckerchief', 'name': 'bandanna'}, {'frequency': 'r', 'synset': 'banjo.n.01', 'synonyms': ['banjo'], 'id': 49, 'def': 'a stringed instrument of the guitar family with a long neck and circular body', 'name': 'banjo'}, {'frequency': 'f', 'synset': 'banner.n.01', 'synonyms': ['banner', 'streamer'], 'id': 50, 'def': 'long strip of cloth or paper used for decoration or advertising', 'name': 'banner'}, {'frequency': 'r', 'synset': 'barbell.n.01', 'synonyms': ['barbell'], 'id': 51, 'def': 'a bar to which heavy discs are attached at each end; used in weightlifting', 'name': 'barbell'}, {'frequency': 'r', 'synset': 'barge.n.01', 'synonyms': ['barge'], 'id': 52, 'def': 'a flatbottom boat for carrying heavy loads (especially on canals)', 'name': 'barge'}, {'frequency': 'f', 'synset': 'barrel.n.02', 'synonyms': ['barrel', 'cask'], 'id': 53, 'def': 'a cylindrical container that holds liquids', 'name': 'barrel'}, {'frequency': 'c', 'synset': 'barrette.n.01', 'synonyms': ['barrette'], 'id': 54, 'def': "a pin for holding women's hair in place", 'name': 'barrette'}, {'frequency': 'c', 'synset': 'barrow.n.03', 'synonyms': ['barrow', 'garden_cart', 'lawn_cart', 'wheelbarrow'], 'id': 55, 'def': 'a cart for carrying small loads; has handles and one or more wheels', 'name': 'barrow'}, {'frequency': 'f', 'synset': 'base.n.03', 'synonyms': ['baseball_base'], 'id': 56, 'def': 'a place that the runner must touch before scoring', 'name': 'baseball_base'}, {'frequency': 'f', 'synset': 'baseball.n.02', 'synonyms': ['baseball'], 'id': 57, 'def': 'a ball used in playing baseball', 'name': 'baseball'}, {'frequency': 'f', 'synset': 'baseball_bat.n.01', 'synonyms': ['baseball_bat'], 'id': 58, 'def': 'an implement used in baseball by the batter', 'name': 'baseball_bat'}, {'frequency': 'f', 'synset': 'baseball_cap.n.01', 'synonyms': ['baseball_cap', 'jockey_cap', 'golf_cap'], 'id': 59, 'def': 'a cap with a bill', 'name': 'baseball_cap'}, {'frequency': 'f', 'synset': 'baseball_glove.n.01', 'synonyms': ['baseball_glove', 'baseball_mitt'], 'id': 60, 'def': 'the handwear used by fielders in playing baseball', 'name': 'baseball_glove'}, {'frequency': 'f', 'synset': 'basket.n.01', 'synonyms': ['basket', 'handbasket'], 'id': 61, 'def': 'a container that is usually woven and has handles', 'name': 'basket'}, {'frequency': 'c', 'synset': 'basketball.n.02', 'synonyms': ['basketball'], 'id': 62, 'def': 'an inflated ball used in playing basketball', 'name': 'basketball'}, {'frequency': 'r', 'synset': 'bass_horn.n.01', 'synonyms': ['bass_horn', 'sousaphone', 'tuba'], 'id': 63, 'def': 'the lowest brass wind instrument', 'name': 'bass_horn'}, {'frequency': 'c', 'synset': 'bat.n.01', 'synonyms': ['bat_(animal)'], 'id': 64, 'def': 'nocturnal mouselike mammal with forelimbs modified to form membranous wings', 'name': 'bat_(animal)'}, {'frequency': 'f', 'synset': 'bath_mat.n.01', 'synonyms': ['bath_mat'], 'id': 65, 'def': 'a heavy towel or mat to stand on while drying yourself after a bath', 'name': 'bath_mat'}, {'frequency': 'f', 'synset': 'bath_towel.n.01', 'synonyms': ['bath_towel'], 'id': 66, 'def': 'a large towel; to dry yourself after a bath', 'name': 'bath_towel'}, {'frequency': 'c', 'synset': 'bathrobe.n.01', 'synonyms': ['bathrobe'], 'id': 67, 'def': 'a loose-fitting robe of towelling; worn after a bath or swim', 'name': 'bathrobe'}, {'frequency': 'f', 'synset': 'bathtub.n.01', 'synonyms': ['bathtub', 'bathing_tub'], 'id': 68, 'def': 'a large open container that you fill with water and use to wash the body', 'name': 'bathtub'}, {'frequency': 'r', 'synset': 'batter.n.02', 'synonyms': ['batter_(food)'], 'id': 69, 'def': 'a liquid or semiliquid mixture, as of flour, eggs, and milk, used in cooking', 'name': 'batter_(food)'}, {'frequency': 'c', 'synset': 'battery.n.02', 'synonyms': ['battery'], 'id': 70, 'def': 'a portable device that produces electricity', 'name': 'battery'}, {'frequency': 'r', 'synset': 'beach_ball.n.01', 'synonyms': ['beachball'], 'id': 71, 'def': 'large and light ball; for play at the seaside', 'name': 'beachball'}, {'frequency': 'c', 'synset': 'bead.n.01', 'synonyms': ['bead'], 'id': 72, 'def': 'a small ball with a hole through the middle used for ornamentation, jewellery, etc.', 'name': 'bead'}, {'frequency': 'c', 'synset': 'bean_curd.n.01', 'synonyms': ['bean_curd', 'tofu'], 'id': 73, 'def': 'cheeselike food made of curdled soybean milk', 'name': 'bean_curd'}, {'frequency': 'c', 'synset': 'beanbag.n.01', 'synonyms': ['beanbag'], 'id': 74, 'def': 'a bag filled with dried beans or similar items; used in games or to sit on', 'name': 'beanbag'}, {'frequency': 'f', 'synset': 'beanie.n.01', 'synonyms': ['beanie', 'beany'], 'id': 75, 'def': 'a small skullcap; formerly worn by schoolboys and college freshmen', 'name': 'beanie'}, {'frequency': 'f', 'synset': 'bear.n.01', 'synonyms': ['bear'], 'id': 76, 'def': 'large carnivorous or omnivorous mammals with shaggy coats and claws', 'name': 'bear'}, {'frequency': 'f', 'synset': 'bed.n.01', 'synonyms': ['bed'], 'id': 77, 'def': 'a piece of furniture that provides a place to sleep', 'name': 'bed'}, {'frequency': 'r', 'synset': 'bedpan.n.01', 'synonyms': ['bedpan'], 'id': 78, 'def': 'a shallow vessel used by a bedridden patient for defecation and urination', 'name': 'bedpan'}, {'frequency': 'f', 'synset': 'bedspread.n.01', 'synonyms': ['bedspread', 'bedcover', 'bed_covering', 'counterpane', 'spread'], 'id': 79, 'def': 'decorative cover for a bed', 'name': 'bedspread'}, {'frequency': 'f', 'synset': 'beef.n.01', 'synonyms': ['cow'], 'id': 80, 'def': 'cattle/cow', 'name': 'cow'}, {'frequency': 'f', 'synset': 'beef.n.02', 'synonyms': ['beef_(food)', 'boeuf_(food)'], 'id': 81, 'def': 'meat from an adult domestic bovine', 'name': 'beef_(food)'}, {'frequency': 'r', 'synset': 'beeper.n.01', 'synonyms': ['beeper', 'pager'], 'id': 82, 'def': 'an device that beeps when the person carrying it is being paged', 'name': 'beeper'}, {'frequency': 'f', 'synset': 'beer_bottle.n.01', 'synonyms': ['beer_bottle'], 'id': 83, 'def': 'a bottle that holds beer', 'name': 'beer_bottle'}, {'frequency': 'c', 'synset': 'beer_can.n.01', 'synonyms': ['beer_can'], 'id': 84, 'def': 'a can that holds beer', 'name': 'beer_can'}, {'frequency': 'r', 'synset': 'beetle.n.01', 'synonyms': ['beetle'], 'id': 85, 'def': 'insect with hard wing covers', 'name': 'beetle'}, {'frequency': 'f', 'synset': 'bell.n.01', 'synonyms': ['bell'], 'id': 86, 'def': 'a hollow device made of metal that makes a ringing sound when struck', 'name': 'bell'}, {'frequency': 'f', 'synset': 'bell_pepper.n.02', 'synonyms': ['bell_pepper', 'capsicum'], 'id': 87, 'def': 'large bell-shaped sweet pepper in green or red or yellow or orange or black varieties', 'name': 'bell_pepper'}, {'frequency': 'f', 'synset': 'belt.n.02', 'synonyms': ['belt'], 'id': 88, 'def': 'a band to tie or buckle around the body (usually at the waist)', 'name': 'belt'}, {'frequency': 'f', 'synset': 'belt_buckle.n.01', 'synonyms': ['belt_buckle'], 'id': 89, 'def': 'the buckle used to fasten a belt', 'name': 'belt_buckle'}, {'frequency': 'f', 'synset': 'bench.n.01', 'synonyms': ['bench'], 'id': 90, 'def': 'a long seat for more than one person', 'name': 'bench'}, {'frequency': 'c', 'synset': 'beret.n.01', 'synonyms': ['beret'], 'id': 91, 'def': 'a cap with no brim or bill; made of soft cloth', 'name': 'beret'}, {'frequency': 'c', 'synset': 'bib.n.02', 'synonyms': ['bib'], 'id': 92, 'def': 'a napkin tied under the chin of a child while eating', 'name': 'bib'}, {'frequency': 'r', 'synset': 'bible.n.01', 'synonyms': ['Bible'], 'id': 93, 'def': 'the sacred writings of the Christian religions', 'name': 'Bible'}, {'frequency': 'f', 'synset': 'bicycle.n.01', 'synonyms': ['bicycle', 'bike_(bicycle)'], 'id': 94, 'def': 'a wheeled vehicle that has two wheels and is moved by foot pedals', 'name': 'bicycle'}, {'frequency': 'f', 'synset': 'bill.n.09', 'synonyms': ['visor', 'vizor'], 'id': 95, 'def': 'a brim that projects to the front to shade the eyes', 'name': 'visor'}, {'frequency': 'f', 'synset': 'billboard.n.01', 'synonyms': ['billboard'], 'id': 96, 'def': 'large outdoor signboard', 'name': 'billboard'}, {'frequency': 'c', 'synset': 'binder.n.03', 'synonyms': ['binder', 'ring-binder'], 'id': 97, 'def': 'holds loose papers or magazines', 'name': 'binder'}, {'frequency': 'c', 'synset': 'binoculars.n.01', 'synonyms': ['binoculars', 'field_glasses', 'opera_glasses'], 'id': 98, 'def': 'an optical instrument designed for simultaneous use by both eyes', 'name': 'binoculars'}, {'frequency': 'f', 'synset': 'bird.n.01', 'synonyms': ['bird'], 'id': 99, 'def': 'animal characterized by feathers and wings', 'name': 'bird'}, {'frequency': 'c', 'synset': 'bird_feeder.n.01', 'synonyms': ['birdfeeder'], 'id': 100, 'def': 'an outdoor device that supplies food for wild birds', 'name': 'birdfeeder'}, {'frequency': 'c', 'synset': 'birdbath.n.01', 'synonyms': ['birdbath'], 'id': 101, 'def': 'an ornamental basin (usually in a garden) for birds to bathe in', 'name': 'birdbath'}, {'frequency': 'c', 'synset': 'birdcage.n.01', 'synonyms': ['birdcage'], 'id': 102, 'def': 'a cage in which a bird can be kept', 'name': 'birdcage'}, {'frequency': 'c', 'synset': 'birdhouse.n.01', 'synonyms': ['birdhouse'], 'id': 103, 'def': 'a shelter for birds', 'name': 'birdhouse'}, {'frequency': 'f', 'synset': 'birthday_cake.n.01', 'synonyms': ['birthday_cake'], 'id': 104, 'def': 'decorated cake served at a birthday party', 'name': 'birthday_cake'}, {'frequency': 'r', 'synset': 'birthday_card.n.01', 'synonyms': ['birthday_card'], 'id': 105, 'def': 'a card expressing a birthday greeting', 'name': 'birthday_card'}, {'frequency': 'r', 'synset': 'black_flag.n.01', 'synonyms': ['pirate_flag'], 'id': 106, 'def': 'a flag usually bearing a white skull and crossbones on a black background', 'name': 'pirate_flag'}, {'frequency': 'c', 'synset': 'black_sheep.n.02', 'synonyms': ['black_sheep'], 'id': 107, 'def': 'sheep with a black coat', 'name': 'black_sheep'}, {'frequency': 'c', 'synset': 'blackberry.n.01', 'synonyms': ['blackberry'], 'id': 108, 'def': 'large sweet black or very dark purple edible aggregate fruit', 'name': 'blackberry'}, {'frequency': 'f', 'synset': 'blackboard.n.01', 'synonyms': ['blackboard', 'chalkboard'], 'id': 109, 'def': 'sheet of slate; for writing with chalk', 'name': 'blackboard'}, {'frequency': 'f', 'synset': 'blanket.n.01', 'synonyms': ['blanket'], 'id': 110, 'def': 'bedding that keeps a person warm in bed', 'name': 'blanket'}, {'frequency': 'c', 'synset': 'blazer.n.01', 'synonyms': ['blazer', 'sport_jacket', 'sport_coat', 'sports_jacket', 'sports_coat'], 'id': 111, 'def': 'lightweight jacket; often striped in the colors of a club or school', 'name': 'blazer'}, {'frequency': 'f', 'synset': 'blender.n.01', 'synonyms': ['blender', 'liquidizer', 'liquidiser'], 'id': 112, 'def': 'an electrically powered mixer that mix or chop or liquefy foods', 'name': 'blender'}, {'frequency': 'r', 'synset': 'blimp.n.02', 'synonyms': ['blimp'], 'id': 113, 'def': 'a small nonrigid airship used for observation or as a barrage balloon', 'name': 'blimp'}, {'frequency': 'f', 'synset': 'blinker.n.01', 'synonyms': ['blinker', 'flasher'], 'id': 114, 'def': 'a light that flashes on and off; used as a signal or to send messages', 'name': 'blinker'}, {'frequency': 'f', 'synset': 'blouse.n.01', 'synonyms': ['blouse'], 'id': 115, 'def': 'a top worn by women', 'name': 'blouse'}, {'frequency': 'f', 'synset': 'blueberry.n.02', 'synonyms': ['blueberry'], 'id': 116, 'def': 'sweet edible dark-blue berries of blueberry plants', 'name': 'blueberry'}, {'frequency': 'r', 'synset': 'board.n.09', 'synonyms': ['gameboard'], 'id': 117, 'def': 'a flat portable surface (usually rectangular) designed for board games', 'name': 'gameboard'}, {'frequency': 'f', 'synset': 'boat.n.01', 'synonyms': ['boat', 'ship_(boat)'], 'id': 118, 'def': 'a vessel for travel on water', 'name': 'boat'}, {'frequency': 'r', 'synset': 'bob.n.05', 'synonyms': ['bob', 'bobber', 'bobfloat'], 'id': 119, 'def': 'a small float usually made of cork; attached to a fishing line', 'name': 'bob'}, {'frequency': 'c', 'synset': 'bobbin.n.01', 'synonyms': ['bobbin', 'spool', 'reel'], 'id': 120, 'def': 'a thing around which thread/tape/film or other flexible materials can be wound', 'name': 'bobbin'}, {'frequency': 'c', 'synset': 'bobby_pin.n.01', 'synonyms': ['bobby_pin', 'hairgrip'], 'id': 121, 'def': 'a flat wire hairpin used to hold bobbed hair in place', 'name': 'bobby_pin'}, {'frequency': 'c', 'synset': 'boiled_egg.n.01', 'synonyms': ['boiled_egg', 'coddled_egg'], 'id': 122, 'def': 'egg cooked briefly in the shell in gently boiling water', 'name': 'boiled_egg'}, {'frequency': 'r', 'synset': 'bolo_tie.n.01', 'synonyms': ['bolo_tie', 'bolo', 'bola_tie', 'bola'], 'id': 123, 'def': 'a cord fastened around the neck with an ornamental clasp and worn as a necktie', 'name': 'bolo_tie'}, {'frequency': 'c', 'synset': 'bolt.n.03', 'synonyms': ['deadbolt'], 'id': 124, 'def': 'the part of a lock that is engaged or withdrawn with a key', 'name': 'deadbolt'}, {'frequency': 'f', 'synset': 'bolt.n.06', 'synonyms': ['bolt'], 'id': 125, 'def': 'a screw that screws into a nut to form a fastener', 'name': 'bolt'}, {'frequency': 'r', 'synset': 'bonnet.n.01', 'synonyms': ['bonnet'], 'id': 126, 'def': 'a hat tied under the chin', 'name': 'bonnet'}, {'frequency': 'f', 'synset': 'book.n.01', 'synonyms': ['book'], 'id': 127, 'def': 'a written work or composition that has been published', 'name': 'book'}, {'frequency': 'c', 'synset': 'bookcase.n.01', 'synonyms': ['bookcase'], 'id': 128, 'def': 'a piece of furniture with shelves for storing books', 'name': 'bookcase'}, {'frequency': 'c', 'synset': 'booklet.n.01', 'synonyms': ['booklet', 'brochure', 'leaflet', 'pamphlet'], 'id': 129, 'def': 'a small book usually having a paper cover', 'name': 'booklet'}, {'frequency': 'r', 'synset': 'bookmark.n.01', 'synonyms': ['bookmark', 'bookmarker'], 'id': 130, 'def': 'a marker (a piece of paper or ribbon) placed between the pages of a book', 'name': 'bookmark'}, {'frequency': 'r', 'synset': 'boom.n.04', 'synonyms': ['boom_microphone', 'microphone_boom'], 'id': 131, 'def': 'a pole carrying an overhead microphone projected over a film or tv set', 'name': 'boom_microphone'}, {'frequency': 'f', 'synset': 'boot.n.01', 'synonyms': ['boot'], 'id': 132, 'def': 'footwear that covers the whole foot and lower leg', 'name': 'boot'}, {'frequency': 'f', 'synset': 'bottle.n.01', 'synonyms': ['bottle'], 'id': 133, 'def': 'a glass or plastic vessel used for storing drinks or other liquids', 'name': 'bottle'}, {'frequency': 'c', 'synset': 'bottle_opener.n.01', 'synonyms': ['bottle_opener'], 'id': 134, 'def': 'an opener for removing caps or corks from bottles', 'name': 'bottle_opener'}, {'frequency': 'c', 'synset': 'bouquet.n.01', 'synonyms': ['bouquet'], 'id': 135, 'def': 'an arrangement of flowers that is usually given as a present', 'name': 'bouquet'}, {'frequency': 'r', 'synset': 'bow.n.04', 'synonyms': ['bow_(weapon)'], 'id': 136, 'def': 'a weapon for shooting arrows', 'name': 'bow_(weapon)'}, {'frequency': 'f', 'synset': 'bow.n.08', 'synonyms': ['bow_(decorative_ribbons)'], 'id': 137, 'def': 'a decorative interlacing of ribbons', 'name': 'bow_(decorative_ribbons)'}, {'frequency': 'f', 'synset': 'bow_tie.n.01', 'synonyms': ['bow-tie', 'bowtie'], 'id': 138, 'def': "a man's tie that ties in a bow", 'name': 'bow-tie'}, {'frequency': 'f', 'synset': 'bowl.n.03', 'synonyms': ['bowl'], 'id': 139, 'def': 'a dish that is round and open at the top for serving foods', 'name': 'bowl'}, {'frequency': 'r', 'synset': 'bowl.n.08', 'synonyms': ['pipe_bowl'], 'id': 140, 'def': 'a small round container that is open at the top for holding tobacco', 'name': 'pipe_bowl'}, {'frequency': 'c', 'synset': 'bowler_hat.n.01', 'synonyms': ['bowler_hat', 'bowler', 'derby_hat', 'derby', 'plug_hat'], 'id': 141, 'def': 'a felt hat that is round and hard with a narrow brim', 'name': 'bowler_hat'}, {'frequency': 'r', 'synset': 'bowling_ball.n.01', 'synonyms': ['bowling_ball'], 'id': 142, 'def': 'a large ball with finger holes used in the sport of bowling', 'name': 'bowling_ball'}, {'frequency': 'f', 'synset': 'box.n.01', 'synonyms': ['box'], 'id': 143, 'def': 'a (usually rectangular) container; may have a lid', 'name': 'box'}, {'frequency': 'r', 'synset': 'boxing_glove.n.01', 'synonyms': ['boxing_glove'], 'id': 144, 'def': 'large glove coverings the fists of a fighter worn for the sport of boxing', 'name': 'boxing_glove'}, {'frequency': 'c', 'synset': 'brace.n.06', 'synonyms': ['suspenders'], 'id': 145, 'def': 'elastic straps that hold trousers up (usually used in the plural)', 'name': 'suspenders'}, {'frequency': 'f', 'synset': 'bracelet.n.02', 'synonyms': ['bracelet', 'bangle'], 'id': 146, 'def': 'jewelry worn around the wrist for decoration', 'name': 'bracelet'}, {'frequency': 'r', 'synset': 'brass.n.07', 'synonyms': ['brass_plaque'], 'id': 147, 'def': 'a memorial made of brass', 'name': 'brass_plaque'}, {'frequency': 'c', 'synset': 'brassiere.n.01', 'synonyms': ['brassiere', 'bra', 'bandeau'], 'id': 148, 'def': 'an undergarment worn by women to support their breasts', 'name': 'brassiere'}, {'frequency': 'c', 'synset': 'bread-bin.n.01', 'synonyms': ['bread-bin', 'breadbox'], 'id': 149, 'def': 'a container used to keep bread or cake in', 'name': 'bread-bin'}, {'frequency': 'f', 'synset': 'bread.n.01', 'synonyms': ['bread'], 'id': 150, 'def': 'food made from dough of flour or meal and usually raised with yeast or baking powder and then baked', 'name': 'bread'}, {'frequency': 'r', 'synset': 'breechcloth.n.01', 'synonyms': ['breechcloth', 'breechclout', 'loincloth'], 'id': 151, 'def': 'a garment that provides covering for the loins', 'name': 'breechcloth'}, {'frequency': 'f', 'synset': 'bridal_gown.n.01', 'synonyms': ['bridal_gown', 'wedding_gown', 'wedding_dress'], 'id': 152, 'def': 'a gown worn by the bride at a wedding', 'name': 'bridal_gown'}, {'frequency': 'c', 'synset': 'briefcase.n.01', 'synonyms': ['briefcase'], 'id': 153, 'def': 'a case with a handle; for carrying papers or files or books', 'name': 'briefcase'}, {'frequency': 'f', 'synset': 'broccoli.n.01', 'synonyms': ['broccoli'], 'id': 154, 'def': 'plant with dense clusters of tight green flower buds', 'name': 'broccoli'}, {'frequency': 'r', 'synset': 'brooch.n.01', 'synonyms': ['broach'], 'id': 155, 'def': 'a decorative pin worn by women', 'name': 'broach'}, {'frequency': 'c', 'synset': 'broom.n.01', 'synonyms': ['broom'], 'id': 156, 'def': 'bundle of straws or twigs attached to a long handle; used for cleaning', 'name': 'broom'}, {'frequency': 'c', 'synset': 'brownie.n.03', 'synonyms': ['brownie'], 'id': 157, 'def': 'square or bar of very rich chocolate cake usually with nuts', 'name': 'brownie'}, {'frequency': 'c', 'synset': 'brussels_sprouts.n.01', 'synonyms': ['brussels_sprouts'], 'id': 158, 'def': 'the small edible cabbage-like buds growing along a stalk', 'name': 'brussels_sprouts'}, {'frequency': 'r', 'synset': 'bubble_gum.n.01', 'synonyms': ['bubble_gum'], 'id': 159, 'def': 'a kind of chewing gum that can be blown into bubbles', 'name': 'bubble_gum'}, {'frequency': 'f', 'synset': 'bucket.n.01', 'synonyms': ['bucket', 'pail'], 'id': 160, 'def': 'a roughly cylindrical vessel that is open at the top', 'name': 'bucket'}, {'frequency': 'r', 'synset': 'buggy.n.01', 'synonyms': ['horse_buggy'], 'id': 161, 'def': 'a small lightweight carriage; drawn by a single horse', 'name': 'horse_buggy'}, {'frequency': 'c', 'synset': 'bull.n.11', 'synonyms': ['horned_cow'], 'id': 162, 'def': 'a cow with horns', 'name': 'bull'}, {'frequency': 'c', 'synset': 'bulldog.n.01', 'synonyms': ['bulldog'], 'id': 163, 'def': 'a thickset short-haired dog with a large head and strong undershot lower jaw', 'name': 'bulldog'}, {'frequency': 'r', 'synset': 'bulldozer.n.01', 'synonyms': ['bulldozer', 'dozer'], 'id': 164, 'def': 'large powerful tractor; a large blade in front flattens areas of ground', 'name': 'bulldozer'}, {'frequency': 'c', 'synset': 'bullet_train.n.01', 'synonyms': ['bullet_train'], 'id': 165, 'def': 'a high-speed passenger train', 'name': 'bullet_train'}, {'frequency': 'c', 'synset': 'bulletin_board.n.02', 'synonyms': ['bulletin_board', 'notice_board'], 'id': 166, 'def': 'a board that hangs on a wall; displays announcements', 'name': 'bulletin_board'}, {'frequency': 'r', 'synset': 'bulletproof_vest.n.01', 'synonyms': ['bulletproof_vest'], 'id': 167, 'def': 'a vest capable of resisting the impact of a bullet', 'name': 'bulletproof_vest'}, {'frequency': 'c', 'synset': 'bullhorn.n.01', 'synonyms': ['bullhorn', 'megaphone'], 'id': 168, 'def': 'a portable loudspeaker with built-in microphone and amplifier', 'name': 'bullhorn'}, {'frequency': 'f', 'synset': 'bun.n.01', 'synonyms': ['bun', 'roll'], 'id': 169, 'def': 'small rounded bread either plain or sweet', 'name': 'bun'}, {'frequency': 'c', 'synset': 'bunk_bed.n.01', 'synonyms': ['bunk_bed'], 'id': 170, 'def': 'beds built one above the other', 'name': 'bunk_bed'}, {'frequency': 'f', 'synset': 'buoy.n.01', 'synonyms': ['buoy'], 'id': 171, 'def': 'a float attached by rope to the seabed to mark channels in a harbor or underwater hazards', 'name': 'buoy'}, {'frequency': 'r', 'synset': 'burrito.n.01', 'synonyms': ['burrito'], 'id': 172, 'def': 'a flour tortilla folded around a filling', 'name': 'burrito'}, {'frequency': 'f', 'synset': 'bus.n.01', 'synonyms': ['bus_(vehicle)', 'autobus', 'charabanc', 'double-decker', 'motorbus', 'motorcoach'], 'id': 173, 'def': 'a vehicle carrying many passengers; used for public transport', 'name': 'bus_(vehicle)'}, {'frequency': 'c', 'synset': 'business_card.n.01', 'synonyms': ['business_card'], 'id': 174, 'def': "a card on which are printed the person's name and business affiliation", 'name': 'business_card'}, {'frequency': 'f', 'synset': 'butter.n.01', 'synonyms': ['butter'], 'id': 175, 'def': 'an edible emulsion of fat globules made by churning milk or cream; for cooking and table use', 'name': 'butter'}, {'frequency': 'c', 'synset': 'butterfly.n.01', 'synonyms': ['butterfly'], 'id': 176, 'def': 'insect typically having a slender body with knobbed antennae and broad colorful wings', 'name': 'butterfly'}, {'frequency': 'f', 'synset': 'button.n.01', 'synonyms': ['button'], 'id': 177, 'def': 'a round fastener sewn to shirts and coats etc to fit through buttonholes', 'name': 'button'}, {'frequency': 'f', 'synset': 'cab.n.03', 'synonyms': ['cab_(taxi)', 'taxi', 'taxicab'], 'id': 178, 'def': 'a car that takes passengers where they want to go in exchange for money', 'name': 'cab_(taxi)'}, {'frequency': 'r', 'synset': 'cabana.n.01', 'synonyms': ['cabana'], 'id': 179, 'def': 'a small tent used as a dressing room beside the sea or a swimming pool', 'name': 'cabana'}, {'frequency': 'c', 'synset': 'cabin_car.n.01', 'synonyms': ['cabin_car', 'caboose'], 'id': 180, 'def': 'a car on a freight train for use of the train crew; usually the last car on the train', 'name': 'cabin_car'}, {'frequency': 'f', 'synset': 'cabinet.n.01', 'synonyms': ['cabinet'], 'id': 181, 'def': 'a piece of furniture resembling a cupboard with doors and shelves and drawers', 'name': 'cabinet'}, {'frequency': 'r', 'synset': 'cabinet.n.03', 'synonyms': ['locker', 'storage_locker'], 'id': 182, 'def': 'a storage compartment for clothes and valuables; usually it has a lock', 'name': 'locker'}, {'frequency': 'f', 'synset': 'cake.n.03', 'synonyms': ['cake'], 'id': 183, 'def': 'baked goods made from or based on a mixture of flour, sugar, eggs, and fat', 'name': 'cake'}, {'frequency': 'c', 'synset': 'calculator.n.02', 'synonyms': ['calculator'], 'id': 184, 'def': 'a small machine that is used for mathematical calculations', 'name': 'calculator'}, {'frequency': 'f', 'synset': 'calendar.n.02', 'synonyms': ['calendar'], 'id': 185, 'def': 'a list or register of events (appointments/social events/court cases, etc)', 'name': 'calendar'}, {'frequency': 'c', 'synset': 'calf.n.01', 'synonyms': ['calf'], 'id': 186, 'def': 'young of domestic cattle', 'name': 'calf'}, {'frequency': 'c', 'synset': 'camcorder.n.01', 'synonyms': ['camcorder'], 'id': 187, 'def': 'a portable television camera and videocassette recorder', 'name': 'camcorder'}, {'frequency': 'c', 'synset': 'camel.n.01', 'synonyms': ['camel'], 'id': 188, 'def': 'cud-chewing mammal used as a draft or saddle animal in desert regions', 'name': 'camel'}, {'frequency': 'f', 'synset': 'camera.n.01', 'synonyms': ['camera'], 'id': 189, 'def': 'equipment for taking photographs', 'name': 'camera'}, {'frequency': 'c', 'synset': 'camera_lens.n.01', 'synonyms': ['camera_lens'], 'id': 190, 'def': 'a lens that focuses the image in a camera', 'name': 'camera_lens'}, {'frequency': 'c', 'synset': 'camper.n.02', 'synonyms': ['camper_(vehicle)', 'camping_bus', 'motor_home'], 'id': 191, 'def': 'a recreational vehicle equipped for camping out while traveling', 'name': 'camper_(vehicle)'}, {'frequency': 'f', 'synset': 'can.n.01', 'synonyms': ['can', 'tin_can'], 'id': 192, 'def': 'airtight sealed metal container for food or drink or paint etc.', 'name': 'can'}, {'frequency': 'c', 'synset': 'can_opener.n.01', 'synonyms': ['can_opener', 'tin_opener'], 'id': 193, 'def': 'a device for cutting cans open', 'name': 'can_opener'}, {'frequency': 'f', 'synset': 'candle.n.01', 'synonyms': ['candle', 'candlestick'], 'id': 194, 'def': 'stick of wax with a wick in the middle', 'name': 'candle'}, {'frequency': 'f', 'synset': 'candlestick.n.01', 'synonyms': ['candle_holder'], 'id': 195, 'def': 'a holder with sockets for candles', 'name': 'candle_holder'}, {'frequency': 'r', 'synset': 'candy_bar.n.01', 'synonyms': ['candy_bar'], 'id': 196, 'def': 'a candy shaped as a bar', 'name': 'candy_bar'}, {'frequency': 'c', 'synset': 'candy_cane.n.01', 'synonyms': ['candy_cane'], 'id': 197, 'def': 'a hard candy in the shape of a rod (usually with stripes)', 'name': 'candy_cane'}, {'frequency': 'c', 'synset': 'cane.n.01', 'synonyms': ['walking_cane'], 'id': 198, 'def': 'a stick that people can lean on to help them walk', 'name': 'walking_cane'}, {'frequency': 'c', 'synset': 'canister.n.02', 'synonyms': ['canister', 'cannister'], 'id': 199, 'def': 'metal container for storing dry foods such as tea or flour', 'name': 'canister'}, {'frequency': 'c', 'synset': 'canoe.n.01', 'synonyms': ['canoe'], 'id': 200, 'def': 'small and light boat; pointed at both ends; propelled with a paddle', 'name': 'canoe'}, {'frequency': 'c', 'synset': 'cantaloup.n.02', 'synonyms': ['cantaloup', 'cantaloupe'], 'id': 201, 'def': 'the fruit of a cantaloup vine; small to medium-sized melon with yellowish flesh', 'name': 'cantaloup'}, {'frequency': 'r', 'synset': 'canteen.n.01', 'synonyms': ['canteen'], 'id': 202, 'def': 'a flask for carrying water; used by soldiers or travelers', 'name': 'canteen'}, {'frequency': 'f', 'synset': 'cap.n.01', 'synonyms': ['cap_(headwear)'], 'id': 203, 'def': 'a tight-fitting headwear', 'name': 'cap_(headwear)'}, {'frequency': 'f', 'synset': 'cap.n.02', 'synonyms': ['bottle_cap', 'cap_(container_lid)'], 'id': 204, 'def': 'a top (as for a bottle)', 'name': 'bottle_cap'}, {'frequency': 'c', 'synset': 'cape.n.02', 'synonyms': ['cape'], 'id': 205, 'def': 'a sleeveless garment like a cloak but shorter', 'name': 'cape'}, {'frequency': 'c', 'synset': 'cappuccino.n.01', 'synonyms': ['cappuccino', 'coffee_cappuccino'], 'id': 206, 'def': 'equal parts of espresso and steamed milk', 'name': 'cappuccino'}, {'frequency': 'f', 'synset': 'car.n.01', 'synonyms': ['car_(automobile)', 'auto_(automobile)', 'automobile'], 'id': 207, 'def': 'a motor vehicle with four wheels', 'name': 'car_(automobile)'}, {'frequency': 'f', 'synset': 'car.n.02', 'synonyms': ['railcar_(part_of_a_train)', 'railway_car_(part_of_a_train)', 'railroad_car_(part_of_a_train)'], 'id': 208, 'def': 'a wheeled vehicle adapted to the rails of railroad (mark each individual railcar separately)', 'name': 'railcar_(part_of_a_train)'}, {'frequency': 'r', 'synset': 'car.n.04', 'synonyms': ['elevator_car'], 'id': 209, 'def': 'where passengers ride up and down', 'name': 'elevator_car'}, {'frequency': 'r', 'synset': 'car_battery.n.01', 'synonyms': ['car_battery', 'automobile_battery'], 'id': 210, 'def': 'a battery in a motor vehicle', 'name': 'car_battery'}, {'frequency': 'c', 'synset': 'card.n.02', 'synonyms': ['identity_card'], 'id': 211, 'def': 'a card certifying the identity of the bearer', 'name': 'identity_card'}, {'frequency': 'c', 'synset': 'card.n.03', 'synonyms': ['card'], 'id': 212, 'def': 'a rectangular piece of paper used to send messages (e.g. greetings or pictures)', 'name': 'card'}, {'frequency': 'c', 'synset': 'cardigan.n.01', 'synonyms': ['cardigan'], 'id': 213, 'def': 'knitted jacket that is fastened up the front with buttons or a zipper', 'name': 'cardigan'}, {'frequency': 'r', 'synset': 'cargo_ship.n.01', 'synonyms': ['cargo_ship', 'cargo_vessel'], 'id': 214, 'def': 'a ship designed to carry cargo', 'name': 'cargo_ship'}, {'frequency': 'r', 'synset': 'carnation.n.01', 'synonyms': ['carnation'], 'id': 215, 'def': 'plant with pink to purple-red spice-scented usually double flowers', 'name': 'carnation'}, {'frequency': 'c', 'synset': 'carriage.n.02', 'synonyms': ['horse_carriage'], 'id': 216, 'def': 'a vehicle with wheels drawn by one or more horses', 'name': 'horse_carriage'}, {'frequency': 'f', 'synset': 'carrot.n.01', 'synonyms': ['carrot'], 'id': 217, 'def': 'deep orange edible root of the cultivated carrot plant', 'name': 'carrot'}, {'frequency': 'f', 'synset': 'carryall.n.01', 'synonyms': ['tote_bag'], 'id': 218, 'def': 'a capacious bag or basket', 'name': 'tote_bag'}, {'frequency': 'c', 'synset': 'cart.n.01', 'synonyms': ['cart'], 'id': 219, 'def': 'a heavy open wagon usually having two wheels and drawn by an animal', 'name': 'cart'}, {'frequency': 'c', 'synset': 'carton.n.02', 'synonyms': ['carton'], 'id': 220, 'def': 'a container made of cardboard for holding food or drink', 'name': 'carton'}, {'frequency': 'c', 'synset': 'cash_register.n.01', 'synonyms': ['cash_register', 'register_(for_cash_transactions)'], 'id': 221, 'def': 'a cashbox with an adding machine to register transactions', 'name': 'cash_register'}, {'frequency': 'r', 'synset': 'casserole.n.01', 'synonyms': ['casserole'], 'id': 222, 'def': 'food cooked and served in a casserole', 'name': 'casserole'}, {'frequency': 'r', 'synset': 'cassette.n.01', 'synonyms': ['cassette'], 'id': 223, 'def': 'a container that holds a magnetic tape used for recording or playing sound or video', 'name': 'cassette'}, {'frequency': 'c', 'synset': 'cast.n.05', 'synonyms': ['cast', 'plaster_cast', 'plaster_bandage'], 'id': 224, 'def': 'bandage consisting of a firm covering that immobilizes broken bones while they heal', 'name': 'cast'}, {'frequency': 'f', 'synset': 'cat.n.01', 'synonyms': ['cat'], 'id': 225, 'def': 'a domestic house cat', 'name': 'cat'}, {'frequency': 'f', 'synset': 'cauliflower.n.02', 'synonyms': ['cauliflower'], 'id': 226, 'def': 'edible compact head of white undeveloped flowers', 'name': 'cauliflower'}, {'frequency': 'c', 'synset': 'cayenne.n.02', 'synonyms': ['cayenne_(spice)', 'cayenne_pepper_(spice)', 'red_pepper_(spice)'], 'id': 227, 'def': 'ground pods and seeds of pungent red peppers of the genus Capsicum', 'name': 'cayenne_(spice)'}, {'frequency': 'c', 'synset': 'cd_player.n.01', 'synonyms': ['CD_player'], 'id': 228, 'def': 'electronic equipment for playing compact discs (CDs)', 'name': 'CD_player'}, {'frequency': 'f', 'synset': 'celery.n.01', 'synonyms': ['celery'], 'id': 229, 'def': 'widely cultivated herb with aromatic leaf stalks that are eaten raw or cooked', 'name': 'celery'}, {'frequency': 'f', 'synset': 'cellular_telephone.n.01', 'synonyms': ['cellular_telephone', 'cellular_phone', 'cellphone', 'mobile_phone', 'smart_phone'], 'id': 230, 'def': 'a hand-held mobile telephone', 'name': 'cellular_telephone'}, {'frequency': 'r', 'synset': 'chain_mail.n.01', 'synonyms': ['chain_mail', 'ring_mail', 'chain_armor', 'chain_armour', 'ring_armor', 'ring_armour'], 'id': 231, 'def': '(Middle Ages) flexible armor made of interlinked metal rings', 'name': 'chain_mail'}, {'frequency': 'f', 'synset': 'chair.n.01', 'synonyms': ['chair'], 'id': 232, 'def': 'a seat for one person, with a support for the back', 'name': 'chair'}, {'frequency': 'r', 'synset': 'chaise_longue.n.01', 'synonyms': ['chaise_longue', 'chaise', 'daybed'], 'id': 233, 'def': 'a long chair; for reclining', 'name': 'chaise_longue'}, {'frequency': 'r', 'synset': 'chalice.n.01', 'synonyms': ['chalice'], 'id': 234, 'def': 'a bowl-shaped drinking vessel; especially the Eucharistic cup', 'name': 'chalice'}, {'frequency': 'f', 'synset': 'chandelier.n.01', 'synonyms': ['chandelier'], 'id': 235, 'def': 'branched lighting fixture; often ornate; hangs from the ceiling', 'name': 'chandelier'}, {'frequency': 'r', 'synset': 'chap.n.04', 'synonyms': ['chap'], 'id': 236, 'def': 'leather leggings without a seat; worn over trousers by cowboys to protect their legs', 'name': 'chap'}, {'frequency': 'r', 'synset': 'checkbook.n.01', 'synonyms': ['checkbook', 'chequebook'], 'id': 237, 'def': 'a book issued to holders of checking accounts', 'name': 'checkbook'}, {'frequency': 'r', 'synset': 'checkerboard.n.01', 'synonyms': ['checkerboard'], 'id': 238, 'def': 'a board having 64 squares of two alternating colors', 'name': 'checkerboard'}, {'frequency': 'c', 'synset': 'cherry.n.03', 'synonyms': ['cherry'], 'id': 239, 'def': 'a red fruit with a single hard stone', 'name': 'cherry'}, {'frequency': 'r', 'synset': 'chessboard.n.01', 'synonyms': ['chessboard'], 'id': 240, 'def': 'a checkerboard used to play chess', 'name': 'chessboard'}, {'frequency': 'c', 'synset': 'chicken.n.02', 'synonyms': ['chicken_(animal)'], 'id': 241, 'def': 'a domestic fowl bred for flesh or eggs', 'name': 'chicken_(animal)'}, {'frequency': 'c', 'synset': 'chickpea.n.01', 'synonyms': ['chickpea', 'garbanzo'], 'id': 242, 'def': 'the seed of the chickpea plant; usually dried', 'name': 'chickpea'}, {'frequency': 'c', 'synset': 'chili.n.02', 'synonyms': ['chili_(vegetable)', 'chili_pepper_(vegetable)', 'chilli_(vegetable)', 'chilly_(vegetable)', 'chile_(vegetable)'], 'id': 243, 'def': 'very hot and finely tapering pepper of special pungency', 'name': 'chili_(vegetable)'}, {'frequency': 'r', 'synset': 'chime.n.01', 'synonyms': ['chime', 'gong'], 'id': 244, 'def': 'an instrument consisting of a set of bells that are struck with a hammer', 'name': 'chime'}, {'frequency': 'r', 'synset': 'chinaware.n.01', 'synonyms': ['chinaware'], 'id': 245, 'def': 'dishware made of high quality porcelain', 'name': 'chinaware'}, {'frequency': 'c', 'synset': 'chip.n.04', 'synonyms': ['crisp_(potato_chip)', 'potato_chip'], 'id': 246, 'def': 'a thin crisp slice of potato fried in deep fat', 'name': 'crisp_(potato_chip)'}, {'frequency': 'r', 'synset': 'chip.n.06', 'synonyms': ['poker_chip'], 'id': 247, 'def': 'a small disk-shaped counter used to represent money when gambling', 'name': 'poker_chip'}, {'frequency': 'c', 'synset': 'chocolate_bar.n.01', 'synonyms': ['chocolate_bar'], 'id': 248, 'def': 'a bar of chocolate candy', 'name': 'chocolate_bar'}, {'frequency': 'c', 'synset': 'chocolate_cake.n.01', 'synonyms': ['chocolate_cake'], 'id': 249, 'def': 'cake containing chocolate', 'name': 'chocolate_cake'}, {'frequency': 'r', 'synset': 'chocolate_milk.n.01', 'synonyms': ['chocolate_milk'], 'id': 250, 'def': 'milk flavored with chocolate syrup', 'name': 'chocolate_milk'}, {'frequency': 'r', 'synset': 'chocolate_mousse.n.01', 'synonyms': ['chocolate_mousse'], 'id': 251, 'def': 'dessert mousse made with chocolate', 'name': 'chocolate_mousse'}, {'frequency': 'f', 'synset': 'choker.n.03', 'synonyms': ['choker', 'collar', 'neckband'], 'id': 252, 'def': 'shirt collar, animal collar, or tight-fitting necklace', 'name': 'choker'}, {'frequency': 'f', 'synset': 'chopping_board.n.01', 'synonyms': ['chopping_board', 'cutting_board', 'chopping_block'], 'id': 253, 'def': 'a wooden board where meats or vegetables can be cut', 'name': 'chopping_board'}, {'frequency': 'f', 'synset': 'chopstick.n.01', 'synonyms': ['chopstick'], 'id': 254, 'def': 'one of a pair of slender sticks used as oriental tableware to eat food with', 'name': 'chopstick'}, {'frequency': 'f', 'synset': 'christmas_tree.n.05', 'synonyms': ['Christmas_tree'], 'id': 255, 'def': 'an ornamented evergreen used as a Christmas decoration', 'name': 'Christmas_tree'}, {'frequency': 'c', 'synset': 'chute.n.02', 'synonyms': ['slide'], 'id': 256, 'def': 'sloping channel through which things can descend', 'name': 'slide'}, {'frequency': 'r', 'synset': 'cider.n.01', 'synonyms': ['cider', 'cyder'], 'id': 257, 'def': 'a beverage made from juice pressed from apples', 'name': 'cider'}, {'frequency': 'r', 'synset': 'cigar_box.n.01', 'synonyms': ['cigar_box'], 'id': 258, 'def': 'a box for holding cigars', 'name': 'cigar_box'}, {'frequency': 'f', 'synset': 'cigarette.n.01', 'synonyms': ['cigarette'], 'id': 259, 'def': 'finely ground tobacco wrapped in paper; for smoking', 'name': 'cigarette'}, {'frequency': 'c', 'synset': 'cigarette_case.n.01', 'synonyms': ['cigarette_case', 'cigarette_pack'], 'id': 260, 'def': 'a small flat case for holding cigarettes', 'name': 'cigarette_case'}, {'frequency': 'f', 'synset': 'cistern.n.02', 'synonyms': ['cistern', 'water_tank'], 'id': 261, 'def': 'a tank that holds the water used to flush a toilet', 'name': 'cistern'}, {'frequency': 'r', 'synset': 'clarinet.n.01', 'synonyms': ['clarinet'], 'id': 262, 'def': 'a single-reed instrument with a straight tube', 'name': 'clarinet'}, {'frequency': 'c', 'synset': 'clasp.n.01', 'synonyms': ['clasp'], 'id': 263, 'def': 'a fastener (as a buckle or hook) that is used to hold two things together', 'name': 'clasp'}, {'frequency': 'c', 'synset': 'cleansing_agent.n.01', 'synonyms': ['cleansing_agent', 'cleanser', 'cleaner'], 'id': 264, 'def': 'a preparation used in cleaning something', 'name': 'cleansing_agent'}, {'frequency': 'r', 'synset': 'cleat.n.02', 'synonyms': ['cleat_(for_securing_rope)'], 'id': 265, 'def': 'a fastener (usually with two projecting horns) around which a rope can be secured', 'name': 'cleat_(for_securing_rope)'}, {'frequency': 'r', 'synset': 'clementine.n.01', 'synonyms': ['clementine'], 'id': 266, 'def': 'a variety of mandarin orange', 'name': 'clementine'}, {'frequency': 'c', 'synset': 'clip.n.03', 'synonyms': ['clip'], 'id': 267, 'def': 'any of various small fasteners used to hold loose articles together', 'name': 'clip'}, {'frequency': 'c', 'synset': 'clipboard.n.01', 'synonyms': ['clipboard'], 'id': 268, 'def': 'a small writing board with a clip at the top for holding papers', 'name': 'clipboard'}, {'frequency': 'r', 'synset': 'clipper.n.03', 'synonyms': ['clippers_(for_plants)'], 'id': 269, 'def': 'shears for cutting grass or shrubbery (often used in the plural)', 'name': 'clippers_(for_plants)'}, {'frequency': 'r', 'synset': 'cloak.n.02', 'synonyms': ['cloak'], 'id': 270, 'def': 'a loose outer garment', 'name': 'cloak'}, {'frequency': 'f', 'synset': 'clock.n.01', 'synonyms': ['clock', 'timepiece', 'timekeeper'], 'id': 271, 'def': 'a timepiece that shows the time of day', 'name': 'clock'}, {'frequency': 'f', 'synset': 'clock_tower.n.01', 'synonyms': ['clock_tower'], 'id': 272, 'def': 'a tower with a large clock visible high up on an outside face', 'name': 'clock_tower'}, {'frequency': 'c', 'synset': 'clothes_hamper.n.01', 'synonyms': ['clothes_hamper', 'laundry_basket', 'clothes_basket'], 'id': 273, 'def': 'a hamper that holds dirty clothes to be washed or wet clothes to be dried', 'name': 'clothes_hamper'}, {'frequency': 'c', 'synset': 'clothespin.n.01', 'synonyms': ['clothespin', 'clothes_peg'], 'id': 274, 'def': 'wood or plastic fastener; for holding clothes on a clothesline', 'name': 'clothespin'}, {'frequency': 'r', 'synset': 'clutch_bag.n.01', 'synonyms': ['clutch_bag'], 'id': 275, 'def': "a woman's strapless purse that is carried in the hand", 'name': 'clutch_bag'}, {'frequency': 'f', 'synset': 'coaster.n.03', 'synonyms': ['coaster'], 'id': 276, 'def': 'a covering (plate or mat) that protects the surface of a table', 'name': 'coaster'}, {'frequency': 'f', 'synset': 'coat.n.01', 'synonyms': ['coat'], 'id': 277, 'def': 'an outer garment that has sleeves and covers the body from shoulder down', 'name': 'coat'}, {'frequency': 'c', 'synset': 'coat_hanger.n.01', 'synonyms': ['coat_hanger', 'clothes_hanger', 'dress_hanger'], 'id': 278, 'def': "a hanger that is shaped like a person's shoulders", 'name': 'coat_hanger'}, {'frequency': 'c', 'synset': 'coatrack.n.01', 'synonyms': ['coatrack', 'hatrack'], 'id': 279, 'def': 'a rack with hooks for temporarily holding coats and hats', 'name': 'coatrack'}, {'frequency': 'c', 'synset': 'cock.n.04', 'synonyms': ['cock', 'rooster'], 'id': 280, 'def': 'adult male chicken', 'name': 'cock'}, {'frequency': 'r', 'synset': 'cockroach.n.01', 'synonyms': ['cockroach'], 'id': 281, 'def': 'any of numerous chiefly nocturnal insects; some are domestic pests', 'name': 'cockroach'}, {'frequency': 'r', 'synset': 'cocoa.n.01', 'synonyms': ['cocoa_(beverage)', 'hot_chocolate_(beverage)', 'drinking_chocolate'], 'id': 282, 'def': 'a beverage made from cocoa powder and milk and sugar; usually drunk hot', 'name': 'cocoa_(beverage)'}, {'frequency': 'c', 'synset': 'coconut.n.02', 'synonyms': ['coconut', 'cocoanut'], 'id': 283, 'def': 'large hard-shelled brown oval nut with a fibrous husk', 'name': 'coconut'}, {'frequency': 'f', 'synset': 'coffee_maker.n.01', 'synonyms': ['coffee_maker', 'coffee_machine'], 'id': 284, 'def': 'a kitchen appliance for brewing coffee automatically', 'name': 'coffee_maker'}, {'frequency': 'f', 'synset': 'coffee_table.n.01', 'synonyms': ['coffee_table', 'cocktail_table'], 'id': 285, 'def': 'low table where magazines can be placed and coffee or cocktails are served', 'name': 'coffee_table'}, {'frequency': 'c', 'synset': 'coffeepot.n.01', 'synonyms': ['coffeepot'], 'id': 286, 'def': 'tall pot in which coffee is brewed', 'name': 'coffeepot'}, {'frequency': 'r', 'synset': 'coil.n.05', 'synonyms': ['coil'], 'id': 287, 'def': 'tubing that is wound in a spiral', 'name': 'coil'}, {'frequency': 'c', 'synset': 'coin.n.01', 'synonyms': ['coin'], 'id': 288, 'def': 'a flat metal piece (usually a disc) used as money', 'name': 'coin'}, {'frequency': 'c', 'synset': 'colander.n.01', 'synonyms': ['colander', 'cullender'], 'id': 289, 'def': 'bowl-shaped strainer; used to wash or drain foods', 'name': 'colander'}, {'frequency': 'c', 'synset': 'coleslaw.n.01', 'synonyms': ['coleslaw', 'slaw'], 'id': 290, 'def': 'basically shredded cabbage', 'name': 'coleslaw'}, {'frequency': 'r', 'synset': 'coloring_material.n.01', 'synonyms': ['coloring_material', 'colouring_material'], 'id': 291, 'def': 'any material used for its color', 'name': 'coloring_material'}, {'frequency': 'r', 'synset': 'combination_lock.n.01', 'synonyms': ['combination_lock'], 'id': 292, 'def': 'lock that can be opened only by turning dials in a special sequence', 'name': 'combination_lock'}, {'frequency': 'c', 'synset': 'comforter.n.04', 'synonyms': ['pacifier', 'teething_ring'], 'id': 293, 'def': 'device used for an infant to suck or bite on', 'name': 'pacifier'}, {'frequency': 'r', 'synset': 'comic_book.n.01', 'synonyms': ['comic_book'], 'id': 294, 'def': 'a magazine devoted to comic strips', 'name': 'comic_book'}, {'frequency': 'r', 'synset': 'compass.n.01', 'synonyms': ['compass'], 'id': 295, 'def': 'navigational instrument for finding directions', 'name': 'compass'}, {'frequency': 'f', 'synset': 'computer_keyboard.n.01', 'synonyms': ['computer_keyboard', 'keyboard_(computer)'], 'id': 296, 'def': 'a keyboard that is a data input device for computers', 'name': 'computer_keyboard'}, {'frequency': 'f', 'synset': 'condiment.n.01', 'synonyms': ['condiment'], 'id': 297, 'def': 'a preparation (a sauce or relish or spice) to enhance flavor or enjoyment', 'name': 'condiment'}, {'frequency': 'f', 'synset': 'cone.n.01', 'synonyms': ['cone', 'traffic_cone'], 'id': 298, 'def': 'a cone-shaped object used to direct traffic', 'name': 'cone'}, {'frequency': 'f', 'synset': 'control.n.09', 'synonyms': ['control', 'controller'], 'id': 299, 'def': 'a mechanism that controls the operation of a machine', 'name': 'control'}, {'frequency': 'r', 'synset': 'convertible.n.01', 'synonyms': ['convertible_(automobile)'], 'id': 300, 'def': 'a car that has top that can be folded or removed', 'name': 'convertible_(automobile)'}, {'frequency': 'r', 'synset': 'convertible.n.03', 'synonyms': ['sofa_bed'], 'id': 301, 'def': 'a sofa that can be converted into a bed', 'name': 'sofa_bed'}, {'frequency': 'r', 'synset': 'cooker.n.01', 'synonyms': ['cooker'], 'id': 302, 'def': 'a utensil for cooking', 'name': 'cooker'}, {'frequency': 'f', 'synset': 'cookie.n.01', 'synonyms': ['cookie', 'cooky', 'biscuit_(cookie)'], 'id': 303, 'def': "any of various small flat sweet cakes (`biscuit' is the British term)", 'name': 'cookie'}, {'frequency': 'r', 'synset': 'cooking_utensil.n.01', 'synonyms': ['cooking_utensil'], 'id': 304, 'def': 'a kitchen utensil made of material that does not melt easily; used for cooking', 'name': 'cooking_utensil'}, {'frequency': 'f', 'synset': 'cooler.n.01', 'synonyms': ['cooler_(for_food)', 'ice_chest'], 'id': 305, 'def': 'an insulated box for storing food often with ice', 'name': 'cooler_(for_food)'}, {'frequency': 'f', 'synset': 'cork.n.04', 'synonyms': ['cork_(bottle_plug)', 'bottle_cork'], 'id': 306, 'def': 'the plug in the mouth of a bottle (especially a wine bottle)', 'name': 'cork_(bottle_plug)'}, {'frequency': 'r', 'synset': 'corkboard.n.01', 'synonyms': ['corkboard'], 'id': 307, 'def': 'a sheet consisting of cork granules', 'name': 'corkboard'}, {'frequency': 'c', 'synset': 'corkscrew.n.01', 'synonyms': ['corkscrew', 'bottle_screw'], 'id': 308, 'def': 'a bottle opener that pulls corks', 'name': 'corkscrew'}, {'frequency': 'f', 'synset': 'corn.n.03', 'synonyms': ['edible_corn', 'corn', 'maize'], 'id': 309, 'def': 'ears or kernels of corn that can be prepared and served for human food (only mark individual ears or kernels)', 'name': 'edible_corn'}, {'frequency': 'r', 'synset': 'cornbread.n.01', 'synonyms': ['cornbread'], 'id': 310, 'def': 'bread made primarily of cornmeal', 'name': 'cornbread'}, {'frequency': 'c', 'synset': 'cornet.n.01', 'synonyms': ['cornet', 'horn', 'trumpet'], 'id': 311, 'def': 'a brass musical instrument with a narrow tube and a flared bell and many valves', 'name': 'cornet'}, {'frequency': 'c', 'synset': 'cornice.n.01', 'synonyms': ['cornice', 'valance', 'valance_board', 'pelmet'], 'id': 312, 'def': 'a decorative framework to conceal curtain fixtures at the top of a window casing', 'name': 'cornice'}, {'frequency': 'r', 'synset': 'cornmeal.n.01', 'synonyms': ['cornmeal'], 'id': 313, 'def': 'coarsely ground corn', 'name': 'cornmeal'}, {'frequency': 'c', 'synset': 'corset.n.01', 'synonyms': ['corset', 'girdle'], 'id': 314, 'def': "a woman's close-fitting foundation garment", 'name': 'corset'}, {'frequency': 'c', 'synset': 'costume.n.04', 'synonyms': ['costume'], 'id': 315, 'def': 'the attire characteristic of a country or a time or a social class', 'name': 'costume'}, {'frequency': 'r', 'synset': 'cougar.n.01', 'synonyms': ['cougar', 'puma', 'catamount', 'mountain_lion', 'panther'], 'id': 316, 'def': 'large American feline resembling a lion', 'name': 'cougar'}, {'frequency': 'r', 'synset': 'coverall.n.01', 'synonyms': ['coverall'], 'id': 317, 'def': 'a loose-fitting protective garment that is worn over other clothing', 'name': 'coverall'}, {'frequency': 'c', 'synset': 'cowbell.n.01', 'synonyms': ['cowbell'], 'id': 318, 'def': 'a bell hung around the neck of cow so that the cow can be easily located', 'name': 'cowbell'}, {'frequency': 'f', 'synset': 'cowboy_hat.n.01', 'synonyms': ['cowboy_hat', 'ten-gallon_hat'], 'id': 319, 'def': 'a hat with a wide brim and a soft crown; worn by American ranch hands', 'name': 'cowboy_hat'}, {'frequency': 'c', 'synset': 'crab.n.01', 'synonyms': ['crab_(animal)'], 'id': 320, 'def': 'decapod having eyes on short stalks and a broad flattened shell and pincers', 'name': 'crab_(animal)'}, {'frequency': 'r', 'synset': 'crab.n.05', 'synonyms': ['crabmeat'], 'id': 321, 'def': 'the edible flesh of any of various crabs', 'name': 'crabmeat'}, {'frequency': 'c', 'synset': 'cracker.n.01', 'synonyms': ['cracker'], 'id': 322, 'def': 'a thin crisp wafer', 'name': 'cracker'}, {'frequency': 'r', 'synset': 'crape.n.01', 'synonyms': ['crape', 'crepe', 'French_pancake'], 'id': 323, 'def': 'small very thin pancake', 'name': 'crape'}, {'frequency': 'f', 'synset': 'crate.n.01', 'synonyms': ['crate'], 'id': 324, 'def': 'a rugged box (usually made of wood); used for shipping', 'name': 'crate'}, {'frequency': 'c', 'synset': 'crayon.n.01', 'synonyms': ['crayon', 'wax_crayon'], 'id': 325, 'def': 'writing or drawing implement made of a colored stick of composition wax', 'name': 'crayon'}, {'frequency': 'r', 'synset': 'cream_pitcher.n.01', 'synonyms': ['cream_pitcher'], 'id': 326, 'def': 'a small pitcher for serving cream', 'name': 'cream_pitcher'}, {'frequency': 'c', 'synset': 'crescent_roll.n.01', 'synonyms': ['crescent_roll', 'croissant'], 'id': 327, 'def': 'very rich flaky crescent-shaped roll', 'name': 'crescent_roll'}, {'frequency': 'c', 'synset': 'crib.n.01', 'synonyms': ['crib', 'cot'], 'id': 328, 'def': 'baby bed with high sides made of slats', 'name': 'crib'}, {'frequency': 'c', 'synset': 'crock.n.03', 'synonyms': ['crock_pot', 'earthenware_jar'], 'id': 329, 'def': 'an earthen jar (made of baked clay) or a modern electric crockpot', 'name': 'crock_pot'}, {'frequency': 'f', 'synset': 'crossbar.n.01', 'synonyms': ['crossbar'], 'id': 330, 'def': 'a horizontal bar that goes across something', 'name': 'crossbar'}, {'frequency': 'r', 'synset': 'crouton.n.01', 'synonyms': ['crouton'], 'id': 331, 'def': 'a small piece of toasted or fried bread; served in soup or salads', 'name': 'crouton'}, {'frequency': 'c', 'synset': 'crow.n.01', 'synonyms': ['crow'], 'id': 332, 'def': 'black birds having a raucous call', 'name': 'crow'}, {'frequency': 'r', 'synset': 'crowbar.n.01', 'synonyms': ['crowbar', 'wrecking_bar', 'pry_bar'], 'id': 333, 'def': 'a heavy iron lever with one end forged into a wedge', 'name': 'crowbar'}, {'frequency': 'c', 'synset': 'crown.n.04', 'synonyms': ['crown'], 'id': 334, 'def': 'an ornamental jeweled headdress signifying sovereignty', 'name': 'crown'}, {'frequency': 'c', 'synset': 'crucifix.n.01', 'synonyms': ['crucifix'], 'id': 335, 'def': 'representation of the cross on which Jesus died', 'name': 'crucifix'}, {'frequency': 'c', 'synset': 'cruise_ship.n.01', 'synonyms': ['cruise_ship', 'cruise_liner'], 'id': 336, 'def': 'a passenger ship used commercially for pleasure cruises', 'name': 'cruise_ship'}, {'frequency': 'c', 'synset': 'cruiser.n.01', 'synonyms': ['police_cruiser', 'patrol_car', 'police_car', 'squad_car'], 'id': 337, 'def': 'a car in which policemen cruise the streets', 'name': 'police_cruiser'}, {'frequency': 'f', 'synset': 'crumb.n.03', 'synonyms': ['crumb'], 'id': 338, 'def': 'small piece of e.g. bread or cake', 'name': 'crumb'}, {'frequency': 'c', 'synset': 'crutch.n.01', 'synonyms': ['crutch'], 'id': 339, 'def': 'a wooden or metal staff that fits under the armpit and reaches to the ground', 'name': 'crutch'}, {'frequency': 'c', 'synset': 'cub.n.03', 'synonyms': ['cub_(animal)'], 'id': 340, 'def': 'the young of certain carnivorous mammals such as the bear or wolf or lion', 'name': 'cub_(animal)'}, {'frequency': 'c', 'synset': 'cube.n.05', 'synonyms': ['cube', 'square_block'], 'id': 341, 'def': 'a block in the (approximate) shape of a cube', 'name': 'cube'}, {'frequency': 'f', 'synset': 'cucumber.n.02', 'synonyms': ['cucumber', 'cuke'], 'id': 342, 'def': 'cylindrical green fruit with thin green rind and white flesh eaten as a vegetable', 'name': 'cucumber'}, {'frequency': 'c', 'synset': 'cufflink.n.01', 'synonyms': ['cufflink'], 'id': 343, 'def': 'jewelry consisting of linked buttons used to fasten the cuffs of a shirt', 'name': 'cufflink'}, {'frequency': 'f', 'synset': 'cup.n.01', 'synonyms': ['cup'], 'id': 344, 'def': 'a small open container usually used for drinking; usually has a handle', 'name': 'cup'}, {'frequency': 'c', 'synset': 'cup.n.08', 'synonyms': ['trophy_cup'], 'id': 345, 'def': 'a metal award or cup-shaped vessel with handles that is awarded as a trophy to a competition winner', 'name': 'trophy_cup'}, {'frequency': 'f', 'synset': 'cupboard.n.01', 'synonyms': ['cupboard', 'closet'], 'id': 346, 'def': 'a small room (or recess) or cabinet used for storage space', 'name': 'cupboard'}, {'frequency': 'f', 'synset': 'cupcake.n.01', 'synonyms': ['cupcake'], 'id': 347, 'def': 'small cake baked in a muffin tin', 'name': 'cupcake'}, {'frequency': 'r', 'synset': 'curler.n.01', 'synonyms': ['hair_curler', 'hair_roller', 'hair_crimper'], 'id': 348, 'def': 'a cylindrical tube around which the hair is wound to curl it', 'name': 'hair_curler'}, {'frequency': 'r', 'synset': 'curling_iron.n.01', 'synonyms': ['curling_iron'], 'id': 349, 'def': 'a cylindrical home appliance that heats hair that has been curled around it', 'name': 'curling_iron'}, {'frequency': 'f', 'synset': 'curtain.n.01', 'synonyms': ['curtain', 'drapery'], 'id': 350, 'def': 'hanging cloth used as a blind (especially for a window)', 'name': 'curtain'}, {'frequency': 'f', 'synset': 'cushion.n.03', 'synonyms': ['cushion'], 'id': 351, 'def': 'a soft bag filled with air or padding such as feathers or foam rubber', 'name': 'cushion'}, {'frequency': 'r', 'synset': 'cylinder.n.04', 'synonyms': ['cylinder'], 'id': 352, 'def': 'a cylindrical container', 'name': 'cylinder'}, {'frequency': 'r', 'synset': 'cymbal.n.01', 'synonyms': ['cymbal'], 'id': 353, 'def': 'a percussion instrument consisting of a concave brass disk', 'name': 'cymbal'}, {'frequency': 'r', 'synset': 'dagger.n.01', 'synonyms': ['dagger'], 'id': 354, 'def': 'a short knife with a pointed blade used for piercing or stabbing', 'name': 'dagger'}, {'frequency': 'r', 'synset': 'dalmatian.n.02', 'synonyms': ['dalmatian'], 'id': 355, 'def': 'a large breed having a smooth white coat with black or brown spots', 'name': 'dalmatian'}, {'frequency': 'c', 'synset': 'dartboard.n.01', 'synonyms': ['dartboard'], 'id': 356, 'def': 'a circular board of wood or cork used as the target in the game of darts', 'name': 'dartboard'}, {'frequency': 'r', 'synset': 'date.n.08', 'synonyms': ['date_(fruit)'], 'id': 357, 'def': 'sweet edible fruit of the date palm with a single long woody seed', 'name': 'date_(fruit)'}, {'frequency': 'f', 'synset': 'deck_chair.n.01', 'synonyms': ['deck_chair', 'beach_chair'], 'id': 358, 'def': 'a folding chair for use outdoors; a wooden frame supports a length of canvas', 'name': 'deck_chair'}, {'frequency': 'c', 'synset': 'deer.n.01', 'synonyms': ['deer', 'cervid'], 'id': 359, 'def': "distinguished from Bovidae by the male's having solid deciduous antlers", 'name': 'deer'}, {'frequency': 'c', 'synset': 'dental_floss.n.01', 'synonyms': ['dental_floss', 'floss'], 'id': 360, 'def': 'a soft thread for cleaning the spaces between the teeth', 'name': 'dental_floss'}, {'frequency': 'f', 'synset': 'desk.n.01', 'synonyms': ['desk'], 'id': 361, 'def': 'a piece of furniture with a writing surface and usually drawers or other compartments', 'name': 'desk'}, {'frequency': 'r', 'synset': 'detergent.n.01', 'synonyms': ['detergent'], 'id': 362, 'def': 'a surface-active chemical widely used in industry and laundering', 'name': 'detergent'}, {'frequency': 'c', 'synset': 'diaper.n.01', 'synonyms': ['diaper'], 'id': 363, 'def': 'garment consisting of a folded cloth drawn up between the legs and fastened at the waist', 'name': 'diaper'}, {'frequency': 'r', 'synset': 'diary.n.01', 'synonyms': ['diary', 'journal'], 'id': 364, 'def': 'yearly planner book', 'name': 'diary'}, {'frequency': 'r', 'synset': 'die.n.01', 'synonyms': ['die', 'dice'], 'id': 365, 'def': 'a small cube with 1 to 6 spots on the six faces; used in gambling', 'name': 'die'}, {'frequency': 'r', 'synset': 'dinghy.n.01', 'synonyms': ['dinghy', 'dory', 'rowboat'], 'id': 366, 'def': 'a small boat of shallow draft with seats and oars with which it is propelled', 'name': 'dinghy'}, {'frequency': 'f', 'synset': 'dining_table.n.01', 'synonyms': ['dining_table'], 'id': 367, 'def': 'a table at which meals are served', 'name': 'dining_table'}, {'frequency': 'r', 'synset': 'dinner_jacket.n.01', 'synonyms': ['tux', 'tuxedo'], 'id': 368, 'def': 'semiformal evening dress for men', 'name': 'tux'}, {'frequency': 'f', 'synset': 'dish.n.01', 'synonyms': ['dish'], 'id': 369, 'def': 'a piece of dishware normally used as a container for holding or serving food', 'name': 'dish'}, {'frequency': 'c', 'synset': 'dish.n.05', 'synonyms': ['dish_antenna'], 'id': 370, 'def': 'directional antenna consisting of a parabolic reflector', 'name': 'dish_antenna'}, {'frequency': 'c', 'synset': 'dishrag.n.01', 'synonyms': ['dishrag', 'dishcloth'], 'id': 371, 'def': 'a cloth for washing dishes or cleaning in general', 'name': 'dishrag'}, {'frequency': 'f', 'synset': 'dishtowel.n.01', 'synonyms': ['dishtowel', 'tea_towel'], 'id': 372, 'def': 'a towel for drying dishes', 'name': 'dishtowel'}, {'frequency': 'f', 'synset': 'dishwasher.n.01', 'synonyms': ['dishwasher', 'dishwashing_machine'], 'id': 373, 'def': 'a machine for washing dishes', 'name': 'dishwasher'}, {'frequency': 'r', 'synset': 'dishwasher_detergent.n.01', 'synonyms': ['dishwasher_detergent', 'dishwashing_detergent', 'dishwashing_liquid', 'dishsoap'], 'id': 374, 'def': 'dishsoap or dish detergent designed for use in dishwashers', 'name': 'dishwasher_detergent'}, {'frequency': 'f', 'synset': 'dispenser.n.01', 'synonyms': ['dispenser'], 'id': 375, 'def': 'a container so designed that the contents can be used in prescribed amounts', 'name': 'dispenser'}, {'frequency': 'r', 'synset': 'diving_board.n.01', 'synonyms': ['diving_board'], 'id': 376, 'def': 'a springboard from which swimmers can dive', 'name': 'diving_board'}, {'frequency': 'f', 'synset': 'dixie_cup.n.01', 'synonyms': ['Dixie_cup', 'paper_cup'], 'id': 377, 'def': 'a disposable cup made of paper; for holding drinks', 'name': 'Dixie_cup'}, {'frequency': 'f', 'synset': 'dog.n.01', 'synonyms': ['dog'], 'id': 378, 'def': 'a common domesticated dog', 'name': 'dog'}, {'frequency': 'f', 'synset': 'dog_collar.n.01', 'synonyms': ['dog_collar'], 'id': 379, 'def': 'a collar for a dog', 'name': 'dog_collar'}, {'frequency': 'f', 'synset': 'doll.n.01', 'synonyms': ['doll'], 'id': 380, 'def': 'a toy replica of a HUMAN (NOT AN ANIMAL)', 'name': 'doll'}, {'frequency': 'r', 'synset': 'dollar.n.02', 'synonyms': ['dollar', 'dollar_bill', 'one_dollar_bill'], 'id': 381, 'def': 'a piece of paper money worth one dollar', 'name': 'dollar'}, {'frequency': 'r', 'synset': 'dollhouse.n.01', 'synonyms': ['dollhouse', "doll's_house"], 'id': 382, 'def': "a house so small that it is likened to a child's plaything", 'name': 'dollhouse'}, {'frequency': 'c', 'synset': 'dolphin.n.02', 'synonyms': ['dolphin'], 'id': 383, 'def': 'any of various small toothed whales with a beaklike snout; larger than porpoises', 'name': 'dolphin'}, {'frequency': 'c', 'synset': 'domestic_ass.n.01', 'synonyms': ['domestic_ass', 'donkey'], 'id': 384, 'def': 'domestic beast of burden descended from the African wild ass; patient but stubborn', 'name': 'domestic_ass'}, {'frequency': 'f', 'synset': 'doorknob.n.01', 'synonyms': ['doorknob', 'doorhandle'], 'id': 385, 'def': "a knob used to open a door (often called `doorhandle' in Great Britain)", 'name': 'doorknob'}, {'frequency': 'c', 'synset': 'doormat.n.02', 'synonyms': ['doormat', 'welcome_mat'], 'id': 386, 'def': 'a mat placed outside an exterior door for wiping the shoes before entering', 'name': 'doormat'}, {'frequency': 'f', 'synset': 'doughnut.n.02', 'synonyms': ['doughnut', 'donut'], 'id': 387, 'def': 'a small ring-shaped friedcake', 'name': 'doughnut'}, {'frequency': 'r', 'synset': 'dove.n.01', 'synonyms': ['dove'], 'id': 388, 'def': 'any of numerous small pigeons', 'name': 'dove'}, {'frequency': 'r', 'synset': 'dragonfly.n.01', 'synonyms': ['dragonfly'], 'id': 389, 'def': 'slender-bodied non-stinging insect having iridescent wings that are outspread at rest', 'name': 'dragonfly'}, {'frequency': 'f', 'synset': 'drawer.n.01', 'synonyms': ['drawer'], 'id': 390, 'def': 'a boxlike container in a piece of furniture; made so as to slide in and out', 'name': 'drawer'}, {'frequency': 'c', 'synset': 'drawers.n.01', 'synonyms': ['underdrawers', 'boxers', 'boxershorts'], 'id': 391, 'def': 'underpants worn by men', 'name': 'underdrawers'}, {'frequency': 'f', 'synset': 'dress.n.01', 'synonyms': ['dress', 'frock'], 'id': 392, 'def': 'a one-piece garment for a woman; has skirt and bodice', 'name': 'dress'}, {'frequency': 'c', 'synset': 'dress_hat.n.01', 'synonyms': ['dress_hat', 'high_hat', 'opera_hat', 'silk_hat', 'top_hat'], 'id': 393, 'def': "a man's hat with a tall crown; usually covered with silk or with beaver fur", 'name': 'dress_hat'}, {'frequency': 'f', 'synset': 'dress_suit.n.01', 'synonyms': ['dress_suit'], 'id': 394, 'def': 'formalwear consisting of full evening dress for men', 'name': 'dress_suit'}, {'frequency': 'f', 'synset': 'dresser.n.05', 'synonyms': ['dresser'], 'id': 395, 'def': 'a cabinet with shelves', 'name': 'dresser'}, {'frequency': 'c', 'synset': 'drill.n.01', 'synonyms': ['drill'], 'id': 396, 'def': 'a tool with a sharp rotating point for making holes in hard materials', 'name': 'drill'}, {'frequency': 'r', 'synset': 'drone.n.04', 'synonyms': ['drone'], 'id': 397, 'def': 'an aircraft without a pilot that is operated by remote control', 'name': 'drone'}, {'frequency': 'r', 'synset': 'dropper.n.01', 'synonyms': ['dropper', 'eye_dropper'], 'id': 398, 'def': 'pipet consisting of a small tube with a vacuum bulb at one end for drawing liquid in and releasing it a drop at a time', 'name': 'dropper'}, {'frequency': 'c', 'synset': 'drum.n.01', 'synonyms': ['drum_(musical_instrument)'], 'id': 399, 'def': 'a musical percussion instrument; usually consists of a hollow cylinder with a membrane stretched across each end', 'name': 'drum_(musical_instrument)'}, {'frequency': 'r', 'synset': 'drumstick.n.02', 'synonyms': ['drumstick'], 'id': 400, 'def': 'a stick used for playing a drum', 'name': 'drumstick'}, {'frequency': 'f', 'synset': 'duck.n.01', 'synonyms': ['duck'], 'id': 401, 'def': 'small web-footed broad-billed swimming bird', 'name': 'duck'}, {'frequency': 'c', 'synset': 'duckling.n.02', 'synonyms': ['duckling'], 'id': 402, 'def': 'young duck', 'name': 'duckling'}, {'frequency': 'c', 'synset': 'duct_tape.n.01', 'synonyms': ['duct_tape'], 'id': 403, 'def': 'a wide silvery adhesive tape', 'name': 'duct_tape'}, {'frequency': 'f', 'synset': 'duffel_bag.n.01', 'synonyms': ['duffel_bag', 'duffle_bag', 'duffel', 'duffle'], 'id': 404, 'def': 'a large cylindrical bag of heavy cloth (does not include suitcases)', 'name': 'duffel_bag'}, {'frequency': 'r', 'synset': 'dumbbell.n.01', 'synonyms': ['dumbbell'], 'id': 405, 'def': 'an exercising weight with two ball-like ends connected by a short handle', 'name': 'dumbbell'}, {'frequency': 'c', 'synset': 'dumpster.n.01', 'synonyms': ['dumpster'], 'id': 406, 'def': 'a container designed to receive and transport and dump waste', 'name': 'dumpster'}, {'frequency': 'r', 'synset': 'dustpan.n.02', 'synonyms': ['dustpan'], 'id': 407, 'def': 'a short-handled receptacle into which dust can be swept', 'name': 'dustpan'}, {'frequency': 'c', 'synset': 'eagle.n.01', 'synonyms': ['eagle'], 'id': 408, 'def': 'large birds of prey noted for their broad wings and strong soaring flight', 'name': 'eagle'}, {'frequency': 'f', 'synset': 'earphone.n.01', 'synonyms': ['earphone', 'earpiece', 'headphone'], 'id': 409, 'def': 'device for listening to audio that is held over or inserted into the ear', 'name': 'earphone'}, {'frequency': 'r', 'synset': 'earplug.n.01', 'synonyms': ['earplug'], 'id': 410, 'def': 'a soft plug that is inserted into the ear canal to block sound', 'name': 'earplug'}, {'frequency': 'f', 'synset': 'earring.n.01', 'synonyms': ['earring'], 'id': 411, 'def': 'jewelry to ornament the ear', 'name': 'earring'}, {'frequency': 'c', 'synset': 'easel.n.01', 'synonyms': ['easel'], 'id': 412, 'def': "an upright tripod for displaying something (usually an artist's canvas)", 'name': 'easel'}, {'frequency': 'r', 'synset': 'eclair.n.01', 'synonyms': ['eclair'], 'id': 413, 'def': 'oblong cream puff', 'name': 'eclair'}, {'frequency': 'r', 'synset': 'eel.n.01', 'synonyms': ['eel'], 'id': 414, 'def': 'an elongate fish with fatty flesh', 'name': 'eel'}, {'frequency': 'f', 'synset': 'egg.n.02', 'synonyms': ['egg', 'eggs'], 'id': 415, 'def': 'oval reproductive body of a fowl (especially a hen) used as food', 'name': 'egg'}, {'frequency': 'r', 'synset': 'egg_roll.n.01', 'synonyms': ['egg_roll', 'spring_roll'], 'id': 416, 'def': 'minced vegetables and meat wrapped in a pancake and fried', 'name': 'egg_roll'}, {'frequency': 'c', 'synset': 'egg_yolk.n.01', 'synonyms': ['egg_yolk', 'yolk_(egg)'], 'id': 417, 'def': 'the yellow spherical part of an egg', 'name': 'egg_yolk'}, {'frequency': 'c', 'synset': 'eggbeater.n.02', 'synonyms': ['eggbeater', 'eggwhisk'], 'id': 418, 'def': 'a mixer for beating eggs or whipping cream', 'name': 'eggbeater'}, {'frequency': 'c', 'synset': 'eggplant.n.01', 'synonyms': ['eggplant', 'aubergine'], 'id': 419, 'def': 'egg-shaped vegetable having a shiny skin typically dark purple', 'name': 'eggplant'}, {'frequency': 'r', 'synset': 'electric_chair.n.01', 'synonyms': ['electric_chair'], 'id': 420, 'def': 'a chair-shaped instrument of execution by electrocution', 'name': 'electric_chair'}, {'frequency': 'f', 'synset': 'electric_refrigerator.n.01', 'synonyms': ['refrigerator'], 'id': 421, 'def': 'a refrigerator in which the coolant is pumped around by an electric motor', 'name': 'refrigerator'}, {'frequency': 'f', 'synset': 'elephant.n.01', 'synonyms': ['elephant'], 'id': 422, 'def': 'a common elephant', 'name': 'elephant'}, {'frequency': 'c', 'synset': 'elk.n.01', 'synonyms': ['elk', 'moose'], 'id': 423, 'def': 'large northern deer with enormous flattened antlers in the male', 'name': 'elk'}, {'frequency': 'c', 'synset': 'envelope.n.01', 'synonyms': ['envelope'], 'id': 424, 'def': 'a flat (usually rectangular) container for a letter, thin package, etc.', 'name': 'envelope'}, {'frequency': 'c', 'synset': 'eraser.n.01', 'synonyms': ['eraser'], 'id': 425, 'def': 'an implement used to erase something', 'name': 'eraser'}, {'frequency': 'r', 'synset': 'escargot.n.01', 'synonyms': ['escargot'], 'id': 426, 'def': 'edible snail usually served in the shell with a sauce of melted butter and garlic', 'name': 'escargot'}, {'frequency': 'r', 'synset': 'eyepatch.n.01', 'synonyms': ['eyepatch'], 'id': 427, 'def': 'a protective cloth covering for an injured eye', 'name': 'eyepatch'}, {'frequency': 'r', 'synset': 'falcon.n.01', 'synonyms': ['falcon'], 'id': 428, 'def': 'birds of prey having long pointed powerful wings adapted for swift flight', 'name': 'falcon'}, {'frequency': 'f', 'synset': 'fan.n.01', 'synonyms': ['fan'], 'id': 429, 'def': 'a device for creating a current of air by movement of a surface or surfaces', 'name': 'fan'}, {'frequency': 'f', 'synset': 'faucet.n.01', 'synonyms': ['faucet', 'spigot', 'tap'], 'id': 430, 'def': 'a regulator for controlling the flow of a liquid from a reservoir', 'name': 'faucet'}, {'frequency': 'r', 'synset': 'fedora.n.01', 'synonyms': ['fedora'], 'id': 431, 'def': 'a hat made of felt with a creased crown', 'name': 'fedora'}, {'frequency': 'r', 'synset': 'ferret.n.02', 'synonyms': ['ferret'], 'id': 432, 'def': 'domesticated albino variety of the European polecat bred for hunting rats and rabbits', 'name': 'ferret'}, {'frequency': 'c', 'synset': 'ferris_wheel.n.01', 'synonyms': ['Ferris_wheel'], 'id': 433, 'def': 'a large wheel with suspended seats that remain upright as the wheel rotates', 'name': 'Ferris_wheel'}, {'frequency': 'c', 'synset': 'ferry.n.01', 'synonyms': ['ferry', 'ferryboat'], 'id': 434, 'def': 'a boat that transports people or vehicles across a body of water and operates on a regular schedule', 'name': 'ferry'}, {'frequency': 'r', 'synset': 'fig.n.04', 'synonyms': ['fig_(fruit)'], 'id': 435, 'def': 'fleshy sweet pear-shaped yellowish or purple fruit eaten fresh or preserved or dried', 'name': 'fig_(fruit)'}, {'frequency': 'c', 'synset': 'fighter.n.02', 'synonyms': ['fighter_jet', 'fighter_aircraft', 'attack_aircraft'], 'id': 436, 'def': 'a high-speed military or naval airplane designed to destroy enemy targets', 'name': 'fighter_jet'}, {'frequency': 'f', 'synset': 'figurine.n.01', 'synonyms': ['figurine'], 'id': 437, 'def': 'a small carved or molded figure', 'name': 'figurine'}, {'frequency': 'c', 'synset': 'file.n.03', 'synonyms': ['file_cabinet', 'filing_cabinet'], 'id': 438, 'def': 'office furniture consisting of a container for keeping papers in order', 'name': 'file_cabinet'}, {'frequency': 'r', 'synset': 'file.n.04', 'synonyms': ['file_(tool)'], 'id': 439, 'def': 'a steel hand tool with small sharp teeth on some or all of its surfaces; used for smoothing wood or metal', 'name': 'file_(tool)'}, {'frequency': 'f', 'synset': 'fire_alarm.n.02', 'synonyms': ['fire_alarm', 'smoke_alarm'], 'id': 440, 'def': 'an alarm that is tripped off by fire or smoke', 'name': 'fire_alarm'}, {'frequency': 'f', 'synset': 'fire_engine.n.01', 'synonyms': ['fire_engine', 'fire_truck'], 'id': 441, 'def': 'large trucks that carry firefighters and equipment to the site of a fire', 'name': 'fire_engine'}, {'frequency': 'f', 'synset': 'fire_extinguisher.n.01', 'synonyms': ['fire_extinguisher', 'extinguisher'], 'id': 442, 'def': 'a manually operated device for extinguishing small fires', 'name': 'fire_extinguisher'}, {'frequency': 'c', 'synset': 'fire_hose.n.01', 'synonyms': ['fire_hose'], 'id': 443, 'def': 'a large hose that carries water from a fire hydrant to the site of the fire', 'name': 'fire_hose'}, {'frequency': 'f', 'synset': 'fireplace.n.01', 'synonyms': ['fireplace'], 'id': 444, 'def': 'an open recess in a wall at the base of a chimney where a fire can be built', 'name': 'fireplace'}, {'frequency': 'f', 'synset': 'fireplug.n.01', 'synonyms': ['fireplug', 'fire_hydrant', 'hydrant'], 'id': 445, 'def': 'an upright hydrant for drawing water to use in fighting a fire', 'name': 'fireplug'}, {'frequency': 'r', 'synset': 'first-aid_kit.n.01', 'synonyms': ['first-aid_kit'], 'id': 446, 'def': 'kit consisting of a set of bandages and medicines for giving first aid', 'name': 'first-aid_kit'}, {'frequency': 'f', 'synset': 'fish.n.01', 'synonyms': ['fish'], 'id': 447, 'def': 'any of various mostly cold-blooded aquatic vertebrates usually having scales and breathing through gills', 'name': 'fish'}, {'frequency': 'c', 'synset': 'fish.n.02', 'synonyms': ['fish_(food)'], 'id': 448, 'def': 'the flesh of fish used as food', 'name': 'fish_(food)'}, {'frequency': 'r', 'synset': 'fishbowl.n.02', 'synonyms': ['fishbowl', 'goldfish_bowl'], 'id': 449, 'def': 'a transparent bowl in which small fish are kept', 'name': 'fishbowl'}, {'frequency': 'c', 'synset': 'fishing_rod.n.01', 'synonyms': ['fishing_rod', 'fishing_pole'], 'id': 450, 'def': 'a rod that is used in fishing to extend the fishing line', 'name': 'fishing_rod'}, {'frequency': 'f', 'synset': 'flag.n.01', 'synonyms': ['flag'], 'id': 451, 'def': 'emblem usually consisting of a rectangular piece of cloth of distinctive design (do not include pole)', 'name': 'flag'}, {'frequency': 'f', 'synset': 'flagpole.n.02', 'synonyms': ['flagpole', 'flagstaff'], 'id': 452, 'def': 'a tall staff or pole on which a flag is raised', 'name': 'flagpole'}, {'frequency': 'c', 'synset': 'flamingo.n.01', 'synonyms': ['flamingo'], 'id': 453, 'def': 'large pink web-footed bird with down-bent bill', 'name': 'flamingo'}, {'frequency': 'c', 'synset': 'flannel.n.01', 'synonyms': ['flannel'], 'id': 454, 'def': 'a soft light woolen fabric; used for clothing', 'name': 'flannel'}, {'frequency': 'c', 'synset': 'flap.n.01', 'synonyms': ['flap'], 'id': 455, 'def': 'any broad thin covering attached at one edge, such as a mud flap next to a wheel or a flap on an airplane wing', 'name': 'flap'}, {'frequency': 'r', 'synset': 'flash.n.10', 'synonyms': ['flash', 'flashbulb'], 'id': 456, 'def': 'a lamp for providing momentary light to take a photograph', 'name': 'flash'}, {'frequency': 'c', 'synset': 'flashlight.n.01', 'synonyms': ['flashlight', 'torch'], 'id': 457, 'def': 'a small portable battery-powered electric lamp', 'name': 'flashlight'}, {'frequency': 'r', 'synset': 'fleece.n.03', 'synonyms': ['fleece'], 'id': 458, 'def': 'a soft bulky fabric with deep pile; used chiefly for clothing', 'name': 'fleece'}, {'frequency': 'f', 'synset': 'flip-flop.n.02', 'synonyms': ['flip-flop_(sandal)'], 'id': 459, 'def': 'a backless sandal held to the foot by a thong between two toes', 'name': 'flip-flop_(sandal)'}, {'frequency': 'c', 'synset': 'flipper.n.01', 'synonyms': ['flipper_(footwear)', 'fin_(footwear)'], 'id': 460, 'def': 'a shoe to aid a person in swimming', 'name': 'flipper_(footwear)'}, {'frequency': 'f', 'synset': 'flower_arrangement.n.01', 'synonyms': ['flower_arrangement', 'floral_arrangement'], 'id': 461, 'def': 'a decorative arrangement of flowers', 'name': 'flower_arrangement'}, {'frequency': 'c', 'synset': 'flute.n.02', 'synonyms': ['flute_glass', 'champagne_flute'], 'id': 462, 'def': 'a tall narrow wineglass', 'name': 'flute_glass'}, {'frequency': 'c', 'synset': 'foal.n.01', 'synonyms': ['foal'], 'id': 463, 'def': 'a young horse', 'name': 'foal'}, {'frequency': 'c', 'synset': 'folding_chair.n.01', 'synonyms': ['folding_chair'], 'id': 464, 'def': 'a chair that can be folded flat for storage', 'name': 'folding_chair'}, {'frequency': 'c', 'synset': 'food_processor.n.01', 'synonyms': ['food_processor'], 'id': 465, 'def': 'a kitchen appliance for shredding, blending, chopping, or slicing food', 'name': 'food_processor'}, {'frequency': 'c', 'synset': 'football.n.02', 'synonyms': ['football_(American)'], 'id': 466, 'def': 'the inflated oblong ball used in playing American football', 'name': 'football_(American)'}, {'frequency': 'r', 'synset': 'football_helmet.n.01', 'synonyms': ['football_helmet'], 'id': 467, 'def': 'a padded helmet with a face mask to protect the head of football players', 'name': 'football_helmet'}, {'frequency': 'c', 'synset': 'footstool.n.01', 'synonyms': ['footstool', 'footrest'], 'id': 468, 'def': 'a low seat or a stool to rest the feet of a seated person', 'name': 'footstool'}, {'frequency': 'f', 'synset': 'fork.n.01', 'synonyms': ['fork'], 'id': 469, 'def': 'cutlery used for serving and eating food', 'name': 'fork'}, {'frequency': 'c', 'synset': 'forklift.n.01', 'synonyms': ['forklift'], 'id': 470, 'def': 'an industrial vehicle with a power operated fork in front that can be inserted under loads to lift and move them', 'name': 'forklift'}, {'frequency': 'c', 'synset': 'freight_car.n.01', 'synonyms': ['freight_car'], 'id': 471, 'def': 'a railway car that carries freight', 'name': 'freight_car'}, {'frequency': 'c', 'synset': 'french_toast.n.01', 'synonyms': ['French_toast'], 'id': 472, 'def': 'bread slice dipped in egg and milk and fried', 'name': 'French_toast'}, {'frequency': 'c', 'synset': 'freshener.n.01', 'synonyms': ['freshener', 'air_freshener'], 'id': 473, 'def': 'anything that freshens air by removing or covering odor', 'name': 'freshener'}, {'frequency': 'f', 'synset': 'frisbee.n.01', 'synonyms': ['frisbee'], 'id': 474, 'def': 'a light, plastic disk propelled with a flip of the wrist for recreation or competition', 'name': 'frisbee'}, {'frequency': 'c', 'synset': 'frog.n.01', 'synonyms': ['frog', 'toad', 'toad_frog'], 'id': 475, 'def': 'a tailless stout-bodied amphibians with long hind limbs for leaping', 'name': 'frog'}, {'frequency': 'c', 'synset': 'fruit_juice.n.01', 'synonyms': ['fruit_juice'], 'id': 476, 'def': 'drink produced by squeezing or crushing fruit', 'name': 'fruit_juice'}, {'frequency': 'f', 'synset': 'frying_pan.n.01', 'synonyms': ['frying_pan', 'frypan', 'skillet'], 'id': 477, 'def': 'a pan used for frying foods', 'name': 'frying_pan'}, {'frequency': 'r', 'synset': 'fudge.n.01', 'synonyms': ['fudge'], 'id': 478, 'def': 'soft creamy candy', 'name': 'fudge'}, {'frequency': 'r', 'synset': 'funnel.n.02', 'synonyms': ['funnel'], 'id': 479, 'def': 'a cone-shaped utensil used to channel a substance into a container with a small mouth', 'name': 'funnel'}, {'frequency': 'r', 'synset': 'futon.n.01', 'synonyms': ['futon'], 'id': 480, 'def': 'a pad that is used for sleeping on the floor or on a raised frame', 'name': 'futon'}, {'frequency': 'r', 'synset': 'gag.n.02', 'synonyms': ['gag', 'muzzle'], 'id': 481, 'def': "restraint put into a person's mouth to prevent speaking or shouting", 'name': 'gag'}, {'frequency': 'r', 'synset': 'garbage.n.03', 'synonyms': ['garbage'], 'id': 482, 'def': 'a receptacle where waste can be discarded', 'name': 'garbage'}, {'frequency': 'c', 'synset': 'garbage_truck.n.01', 'synonyms': ['garbage_truck'], 'id': 483, 'def': 'a truck for collecting domestic refuse', 'name': 'garbage_truck'}, {'frequency': 'c', 'synset': 'garden_hose.n.01', 'synonyms': ['garden_hose'], 'id': 484, 'def': 'a hose used for watering a lawn or garden', 'name': 'garden_hose'}, {'frequency': 'c', 'synset': 'gargle.n.01', 'synonyms': ['gargle', 'mouthwash'], 'id': 485, 'def': 'a medicated solution used for gargling and rinsing the mouth', 'name': 'gargle'}, {'frequency': 'r', 'synset': 'gargoyle.n.02', 'synonyms': ['gargoyle'], 'id': 486, 'def': 'an ornament consisting of a grotesquely carved figure of a person or animal', 'name': 'gargoyle'}, {'frequency': 'c', 'synset': 'garlic.n.02', 'synonyms': ['garlic', 'ail'], 'id': 487, 'def': 'aromatic bulb used as seasoning', 'name': 'garlic'}, {'frequency': 'r', 'synset': 'gasmask.n.01', 'synonyms': ['gasmask', 'respirator', 'gas_helmet'], 'id': 488, 'def': 'a protective face mask with a filter', 'name': 'gasmask'}, {'frequency': 'c', 'synset': 'gazelle.n.01', 'synonyms': ['gazelle'], 'id': 489, 'def': 'small swift graceful antelope of Africa and Asia having lustrous eyes', 'name': 'gazelle'}, {'frequency': 'c', 'synset': 'gelatin.n.02', 'synonyms': ['gelatin', 'jelly'], 'id': 490, 'def': 'an edible jelly made with gelatin and used as a dessert or salad base or a coating for foods', 'name': 'gelatin'}, {'frequency': 'r', 'synset': 'gem.n.02', 'synonyms': ['gemstone'], 'id': 491, 'def': 'a crystalline rock that can be cut and polished for jewelry', 'name': 'gemstone'}, {'frequency': 'r', 'synset': 'generator.n.02', 'synonyms': ['generator'], 'id': 492, 'def': 'engine that converts mechanical energy into electrical energy by electromagnetic induction', 'name': 'generator'}, {'frequency': 'c', 'synset': 'giant_panda.n.01', 'synonyms': ['giant_panda', 'panda', 'panda_bear'], 'id': 493, 'def': 'large black-and-white herbivorous mammal of bamboo forests of China and Tibet', 'name': 'giant_panda'}, {'frequency': 'c', 'synset': 'gift_wrap.n.01', 'synonyms': ['gift_wrap'], 'id': 494, 'def': 'attractive wrapping paper suitable for wrapping gifts', 'name': 'gift_wrap'}, {'frequency': 'c', 'synset': 'ginger.n.03', 'synonyms': ['ginger', 'gingerroot'], 'id': 495, 'def': 'the root of the common ginger plant; used fresh as a seasoning', 'name': 'ginger'}, {'frequency': 'f', 'synset': 'giraffe.n.01', 'synonyms': ['giraffe'], 'id': 496, 'def': 'tall animal having a spotted coat and small horns and very long neck and legs', 'name': 'giraffe'}, {'frequency': 'c', 'synset': 'girdle.n.02', 'synonyms': ['cincture', 'sash', 'waistband', 'waistcloth'], 'id': 497, 'def': 'a band of material around the waist that strengthens a skirt or trousers', 'name': 'cincture'}, {'frequency': 'f', 'synset': 'glass.n.02', 'synonyms': ['glass_(drink_container)', 'drinking_glass'], 'id': 498, 'def': 'a container for holding liquids while drinking', 'name': 'glass_(drink_container)'}, {'frequency': 'c', 'synset': 'globe.n.03', 'synonyms': ['globe'], 'id': 499, 'def': 'a sphere on which a map (especially of the earth) is represented', 'name': 'globe'}, {'frequency': 'f', 'synset': 'glove.n.02', 'synonyms': ['glove'], 'id': 500, 'def': 'handwear covering the hand', 'name': 'glove'}, {'frequency': 'c', 'synset': 'goat.n.01', 'synonyms': ['goat'], 'id': 501, 'def': 'a common goat', 'name': 'goat'}, {'frequency': 'f', 'synset': 'goggles.n.01', 'synonyms': ['goggles'], 'id': 502, 'def': 'tight-fitting spectacles worn to protect the eyes', 'name': 'goggles'}, {'frequency': 'r', 'synset': 'goldfish.n.01', 'synonyms': ['goldfish'], 'id': 503, 'def': 'small golden or orange-red freshwater fishes used as pond or aquarium pets', 'name': 'goldfish'}, {'frequency': 'c', 'synset': 'golf_club.n.02', 'synonyms': ['golf_club', 'golf-club'], 'id': 504, 'def': 'golf equipment used by a golfer to hit a golf ball', 'name': 'golf_club'}, {'frequency': 'c', 'synset': 'golfcart.n.01', 'synonyms': ['golfcart'], 'id': 505, 'def': 'a small motor vehicle in which golfers can ride between shots', 'name': 'golfcart'}, {'frequency': 'r', 'synset': 'gondola.n.02', 'synonyms': ['gondola_(boat)'], 'id': 506, 'def': 'long narrow flat-bottomed boat propelled by sculling; traditionally used on canals of Venice', 'name': 'gondola_(boat)'}, {'frequency': 'c', 'synset': 'goose.n.01', 'synonyms': ['goose'], 'id': 507, 'def': 'loud, web-footed long-necked aquatic birds usually larger than ducks', 'name': 'goose'}, {'frequency': 'r', 'synset': 'gorilla.n.01', 'synonyms': ['gorilla'], 'id': 508, 'def': 'largest ape', 'name': 'gorilla'}, {'frequency': 'r', 'synset': 'gourd.n.02', 'synonyms': ['gourd'], 'id': 509, 'def': 'any of numerous inedible fruits with hard rinds', 'name': 'gourd'}, {'frequency': 'f', 'synset': 'grape.n.01', 'synonyms': ['grape'], 'id': 510, 'def': 'any of various juicy fruit with green or purple skins; grow in clusters', 'name': 'grape'}, {'frequency': 'c', 'synset': 'grater.n.01', 'synonyms': ['grater'], 'id': 511, 'def': 'utensil with sharp perforations for shredding foods (as vegetables or cheese)', 'name': 'grater'}, {'frequency': 'c', 'synset': 'gravestone.n.01', 'synonyms': ['gravestone', 'headstone', 'tombstone'], 'id': 512, 'def': 'a stone that is used to mark a grave', 'name': 'gravestone'}, {'frequency': 'r', 'synset': 'gravy_boat.n.01', 'synonyms': ['gravy_boat', 'gravy_holder'], 'id': 513, 'def': 'a dish (often boat-shaped) for serving gravy or sauce', 'name': 'gravy_boat'}, {'frequency': 'f', 'synset': 'green_bean.n.02', 'synonyms': ['green_bean'], 'id': 514, 'def': 'a common bean plant cultivated for its slender green edible pods', 'name': 'green_bean'}, {'frequency': 'f', 'synset': 'green_onion.n.01', 'synonyms': ['green_onion', 'spring_onion', 'scallion'], 'id': 515, 'def': 'a young onion before the bulb has enlarged', 'name': 'green_onion'}, {'frequency': 'r', 'synset': 'griddle.n.01', 'synonyms': ['griddle'], 'id': 516, 'def': 'cooking utensil consisting of a flat heated surface on which food is cooked', 'name': 'griddle'}, {'frequency': 'f', 'synset': 'grill.n.02', 'synonyms': ['grill', 'grille', 'grillwork', 'radiator_grille'], 'id': 517, 'def': 'a framework of metal bars used as a partition or a grate', 'name': 'grill'}, {'frequency': 'r', 'synset': 'grits.n.01', 'synonyms': ['grits', 'hominy_grits'], 'id': 518, 'def': 'coarsely ground corn boiled as a breakfast dish', 'name': 'grits'}, {'frequency': 'c', 'synset': 'grizzly.n.01', 'synonyms': ['grizzly', 'grizzly_bear'], 'id': 519, 'def': 'powerful brownish-yellow bear of the uplands of western North America', 'name': 'grizzly'}, {'frequency': 'c', 'synset': 'grocery_bag.n.01', 'synonyms': ['grocery_bag'], 'id': 520, 'def': "a sack for holding customer's groceries", 'name': 'grocery_bag'}, {'frequency': 'f', 'synset': 'guitar.n.01', 'synonyms': ['guitar'], 'id': 521, 'def': 'a stringed instrument usually having six strings; played by strumming or plucking', 'name': 'guitar'}, {'frequency': 'c', 'synset': 'gull.n.02', 'synonyms': ['gull', 'seagull'], 'id': 522, 'def': 'mostly white aquatic bird having long pointed wings and short legs', 'name': 'gull'}, {'frequency': 'c', 'synset': 'gun.n.01', 'synonyms': ['gun'], 'id': 523, 'def': 'a weapon that discharges a bullet at high velocity from a metal tube', 'name': 'gun'}, {'frequency': 'f', 'synset': 'hairbrush.n.01', 'synonyms': ['hairbrush'], 'id': 524, 'def': "a brush used to groom a person's hair", 'name': 'hairbrush'}, {'frequency': 'c', 'synset': 'hairnet.n.01', 'synonyms': ['hairnet'], 'id': 525, 'def': 'a small net that someone wears over their hair to keep it in place', 'name': 'hairnet'}, {'frequency': 'c', 'synset': 'hairpin.n.01', 'synonyms': ['hairpin'], 'id': 526, 'def': "a double pronged pin used to hold women's hair in place", 'name': 'hairpin'}, {'frequency': 'r', 'synset': 'halter.n.03', 'synonyms': ['halter_top'], 'id': 527, 'def': "a woman's top that fastens behind the back and neck leaving the back and arms uncovered", 'name': 'halter_top'}, {'frequency': 'f', 'synset': 'ham.n.01', 'synonyms': ['ham', 'jambon', 'gammon'], 'id': 528, 'def': 'meat cut from the thigh of a hog (usually smoked)', 'name': 'ham'}, {'frequency': 'c', 'synset': 'hamburger.n.01', 'synonyms': ['hamburger', 'beefburger', 'burger'], 'id': 529, 'def': 'a sandwich consisting of a patty of minced beef served on a bun', 'name': 'hamburger'}, {'frequency': 'c', 'synset': 'hammer.n.02', 'synonyms': ['hammer'], 'id': 530, 'def': 'a hand tool with a heavy head and a handle; used to deliver an impulsive force by striking', 'name': 'hammer'}, {'frequency': 'c', 'synset': 'hammock.n.02', 'synonyms': ['hammock'], 'id': 531, 'def': 'a hanging bed of canvas or rope netting (usually suspended between two trees)', 'name': 'hammock'}, {'frequency': 'r', 'synset': 'hamper.n.02', 'synonyms': ['hamper'], 'id': 532, 'def': 'a basket usually with a cover', 'name': 'hamper'}, {'frequency': 'c', 'synset': 'hamster.n.01', 'synonyms': ['hamster'], 'id': 533, 'def': 'short-tailed burrowing rodent with large cheek pouches', 'name': 'hamster'}, {'frequency': 'f', 'synset': 'hand_blower.n.01', 'synonyms': ['hair_dryer'], 'id': 534, 'def': 'a hand-held electric blower that can blow warm air onto the hair', 'name': 'hair_dryer'}, {'frequency': 'r', 'synset': 'hand_glass.n.01', 'synonyms': ['hand_glass', 'hand_mirror'], 'id': 535, 'def': 'a mirror intended to be held in the hand', 'name': 'hand_glass'}, {'frequency': 'f', 'synset': 'hand_towel.n.01', 'synonyms': ['hand_towel', 'face_towel'], 'id': 536, 'def': 'a small towel used to dry the hands or face', 'name': 'hand_towel'}, {'frequency': 'c', 'synset': 'handcart.n.01', 'synonyms': ['handcart', 'pushcart', 'hand_truck'], 'id': 537, 'def': 'wheeled vehicle that can be pushed by a person', 'name': 'handcart'}, {'frequency': 'r', 'synset': 'handcuff.n.01', 'synonyms': ['handcuff'], 'id': 538, 'def': 'shackle that consists of a metal loop that can be locked around the wrist', 'name': 'handcuff'}, {'frequency': 'c', 'synset': 'handkerchief.n.01', 'synonyms': ['handkerchief'], 'id': 539, 'def': 'a square piece of cloth used for wiping the eyes or nose or as a costume accessory', 'name': 'handkerchief'}, {'frequency': 'f', 'synset': 'handle.n.01', 'synonyms': ['handle', 'grip', 'handgrip'], 'id': 540, 'def': 'the appendage to an object that is designed to be held in order to use or move it', 'name': 'handle'}, {'frequency': 'r', 'synset': 'handsaw.n.01', 'synonyms': ['handsaw', "carpenter's_saw"], 'id': 541, 'def': 'a saw used with one hand for cutting wood', 'name': 'handsaw'}, {'frequency': 'r', 'synset': 'hardback.n.01', 'synonyms': ['hardback_book', 'hardcover_book'], 'id': 542, 'def': 'a book with cardboard or cloth or leather covers', 'name': 'hardback_book'}, {'frequency': 'r', 'synset': 'harmonium.n.01', 'synonyms': ['harmonium', 'organ_(musical_instrument)', 'reed_organ_(musical_instrument)'], 'id': 543, 'def': 'a free-reed instrument in which air is forced through the reeds by bellows', 'name': 'harmonium'}, {'frequency': 'f', 'synset': 'hat.n.01', 'synonyms': ['hat'], 'id': 544, 'def': 'headwear that protects the head from bad weather, sun, or worn for fashion', 'name': 'hat'}, {'frequency': 'r', 'synset': 'hatbox.n.01', 'synonyms': ['hatbox'], 'id': 545, 'def': 'a round piece of luggage for carrying hats', 'name': 'hatbox'}, {'frequency': 'c', 'synset': 'head_covering.n.01', 'synonyms': ['veil'], 'id': 546, 'def': 'a garment that covers the head OR face', 'name': 'veil'}, {'frequency': 'f', 'synset': 'headband.n.01', 'synonyms': ['headband'], 'id': 547, 'def': 'a band worn around or over the head', 'name': 'headband'}, {'frequency': 'f', 'synset': 'headboard.n.01', 'synonyms': ['headboard'], 'id': 548, 'def': 'a vertical board or panel forming the head of a bedstead', 'name': 'headboard'}, {'frequency': 'f', 'synset': 'headlight.n.01', 'synonyms': ['headlight', 'headlamp'], 'id': 549, 'def': 'a powerful light with reflector; attached to the front of an automobile or locomotive', 'name': 'headlight'}, {'frequency': 'c', 'synset': 'headscarf.n.01', 'synonyms': ['headscarf'], 'id': 550, 'def': 'a kerchief worn over the head and tied under the chin', 'name': 'headscarf'}, {'frequency': 'r', 'synset': 'headset.n.01', 'synonyms': ['headset'], 'id': 551, 'def': 'receiver consisting of a pair of headphones', 'name': 'headset'}, {'frequency': 'c', 'synset': 'headstall.n.01', 'synonyms': ['headstall_(for_horses)', 'headpiece_(for_horses)'], 'id': 552, 'def': "the band that is the part of a bridle that fits around a horse's head", 'name': 'headstall_(for_horses)'}, {'frequency': 'c', 'synset': 'heart.n.02', 'synonyms': ['heart'], 'id': 553, 'def': 'a muscular organ; its contractions move the blood through the body', 'name': 'heart'}, {'frequency': 'c', 'synset': 'heater.n.01', 'synonyms': ['heater', 'warmer'], 'id': 554, 'def': 'device that heats water or supplies warmth to a room', 'name': 'heater'}, {'frequency': 'c', 'synset': 'helicopter.n.01', 'synonyms': ['helicopter'], 'id': 555, 'def': 'an aircraft without wings that obtains its lift from the rotation of overhead blades', 'name': 'helicopter'}, {'frequency': 'f', 'synset': 'helmet.n.02', 'synonyms': ['helmet'], 'id': 556, 'def': 'a protective headgear made of hard material to resist blows', 'name': 'helmet'}, {'frequency': 'r', 'synset': 'heron.n.02', 'synonyms': ['heron'], 'id': 557, 'def': 'grey or white wading bird with long neck and long legs and (usually) long bill', 'name': 'heron'}, {'frequency': 'c', 'synset': 'highchair.n.01', 'synonyms': ['highchair', 'feeding_chair'], 'id': 558, 'def': 'a chair for feeding a very young child', 'name': 'highchair'}, {'frequency': 'f', 'synset': 'hinge.n.01', 'synonyms': ['hinge'], 'id': 559, 'def': 'a joint that holds two parts together so that one can swing relative to the other', 'name': 'hinge'}, {'frequency': 'r', 'synset': 'hippopotamus.n.01', 'synonyms': ['hippopotamus'], 'id': 560, 'def': 'massive thick-skinned animal living in or around rivers of tropical Africa', 'name': 'hippopotamus'}, {'frequency': 'r', 'synset': 'hockey_stick.n.01', 'synonyms': ['hockey_stick'], 'id': 561, 'def': 'sports implement consisting of a stick used by hockey players to move the puck', 'name': 'hockey_stick'}, {'frequency': 'c', 'synset': 'hog.n.03', 'synonyms': ['hog', 'pig'], 'id': 562, 'def': 'domestic swine', 'name': 'hog'}, {'frequency': 'f', 'synset': 'home_plate.n.01', 'synonyms': ['home_plate_(baseball)', 'home_base_(baseball)'], 'id': 563, 'def': '(baseball) a rubber slab where the batter stands; it must be touched by a base runner in order to score', 'name': 'home_plate_(baseball)'}, {'frequency': 'c', 'synset': 'honey.n.01', 'synonyms': ['honey'], 'id': 564, 'def': 'a sweet yellow liquid produced by bees', 'name': 'honey'}, {'frequency': 'f', 'synset': 'hood.n.06', 'synonyms': ['fume_hood', 'exhaust_hood'], 'id': 565, 'def': 'metal covering leading to a vent that exhausts smoke or fumes', 'name': 'fume_hood'}, {'frequency': 'f', 'synset': 'hook.n.05', 'synonyms': ['hook'], 'id': 566, 'def': 'a curved or bent implement for suspending or pulling something', 'name': 'hook'}, {'frequency': 'r', 'synset': 'hookah.n.01', 'synonyms': ['hookah', 'narghile', 'nargileh', 'sheesha', 'shisha', 'water_pipe'], 'id': 567, 'def': 'a tobacco pipe with a long flexible tube connected to a container where the smoke is cooled by passing through water', 'name': 'hookah'}, {'frequency': 'r', 'synset': 'hornet.n.01', 'synonyms': ['hornet'], 'id': 568, 'def': 'large stinging wasp', 'name': 'hornet'}, {'frequency': 'f', 'synset': 'horse.n.01', 'synonyms': ['horse'], 'id': 569, 'def': 'a common horse', 'name': 'horse'}, {'frequency': 'f', 'synset': 'hose.n.03', 'synonyms': ['hose', 'hosepipe'], 'id': 570, 'def': 'a flexible pipe for conveying a liquid or gas', 'name': 'hose'}, {'frequency': 'r', 'synset': 'hot-air_balloon.n.01', 'synonyms': ['hot-air_balloon'], 'id': 571, 'def': 'balloon for travel through the air in a basket suspended below a large bag of heated air', 'name': 'hot-air_balloon'}, {'frequency': 'r', 'synset': 'hot_plate.n.01', 'synonyms': ['hotplate'], 'id': 572, 'def': 'a portable electric appliance for heating or cooking or keeping food warm', 'name': 'hotplate'}, {'frequency': 'c', 'synset': 'hot_sauce.n.01', 'synonyms': ['hot_sauce'], 'id': 573, 'def': 'a pungent peppery sauce', 'name': 'hot_sauce'}, {'frequency': 'r', 'synset': 'hourglass.n.01', 'synonyms': ['hourglass'], 'id': 574, 'def': 'a sandglass timer that runs for sixty minutes', 'name': 'hourglass'}, {'frequency': 'r', 'synset': 'houseboat.n.01', 'synonyms': ['houseboat'], 'id': 575, 'def': 'a barge that is designed and equipped for use as a dwelling', 'name': 'houseboat'}, {'frequency': 'c', 'synset': 'hummingbird.n.01', 'synonyms': ['hummingbird'], 'id': 576, 'def': 'tiny American bird having brilliant iridescent plumage and long slender bills', 'name': 'hummingbird'}, {'frequency': 'r', 'synset': 'hummus.n.01', 'synonyms': ['hummus', 'humus', 'hommos', 'hoummos', 'humous'], 'id': 577, 'def': 'a thick spread made from mashed chickpeas', 'name': 'hummus'}, {'frequency': 'f', 'synset': 'ice_bear.n.01', 'synonyms': ['polar_bear'], 'id': 578, 'def': 'white bear of Arctic regions', 'name': 'polar_bear'}, {'frequency': 'c', 'synset': 'ice_cream.n.01', 'synonyms': ['icecream'], 'id': 579, 'def': 'frozen dessert containing cream and sugar and flavoring', 'name': 'icecream'}, {'frequency': 'r', 'synset': 'ice_lolly.n.01', 'synonyms': ['popsicle'], 'id': 580, 'def': 'ice cream or water ice on a small wooden stick', 'name': 'popsicle'}, {'frequency': 'c', 'synset': 'ice_maker.n.01', 'synonyms': ['ice_maker'], 'id': 581, 'def': 'an appliance included in some electric refrigerators for making ice cubes', 'name': 'ice_maker'}, {'frequency': 'r', 'synset': 'ice_pack.n.01', 'synonyms': ['ice_pack', 'ice_bag'], 'id': 582, 'def': 'a waterproof bag filled with ice: applied to the body (especially the head) to cool or reduce swelling', 'name': 'ice_pack'}, {'frequency': 'r', 'synset': 'ice_skate.n.01', 'synonyms': ['ice_skate'], 'id': 583, 'def': 'skate consisting of a boot with a steel blade fitted to the sole', 'name': 'ice_skate'}, {'frequency': 'c', 'synset': 'igniter.n.01', 'synonyms': ['igniter', 'ignitor', 'lighter'], 'id': 584, 'def': 'a substance or device used to start a fire', 'name': 'igniter'}, {'frequency': 'r', 'synset': 'inhaler.n.01', 'synonyms': ['inhaler', 'inhalator'], 'id': 585, 'def': 'a dispenser that produces a chemical vapor to be inhaled through mouth or nose', 'name': 'inhaler'}, {'frequency': 'f', 'synset': 'ipod.n.01', 'synonyms': ['iPod'], 'id': 586, 'def': 'a pocket-sized device used to play music files', 'name': 'iPod'}, {'frequency': 'c', 'synset': 'iron.n.04', 'synonyms': ['iron_(for_clothing)', 'smoothing_iron_(for_clothing)'], 'id': 587, 'def': 'home appliance consisting of a flat metal base that is heated and used to smooth cloth', 'name': 'iron_(for_clothing)'}, {'frequency': 'c', 'synset': 'ironing_board.n.01', 'synonyms': ['ironing_board'], 'id': 588, 'def': 'narrow padded board on collapsible supports; used for ironing clothes', 'name': 'ironing_board'}, {'frequency': 'f', 'synset': 'jacket.n.01', 'synonyms': ['jacket'], 'id': 589, 'def': 'a waist-length coat', 'name': 'jacket'}, {'frequency': 'c', 'synset': 'jam.n.01', 'synonyms': ['jam'], 'id': 590, 'def': 'preserve of crushed fruit', 'name': 'jam'}, {'frequency': 'f', 'synset': 'jar.n.01', 'synonyms': ['jar'], 'id': 591, 'def': 'a vessel (usually cylindrical) with a wide mouth and without handles', 'name': 'jar'}, {'frequency': 'f', 'synset': 'jean.n.01', 'synonyms': ['jean', 'blue_jean', 'denim'], 'id': 592, 'def': '(usually plural) close-fitting trousers of heavy denim for manual work or casual wear', 'name': 'jean'}, {'frequency': 'c', 'synset': 'jeep.n.01', 'synonyms': ['jeep', 'landrover'], 'id': 593, 'def': 'a car suitable for traveling over rough terrain', 'name': 'jeep'}, {'frequency': 'r', 'synset': 'jelly_bean.n.01', 'synonyms': ['jelly_bean', 'jelly_egg'], 'id': 594, 'def': 'sugar-glazed jellied candy', 'name': 'jelly_bean'}, {'frequency': 'f', 'synset': 'jersey.n.03', 'synonyms': ['jersey', 'T-shirt', 'tee_shirt'], 'id': 595, 'def': 'a close-fitting pullover shirt', 'name': 'jersey'}, {'frequency': 'c', 'synset': 'jet.n.01', 'synonyms': ['jet_plane', 'jet-propelled_plane'], 'id': 596, 'def': 'an airplane powered by one or more jet engines', 'name': 'jet_plane'}, {'frequency': 'r', 'synset': 'jewel.n.01', 'synonyms': ['jewel', 'gem', 'precious_stone'], 'id': 597, 'def': 'a precious or semiprecious stone incorporated into a piece of jewelry', 'name': 'jewel'}, {'frequency': 'c', 'synset': 'jewelry.n.01', 'synonyms': ['jewelry', 'jewellery'], 'id': 598, 'def': 'an adornment (as a bracelet or ring or necklace) made of precious metals and set with gems (or imitation gems)', 'name': 'jewelry'}, {'frequency': 'r', 'synset': 'joystick.n.02', 'synonyms': ['joystick'], 'id': 599, 'def': 'a control device for computers consisting of a vertical handle that can move freely in two directions', 'name': 'joystick'}, {'frequency': 'c', 'synset': 'jump_suit.n.01', 'synonyms': ['jumpsuit'], 'id': 600, 'def': "one-piece garment fashioned after a parachutist's uniform", 'name': 'jumpsuit'}, {'frequency': 'c', 'synset': 'kayak.n.01', 'synonyms': ['kayak'], 'id': 601, 'def': 'a small canoe consisting of a light frame made watertight with animal skins', 'name': 'kayak'}, {'frequency': 'r', 'synset': 'keg.n.02', 'synonyms': ['keg'], 'id': 602, 'def': 'small cask or barrel', 'name': 'keg'}, {'frequency': 'r', 'synset': 'kennel.n.01', 'synonyms': ['kennel', 'doghouse'], 'id': 603, 'def': 'outbuilding that serves as a shelter for a dog', 'name': 'kennel'}, {'frequency': 'c', 'synset': 'kettle.n.01', 'synonyms': ['kettle', 'boiler'], 'id': 604, 'def': 'a metal pot for stewing or boiling; usually has a lid', 'name': 'kettle'}, {'frequency': 'f', 'synset': 'key.n.01', 'synonyms': ['key'], 'id': 605, 'def': 'metal instrument used to unlock a lock', 'name': 'key'}, {'frequency': 'r', 'synset': 'keycard.n.01', 'synonyms': ['keycard'], 'id': 606, 'def': 'a plastic card used to gain access typically to a door', 'name': 'keycard'}, {'frequency': 'c', 'synset': 'kilt.n.01', 'synonyms': ['kilt'], 'id': 607, 'def': 'a knee-length pleated tartan skirt worn by men as part of the traditional dress in the Highlands of northern Scotland', 'name': 'kilt'}, {'frequency': 'c', 'synset': 'kimono.n.01', 'synonyms': ['kimono'], 'id': 608, 'def': 'a loose robe; imitated from robes originally worn by Japanese', 'name': 'kimono'}, {'frequency': 'f', 'synset': 'kitchen_sink.n.01', 'synonyms': ['kitchen_sink'], 'id': 609, 'def': 'a sink in a kitchen', 'name': 'kitchen_sink'}, {'frequency': 'r', 'synset': 'kitchen_table.n.01', 'synonyms': ['kitchen_table'], 'id': 610, 'def': 'a table in the kitchen', 'name': 'kitchen_table'}, {'frequency': 'f', 'synset': 'kite.n.03', 'synonyms': ['kite'], 'id': 611, 'def': 'plaything consisting of a light frame covered with tissue paper; flown in wind at end of a string', 'name': 'kite'}, {'frequency': 'c', 'synset': 'kitten.n.01', 'synonyms': ['kitten', 'kitty'], 'id': 612, 'def': 'young domestic cat', 'name': 'kitten'}, {'frequency': 'c', 'synset': 'kiwi.n.03', 'synonyms': ['kiwi_fruit'], 'id': 613, 'def': 'fuzzy brown egg-shaped fruit with slightly tart green flesh', 'name': 'kiwi_fruit'}, {'frequency': 'f', 'synset': 'knee_pad.n.01', 'synonyms': ['knee_pad'], 'id': 614, 'def': 'protective garment consisting of a pad worn by football or baseball or hockey players', 'name': 'knee_pad'}, {'frequency': 'f', 'synset': 'knife.n.01', 'synonyms': ['knife'], 'id': 615, 'def': 'tool with a blade and point used as a cutting instrument', 'name': 'knife'}, {'frequency': 'r', 'synset': 'knitting_needle.n.01', 'synonyms': ['knitting_needle'], 'id': 616, 'def': 'needle consisting of a slender rod with pointed ends; usually used in pairs', 'name': 'knitting_needle'}, {'frequency': 'f', 'synset': 'knob.n.02', 'synonyms': ['knob'], 'id': 617, 'def': 'a round handle often found on a door', 'name': 'knob'}, {'frequency': 'r', 'synset': 'knocker.n.05', 'synonyms': ['knocker_(on_a_door)', 'doorknocker'], 'id': 618, 'def': 'a device (usually metal and ornamental) attached by a hinge to a door', 'name': 'knocker_(on_a_door)'}, {'frequency': 'r', 'synset': 'koala.n.01', 'synonyms': ['koala', 'koala_bear'], 'id': 619, 'def': 'sluggish tailless Australian marsupial with grey furry ears and coat', 'name': 'koala'}, {'frequency': 'r', 'synset': 'lab_coat.n.01', 'synonyms': ['lab_coat', 'laboratory_coat'], 'id': 620, 'def': 'a light coat worn to protect clothing from substances used while working in a laboratory', 'name': 'lab_coat'}, {'frequency': 'f', 'synset': 'ladder.n.01', 'synonyms': ['ladder'], 'id': 621, 'def': 'steps consisting of two parallel members connected by rungs', 'name': 'ladder'}, {'frequency': 'c', 'synset': 'ladle.n.01', 'synonyms': ['ladle'], 'id': 622, 'def': 'a spoon-shaped vessel with a long handle frequently used to transfer liquids', 'name': 'ladle'}, {'frequency': 'c', 'synset': 'ladybug.n.01', 'synonyms': ['ladybug', 'ladybeetle', 'ladybird_beetle'], 'id': 623, 'def': 'small round bright-colored and spotted beetle, typically red and black', 'name': 'ladybug'}, {'frequency': 'f', 'synset': 'lamb.n.01', 'synonyms': ['lamb_(animal)'], 'id': 624, 'def': 'young sheep', 'name': 'lamb_(animal)'}, {'frequency': 'r', 'synset': 'lamb_chop.n.01', 'synonyms': ['lamb-chop', 'lambchop'], 'id': 625, 'def': 'chop cut from a lamb', 'name': 'lamb-chop'}, {'frequency': 'f', 'synset': 'lamp.n.02', 'synonyms': ['lamp'], 'id': 626, 'def': 'a piece of furniture holding one or more electric light bulbs', 'name': 'lamp'}, {'frequency': 'f', 'synset': 'lamppost.n.01', 'synonyms': ['lamppost'], 'id': 627, 'def': 'a metal post supporting an outdoor lamp (such as a streetlight)', 'name': 'lamppost'}, {'frequency': 'f', 'synset': 'lampshade.n.01', 'synonyms': ['lampshade'], 'id': 628, 'def': 'a protective ornamental shade used to screen a light bulb from direct view', 'name': 'lampshade'}, {'frequency': 'c', 'synset': 'lantern.n.01', 'synonyms': ['lantern'], 'id': 629, 'def': 'light in a transparent protective case', 'name': 'lantern'}, {'frequency': 'f', 'synset': 'lanyard.n.02', 'synonyms': ['lanyard', 'laniard'], 'id': 630, 'def': 'a cord worn around the neck to hold a knife or whistle, etc.', 'name': 'lanyard'}, {'frequency': 'f', 'synset': 'laptop.n.01', 'synonyms': ['laptop_computer', 'notebook_computer'], 'id': 631, 'def': 'a portable computer small enough to use in your lap', 'name': 'laptop_computer'}, {'frequency': 'r', 'synset': 'lasagna.n.01', 'synonyms': ['lasagna', 'lasagne'], 'id': 632, 'def': 'baked dish of layers of lasagna pasta with sauce and cheese and meat or vegetables', 'name': 'lasagna'}, {'frequency': 'f', 'synset': 'latch.n.02', 'synonyms': ['latch'], 'id': 633, 'def': 'a bar that can be lowered or slid into a groove to fasten a door or gate', 'name': 'latch'}, {'frequency': 'r', 'synset': 'lawn_mower.n.01', 'synonyms': ['lawn_mower'], 'id': 634, 'def': 'garden tool for mowing grass on lawns', 'name': 'lawn_mower'}, {'frequency': 'r', 'synset': 'leather.n.01', 'synonyms': ['leather'], 'id': 635, 'def': 'an animal skin made smooth and flexible by removing the hair and then tanning', 'name': 'leather'}, {'frequency': 'c', 'synset': 'legging.n.01', 'synonyms': ['legging_(clothing)', 'leging_(clothing)', 'leg_covering'], 'id': 636, 'def': 'a garment covering the leg (usually extending from the knee to the ankle)', 'name': 'legging_(clothing)'}, {'frequency': 'c', 'synset': 'lego.n.01', 'synonyms': ['Lego', 'Lego_set'], 'id': 637, 'def': "a child's plastic construction set for making models from blocks", 'name': 'Lego'}, {'frequency': 'r', 'synset': 'legume.n.02', 'synonyms': ['legume'], 'id': 638, 'def': 'the fruit or seed of bean or pea plants', 'name': 'legume'}, {'frequency': 'f', 'synset': 'lemon.n.01', 'synonyms': ['lemon'], 'id': 639, 'def': 'yellow oval fruit with juicy acidic flesh', 'name': 'lemon'}, {'frequency': 'r', 'synset': 'lemonade.n.01', 'synonyms': ['lemonade'], 'id': 640, 'def': 'sweetened beverage of diluted lemon juice', 'name': 'lemonade'}, {'frequency': 'f', 'synset': 'lettuce.n.02', 'synonyms': ['lettuce'], 'id': 641, 'def': 'leafy plant commonly eaten in salad or on sandwiches', 'name': 'lettuce'}, {'frequency': 'f', 'synset': 'license_plate.n.01', 'synonyms': ['license_plate', 'numberplate'], 'id': 642, 'def': "a plate mounted on the front and back of car and bearing the car's registration number", 'name': 'license_plate'}, {'frequency': 'f', 'synset': 'life_buoy.n.01', 'synonyms': ['life_buoy', 'lifesaver', 'life_belt', 'life_ring'], 'id': 643, 'def': 'a ring-shaped life preserver used to prevent drowning (NOT a life-jacket or vest)', 'name': 'life_buoy'}, {'frequency': 'f', 'synset': 'life_jacket.n.01', 'synonyms': ['life_jacket', 'life_vest'], 'id': 644, 'def': 'life preserver consisting of a sleeveless jacket of buoyant or inflatable design', 'name': 'life_jacket'}, {'frequency': 'f', 'synset': 'light_bulb.n.01', 'synonyms': ['lightbulb'], 'id': 645, 'def': 'lightblub/source of light', 'name': 'lightbulb'}, {'frequency': 'r', 'synset': 'lightning_rod.n.02', 'synonyms': ['lightning_rod', 'lightning_conductor'], 'id': 646, 'def': 'a metallic conductor that is attached to a high point and leads to the ground', 'name': 'lightning_rod'}, {'frequency': 'f', 'synset': 'lime.n.06', 'synonyms': ['lime'], 'id': 647, 'def': 'the green acidic fruit of any of various lime trees', 'name': 'lime'}, {'frequency': 'r', 'synset': 'limousine.n.01', 'synonyms': ['limousine'], 'id': 648, 'def': 'long luxurious car; usually driven by a chauffeur', 'name': 'limousine'}, {'frequency': 'c', 'synset': 'lion.n.01', 'synonyms': ['lion'], 'id': 649, 'def': 'large gregarious predatory cat of Africa and India', 'name': 'lion'}, {'frequency': 'c', 'synset': 'lip_balm.n.01', 'synonyms': ['lip_balm'], 'id': 650, 'def': 'a balm applied to the lips', 'name': 'lip_balm'}, {'frequency': 'r', 'synset': 'liquor.n.01', 'synonyms': ['liquor', 'spirits', 'hard_liquor', 'liqueur', 'cordial'], 'id': 651, 'def': 'liquor or beer', 'name': 'liquor'}, {'frequency': 'c', 'synset': 'lizard.n.01', 'synonyms': ['lizard'], 'id': 652, 'def': 'a reptile with usually two pairs of legs and a tapering tail', 'name': 'lizard'}, {'frequency': 'f', 'synset': 'log.n.01', 'synonyms': ['log'], 'id': 653, 'def': 'a segment of the trunk of a tree when stripped of branches', 'name': 'log'}, {'frequency': 'c', 'synset': 'lollipop.n.02', 'synonyms': ['lollipop'], 'id': 654, 'def': 'hard candy on a stick', 'name': 'lollipop'}, {'frequency': 'f', 'synset': 'loudspeaker.n.01', 'synonyms': ['speaker_(stero_equipment)'], 'id': 655, 'def': 'electronic device that produces sound often as part of a stereo system', 'name': 'speaker_(stero_equipment)'}, {'frequency': 'c', 'synset': 'love_seat.n.01', 'synonyms': ['loveseat'], 'id': 656, 'def': 'small sofa that seats two people', 'name': 'loveseat'}, {'frequency': 'r', 'synset': 'machine_gun.n.01', 'synonyms': ['machine_gun'], 'id': 657, 'def': 'a rapidly firing automatic gun', 'name': 'machine_gun'}, {'frequency': 'f', 'synset': 'magazine.n.02', 'synonyms': ['magazine'], 'id': 658, 'def': 'a paperback periodic publication', 'name': 'magazine'}, {'frequency': 'f', 'synset': 'magnet.n.01', 'synonyms': ['magnet'], 'id': 659, 'def': 'a device that attracts iron and produces a magnetic field', 'name': 'magnet'}, {'frequency': 'c', 'synset': 'mail_slot.n.01', 'synonyms': ['mail_slot'], 'id': 660, 'def': 'a slot (usually in a door) through which mail can be delivered', 'name': 'mail_slot'}, {'frequency': 'f', 'synset': 'mailbox.n.01', 'synonyms': ['mailbox_(at_home)', 'letter_box_(at_home)'], 'id': 661, 'def': 'a private box for delivery of mail', 'name': 'mailbox_(at_home)'}, {'frequency': 'r', 'synset': 'mallard.n.01', 'synonyms': ['mallard'], 'id': 662, 'def': 'wild dabbling duck from which domestic ducks are descended', 'name': 'mallard'}, {'frequency': 'r', 'synset': 'mallet.n.01', 'synonyms': ['mallet'], 'id': 663, 'def': 'a sports implement with a long handle and a hammer-like head used to hit a ball', 'name': 'mallet'}, {'frequency': 'r', 'synset': 'mammoth.n.01', 'synonyms': ['mammoth'], 'id': 664, 'def': 'any of numerous extinct elephants widely distributed in the Pleistocene', 'name': 'mammoth'}, {'frequency': 'r', 'synset': 'manatee.n.01', 'synonyms': ['manatee'], 'id': 665, 'def': 'sirenian mammal of tropical coastal waters of America', 'name': 'manatee'}, {'frequency': 'c', 'synset': 'mandarin.n.05', 'synonyms': ['mandarin_orange'], 'id': 666, 'def': 'a somewhat flat reddish-orange loose skinned citrus of China', 'name': 'mandarin_orange'}, {'frequency': 'c', 'synset': 'manger.n.01', 'synonyms': ['manger', 'trough'], 'id': 667, 'def': 'a container (usually in a barn or stable) from which cattle or horses feed', 'name': 'manger'}, {'frequency': 'f', 'synset': 'manhole.n.01', 'synonyms': ['manhole'], 'id': 668, 'def': 'a hole (usually with a flush cover) through which a person can gain access to an underground structure', 'name': 'manhole'}, {'frequency': 'f', 'synset': 'map.n.01', 'synonyms': ['map'], 'id': 669, 'def': "a diagrammatic representation of the earth's surface (or part of it)", 'name': 'map'}, {'frequency': 'f', 'synset': 'marker.n.03', 'synonyms': ['marker'], 'id': 670, 'def': 'a writing implement for making a mark', 'name': 'marker'}, {'frequency': 'r', 'synset': 'martini.n.01', 'synonyms': ['martini'], 'id': 671, 'def': 'a cocktail made of gin (or vodka) with dry vermouth', 'name': 'martini'}, {'frequency': 'r', 'synset': 'mascot.n.01', 'synonyms': ['mascot'], 'id': 672, 'def': 'a person or animal that is adopted by a team or other group as a symbolic figure', 'name': 'mascot'}, {'frequency': 'c', 'synset': 'mashed_potato.n.01', 'synonyms': ['mashed_potato'], 'id': 673, 'def': 'potato that has been peeled and boiled and then mashed', 'name': 'mashed_potato'}, {'frequency': 'r', 'synset': 'masher.n.02', 'synonyms': ['masher'], 'id': 674, 'def': 'a kitchen utensil used for mashing (e.g. potatoes)', 'name': 'masher'}, {'frequency': 'f', 'synset': 'mask.n.04', 'synonyms': ['mask', 'facemask'], 'id': 675, 'def': 'a protective covering worn over the face', 'name': 'mask'}, {'frequency': 'f', 'synset': 'mast.n.01', 'synonyms': ['mast'], 'id': 676, 'def': 'a vertical spar for supporting sails', 'name': 'mast'}, {'frequency': 'c', 'synset': 'mat.n.03', 'synonyms': ['mat_(gym_equipment)', 'gym_mat'], 'id': 677, 'def': 'sports equipment consisting of a piece of thick padding on the floor for gymnastics', 'name': 'mat_(gym_equipment)'}, {'frequency': 'r', 'synset': 'matchbox.n.01', 'synonyms': ['matchbox'], 'id': 678, 'def': 'a box for holding matches', 'name': 'matchbox'}, {'frequency': 'f', 'synset': 'mattress.n.01', 'synonyms': ['mattress'], 'id': 679, 'def': 'a thick pad filled with resilient material used as a bed or part of a bed', 'name': 'mattress'}, {'frequency': 'c', 'synset': 'measuring_cup.n.01', 'synonyms': ['measuring_cup'], 'id': 680, 'def': 'graduated cup used to measure liquid or granular ingredients', 'name': 'measuring_cup'}, {'frequency': 'c', 'synset': 'measuring_stick.n.01', 'synonyms': ['measuring_stick', 'ruler_(measuring_stick)', 'measuring_rod'], 'id': 681, 'def': 'measuring instrument having a sequence of marks at regular intervals', 'name': 'measuring_stick'}, {'frequency': 'c', 'synset': 'meatball.n.01', 'synonyms': ['meatball'], 'id': 682, 'def': 'ground meat formed into a ball and fried or simmered in broth', 'name': 'meatball'}, {'frequency': 'c', 'synset': 'medicine.n.02', 'synonyms': ['medicine'], 'id': 683, 'def': 'something that treats or prevents or alleviates the symptoms of disease', 'name': 'medicine'}, {'frequency': 'c', 'synset': 'melon.n.01', 'synonyms': ['melon'], 'id': 684, 'def': 'fruit of the gourd family having a hard rind and sweet juicy flesh', 'name': 'melon'}, {'frequency': 'f', 'synset': 'microphone.n.01', 'synonyms': ['microphone'], 'id': 685, 'def': 'device for converting sound waves into electrical energy', 'name': 'microphone'}, {'frequency': 'r', 'synset': 'microscope.n.01', 'synonyms': ['microscope'], 'id': 686, 'def': 'magnifier of the image of small objects', 'name': 'microscope'}, {'frequency': 'f', 'synset': 'microwave.n.02', 'synonyms': ['microwave_oven'], 'id': 687, 'def': 'kitchen appliance that cooks food by passing an electromagnetic wave through it', 'name': 'microwave_oven'}, {'frequency': 'r', 'synset': 'milestone.n.01', 'synonyms': ['milestone', 'milepost'], 'id': 688, 'def': 'stone post at side of a road to show distances', 'name': 'milestone'}, {'frequency': 'f', 'synset': 'milk.n.01', 'synonyms': ['milk'], 'id': 689, 'def': 'a white nutritious liquid secreted by mammals and used as food by human beings', 'name': 'milk'}, {'frequency': 'r', 'synset': 'milk_can.n.01', 'synonyms': ['milk_can'], 'id': 690, 'def': 'can for transporting milk', 'name': 'milk_can'}, {'frequency': 'r', 'synset': 'milkshake.n.01', 'synonyms': ['milkshake'], 'id': 691, 'def': 'frothy drink of milk and flavoring and sometimes fruit or ice cream', 'name': 'milkshake'}, {'frequency': 'f', 'synset': 'minivan.n.01', 'synonyms': ['minivan'], 'id': 692, 'def': 'a small box-shaped passenger van', 'name': 'minivan'}, {'frequency': 'r', 'synset': 'mint.n.05', 'synonyms': ['mint_candy'], 'id': 693, 'def': 'a candy that is flavored with a mint oil', 'name': 'mint_candy'}, {'frequency': 'f', 'synset': 'mirror.n.01', 'synonyms': ['mirror'], 'id': 694, 'def': 'polished surface that forms images by reflecting light', 'name': 'mirror'}, {'frequency': 'c', 'synset': 'mitten.n.01', 'synonyms': ['mitten'], 'id': 695, 'def': 'glove that encases the thumb separately and the other four fingers together', 'name': 'mitten'}, {'frequency': 'c', 'synset': 'mixer.n.04', 'synonyms': ['mixer_(kitchen_tool)', 'stand_mixer'], 'id': 696, 'def': 'a kitchen utensil that is used for mixing foods', 'name': 'mixer_(kitchen_tool)'}, {'frequency': 'c', 'synset': 'money.n.03', 'synonyms': ['money'], 'id': 697, 'def': 'the official currency issued by a government or national bank', 'name': 'money'}, {'frequency': 'f', 'synset': 'monitor.n.04', 'synonyms': ['monitor_(computer_equipment) computer_monitor'], 'id': 698, 'def': 'a computer monitor', 'name': 'monitor_(computer_equipment) computer_monitor'}, {'frequency': 'c', 'synset': 'monkey.n.01', 'synonyms': ['monkey'], 'id': 699, 'def': 'any of various long-tailed primates', 'name': 'monkey'}, {'frequency': 'f', 'synset': 'motor.n.01', 'synonyms': ['motor'], 'id': 700, 'def': 'machine that converts other forms of energy into mechanical energy and so imparts motion', 'name': 'motor'}, {'frequency': 'f', 'synset': 'motor_scooter.n.01', 'synonyms': ['motor_scooter', 'scooter'], 'id': 701, 'def': 'a wheeled vehicle with small wheels and a low-powered engine', 'name': 'motor_scooter'}, {'frequency': 'r', 'synset': 'motor_vehicle.n.01', 'synonyms': ['motor_vehicle', 'automotive_vehicle'], 'id': 702, 'def': 'a self-propelled wheeled vehicle that does not run on rails', 'name': 'motor_vehicle'}, {'frequency': 'f', 'synset': 'motorcycle.n.01', 'synonyms': ['motorcycle'], 'id': 703, 'def': 'a motor vehicle with two wheels and a strong frame', 'name': 'motorcycle'}, {'frequency': 'f', 'synset': 'mound.n.01', 'synonyms': ['mound_(baseball)', "pitcher's_mound"], 'id': 704, 'def': '(baseball) the slight elevation on which the pitcher stands', 'name': 'mound_(baseball)'}, {'frequency': 'f', 'synset': 'mouse.n.04', 'synonyms': ['mouse_(computer_equipment)', 'computer_mouse'], 'id': 705, 'def': 'a computer input device that controls an on-screen pointer (does not include trackpads / touchpads)', 'name': 'mouse_(computer_equipment)'}, {'frequency': 'f', 'synset': 'mousepad.n.01', 'synonyms': ['mousepad'], 'id': 706, 'def': 'a small portable pad that provides an operating surface for a computer mouse', 'name': 'mousepad'}, {'frequency': 'c', 'synset': 'muffin.n.01', 'synonyms': ['muffin'], 'id': 707, 'def': 'a sweet quick bread baked in a cup-shaped pan', 'name': 'muffin'}, {'frequency': 'f', 'synset': 'mug.n.04', 'synonyms': ['mug'], 'id': 708, 'def': 'with handle and usually cylindrical', 'name': 'mug'}, {'frequency': 'f', 'synset': 'mushroom.n.02', 'synonyms': ['mushroom'], 'id': 709, 'def': 'a common mushroom', 'name': 'mushroom'}, {'frequency': 'r', 'synset': 'music_stool.n.01', 'synonyms': ['music_stool', 'piano_stool'], 'id': 710, 'def': 'a stool for piano players; usually adjustable in height', 'name': 'music_stool'}, {'frequency': 'c', 'synset': 'musical_instrument.n.01', 'synonyms': ['musical_instrument', 'instrument_(musical)'], 'id': 711, 'def': 'any of various devices or contrivances that can be used to produce musical tones or sounds', 'name': 'musical_instrument'}, {'frequency': 'r', 'synset': 'nailfile.n.01', 'synonyms': ['nailfile'], 'id': 712, 'def': 'a small flat file for shaping the nails', 'name': 'nailfile'}, {'frequency': 'f', 'synset': 'napkin.n.01', 'synonyms': ['napkin', 'table_napkin', 'serviette'], 'id': 713, 'def': 'a small piece of table linen or paper that is used to wipe the mouth and to cover the lap in order to protect clothing', 'name': 'napkin'}, {'frequency': 'r', 'synset': 'neckerchief.n.01', 'synonyms': ['neckerchief'], 'id': 714, 'def': 'a kerchief worn around the neck', 'name': 'neckerchief'}, {'frequency': 'f', 'synset': 'necklace.n.01', 'synonyms': ['necklace'], 'id': 715, 'def': 'jewelry consisting of a cord or chain (often bearing gems) worn about the neck as an ornament', 'name': 'necklace'}, {'frequency': 'f', 'synset': 'necktie.n.01', 'synonyms': ['necktie', 'tie_(necktie)'], 'id': 716, 'def': 'neckwear consisting of a long narrow piece of material worn under a collar and tied in knot at the front', 'name': 'necktie'}, {'frequency': 'c', 'synset': 'needle.n.03', 'synonyms': ['needle'], 'id': 717, 'def': 'a sharp pointed implement (usually metal)', 'name': 'needle'}, {'frequency': 'c', 'synset': 'nest.n.01', 'synonyms': ['nest'], 'id': 718, 'def': 'a structure in which animals lay eggs or give birth to their young', 'name': 'nest'}, {'frequency': 'f', 'synset': 'newspaper.n.01', 'synonyms': ['newspaper', 'paper_(newspaper)'], 'id': 719, 'def': 'a daily or weekly publication on folded sheets containing news, articles, and advertisements', 'name': 'newspaper'}, {'frequency': 'c', 'synset': 'newsstand.n.01', 'synonyms': ['newsstand'], 'id': 720, 'def': 'a stall where newspapers and other periodicals are sold', 'name': 'newsstand'}, {'frequency': 'c', 'synset': 'nightwear.n.01', 'synonyms': ['nightshirt', 'nightwear', 'sleepwear', 'nightclothes'], 'id': 721, 'def': 'garments designed to be worn in bed', 'name': 'nightshirt'}, {'frequency': 'r', 'synset': 'nosebag.n.01', 'synonyms': ['nosebag_(for_animals)', 'feedbag'], 'id': 722, 'def': 'a canvas bag that is used to feed an animal (such as a horse); covers the muzzle and fastens at the top of the head', 'name': 'nosebag_(for_animals)'}, {'frequency': 'c', 'synset': 'noseband.n.01', 'synonyms': ['noseband_(for_animals)', 'nosepiece_(for_animals)'], 'id': 723, 'def': "a strap that is the part of a bridle that goes over the animal's nose", 'name': 'noseband_(for_animals)'}, {'frequency': 'f', 'synset': 'notebook.n.01', 'synonyms': ['notebook'], 'id': 724, 'def': 'a book with blank pages for recording notes or memoranda', 'name': 'notebook'}, {'frequency': 'c', 'synset': 'notepad.n.01', 'synonyms': ['notepad'], 'id': 725, 'def': 'a pad of paper for keeping notes', 'name': 'notepad'}, {'frequency': 'f', 'synset': 'nut.n.03', 'synonyms': ['nut'], 'id': 726, 'def': 'a small metal block (usually square or hexagonal) with internal screw thread to be fitted onto a bolt', 'name': 'nut'}, {'frequency': 'r', 'synset': 'nutcracker.n.01', 'synonyms': ['nutcracker'], 'id': 727, 'def': 'a hand tool used to crack nuts open', 'name': 'nutcracker'}, {'frequency': 'f', 'synset': 'oar.n.01', 'synonyms': ['oar'], 'id': 728, 'def': 'an implement used to propel or steer a boat', 'name': 'oar'}, {'frequency': 'r', 'synset': 'octopus.n.01', 'synonyms': ['octopus_(food)'], 'id': 729, 'def': 'tentacles of octopus prepared as food', 'name': 'octopus_(food)'}, {'frequency': 'r', 'synset': 'octopus.n.02', 'synonyms': ['octopus_(animal)'], 'id': 730, 'def': 'bottom-living cephalopod having a soft oval body with eight long tentacles', 'name': 'octopus_(animal)'}, {'frequency': 'c', 'synset': 'oil_lamp.n.01', 'synonyms': ['oil_lamp', 'kerosene_lamp', 'kerosine_lamp'], 'id': 731, 'def': 'a lamp that burns oil (as kerosine) for light', 'name': 'oil_lamp'}, {'frequency': 'c', 'synset': 'olive_oil.n.01', 'synonyms': ['olive_oil'], 'id': 732, 'def': 'oil from olives', 'name': 'olive_oil'}, {'frequency': 'r', 'synset': 'omelet.n.01', 'synonyms': ['omelet', 'omelette'], 'id': 733, 'def': 'beaten eggs cooked until just set; may be folded around e.g. ham or cheese or jelly', 'name': 'omelet'}, {'frequency': 'f', 'synset': 'onion.n.01', 'synonyms': ['onion'], 'id': 734, 'def': 'the bulb of an onion plant', 'name': 'onion'}, {'frequency': 'f', 'synset': 'orange.n.01', 'synonyms': ['orange_(fruit)'], 'id': 735, 'def': 'orange (FRUIT of an orange tree)', 'name': 'orange_(fruit)'}, {'frequency': 'c', 'synset': 'orange_juice.n.01', 'synonyms': ['orange_juice'], 'id': 736, 'def': 'bottled or freshly squeezed juice of oranges', 'name': 'orange_juice'}, {'frequency': 'c', 'synset': 'ostrich.n.02', 'synonyms': ['ostrich'], 'id': 737, 'def': 'fast-running African flightless bird with two-toed feet; largest living bird', 'name': 'ostrich'}, {'frequency': 'f', 'synset': 'ottoman.n.03', 'synonyms': ['ottoman', 'pouf', 'pouffe', 'hassock'], 'id': 738, 'def': 'a thick standalone cushion used as a seat or footrest, often next to a chair', 'name': 'ottoman'}, {'frequency': 'f', 'synset': 'oven.n.01', 'synonyms': ['oven'], 'id': 739, 'def': 'kitchen appliance used for baking or roasting', 'name': 'oven'}, {'frequency': 'c', 'synset': 'overall.n.01', 'synonyms': ['overalls_(clothing)'], 'id': 740, 'def': 'work clothing consisting of denim trousers usually with a bib and shoulder straps', 'name': 'overalls_(clothing)'}, {'frequency': 'c', 'synset': 'owl.n.01', 'synonyms': ['owl'], 'id': 741, 'def': 'nocturnal bird of prey with hawk-like beak and claws and large head with front-facing eyes', 'name': 'owl'}, {'frequency': 'c', 'synset': 'packet.n.03', 'synonyms': ['packet'], 'id': 742, 'def': 'a small package or bundle', 'name': 'packet'}, {'frequency': 'r', 'synset': 'pad.n.03', 'synonyms': ['inkpad', 'inking_pad', 'stamp_pad'], 'id': 743, 'def': 'absorbent material saturated with ink used to transfer ink evenly to a rubber stamp', 'name': 'inkpad'}, {'frequency': 'c', 'synset': 'pad.n.04', 'synonyms': ['pad'], 'id': 744, 'def': 'mostly arm/knee pads labeled', 'name': 'pad'}, {'frequency': 'f', 'synset': 'paddle.n.04', 'synonyms': ['paddle', 'boat_paddle'], 'id': 745, 'def': 'a short light oar used without an oarlock to propel a canoe or small boat', 'name': 'paddle'}, {'frequency': 'c', 'synset': 'padlock.n.01', 'synonyms': ['padlock'], 'id': 746, 'def': 'a detachable, portable lock', 'name': 'padlock'}, {'frequency': 'c', 'synset': 'paintbrush.n.01', 'synonyms': ['paintbrush'], 'id': 747, 'def': 'a brush used as an applicator to apply paint', 'name': 'paintbrush'}, {'frequency': 'f', 'synset': 'painting.n.01', 'synonyms': ['painting'], 'id': 748, 'def': 'graphic art consisting of an artistic composition made by applying paints to a surface', 'name': 'painting'}, {'frequency': 'f', 'synset': 'pajama.n.02', 'synonyms': ['pajamas', 'pyjamas'], 'id': 749, 'def': 'loose-fitting nightclothes worn for sleeping or lounging', 'name': 'pajamas'}, {'frequency': 'c', 'synset': 'palette.n.02', 'synonyms': ['palette', 'pallet'], 'id': 750, 'def': 'board that provides a flat surface on which artists mix paints and the range of colors used', 'name': 'palette'}, {'frequency': 'f', 'synset': 'pan.n.01', 'synonyms': ['pan_(for_cooking)', 'cooking_pan'], 'id': 751, 'def': 'cooking utensil consisting of a wide metal vessel', 'name': 'pan_(for_cooking)'}, {'frequency': 'r', 'synset': 'pan.n.03', 'synonyms': ['pan_(metal_container)'], 'id': 752, 'def': 'shallow container made of metal', 'name': 'pan_(metal_container)'}, {'frequency': 'c', 'synset': 'pancake.n.01', 'synonyms': ['pancake'], 'id': 753, 'def': 'a flat cake of thin batter fried on both sides on a griddle', 'name': 'pancake'}, {'frequency': 'r', 'synset': 'pantyhose.n.01', 'synonyms': ['pantyhose'], 'id': 754, 'def': "a woman's tights consisting of underpants and stockings", 'name': 'pantyhose'}, {'frequency': 'r', 'synset': 'papaya.n.02', 'synonyms': ['papaya'], 'id': 755, 'def': 'large oval melon-like tropical fruit with yellowish flesh', 'name': 'papaya'}, {'frequency': 'f', 'synset': 'paper_plate.n.01', 'synonyms': ['paper_plate'], 'id': 756, 'def': 'a disposable plate made of cardboard', 'name': 'paper_plate'}, {'frequency': 'f', 'synset': 'paper_towel.n.01', 'synonyms': ['paper_towel'], 'id': 757, 'def': 'a disposable towel made of absorbent paper', 'name': 'paper_towel'}, {'frequency': 'r', 'synset': 'paperback_book.n.01', 'synonyms': ['paperback_book', 'paper-back_book', 'softback_book', 'soft-cover_book'], 'id': 758, 'def': 'a book with paper covers', 'name': 'paperback_book'}, {'frequency': 'r', 'synset': 'paperweight.n.01', 'synonyms': ['paperweight'], 'id': 759, 'def': 'a weight used to hold down a stack of papers', 'name': 'paperweight'}, {'frequency': 'c', 'synset': 'parachute.n.01', 'synonyms': ['parachute'], 'id': 760, 'def': 'rescue equipment consisting of a device that fills with air and retards your fall', 'name': 'parachute'}, {'frequency': 'c', 'synset': 'parakeet.n.01', 'synonyms': ['parakeet', 'parrakeet', 'parroket', 'paraquet', 'paroquet', 'parroquet'], 'id': 761, 'def': 'any of numerous small slender long-tailed parrots', 'name': 'parakeet'}, {'frequency': 'c', 'synset': 'parasail.n.01', 'synonyms': ['parasail_(sports)'], 'id': 762, 'def': 'parachute that will lift a person up into the air when it is towed by a motorboat or a car', 'name': 'parasail_(sports)'}, {'frequency': 'c', 'synset': 'parasol.n.01', 'synonyms': ['parasol', 'sunshade'], 'id': 763, 'def': 'a handheld collapsible source of shade', 'name': 'parasol'}, {'frequency': 'r', 'synset': 'parchment.n.01', 'synonyms': ['parchment'], 'id': 764, 'def': 'a superior paper resembling sheepskin', 'name': 'parchment'}, {'frequency': 'c', 'synset': 'parka.n.01', 'synonyms': ['parka', 'anorak'], 'id': 765, 'def': "a kind of heavy jacket (`windcheater' is a British term)", 'name': 'parka'}, {'frequency': 'f', 'synset': 'parking_meter.n.01', 'synonyms': ['parking_meter'], 'id': 766, 'def': 'a coin-operated timer located next to a parking space', 'name': 'parking_meter'}, {'frequency': 'c', 'synset': 'parrot.n.01', 'synonyms': ['parrot'], 'id': 767, 'def': 'usually brightly colored tropical birds with short hooked beaks and the ability to mimic sounds', 'name': 'parrot'}, {'frequency': 'c', 'synset': 'passenger_car.n.01', 'synonyms': ['passenger_car_(part_of_a_train)', 'coach_(part_of_a_train)'], 'id': 768, 'def': 'a railcar where passengers ride', 'name': 'passenger_car_(part_of_a_train)'}, {'frequency': 'r', 'synset': 'passenger_ship.n.01', 'synonyms': ['passenger_ship'], 'id': 769, 'def': 'a ship built to carry passengers', 'name': 'passenger_ship'}, {'frequency': 'c', 'synset': 'passport.n.02', 'synonyms': ['passport'], 'id': 770, 'def': 'a document issued by a country to a citizen allowing that person to travel abroad and re-enter the home country', 'name': 'passport'}, {'frequency': 'f', 'synset': 'pastry.n.02', 'synonyms': ['pastry'], 'id': 771, 'def': 'any of various baked foods made of dough or batter', 'name': 'pastry'}, {'frequency': 'r', 'synset': 'patty.n.01', 'synonyms': ['patty_(food)'], 'id': 772, 'def': 'small flat mass of chopped food', 'name': 'patty_(food)'}, {'frequency': 'c', 'synset': 'pea.n.01', 'synonyms': ['pea_(food)'], 'id': 773, 'def': 'seed of a pea plant used for food', 'name': 'pea_(food)'}, {'frequency': 'c', 'synset': 'peach.n.03', 'synonyms': ['peach'], 'id': 774, 'def': 'downy juicy fruit with sweet yellowish or whitish flesh', 'name': 'peach'}, {'frequency': 'c', 'synset': 'peanut_butter.n.01', 'synonyms': ['peanut_butter'], 'id': 775, 'def': 'a spread made from ground peanuts', 'name': 'peanut_butter'}, {'frequency': 'f', 'synset': 'pear.n.01', 'synonyms': ['pear'], 'id': 776, 'def': 'sweet juicy gritty-textured fruit available in many varieties', 'name': 'pear'}, {'frequency': 'c', 'synset': 'peeler.n.03', 'synonyms': ['peeler_(tool_for_fruit_and_vegetables)'], 'id': 777, 'def': 'a device for peeling vegetables or fruits', 'name': 'peeler_(tool_for_fruit_and_vegetables)'}, {'frequency': 'r', 'synset': 'peg.n.04', 'synonyms': ['wooden_leg', 'pegleg'], 'id': 778, 'def': 'a prosthesis that replaces a missing leg', 'name': 'wooden_leg'}, {'frequency': 'r', 'synset': 'pegboard.n.01', 'synonyms': ['pegboard'], 'id': 779, 'def': 'a board perforated with regularly spaced holes into which pegs can be fitted', 'name': 'pegboard'}, {'frequency': 'c', 'synset': 'pelican.n.01', 'synonyms': ['pelican'], 'id': 780, 'def': 'large long-winged warm-water seabird having a large bill with a distensible pouch for fish', 'name': 'pelican'}, {'frequency': 'f', 'synset': 'pen.n.01', 'synonyms': ['pen'], 'id': 781, 'def': 'a writing implement with a point from which ink flows', 'name': 'pen'}, {'frequency': 'f', 'synset': 'pencil.n.01', 'synonyms': ['pencil'], 'id': 782, 'def': 'a thin cylindrical pointed writing implement made of wood and graphite', 'name': 'pencil'}, {'frequency': 'r', 'synset': 'pencil_box.n.01', 'synonyms': ['pencil_box', 'pencil_case'], 'id': 783, 'def': 'a box for holding pencils', 'name': 'pencil_box'}, {'frequency': 'r', 'synset': 'pencil_sharpener.n.01', 'synonyms': ['pencil_sharpener'], 'id': 784, 'def': 'a rotary implement for sharpening the point on pencils', 'name': 'pencil_sharpener'}, {'frequency': 'r', 'synset': 'pendulum.n.01', 'synonyms': ['pendulum'], 'id': 785, 'def': 'an apparatus consisting of an object mounted so that it swings freely under the influence of gravity', 'name': 'pendulum'}, {'frequency': 'c', 'synset': 'penguin.n.01', 'synonyms': ['penguin'], 'id': 786, 'def': 'short-legged flightless birds of cold southern regions having webbed feet and wings modified as flippers', 'name': 'penguin'}, {'frequency': 'r', 'synset': 'pennant.n.02', 'synonyms': ['pennant'], 'id': 787, 'def': 'a flag longer than it is wide (and often tapering)', 'name': 'pennant'}, {'frequency': 'r', 'synset': 'penny.n.02', 'synonyms': ['penny_(coin)'], 'id': 788, 'def': 'a coin worth one-hundredth of the value of the basic unit', 'name': 'penny_(coin)'}, {'frequency': 'f', 'synset': 'pepper.n.03', 'synonyms': ['pepper', 'peppercorn'], 'id': 789, 'def': 'pungent seasoning from the berry of the common pepper plant; whole or ground', 'name': 'pepper'}, {'frequency': 'c', 'synset': 'pepper_mill.n.01', 'synonyms': ['pepper_mill', 'pepper_grinder'], 'id': 790, 'def': 'a mill for grinding pepper', 'name': 'pepper_mill'}, {'frequency': 'c', 'synset': 'perfume.n.02', 'synonyms': ['perfume'], 'id': 791, 'def': 'a toiletry that emits and diffuses a fragrant odor', 'name': 'perfume'}, {'frequency': 'r', 'synset': 'persimmon.n.02', 'synonyms': ['persimmon'], 'id': 792, 'def': 'orange fruit resembling a plum; edible when fully ripe', 'name': 'persimmon'}, {'frequency': 'f', 'synset': 'person.n.01', 'synonyms': ['person', 'baby', 'child', 'boy', 'girl', 'man', 'woman', 'human'], 'id': 793, 'def': 'a human being', 'name': 'person'}, {'frequency': 'c', 'synset': 'pet.n.01', 'synonyms': ['pet'], 'id': 794, 'def': 'a domesticated animal kept for companionship or amusement', 'name': 'pet'}, {'frequency': 'c', 'synset': 'pew.n.01', 'synonyms': ['pew_(church_bench)', 'church_bench'], 'id': 795, 'def': 'long bench with backs; used in church by the congregation', 'name': 'pew_(church_bench)'}, {'frequency': 'r', 'synset': 'phonebook.n.01', 'synonyms': ['phonebook', 'telephone_book', 'telephone_directory'], 'id': 796, 'def': 'a directory containing an alphabetical list of telephone subscribers and their telephone numbers', 'name': 'phonebook'}, {'frequency': 'c', 'synset': 'phonograph_record.n.01', 'synonyms': ['phonograph_record', 'phonograph_recording', 'record_(phonograph_recording)'], 'id': 797, 'def': 'sound recording consisting of a typically black disk with a continuous groove', 'name': 'phonograph_record'}, {'frequency': 'f', 'synset': 'piano.n.01', 'synonyms': ['piano'], 'id': 798, 'def': 'a keyboard instrument that is played by depressing keys that cause hammers to strike tuned strings and produce sounds', 'name': 'piano'}, {'frequency': 'f', 'synset': 'pickle.n.01', 'synonyms': ['pickle'], 'id': 799, 'def': 'vegetables (especially cucumbers) preserved in brine or vinegar', 'name': 'pickle'}, {'frequency': 'f', 'synset': 'pickup.n.01', 'synonyms': ['pickup_truck'], 'id': 800, 'def': 'a light truck with an open body and low sides and a tailboard', 'name': 'pickup_truck'}, {'frequency': 'c', 'synset': 'pie.n.01', 'synonyms': ['pie'], 'id': 801, 'def': 'dish baked in pastry-lined pan often with a pastry top', 'name': 'pie'}, {'frequency': 'c', 'synset': 'pigeon.n.01', 'synonyms': ['pigeon'], 'id': 802, 'def': 'wild and domesticated birds having a heavy body and short legs', 'name': 'pigeon'}, {'frequency': 'r', 'synset': 'piggy_bank.n.01', 'synonyms': ['piggy_bank', 'penny_bank'], 'id': 803, 'def': "a child's coin bank (often shaped like a pig)", 'name': 'piggy_bank'}, {'frequency': 'f', 'synset': 'pillow.n.01', 'synonyms': ['pillow'], 'id': 804, 'def': 'a cushion to support the head of a sleeping person', 'name': 'pillow'}, {'frequency': 'r', 'synset': 'pin.n.09', 'synonyms': ['pin_(non_jewelry)'], 'id': 805, 'def': 'a small slender (often pointed) piece of wood or metal used to support or fasten or attach things', 'name': 'pin_(non_jewelry)'}, {'frequency': 'f', 'synset': 'pineapple.n.02', 'synonyms': ['pineapple'], 'id': 806, 'def': 'large sweet fleshy tropical fruit with a tuft of stiff leaves', 'name': 'pineapple'}, {'frequency': 'c', 'synset': 'pinecone.n.01', 'synonyms': ['pinecone'], 'id': 807, 'def': 'the seed-producing cone of a pine tree', 'name': 'pinecone'}, {'frequency': 'r', 'synset': 'ping-pong_ball.n.01', 'synonyms': ['ping-pong_ball'], 'id': 808, 'def': 'light hollow ball used in playing table tennis', 'name': 'ping-pong_ball'}, {'frequency': 'r', 'synset': 'pinwheel.n.03', 'synonyms': ['pinwheel'], 'id': 809, 'def': 'a toy consisting of vanes of colored paper or plastic that is pinned to a stick and spins when it is pointed into the wind', 'name': 'pinwheel'}, {'frequency': 'r', 'synset': 'pipe.n.01', 'synonyms': ['tobacco_pipe'], 'id': 810, 'def': 'a tube with a small bowl at one end; used for smoking tobacco', 'name': 'tobacco_pipe'}, {'frequency': 'f', 'synset': 'pipe.n.02', 'synonyms': ['pipe', 'piping'], 'id': 811, 'def': 'a long tube made of metal or plastic that is used to carry water or oil or gas etc.', 'name': 'pipe'}, {'frequency': 'r', 'synset': 'pistol.n.01', 'synonyms': ['pistol', 'handgun'], 'id': 812, 'def': 'a firearm that is held and fired with one hand', 'name': 'pistol'}, {'frequency': 'c', 'synset': 'pita.n.01', 'synonyms': ['pita_(bread)', 'pocket_bread'], 'id': 813, 'def': 'usually small round bread that can open into a pocket for filling', 'name': 'pita_(bread)'}, {'frequency': 'f', 'synset': 'pitcher.n.02', 'synonyms': ['pitcher_(vessel_for_liquid)', 'ewer'], 'id': 814, 'def': 'an open vessel with a handle and a spout for pouring', 'name': 'pitcher_(vessel_for_liquid)'}, {'frequency': 'r', 'synset': 'pitchfork.n.01', 'synonyms': ['pitchfork'], 'id': 815, 'def': 'a long-handled hand tool with sharp widely spaced prongs for lifting and pitching hay', 'name': 'pitchfork'}, {'frequency': 'f', 'synset': 'pizza.n.01', 'synonyms': ['pizza'], 'id': 816, 'def': 'Italian open pie made of thin bread dough spread with a spiced mixture of e.g. tomato sauce and cheese', 'name': 'pizza'}, {'frequency': 'f', 'synset': 'place_mat.n.01', 'synonyms': ['place_mat'], 'id': 817, 'def': 'a mat placed on a table for an individual place setting', 'name': 'place_mat'}, {'frequency': 'f', 'synset': 'plate.n.04', 'synonyms': ['plate'], 'id': 818, 'def': 'dish on which food is served or from which food is eaten', 'name': 'plate'}, {'frequency': 'c', 'synset': 'platter.n.01', 'synonyms': ['platter'], 'id': 819, 'def': 'a large shallow dish used for serving food', 'name': 'platter'}, {'frequency': 'r', 'synset': 'playpen.n.01', 'synonyms': ['playpen'], 'id': 820, 'def': 'a portable enclosure in which babies may be left to play', 'name': 'playpen'}, {'frequency': 'c', 'synset': 'pliers.n.01', 'synonyms': ['pliers', 'plyers'], 'id': 821, 'def': 'a gripping hand tool with two hinged arms and (usually) serrated jaws', 'name': 'pliers'}, {'frequency': 'r', 'synset': 'plow.n.01', 'synonyms': ['plow_(farm_equipment)', 'plough_(farm_equipment)'], 'id': 822, 'def': 'a farm tool having one or more heavy blades to break the soil and cut a furrow prior to sowing', 'name': 'plow_(farm_equipment)'}, {'frequency': 'r', 'synset': 'plume.n.02', 'synonyms': ['plume'], 'id': 823, 'def': 'a feather or cluster of feathers worn as an ornament', 'name': 'plume'}, {'frequency': 'r', 'synset': 'pocket_watch.n.01', 'synonyms': ['pocket_watch'], 'id': 824, 'def': 'a watch that is carried in a small watch pocket', 'name': 'pocket_watch'}, {'frequency': 'c', 'synset': 'pocketknife.n.01', 'synonyms': ['pocketknife'], 'id': 825, 'def': 'a knife with a blade that folds into the handle; suitable for carrying in the pocket', 'name': 'pocketknife'}, {'frequency': 'c', 'synset': 'poker.n.01', 'synonyms': ['poker_(fire_stirring_tool)', 'stove_poker', 'fire_hook'], 'id': 826, 'def': 'fire iron consisting of a metal rod with a handle; used to stir a fire', 'name': 'poker_(fire_stirring_tool)'}, {'frequency': 'f', 'synset': 'pole.n.01', 'synonyms': ['pole', 'post'], 'id': 827, 'def': 'a long (usually round) rod of wood or metal or plastic', 'name': 'pole'}, {'frequency': 'f', 'synset': 'polo_shirt.n.01', 'synonyms': ['polo_shirt', 'sport_shirt'], 'id': 828, 'def': 'a shirt with short sleeves designed for comfort and casual wear', 'name': 'polo_shirt'}, {'frequency': 'r', 'synset': 'poncho.n.01', 'synonyms': ['poncho'], 'id': 829, 'def': 'a blanket-like cloak with a hole in the center for the head', 'name': 'poncho'}, {'frequency': 'c', 'synset': 'pony.n.05', 'synonyms': ['pony'], 'id': 830, 'def': 'any of various breeds of small gentle horses usually less than five feet high at the shoulder', 'name': 'pony'}, {'frequency': 'r', 'synset': 'pool_table.n.01', 'synonyms': ['pool_table', 'billiard_table', 'snooker_table'], 'id': 831, 'def': 'game equipment consisting of a heavy table on which pool is played', 'name': 'pool_table'}, {'frequency': 'f', 'synset': 'pop.n.02', 'synonyms': ['pop_(soda)', 'soda_(pop)', 'tonic', 'soft_drink'], 'id': 832, 'def': 'a sweet drink containing carbonated water and flavoring', 'name': 'pop_(soda)'}, {'frequency': 'c', 'synset': 'postbox.n.01', 'synonyms': ['postbox_(public)', 'mailbox_(public)'], 'id': 833, 'def': 'public box for deposit of mail', 'name': 'postbox_(public)'}, {'frequency': 'c', 'synset': 'postcard.n.01', 'synonyms': ['postcard', 'postal_card', 'mailing-card'], 'id': 834, 'def': 'a card for sending messages by post without an envelope', 'name': 'postcard'}, {'frequency': 'f', 'synset': 'poster.n.01', 'synonyms': ['poster', 'placard'], 'id': 835, 'def': 'a sign posted in a public place as an advertisement', 'name': 'poster'}, {'frequency': 'f', 'synset': 'pot.n.01', 'synonyms': ['pot'], 'id': 836, 'def': 'metal or earthenware cooking vessel that is usually round and deep; often has a handle and lid', 'name': 'pot'}, {'frequency': 'f', 'synset': 'pot.n.04', 'synonyms': ['flowerpot'], 'id': 837, 'def': 'a container in which plants are cultivated', 'name': 'flowerpot'}, {'frequency': 'f', 'synset': 'potato.n.01', 'synonyms': ['potato'], 'id': 838, 'def': 'an edible tuber native to South America', 'name': 'potato'}, {'frequency': 'c', 'synset': 'potholder.n.01', 'synonyms': ['potholder'], 'id': 839, 'def': 'an insulated pad for holding hot pots', 'name': 'potholder'}, {'frequency': 'c', 'synset': 'pottery.n.01', 'synonyms': ['pottery', 'clayware'], 'id': 840, 'def': 'ceramic ware made from clay and baked in a kiln', 'name': 'pottery'}, {'frequency': 'c', 'synset': 'pouch.n.01', 'synonyms': ['pouch'], 'id': 841, 'def': 'a small or medium size container for holding or carrying things', 'name': 'pouch'}, {'frequency': 'c', 'synset': 'power_shovel.n.01', 'synonyms': ['power_shovel', 'excavator', 'digger'], 'id': 842, 'def': 'a machine for excavating', 'name': 'power_shovel'}, {'frequency': 'c', 'synset': 'prawn.n.01', 'synonyms': ['prawn', 'shrimp'], 'id': 843, 'def': 'any of various edible decapod crustaceans', 'name': 'prawn'}, {'frequency': 'c', 'synset': 'pretzel.n.01', 'synonyms': ['pretzel'], 'id': 844, 'def': 'glazed and salted cracker typically in the shape of a loose knot', 'name': 'pretzel'}, {'frequency': 'f', 'synset': 'printer.n.03', 'synonyms': ['printer', 'printing_machine'], 'id': 845, 'def': 'a machine that prints', 'name': 'printer'}, {'frequency': 'c', 'synset': 'projectile.n.01', 'synonyms': ['projectile_(weapon)', 'missile'], 'id': 846, 'def': 'a weapon that is forcibly thrown or projected at a targets', 'name': 'projectile_(weapon)'}, {'frequency': 'c', 'synset': 'projector.n.02', 'synonyms': ['projector'], 'id': 847, 'def': 'an optical instrument that projects an enlarged image onto a screen', 'name': 'projector'}, {'frequency': 'f', 'synset': 'propeller.n.01', 'synonyms': ['propeller', 'propellor'], 'id': 848, 'def': 'a mechanical device that rotates to push against air or water', 'name': 'propeller'}, {'frequency': 'r', 'synset': 'prune.n.01', 'synonyms': ['prune'], 'id': 849, 'def': 'dried plum', 'name': 'prune'}, {'frequency': 'r', 'synset': 'pudding.n.01', 'synonyms': ['pudding'], 'id': 850, 'def': 'any of various soft thick unsweetened baked dishes', 'name': 'pudding'}, {'frequency': 'r', 'synset': 'puffer.n.02', 'synonyms': ['puffer_(fish)', 'pufferfish', 'blowfish', 'globefish'], 'id': 851, 'def': 'fishes whose elongated spiny body can inflate itself with water or air to form a globe', 'name': 'puffer_(fish)'}, {'frequency': 'r', 'synset': 'puffin.n.01', 'synonyms': ['puffin'], 'id': 852, 'def': 'seabirds having short necks and brightly colored compressed bills', 'name': 'puffin'}, {'frequency': 'r', 'synset': 'pug.n.01', 'synonyms': ['pug-dog'], 'id': 853, 'def': 'small compact smooth-coated breed of Asiatic origin having a tightly curled tail and broad flat wrinkled muzzle', 'name': 'pug-dog'}, {'frequency': 'c', 'synset': 'pumpkin.n.02', 'synonyms': ['pumpkin'], 'id': 854, 'def': 'usually large pulpy deep-yellow round fruit of the squash family maturing in late summer or early autumn', 'name': 'pumpkin'}, {'frequency': 'r', 'synset': 'punch.n.03', 'synonyms': ['puncher'], 'id': 855, 'def': 'a tool for making holes or indentations', 'name': 'puncher'}, {'frequency': 'r', 'synset': 'puppet.n.01', 'synonyms': ['puppet', 'marionette'], 'id': 856, 'def': 'a small figure of a person operated from above with strings by a puppeteer', 'name': 'puppet'}, {'frequency': 'c', 'synset': 'puppy.n.01', 'synonyms': ['puppy'], 'id': 857, 'def': 'a young dog', 'name': 'puppy'}, {'frequency': 'r', 'synset': 'quesadilla.n.01', 'synonyms': ['quesadilla'], 'id': 858, 'def': 'a tortilla that is filled with cheese and heated', 'name': 'quesadilla'}, {'frequency': 'r', 'synset': 'quiche.n.02', 'synonyms': ['quiche'], 'id': 859, 'def': 'a tart filled with rich unsweetened custard; often contains other ingredients (as cheese or ham or seafood or vegetables)', 'name': 'quiche'}, {'frequency': 'f', 'synset': 'quilt.n.01', 'synonyms': ['quilt', 'comforter'], 'id': 860, 'def': 'bedding made of two layers of cloth filled with stuffing and stitched together', 'name': 'quilt'}, {'frequency': 'c', 'synset': 'rabbit.n.01', 'synonyms': ['rabbit'], 'id': 861, 'def': 'any of various burrowing animals of the family Leporidae having long ears and short tails', 'name': 'rabbit'}, {'frequency': 'r', 'synset': 'racer.n.02', 'synonyms': ['race_car', 'racing_car'], 'id': 862, 'def': 'a fast car that competes in races', 'name': 'race_car'}, {'frequency': 'c', 'synset': 'racket.n.04', 'synonyms': ['racket', 'racquet'], 'id': 863, 'def': 'a sports implement used to strike a ball in various games', 'name': 'racket'}, {'frequency': 'r', 'synset': 'radar.n.01', 'synonyms': ['radar'], 'id': 864, 'def': 'measuring instrument in which the echo of a pulse of microwave radiation is used to detect and locate distant objects', 'name': 'radar'}, {'frequency': 'f', 'synset': 'radiator.n.03', 'synonyms': ['radiator'], 'id': 865, 'def': 'a mechanism consisting of a metal honeycomb through which hot fluids circulate', 'name': 'radiator'}, {'frequency': 'c', 'synset': 'radio_receiver.n.01', 'synonyms': ['radio_receiver', 'radio_set', 'radio', 'tuner_(radio)'], 'id': 866, 'def': 'an electronic receiver that detects and demodulates and amplifies transmitted radio signals', 'name': 'radio_receiver'}, {'frequency': 'c', 'synset': 'radish.n.03', 'synonyms': ['radish', 'daikon'], 'id': 867, 'def': 'pungent edible root of any of various cultivated radish plants', 'name': 'radish'}, {'frequency': 'c', 'synset': 'raft.n.01', 'synonyms': ['raft'], 'id': 868, 'def': 'a flat float (usually made of logs or planks) that can be used for transport or as a platform for swimmers', 'name': 'raft'}, {'frequency': 'r', 'synset': 'rag_doll.n.01', 'synonyms': ['rag_doll'], 'id': 869, 'def': 'a cloth doll that is stuffed and (usually) painted', 'name': 'rag_doll'}, {'frequency': 'c', 'synset': 'raincoat.n.01', 'synonyms': ['raincoat', 'waterproof_jacket'], 'id': 870, 'def': 'a water-resistant coat', 'name': 'raincoat'}, {'frequency': 'c', 'synset': 'ram.n.05', 'synonyms': ['ram_(animal)'], 'id': 871, 'def': 'uncastrated adult male sheep', 'name': 'ram_(animal)'}, {'frequency': 'c', 'synset': 'raspberry.n.02', 'synonyms': ['raspberry'], 'id': 872, 'def': 'red or black edible aggregate berries usually smaller than the related blackberries', 'name': 'raspberry'}, {'frequency': 'r', 'synset': 'rat.n.01', 'synonyms': ['rat'], 'id': 873, 'def': 'any of various long-tailed rodents similar to but larger than a mouse', 'name': 'rat'}, {'frequency': 'c', 'synset': 'razorblade.n.01', 'synonyms': ['razorblade'], 'id': 874, 'def': 'a blade that has very sharp edge', 'name': 'razorblade'}, {'frequency': 'c', 'synset': 'reamer.n.01', 'synonyms': ['reamer_(juicer)', 'juicer', 'juice_reamer'], 'id': 875, 'def': 'a squeezer with a conical ridged center that is used for squeezing juice from citrus fruit', 'name': 'reamer_(juicer)'}, {'frequency': 'f', 'synset': 'rearview_mirror.n.01', 'synonyms': ['rearview_mirror'], 'id': 876, 'def': 'vehicle mirror (side or rearview)', 'name': 'rearview_mirror'}, {'frequency': 'c', 'synset': 'receipt.n.02', 'synonyms': ['receipt'], 'id': 877, 'def': 'an acknowledgment (usually tangible) that payment has been made', 'name': 'receipt'}, {'frequency': 'c', 'synset': 'recliner.n.01', 'synonyms': ['recliner', 'reclining_chair', 'lounger_(chair)'], 'id': 878, 'def': 'an armchair whose back can be lowered and foot can be raised to allow the sitter to recline in it', 'name': 'recliner'}, {'frequency': 'c', 'synset': 'record_player.n.01', 'synonyms': ['record_player', 'phonograph_(record_player)', 'turntable'], 'id': 879, 'def': 'machine in which rotating records cause a stylus to vibrate and the vibrations are amplified acoustically or electronically', 'name': 'record_player'}, {'frequency': 'f', 'synset': 'reflector.n.01', 'synonyms': ['reflector'], 'id': 880, 'def': 'device that reflects light, radiation, etc.', 'name': 'reflector'}, {'frequency': 'f', 'synset': 'remote_control.n.01', 'synonyms': ['remote_control'], 'id': 881, 'def': 'a device that can be used to control a machine or apparatus from a distance', 'name': 'remote_control'}, {'frequency': 'c', 'synset': 'rhinoceros.n.01', 'synonyms': ['rhinoceros'], 'id': 882, 'def': 'massive powerful herbivorous odd-toed ungulate of southeast Asia and Africa having very thick skin and one or two horns on the snout', 'name': 'rhinoceros'}, {'frequency': 'r', 'synset': 'rib.n.03', 'synonyms': ['rib_(food)'], 'id': 883, 'def': 'cut of meat including one or more ribs', 'name': 'rib_(food)'}, {'frequency': 'c', 'synset': 'rifle.n.01', 'synonyms': ['rifle'], 'id': 884, 'def': 'a shoulder firearm with a long barrel', 'name': 'rifle'}, {'frequency': 'f', 'synset': 'ring.n.08', 'synonyms': ['ring'], 'id': 885, 'def': 'jewelry consisting of a circlet of precious metal (often set with jewels) worn on the finger', 'name': 'ring'}, {'frequency': 'r', 'synset': 'river_boat.n.01', 'synonyms': ['river_boat'], 'id': 886, 'def': 'a boat used on rivers or to ply a river', 'name': 'river_boat'}, {'frequency': 'r', 'synset': 'road_map.n.02', 'synonyms': ['road_map'], 'id': 887, 'def': '(NOT A ROAD) a MAP showing roads (for automobile travel)', 'name': 'road_map'}, {'frequency': 'c', 'synset': 'robe.n.01', 'synonyms': ['robe'], 'id': 888, 'def': 'any loose flowing garment', 'name': 'robe'}, {'frequency': 'c', 'synset': 'rocking_chair.n.01', 'synonyms': ['rocking_chair'], 'id': 889, 'def': 'a chair mounted on rockers', 'name': 'rocking_chair'}, {'frequency': 'r', 'synset': 'rodent.n.01', 'synonyms': ['rodent'], 'id': 890, 'def': 'relatively small placental mammals having a single pair of constantly growing incisor teeth specialized for gnawing', 'name': 'rodent'}, {'frequency': 'r', 'synset': 'roller_skate.n.01', 'synonyms': ['roller_skate'], 'id': 891, 'def': 'a shoe with pairs of rollers (small hard wheels) fixed to the sole', 'name': 'roller_skate'}, {'frequency': 'r', 'synset': 'rollerblade.n.01', 'synonyms': ['Rollerblade'], 'id': 892, 'def': 'an in-line variant of a roller skate', 'name': 'Rollerblade'}, {'frequency': 'c', 'synset': 'rolling_pin.n.01', 'synonyms': ['rolling_pin'], 'id': 893, 'def': 'utensil consisting of a cylinder (usually of wood) with a handle at each end; used to roll out dough', 'name': 'rolling_pin'}, {'frequency': 'r', 'synset': 'root_beer.n.01', 'synonyms': ['root_beer'], 'id': 894, 'def': 'carbonated drink containing extracts of roots and herbs', 'name': 'root_beer'}, {'frequency': 'c', 'synset': 'router.n.02', 'synonyms': ['router_(computer_equipment)'], 'id': 895, 'def': 'a device that forwards data packets between computer networks', 'name': 'router_(computer_equipment)'}, {'frequency': 'f', 'synset': 'rubber_band.n.01', 'synonyms': ['rubber_band', 'elastic_band'], 'id': 896, 'def': 'a narrow band of elastic rubber used to hold things (such as papers) together', 'name': 'rubber_band'}, {'frequency': 'c', 'synset': 'runner.n.08', 'synonyms': ['runner_(carpet)'], 'id': 897, 'def': 'a long narrow carpet', 'name': 'runner_(carpet)'}, {'frequency': 'f', 'synset': 'sack.n.01', 'synonyms': ['plastic_bag', 'paper_bag'], 'id': 898, 'def': "a bag made of paper or plastic for holding customer's purchases", 'name': 'plastic_bag'}, {'frequency': 'f', 'synset': 'saddle.n.01', 'synonyms': ['saddle_(on_an_animal)'], 'id': 899, 'def': 'a seat for the rider of a horse or camel', 'name': 'saddle_(on_an_animal)'}, {'frequency': 'f', 'synset': 'saddle_blanket.n.01', 'synonyms': ['saddle_blanket', 'saddlecloth', 'horse_blanket'], 'id': 900, 'def': 'stable gear consisting of a blanket placed under the saddle', 'name': 'saddle_blanket'}, {'frequency': 'c', 'synset': 'saddlebag.n.01', 'synonyms': ['saddlebag'], 'id': 901, 'def': 'a large bag (or pair of bags) hung over a saddle', 'name': 'saddlebag'}, {'frequency': 'r', 'synset': 'safety_pin.n.01', 'synonyms': ['safety_pin'], 'id': 902, 'def': 'a pin in the form of a clasp; has a guard so the point of the pin will not stick the user', 'name': 'safety_pin'}, {'frequency': 'f', 'synset': 'sail.n.01', 'synonyms': ['sail'], 'id': 903, 'def': 'a large piece of fabric by means of which wind is used to propel a sailing vessel', 'name': 'sail'}, {'frequency': 'f', 'synset': 'salad.n.01', 'synonyms': ['salad'], 'id': 904, 'def': 'food mixtures either arranged on a plate or tossed and served with a moist dressing; usually consisting of or including greens', 'name': 'salad'}, {'frequency': 'r', 'synset': 'salad_plate.n.01', 'synonyms': ['salad_plate', 'salad_bowl'], 'id': 905, 'def': 'a plate or bowl for individual servings of salad', 'name': 'salad_plate'}, {'frequency': 'c', 'synset': 'salami.n.01', 'synonyms': ['salami'], 'id': 906, 'def': 'highly seasoned fatty sausage of pork and beef usually dried', 'name': 'salami'}, {'frequency': 'c', 'synset': 'salmon.n.01', 'synonyms': ['salmon_(fish)'], 'id': 907, 'def': 'any of various large food and game fishes of northern waters', 'name': 'salmon_(fish)'}, {'frequency': 'r', 'synset': 'salmon.n.03', 'synonyms': ['salmon_(food)'], 'id': 908, 'def': 'flesh of any of various marine or freshwater fish of the family Salmonidae', 'name': 'salmon_(food)'}, {'frequency': 'c', 'synset': 'salsa.n.01', 'synonyms': ['salsa'], 'id': 909, 'def': 'spicy sauce of tomatoes and onions and chili peppers to accompany Mexican foods', 'name': 'salsa'}, {'frequency': 'f', 'synset': 'saltshaker.n.01', 'synonyms': ['saltshaker'], 'id': 910, 'def': 'a shaker with a perforated top for sprinkling salt', 'name': 'saltshaker'}, {'frequency': 'f', 'synset': 'sandal.n.01', 'synonyms': ['sandal_(type_of_shoe)'], 'id': 911, 'def': 'a shoe consisting of a sole fastened by straps to the foot', 'name': 'sandal_(type_of_shoe)'}, {'frequency': 'f', 'synset': 'sandwich.n.01', 'synonyms': ['sandwich'], 'id': 912, 'def': 'two (or more) slices of bread with a filling between them', 'name': 'sandwich'}, {'frequency': 'r', 'synset': 'satchel.n.01', 'synonyms': ['satchel'], 'id': 913, 'def': 'luggage consisting of a small case with a flat bottom and (usually) a shoulder strap', 'name': 'satchel'}, {'frequency': 'r', 'synset': 'saucepan.n.01', 'synonyms': ['saucepan'], 'id': 914, 'def': 'a deep pan with a handle; used for stewing or boiling', 'name': 'saucepan'}, {'frequency': 'f', 'synset': 'saucer.n.02', 'synonyms': ['saucer'], 'id': 915, 'def': 'a small shallow dish for holding a cup at the table', 'name': 'saucer'}, {'frequency': 'f', 'synset': 'sausage.n.01', 'synonyms': ['sausage'], 'id': 916, 'def': 'highly seasoned minced meat stuffed in casings', 'name': 'sausage'}, {'frequency': 'r', 'synset': 'sawhorse.n.01', 'synonyms': ['sawhorse', 'sawbuck'], 'id': 917, 'def': 'a framework for holding wood that is being sawed', 'name': 'sawhorse'}, {'frequency': 'r', 'synset': 'sax.n.02', 'synonyms': ['saxophone'], 'id': 918, 'def': "a wind instrument with a `J'-shaped form typically made of brass", 'name': 'saxophone'}, {'frequency': 'f', 'synset': 'scale.n.07', 'synonyms': ['scale_(measuring_instrument)'], 'id': 919, 'def': 'a measuring instrument for weighing; shows amount of mass', 'name': 'scale_(measuring_instrument)'}, {'frequency': 'r', 'synset': 'scarecrow.n.01', 'synonyms': ['scarecrow', 'strawman'], 'id': 920, 'def': 'an effigy in the shape of a man to frighten birds away from seeds', 'name': 'scarecrow'}, {'frequency': 'f', 'synset': 'scarf.n.01', 'synonyms': ['scarf'], 'id': 921, 'def': 'a garment worn around the head or neck or shoulders for warmth or decoration', 'name': 'scarf'}, {'frequency': 'c', 'synset': 'school_bus.n.01', 'synonyms': ['school_bus'], 'id': 922, 'def': 'a bus used to transport children to or from school', 'name': 'school_bus'}, {'frequency': 'f', 'synset': 'scissors.n.01', 'synonyms': ['scissors'], 'id': 923, 'def': 'a tool having two crossed pivoting blades with looped handles', 'name': 'scissors'}, {'frequency': 'f', 'synset': 'scoreboard.n.01', 'synonyms': ['scoreboard'], 'id': 924, 'def': 'a large board for displaying the score of a contest (and some other information)', 'name': 'scoreboard'}, {'frequency': 'r', 'synset': 'scraper.n.01', 'synonyms': ['scraper'], 'id': 925, 'def': 'any of various hand tools for scraping', 'name': 'scraper'}, {'frequency': 'c', 'synset': 'screwdriver.n.01', 'synonyms': ['screwdriver'], 'id': 926, 'def': 'a hand tool for driving screws; has a tip that fits into the head of a screw', 'name': 'screwdriver'}, {'frequency': 'f', 'synset': 'scrub_brush.n.01', 'synonyms': ['scrubbing_brush'], 'id': 927, 'def': 'a brush with short stiff bristles for heavy cleaning', 'name': 'scrubbing_brush'}, {'frequency': 'c', 'synset': 'sculpture.n.01', 'synonyms': ['sculpture'], 'id': 928, 'def': 'a three-dimensional work of art', 'name': 'sculpture'}, {'frequency': 'c', 'synset': 'seabird.n.01', 'synonyms': ['seabird', 'seafowl'], 'id': 929, 'def': 'a bird that frequents coastal waters and the open ocean: gulls; pelicans; gannets; cormorants; albatrosses; petrels; etc.', 'name': 'seabird'}, {'frequency': 'c', 'synset': 'seahorse.n.02', 'synonyms': ['seahorse'], 'id': 930, 'def': 'small fish with horse-like heads bent sharply downward and curled tails', 'name': 'seahorse'}, {'frequency': 'r', 'synset': 'seaplane.n.01', 'synonyms': ['seaplane', 'hydroplane'], 'id': 931, 'def': 'an airplane that can land on or take off from water', 'name': 'seaplane'}, {'frequency': 'c', 'synset': 'seashell.n.01', 'synonyms': ['seashell'], 'id': 932, 'def': 'the shell of a marine organism', 'name': 'seashell'}, {'frequency': 'c', 'synset': 'sewing_machine.n.01', 'synonyms': ['sewing_machine'], 'id': 933, 'def': 'a textile machine used as a home appliance for sewing', 'name': 'sewing_machine'}, {'frequency': 'c', 'synset': 'shaker.n.03', 'synonyms': ['shaker'], 'id': 934, 'def': 'a container in which something can be shaken', 'name': 'shaker'}, {'frequency': 'c', 'synset': 'shampoo.n.01', 'synonyms': ['shampoo'], 'id': 935, 'def': 'cleansing agent consisting of soaps or detergents used for washing the hair', 'name': 'shampoo'}, {'frequency': 'c', 'synset': 'shark.n.01', 'synonyms': ['shark'], 'id': 936, 'def': 'typically large carnivorous fishes with sharpe teeth', 'name': 'shark'}, {'frequency': 'r', 'synset': 'sharpener.n.01', 'synonyms': ['sharpener'], 'id': 937, 'def': 'any implement that is used to make something (an edge or a point) sharper', 'name': 'sharpener'}, {'frequency': 'r', 'synset': 'sharpie.n.03', 'synonyms': ['Sharpie'], 'id': 938, 'def': 'a pen with indelible ink that will write on any surface', 'name': 'Sharpie'}, {'frequency': 'r', 'synset': 'shaver.n.03', 'synonyms': ['shaver_(electric)', 'electric_shaver', 'electric_razor'], 'id': 939, 'def': 'a razor powered by an electric motor', 'name': 'shaver_(electric)'}, {'frequency': 'c', 'synset': 'shaving_cream.n.01', 'synonyms': ['shaving_cream', 'shaving_soap'], 'id': 940, 'def': 'toiletry consisting that forms a rich lather for softening the beard before shaving', 'name': 'shaving_cream'}, {'frequency': 'r', 'synset': 'shawl.n.01', 'synonyms': ['shawl'], 'id': 941, 'def': 'cloak consisting of an oblong piece of cloth used to cover the head and shoulders', 'name': 'shawl'}, {'frequency': 'r', 'synset': 'shears.n.01', 'synonyms': ['shears'], 'id': 942, 'def': 'large scissors with strong blades', 'name': 'shears'}, {'frequency': 'f', 'synset': 'sheep.n.01', 'synonyms': ['sheep'], 'id': 943, 'def': 'woolly usually horned ruminant mammal related to the goat', 'name': 'sheep'}, {'frequency': 'r', 'synset': 'shepherd_dog.n.01', 'synonyms': ['shepherd_dog', 'sheepdog'], 'id': 944, 'def': 'any of various usually long-haired breeds of dog reared to herd and guard sheep', 'name': 'shepherd_dog'}, {'frequency': 'r', 'synset': 'sherbert.n.01', 'synonyms': ['sherbert', 'sherbet'], 'id': 945, 'def': 'a frozen dessert made primarily of fruit juice and sugar', 'name': 'sherbert'}, {'frequency': 'c', 'synset': 'shield.n.02', 'synonyms': ['shield'], 'id': 946, 'def': 'armor carried on the arm to intercept blows', 'name': 'shield'}, {'frequency': 'f', 'synset': 'shirt.n.01', 'synonyms': ['shirt'], 'id': 947, 'def': 'a garment worn on the upper half of the body', 'name': 'shirt'}, {'frequency': 'f', 'synset': 'shoe.n.01', 'synonyms': ['shoe', 'sneaker_(type_of_shoe)', 'tennis_shoe'], 'id': 948, 'def': 'common footwear covering the foot', 'name': 'shoe'}, {'frequency': 'f', 'synset': 'shopping_bag.n.01', 'synonyms': ['shopping_bag'], 'id': 949, 'def': 'a bag made of plastic or strong paper (often with handles); used to transport goods after shopping', 'name': 'shopping_bag'}, {'frequency': 'c', 'synset': 'shopping_cart.n.01', 'synonyms': ['shopping_cart'], 'id': 950, 'def': 'a handcart that holds groceries or other goods while shopping', 'name': 'shopping_cart'}, {'frequency': 'f', 'synset': 'short_pants.n.01', 'synonyms': ['short_pants', 'shorts_(clothing)', 'trunks_(clothing)'], 'id': 951, 'def': 'trousers that end at or above the knee', 'name': 'short_pants'}, {'frequency': 'r', 'synset': 'shot_glass.n.01', 'synonyms': ['shot_glass'], 'id': 952, 'def': 'a small glass adequate to hold a single swallow of whiskey', 'name': 'shot_glass'}, {'frequency': 'f', 'synset': 'shoulder_bag.n.01', 'synonyms': ['shoulder_bag'], 'id': 953, 'def': 'a large handbag that can be carried by a strap looped over the shoulder', 'name': 'shoulder_bag'}, {'frequency': 'c', 'synset': 'shovel.n.01', 'synonyms': ['shovel'], 'id': 954, 'def': 'a hand tool for lifting loose material such as snow, dirt, etc.', 'name': 'shovel'}, {'frequency': 'f', 'synset': 'shower.n.01', 'synonyms': ['shower_head'], 'id': 955, 'def': 'a plumbing fixture that sprays water over you', 'name': 'shower_head'}, {'frequency': 'r', 'synset': 'shower_cap.n.01', 'synonyms': ['shower_cap'], 'id': 956, 'def': 'a tight cap worn to keep hair dry while showering', 'name': 'shower_cap'}, {'frequency': 'f', 'synset': 'shower_curtain.n.01', 'synonyms': ['shower_curtain'], 'id': 957, 'def': 'a curtain that keeps water from splashing out of the shower area', 'name': 'shower_curtain'}, {'frequency': 'r', 'synset': 'shredder.n.01', 'synonyms': ['shredder_(for_paper)'], 'id': 958, 'def': 'a device that shreds documents', 'name': 'shredder_(for_paper)'}, {'frequency': 'f', 'synset': 'signboard.n.01', 'synonyms': ['signboard'], 'id': 959, 'def': 'structure displaying a board on which advertisements can be posted', 'name': 'signboard'}, {'frequency': 'c', 'synset': 'silo.n.01', 'synonyms': ['silo'], 'id': 960, 'def': 'a cylindrical tower used for storing goods', 'name': 'silo'}, {'frequency': 'f', 'synset': 'sink.n.01', 'synonyms': ['sink'], 'id': 961, 'def': 'plumbing fixture consisting of a water basin fixed to a wall or floor and having a drainpipe', 'name': 'sink'}, {'frequency': 'f', 'synset': 'skateboard.n.01', 'synonyms': ['skateboard'], 'id': 962, 'def': 'a board with wheels that is ridden in a standing or crouching position and propelled by foot', 'name': 'skateboard'}, {'frequency': 'c', 'synset': 'skewer.n.01', 'synonyms': ['skewer'], 'id': 963, 'def': 'a long pin for holding meat in position while it is being roasted', 'name': 'skewer'}, {'frequency': 'f', 'synset': 'ski.n.01', 'synonyms': ['ski'], 'id': 964, 'def': 'sports equipment for skiing on snow', 'name': 'ski'}, {'frequency': 'f', 'synset': 'ski_boot.n.01', 'synonyms': ['ski_boot'], 'id': 965, 'def': 'a stiff boot that is fastened to a ski with a ski binding', 'name': 'ski_boot'}, {'frequency': 'f', 'synset': 'ski_parka.n.01', 'synonyms': ['ski_parka', 'ski_jacket'], 'id': 966, 'def': 'a parka to be worn while skiing', 'name': 'ski_parka'}, {'frequency': 'f', 'synset': 'ski_pole.n.01', 'synonyms': ['ski_pole'], 'id': 967, 'def': 'a pole with metal points used as an aid in skiing', 'name': 'ski_pole'}, {'frequency': 'f', 'synset': 'skirt.n.02', 'synonyms': ['skirt'], 'id': 968, 'def': 'a garment hanging from the waist; worn mainly by girls and women', 'name': 'skirt'}, {'frequency': 'r', 'synset': 'skullcap.n.01', 'synonyms': ['skullcap'], 'id': 969, 'def': 'rounded brimless cap fitting the crown of the head', 'name': 'skullcap'}, {'frequency': 'c', 'synset': 'sled.n.01', 'synonyms': ['sled', 'sledge', 'sleigh'], 'id': 970, 'def': 'a vehicle or flat object for transportation over snow by sliding or pulled by dogs, etc.', 'name': 'sled'}, {'frequency': 'c', 'synset': 'sleeping_bag.n.01', 'synonyms': ['sleeping_bag'], 'id': 971, 'def': 'large padded bag designed to be slept in outdoors', 'name': 'sleeping_bag'}, {'frequency': 'r', 'synset': 'sling.n.05', 'synonyms': ['sling_(bandage)', 'triangular_bandage'], 'id': 972, 'def': 'bandage to support an injured forearm; slung over the shoulder or neck', 'name': 'sling_(bandage)'}, {'frequency': 'c', 'synset': 'slipper.n.01', 'synonyms': ['slipper_(footwear)', 'carpet_slipper_(footwear)'], 'id': 973, 'def': 'low footwear that can be slipped on and off easily; usually worn indoors', 'name': 'slipper_(footwear)'}, {'frequency': 'r', 'synset': 'smoothie.n.02', 'synonyms': ['smoothie'], 'id': 974, 'def': 'a thick smooth drink consisting of fresh fruit pureed with ice cream or yoghurt or milk', 'name': 'smoothie'}, {'frequency': 'r', 'synset': 'snake.n.01', 'synonyms': ['snake', 'serpent'], 'id': 975, 'def': 'limbless scaly elongate reptile; some are venomous', 'name': 'snake'}, {'frequency': 'f', 'synset': 'snowboard.n.01', 'synonyms': ['snowboard'], 'id': 976, 'def': 'a board that resembles a broad ski or a small surfboard; used in a standing position to slide down snow-covered slopes', 'name': 'snowboard'}, {'frequency': 'c', 'synset': 'snowman.n.01', 'synonyms': ['snowman'], 'id': 977, 'def': 'a figure of a person made of packed snow', 'name': 'snowman'}, {'frequency': 'c', 'synset': 'snowmobile.n.01', 'synonyms': ['snowmobile'], 'id': 978, 'def': 'tracked vehicle for travel on snow having skis in front', 'name': 'snowmobile'}, {'frequency': 'f', 'synset': 'soap.n.01', 'synonyms': ['soap'], 'id': 979, 'def': 'a cleansing agent made from the salts of vegetable or animal fats', 'name': 'soap'}, {'frequency': 'f', 'synset': 'soccer_ball.n.01', 'synonyms': ['soccer_ball'], 'id': 980, 'def': "an inflated ball used in playing soccer (called `football' outside of the United States)", 'name': 'soccer_ball'}, {'frequency': 'f', 'synset': 'sock.n.01', 'synonyms': ['sock'], 'id': 981, 'def': 'cloth covering for the foot; worn inside the shoe; reaches to between the ankle and the knee', 'name': 'sock'}, {'frequency': 'f', 'synset': 'sofa.n.01', 'synonyms': ['sofa', 'couch', 'lounge'], 'id': 982, 'def': 'an upholstered seat for more than one person', 'name': 'sofa'}, {'frequency': 'r', 'synset': 'softball.n.01', 'synonyms': ['softball'], 'id': 983, 'def': 'ball used in playing softball', 'name': 'softball'}, {'frequency': 'c', 'synset': 'solar_array.n.01', 'synonyms': ['solar_array', 'solar_battery', 'solar_panel'], 'id': 984, 'def': 'electrical device consisting of a large array of connected solar cells', 'name': 'solar_array'}, {'frequency': 'r', 'synset': 'sombrero.n.02', 'synonyms': ['sombrero'], 'id': 985, 'def': 'a straw hat with a tall crown and broad brim; worn in American southwest and in Mexico', 'name': 'sombrero'}, {'frequency': 'f', 'synset': 'soup.n.01', 'synonyms': ['soup'], 'id': 986, 'def': 'liquid food especially of meat or fish or vegetable stock often containing pieces of solid food', 'name': 'soup'}, {'frequency': 'r', 'synset': 'soup_bowl.n.01', 'synonyms': ['soup_bowl'], 'id': 987, 'def': 'a bowl for serving soup', 'name': 'soup_bowl'}, {'frequency': 'c', 'synset': 'soupspoon.n.01', 'synonyms': ['soupspoon'], 'id': 988, 'def': 'a spoon with a rounded bowl for eating soup', 'name': 'soupspoon'}, {'frequency': 'c', 'synset': 'sour_cream.n.01', 'synonyms': ['sour_cream', 'soured_cream'], 'id': 989, 'def': 'soured light cream', 'name': 'sour_cream'}, {'frequency': 'r', 'synset': 'soya_milk.n.01', 'synonyms': ['soya_milk', 'soybean_milk', 'soymilk'], 'id': 990, 'def': 'a milk substitute containing soybean flour and water; used in some infant formulas and in making tofu', 'name': 'soya_milk'}, {'frequency': 'r', 'synset': 'space_shuttle.n.01', 'synonyms': ['space_shuttle'], 'id': 991, 'def': "a reusable spacecraft with wings for a controlled descent through the Earth's atmosphere", 'name': 'space_shuttle'}, {'frequency': 'r', 'synset': 'sparkler.n.02', 'synonyms': ['sparkler_(fireworks)'], 'id': 992, 'def': 'a firework that burns slowly and throws out a shower of sparks', 'name': 'sparkler_(fireworks)'}, {'frequency': 'f', 'synset': 'spatula.n.02', 'synonyms': ['spatula'], 'id': 993, 'def': 'a hand tool with a thin flexible blade used to mix or spread soft substances', 'name': 'spatula'}, {'frequency': 'r', 'synset': 'spear.n.01', 'synonyms': ['spear', 'lance'], 'id': 994, 'def': 'a long pointed rod used as a tool or weapon', 'name': 'spear'}, {'frequency': 'f', 'synset': 'spectacles.n.01', 'synonyms': ['spectacles', 'specs', 'eyeglasses', 'glasses'], 'id': 995, 'def': 'optical instrument consisting of a frame that holds a pair of lenses for correcting defective vision', 'name': 'spectacles'}, {'frequency': 'c', 'synset': 'spice_rack.n.01', 'synonyms': ['spice_rack'], 'id': 996, 'def': 'a rack for displaying containers filled with spices', 'name': 'spice_rack'}, {'frequency': 'c', 'synset': 'spider.n.01', 'synonyms': ['spider'], 'id': 997, 'def': 'predatory arachnid with eight legs, two poison fangs, two feelers, and usually two silk-spinning organs at the back end of the body', 'name': 'spider'}, {'frequency': 'r', 'synset': 'spiny_lobster.n.02', 'synonyms': ['crawfish', 'crayfish'], 'id': 998, 'def': 'large edible marine crustacean having a spiny carapace but lacking the large pincers of true lobsters', 'name': 'crawfish'}, {'frequency': 'c', 'synset': 'sponge.n.01', 'synonyms': ['sponge'], 'id': 999, 'def': 'a porous mass usable to absorb water typically used for cleaning', 'name': 'sponge'}, {'frequency': 'f', 'synset': 'spoon.n.01', 'synonyms': ['spoon'], 'id': 1000, 'def': 'a piece of cutlery with a shallow bowl-shaped container and a handle', 'name': 'spoon'}, {'frequency': 'c', 'synset': 'sportswear.n.01', 'synonyms': ['sportswear', 'athletic_wear', 'activewear'], 'id': 1001, 'def': 'attire worn for sport or for casual wear', 'name': 'sportswear'}, {'frequency': 'c', 'synset': 'spotlight.n.02', 'synonyms': ['spotlight'], 'id': 1002, 'def': 'a lamp that produces a strong beam of light to illuminate a restricted area; used to focus attention of a stage performer', 'name': 'spotlight'}, {'frequency': 'r', 'synset': 'squid.n.01', 'synonyms': ['squid_(food)', 'calamari', 'calamary'], 'id': 1003, 'def': '(Italian cuisine) squid prepared as food', 'name': 'squid_(food)'}, {'frequency': 'c', 'synset': 'squirrel.n.01', 'synonyms': ['squirrel'], 'id': 1004, 'def': 'a kind of arboreal rodent having a long bushy tail', 'name': 'squirrel'}, {'frequency': 'r', 'synset': 'stagecoach.n.01', 'synonyms': ['stagecoach'], 'id': 1005, 'def': 'a large coach-and-four formerly used to carry passengers and mail on regular routes between towns', 'name': 'stagecoach'}, {'frequency': 'c', 'synset': 'stapler.n.01', 'synonyms': ['stapler_(stapling_machine)'], 'id': 1006, 'def': 'a machine that inserts staples into sheets of paper in order to fasten them together', 'name': 'stapler_(stapling_machine)'}, {'frequency': 'c', 'synset': 'starfish.n.01', 'synonyms': ['starfish', 'sea_star'], 'id': 1007, 'def': 'echinoderms characterized by five arms extending from a central disk', 'name': 'starfish'}, {'frequency': 'f', 'synset': 'statue.n.01', 'synonyms': ['statue_(sculpture)'], 'id': 1008, 'def': 'a sculpture representing a human or animal', 'name': 'statue_(sculpture)'}, {'frequency': 'c', 'synset': 'steak.n.01', 'synonyms': ['steak_(food)'], 'id': 1009, 'def': 'a slice of meat cut from the fleshy part of an animal or large fish', 'name': 'steak_(food)'}, {'frequency': 'r', 'synset': 'steak_knife.n.01', 'synonyms': ['steak_knife'], 'id': 1010, 'def': 'a sharp table knife used in eating steak', 'name': 'steak_knife'}, {'frequency': 'f', 'synset': 'steering_wheel.n.01', 'synonyms': ['steering_wheel'], 'id': 1011, 'def': 'a handwheel that is used for steering', 'name': 'steering_wheel'}, {'frequency': 'r', 'synset': 'step_ladder.n.01', 'synonyms': ['stepladder'], 'id': 1012, 'def': 'a folding portable ladder hinged at the top', 'name': 'stepladder'}, {'frequency': 'c', 'synset': 'step_stool.n.01', 'synonyms': ['step_stool'], 'id': 1013, 'def': 'a stool that has one or two steps that fold under the seat', 'name': 'step_stool'}, {'frequency': 'c', 'synset': 'stereo.n.01', 'synonyms': ['stereo_(sound_system)'], 'id': 1014, 'def': 'electronic device for playing audio', 'name': 'stereo_(sound_system)'}, {'frequency': 'r', 'synset': 'stew.n.02', 'synonyms': ['stew'], 'id': 1015, 'def': 'food prepared by stewing especially meat or fish with vegetables', 'name': 'stew'}, {'frequency': 'r', 'synset': 'stirrer.n.02', 'synonyms': ['stirrer'], 'id': 1016, 'def': 'an implement used for stirring', 'name': 'stirrer'}, {'frequency': 'f', 'synset': 'stirrup.n.01', 'synonyms': ['stirrup'], 'id': 1017, 'def': "support consisting of metal loops into which rider's feet go", 'name': 'stirrup'}, {'frequency': 'f', 'synset': 'stool.n.01', 'synonyms': ['stool'], 'id': 1018, 'def': 'a simple seat without a back or arms', 'name': 'stool'}, {'frequency': 'f', 'synset': 'stop_sign.n.01', 'synonyms': ['stop_sign'], 'id': 1019, 'def': 'a traffic sign to notify drivers that they must come to a complete stop', 'name': 'stop_sign'}, {'frequency': 'f', 'synset': 'stoplight.n.01', 'synonyms': ['brake_light'], 'id': 1020, 'def': 'a red light on the rear of a motor vehicle that signals when the brakes are applied', 'name': 'brake_light'}, {'frequency': 'f', 'synset': 'stove.n.01', 'synonyms': ['stove', 'kitchen_stove', 'range_(kitchen_appliance)', 'kitchen_range', 'cooking_stove'], 'id': 1021, 'def': 'a kitchen appliance used for cooking food', 'name': 'stove'}, {'frequency': 'c', 'synset': 'strainer.n.01', 'synonyms': ['strainer'], 'id': 1022, 'def': 'a filter to retain larger pieces while smaller pieces and liquids pass through', 'name': 'strainer'}, {'frequency': 'f', 'synset': 'strap.n.01', 'synonyms': ['strap'], 'id': 1023, 'def': 'an elongated strip of material for binding things together or holding', 'name': 'strap'}, {'frequency': 'f', 'synset': 'straw.n.04', 'synonyms': ['straw_(for_drinking)', 'drinking_straw'], 'id': 1024, 'def': 'a thin paper or plastic tube used to suck liquids into the mouth', 'name': 'straw_(for_drinking)'}, {'frequency': 'f', 'synset': 'strawberry.n.01', 'synonyms': ['strawberry'], 'id': 1025, 'def': 'sweet fleshy red fruit', 'name': 'strawberry'}, {'frequency': 'f', 'synset': 'street_sign.n.01', 'synonyms': ['street_sign'], 'id': 1026, 'def': 'a sign visible from the street', 'name': 'street_sign'}, {'frequency': 'f', 'synset': 'streetlight.n.01', 'synonyms': ['streetlight', 'street_lamp'], 'id': 1027, 'def': 'a lamp supported on a lamppost; for illuminating a street', 'name': 'streetlight'}, {'frequency': 'r', 'synset': 'string_cheese.n.01', 'synonyms': ['string_cheese'], 'id': 1028, 'def': 'cheese formed in long strings twisted together', 'name': 'string_cheese'}, {'frequency': 'r', 'synset': 'stylus.n.02', 'synonyms': ['stylus'], 'id': 1029, 'def': 'a pointed tool for writing or drawing or engraving, including pens', 'name': 'stylus'}, {'frequency': 'r', 'synset': 'subwoofer.n.01', 'synonyms': ['subwoofer'], 'id': 1030, 'def': 'a loudspeaker that is designed to reproduce very low bass frequencies', 'name': 'subwoofer'}, {'frequency': 'r', 'synset': 'sugar_bowl.n.01', 'synonyms': ['sugar_bowl'], 'id': 1031, 'def': 'a dish in which sugar is served', 'name': 'sugar_bowl'}, {'frequency': 'r', 'synset': 'sugarcane.n.01', 'synonyms': ['sugarcane_(plant)'], 'id': 1032, 'def': 'juicy canes whose sap is a source of molasses and commercial sugar; fresh canes are sometimes chewed for the juice', 'name': 'sugarcane_(plant)'}, {'frequency': 'f', 'synset': 'suit.n.01', 'synonyms': ['suit_(clothing)'], 'id': 1033, 'def': 'a set of garments (usually including a jacket and trousers or skirt) for outerwear all of the same fabric and color', 'name': 'suit_(clothing)'}, {'frequency': 'c', 'synset': 'sunflower.n.01', 'synonyms': ['sunflower'], 'id': 1034, 'def': 'any plant of the genus Helianthus having large flower heads with dark disk florets and showy yellow rays', 'name': 'sunflower'}, {'frequency': 'f', 'synset': 'sunglasses.n.01', 'synonyms': ['sunglasses'], 'id': 1035, 'def': 'spectacles that are darkened or polarized to protect the eyes from the glare of the sun', 'name': 'sunglasses'}, {'frequency': 'c', 'synset': 'sunhat.n.01', 'synonyms': ['sunhat'], 'id': 1036, 'def': 'a hat with a broad brim that protects the face from direct exposure to the sun', 'name': 'sunhat'}, {'frequency': 'f', 'synset': 'surfboard.n.01', 'synonyms': ['surfboard'], 'id': 1037, 'def': 'a narrow buoyant board for riding surf', 'name': 'surfboard'}, {'frequency': 'c', 'synset': 'sushi.n.01', 'synonyms': ['sushi'], 'id': 1038, 'def': 'rice (with raw fish) wrapped in seaweed', 'name': 'sushi'}, {'frequency': 'c', 'synset': 'swab.n.02', 'synonyms': ['mop'], 'id': 1039, 'def': 'cleaning implement consisting of absorbent material fastened to a handle; for cleaning floors', 'name': 'mop'}, {'frequency': 'c', 'synset': 'sweat_pants.n.01', 'synonyms': ['sweat_pants'], 'id': 1040, 'def': 'loose-fitting trousers with elastic cuffs; worn by athletes', 'name': 'sweat_pants'}, {'frequency': 'c', 'synset': 'sweatband.n.02', 'synonyms': ['sweatband'], 'id': 1041, 'def': 'a band of material tied around the forehead or wrist to absorb sweat', 'name': 'sweatband'}, {'frequency': 'f', 'synset': 'sweater.n.01', 'synonyms': ['sweater'], 'id': 1042, 'def': 'a crocheted or knitted garment covering the upper part of the body', 'name': 'sweater'}, {'frequency': 'f', 'synset': 'sweatshirt.n.01', 'synonyms': ['sweatshirt'], 'id': 1043, 'def': 'cotton knit pullover with long sleeves worn during athletic activity', 'name': 'sweatshirt'}, {'frequency': 'c', 'synset': 'sweet_potato.n.02', 'synonyms': ['sweet_potato'], 'id': 1044, 'def': 'the edible tuberous root of the sweet potato vine', 'name': 'sweet_potato'}, {'frequency': 'f', 'synset': 'swimsuit.n.01', 'synonyms': ['swimsuit', 'swimwear', 'bathing_suit', 'swimming_costume', 'bathing_costume', 'swimming_trunks', 'bathing_trunks'], 'id': 1045, 'def': 'garment worn for swimming', 'name': 'swimsuit'}, {'frequency': 'c', 'synset': 'sword.n.01', 'synonyms': ['sword'], 'id': 1046, 'def': 'a cutting or thrusting weapon that has a long metal blade', 'name': 'sword'}, {'frequency': 'r', 'synset': 'syringe.n.01', 'synonyms': ['syringe'], 'id': 1047, 'def': 'a medical instrument used to inject or withdraw fluids', 'name': 'syringe'}, {'frequency': 'r', 'synset': 'tabasco.n.02', 'synonyms': ['Tabasco_sauce'], 'id': 1048, 'def': 'very spicy sauce (trade name Tabasco) made from fully-aged red peppers', 'name': 'Tabasco_sauce'}, {'frequency': 'r', 'synset': 'table-tennis_table.n.01', 'synonyms': ['table-tennis_table', 'ping-pong_table'], 'id': 1049, 'def': 'a table used for playing table tennis', 'name': 'table-tennis_table'}, {'frequency': 'f', 'synset': 'table.n.02', 'synonyms': ['table'], 'id': 1050, 'def': 'a piece of furniture having a smooth flat top that is usually supported by one or more vertical legs', 'name': 'table'}, {'frequency': 'c', 'synset': 'table_lamp.n.01', 'synonyms': ['table_lamp'], 'id': 1051, 'def': 'a lamp that sits on a table', 'name': 'table_lamp'}, {'frequency': 'f', 'synset': 'tablecloth.n.01', 'synonyms': ['tablecloth'], 'id': 1052, 'def': 'a covering spread over a dining table', 'name': 'tablecloth'}, {'frequency': 'r', 'synset': 'tachometer.n.01', 'synonyms': ['tachometer'], 'id': 1053, 'def': 'measuring instrument for indicating speed of rotation', 'name': 'tachometer'}, {'frequency': 'r', 'synset': 'taco.n.02', 'synonyms': ['taco'], 'id': 1054, 'def': 'a small tortilla cupped around a filling', 'name': 'taco'}, {'frequency': 'f', 'synset': 'tag.n.02', 'synonyms': ['tag'], 'id': 1055, 'def': 'a label associated with something for the purpose of identification or information', 'name': 'tag'}, {'frequency': 'f', 'synset': 'taillight.n.01', 'synonyms': ['taillight', 'rear_light'], 'id': 1056, 'def': 'lamp (usually red) mounted at the rear of a motor vehicle', 'name': 'taillight'}, {'frequency': 'r', 'synset': 'tambourine.n.01', 'synonyms': ['tambourine'], 'id': 1057, 'def': 'a shallow drum with a single drumhead and with metallic disks in the sides', 'name': 'tambourine'}, {'frequency': 'r', 'synset': 'tank.n.01', 'synonyms': ['army_tank', 'armored_combat_vehicle', 'armoured_combat_vehicle'], 'id': 1058, 'def': 'an enclosed armored military vehicle; has a cannon and moves on caterpillar treads', 'name': 'army_tank'}, {'frequency': 'f', 'synset': 'tank.n.02', 'synonyms': ['tank_(storage_vessel)', 'storage_tank'], 'id': 1059, 'def': 'a large (usually metallic) vessel for holding gases or liquids', 'name': 'tank_(storage_vessel)'}, {'frequency': 'f', 'synset': 'tank_top.n.01', 'synonyms': ['tank_top_(clothing)'], 'id': 1060, 'def': 'a tight-fitting sleeveless shirt with wide shoulder straps and low neck and no front opening', 'name': 'tank_top_(clothing)'}, {'frequency': 'f', 'synset': 'tape.n.01', 'synonyms': ['tape_(sticky_cloth_or_paper)'], 'id': 1061, 'def': 'a long thin piece of cloth or paper as used for binding or fastening', 'name': 'tape_(sticky_cloth_or_paper)'}, {'frequency': 'c', 'synset': 'tape.n.04', 'synonyms': ['tape_measure', 'measuring_tape'], 'id': 1062, 'def': 'measuring instrument consisting of a narrow strip (cloth or metal) marked in inches or centimeters and used for measuring lengths', 'name': 'tape_measure'}, {'frequency': 'c', 'synset': 'tapestry.n.02', 'synonyms': ['tapestry'], 'id': 1063, 'def': 'a heavy textile with a woven design; used for curtains and upholstery', 'name': 'tapestry'}, {'frequency': 'f', 'synset': 'tarpaulin.n.01', 'synonyms': ['tarp'], 'id': 1064, 'def': 'waterproofed canvas', 'name': 'tarp'}, {'frequency': 'c', 'synset': 'tartan.n.01', 'synonyms': ['tartan', 'plaid'], 'id': 1065, 'def': 'a cloth having a crisscross design', 'name': 'tartan'}, {'frequency': 'c', 'synset': 'tassel.n.01', 'synonyms': ['tassel'], 'id': 1066, 'def': 'adornment consisting of a bunch of cords fastened at one end', 'name': 'tassel'}, {'frequency': 'c', 'synset': 'tea_bag.n.01', 'synonyms': ['tea_bag'], 'id': 1067, 'def': 'a measured amount of tea in a bag for an individual serving of tea', 'name': 'tea_bag'}, {'frequency': 'c', 'synset': 'teacup.n.02', 'synonyms': ['teacup'], 'id': 1068, 'def': 'a cup from which tea is drunk', 'name': 'teacup'}, {'frequency': 'c', 'synset': 'teakettle.n.01', 'synonyms': ['teakettle'], 'id': 1069, 'def': 'kettle for boiling water to make tea', 'name': 'teakettle'}, {'frequency': 'f', 'synset': 'teapot.n.01', 'synonyms': ['teapot'], 'id': 1070, 'def': 'pot for brewing tea; usually has a spout and handle', 'name': 'teapot'}, {'frequency': 'f', 'synset': 'teddy.n.01', 'synonyms': ['teddy_bear'], 'id': 1071, 'def': "plaything consisting of a child's toy bear (usually plush and stuffed with soft materials)", 'name': 'teddy_bear'}, {'frequency': 'f', 'synset': 'telephone.n.01', 'synonyms': ['telephone', 'phone', 'telephone_set'], 'id': 1072, 'def': 'electronic device for communicating by voice over long distances (includes wired and wireless/cell phones)', 'name': 'telephone'}, {'frequency': 'c', 'synset': 'telephone_booth.n.01', 'synonyms': ['telephone_booth', 'phone_booth', 'call_box', 'telephone_box', 'telephone_kiosk'], 'id': 1073, 'def': 'booth for using a telephone', 'name': 'telephone_booth'}, {'frequency': 'f', 'synset': 'telephone_pole.n.01', 'synonyms': ['telephone_pole', 'telegraph_pole', 'telegraph_post'], 'id': 1074, 'def': 'tall pole supporting telephone wires', 'name': 'telephone_pole'}, {'frequency': 'r', 'synset': 'telephoto_lens.n.01', 'synonyms': ['telephoto_lens', 'zoom_lens'], 'id': 1075, 'def': 'a camera lens that magnifies the image', 'name': 'telephoto_lens'}, {'frequency': 'c', 'synset': 'television_camera.n.01', 'synonyms': ['television_camera', 'tv_camera'], 'id': 1076, 'def': 'television equipment for capturing and recording video', 'name': 'television_camera'}, {'frequency': 'f', 'synset': 'television_receiver.n.01', 'synonyms': ['television_set', 'tv', 'tv_set'], 'id': 1077, 'def': 'an electronic device that receives television signals and displays them on a screen', 'name': 'television_set'}, {'frequency': 'f', 'synset': 'tennis_ball.n.01', 'synonyms': ['tennis_ball'], 'id': 1078, 'def': 'ball about the size of a fist used in playing tennis', 'name': 'tennis_ball'}, {'frequency': 'f', 'synset': 'tennis_racket.n.01', 'synonyms': ['tennis_racket'], 'id': 1079, 'def': 'a racket used to play tennis', 'name': 'tennis_racket'}, {'frequency': 'r', 'synset': 'tequila.n.01', 'synonyms': ['tequila'], 'id': 1080, 'def': 'Mexican liquor made from fermented juices of an agave plant', 'name': 'tequila'}, {'frequency': 'c', 'synset': 'thermometer.n.01', 'synonyms': ['thermometer'], 'id': 1081, 'def': 'measuring instrument for measuring temperature', 'name': 'thermometer'}, {'frequency': 'c', 'synset': 'thermos.n.01', 'synonyms': ['thermos_bottle'], 'id': 1082, 'def': 'vacuum flask that preserves temperature of hot or cold drinks', 'name': 'thermos_bottle'}, {'frequency': 'f', 'synset': 'thermostat.n.01', 'synonyms': ['thermostat'], 'id': 1083, 'def': 'a regulator for automatically regulating temperature by starting or stopping the supply of heat', 'name': 'thermostat'}, {'frequency': 'r', 'synset': 'thimble.n.02', 'synonyms': ['thimble'], 'id': 1084, 'def': 'a small metal cap to protect the finger while sewing; can be used as a small container', 'name': 'thimble'}, {'frequency': 'c', 'synset': 'thread.n.01', 'synonyms': ['thread', 'yarn'], 'id': 1085, 'def': 'a fine cord of twisted fibers (of cotton or silk or wool or nylon etc.) used in sewing and weaving', 'name': 'thread'}, {'frequency': 'c', 'synset': 'thumbtack.n.01', 'synonyms': ['thumbtack', 'drawing_pin', 'pushpin'], 'id': 1086, 'def': 'a tack for attaching papers to a bulletin board or drawing board', 'name': 'thumbtack'}, {'frequency': 'c', 'synset': 'tiara.n.01', 'synonyms': ['tiara'], 'id': 1087, 'def': 'a jeweled headdress worn by women on formal occasions', 'name': 'tiara'}, {'frequency': 'c', 'synset': 'tiger.n.02', 'synonyms': ['tiger'], 'id': 1088, 'def': 'large feline of forests in most of Asia having a tawny coat with black stripes', 'name': 'tiger'}, {'frequency': 'c', 'synset': 'tights.n.01', 'synonyms': ['tights_(clothing)', 'leotards'], 'id': 1089, 'def': 'skintight knit hose covering the body from the waist to the feet worn by acrobats and dancers and as stockings by women and girls', 'name': 'tights_(clothing)'}, {'frequency': 'c', 'synset': 'timer.n.01', 'synonyms': ['timer', 'stopwatch'], 'id': 1090, 'def': 'a timepiece that measures a time interval and signals its end', 'name': 'timer'}, {'frequency': 'f', 'synset': 'tinfoil.n.01', 'synonyms': ['tinfoil'], 'id': 1091, 'def': 'foil made of tin or an alloy of tin and lead', 'name': 'tinfoil'}, {'frequency': 'c', 'synset': 'tinsel.n.01', 'synonyms': ['tinsel'], 'id': 1092, 'def': 'a showy decoration that is basically valueless', 'name': 'tinsel'}, {'frequency': 'f', 'synset': 'tissue.n.02', 'synonyms': ['tissue_paper'], 'id': 1093, 'def': 'a soft thin (usually translucent) paper', 'name': 'tissue_paper'}, {'frequency': 'c', 'synset': 'toast.n.01', 'synonyms': ['toast_(food)'], 'id': 1094, 'def': 'slice of bread that has been toasted', 'name': 'toast_(food)'}, {'frequency': 'f', 'synset': 'toaster.n.02', 'synonyms': ['toaster'], 'id': 1095, 'def': 'a kitchen appliance (usually electric) for toasting bread', 'name': 'toaster'}, {'frequency': 'f', 'synset': 'toaster_oven.n.01', 'synonyms': ['toaster_oven'], 'id': 1096, 'def': 'kitchen appliance consisting of a small electric oven for toasting or warming food', 'name': 'toaster_oven'}, {'frequency': 'f', 'synset': 'toilet.n.02', 'synonyms': ['toilet'], 'id': 1097, 'def': 'a plumbing fixture for defecation and urination', 'name': 'toilet'}, {'frequency': 'f', 'synset': 'toilet_tissue.n.01', 'synonyms': ['toilet_tissue', 'toilet_paper', 'bathroom_tissue'], 'id': 1098, 'def': 'a soft thin absorbent paper for use in toilets', 'name': 'toilet_tissue'}, {'frequency': 'f', 'synset': 'tomato.n.01', 'synonyms': ['tomato'], 'id': 1099, 'def': 'mildly acid red or yellow pulpy fruit eaten as a vegetable', 'name': 'tomato'}, {'frequency': 'f', 'synset': 'tongs.n.01', 'synonyms': ['tongs'], 'id': 1100, 'def': 'any of various devices for taking hold of objects; usually have two hinged legs with handles above and pointed hooks below', 'name': 'tongs'}, {'frequency': 'c', 'synset': 'toolbox.n.01', 'synonyms': ['toolbox'], 'id': 1101, 'def': 'a box or chest or cabinet for holding hand tools', 'name': 'toolbox'}, {'frequency': 'f', 'synset': 'toothbrush.n.01', 'synonyms': ['toothbrush'], 'id': 1102, 'def': 'small brush; has long handle; used to clean teeth', 'name': 'toothbrush'}, {'frequency': 'f', 'synset': 'toothpaste.n.01', 'synonyms': ['toothpaste'], 'id': 1103, 'def': 'a dentifrice in the form of a paste', 'name': 'toothpaste'}, {'frequency': 'f', 'synset': 'toothpick.n.01', 'synonyms': ['toothpick'], 'id': 1104, 'def': 'pick consisting of a small strip of wood or plastic; used to pick food from between the teeth', 'name': 'toothpick'}, {'frequency': 'f', 'synset': 'top.n.09', 'synonyms': ['cover'], 'id': 1105, 'def': 'covering for a hole (especially a hole in the top of a container)', 'name': 'cover'}, {'frequency': 'c', 'synset': 'tortilla.n.01', 'synonyms': ['tortilla'], 'id': 1106, 'def': 'thin unleavened pancake made from cornmeal or wheat flour', 'name': 'tortilla'}, {'frequency': 'c', 'synset': 'tow_truck.n.01', 'synonyms': ['tow_truck'], 'id': 1107, 'def': 'a truck equipped to hoist and pull wrecked cars (or to remove cars from no-parking zones)', 'name': 'tow_truck'}, {'frequency': 'f', 'synset': 'towel.n.01', 'synonyms': ['towel'], 'id': 1108, 'def': 'a rectangular piece of absorbent cloth (or paper) for drying or wiping', 'name': 'towel'}, {'frequency': 'f', 'synset': 'towel_rack.n.01', 'synonyms': ['towel_rack', 'towel_rail', 'towel_bar'], 'id': 1109, 'def': 'a rack consisting of one or more bars on which towels can be hung', 'name': 'towel_rack'}, {'frequency': 'f', 'synset': 'toy.n.03', 'synonyms': ['toy'], 'id': 1110, 'def': 'a device regarded as providing amusement', 'name': 'toy'}, {'frequency': 'c', 'synset': 'tractor.n.01', 'synonyms': ['tractor_(farm_equipment)'], 'id': 1111, 'def': 'a wheeled vehicle with large wheels; used in farming and other applications', 'name': 'tractor_(farm_equipment)'}, {'frequency': 'f', 'synset': 'traffic_light.n.01', 'synonyms': ['traffic_light'], 'id': 1112, 'def': 'a device to control vehicle traffic often consisting of three or more lights', 'name': 'traffic_light'}, {'frequency': 'c', 'synset': 'trail_bike.n.01', 'synonyms': ['dirt_bike'], 'id': 1113, 'def': 'a lightweight motorcycle equipped with rugged tires and suspension for off-road use', 'name': 'dirt_bike'}, {'frequency': 'f', 'synset': 'trailer_truck.n.01', 'synonyms': ['trailer_truck', 'tractor_trailer', 'trucking_rig', 'articulated_lorry', 'semi_truck'], 'id': 1114, 'def': 'a truck consisting of a tractor and trailer together', 'name': 'trailer_truck'}, {'frequency': 'f', 'synset': 'train.n.01', 'synonyms': ['train_(railroad_vehicle)', 'railroad_train'], 'id': 1115, 'def': 'public or private transport provided by a line of railway cars coupled together and drawn by a locomotive', 'name': 'train_(railroad_vehicle)'}, {'frequency': 'r', 'synset': 'trampoline.n.01', 'synonyms': ['trampoline'], 'id': 1116, 'def': 'gymnastic apparatus consisting of a strong canvas sheet attached with springs to a metal frame', 'name': 'trampoline'}, {'frequency': 'f', 'synset': 'tray.n.01', 'synonyms': ['tray'], 'id': 1117, 'def': 'an open receptacle for holding or displaying or serving articles or food', 'name': 'tray'}, {'frequency': 'r', 'synset': 'trench_coat.n.01', 'synonyms': ['trench_coat'], 'id': 1118, 'def': 'a military style raincoat; belted with deep pockets', 'name': 'trench_coat'}, {'frequency': 'r', 'synset': 'triangle.n.05', 'synonyms': ['triangle_(musical_instrument)'], 'id': 1119, 'def': 'a percussion instrument consisting of a metal bar bent in the shape of an open triangle', 'name': 'triangle_(musical_instrument)'}, {'frequency': 'c', 'synset': 'tricycle.n.01', 'synonyms': ['tricycle'], 'id': 1120, 'def': 'a vehicle with three wheels that is moved by foot pedals', 'name': 'tricycle'}, {'frequency': 'f', 'synset': 'tripod.n.01', 'synonyms': ['tripod'], 'id': 1121, 'def': 'a three-legged rack used for support', 'name': 'tripod'}, {'frequency': 'f', 'synset': 'trouser.n.01', 'synonyms': ['trousers', 'pants_(clothing)'], 'id': 1122, 'def': 'a garment extending from the waist to the knee or ankle, covering each leg separately', 'name': 'trousers'}, {'frequency': 'f', 'synset': 'truck.n.01', 'synonyms': ['truck'], 'id': 1123, 'def': 'an automotive vehicle suitable for hauling', 'name': 'truck'}, {'frequency': 'r', 'synset': 'truffle.n.03', 'synonyms': ['truffle_(chocolate)', 'chocolate_truffle'], 'id': 1124, 'def': 'creamy chocolate candy', 'name': 'truffle_(chocolate)'}, {'frequency': 'c', 'synset': 'trunk.n.02', 'synonyms': ['trunk'], 'id': 1125, 'def': 'luggage consisting of a large strong case used when traveling or for storage', 'name': 'trunk'}, {'frequency': 'r', 'synset': 'tub.n.02', 'synonyms': ['vat'], 'id': 1126, 'def': 'a large vessel for holding or storing liquids', 'name': 'vat'}, {'frequency': 'c', 'synset': 'turban.n.01', 'synonyms': ['turban'], 'id': 1127, 'def': 'a traditional headdress consisting of a long scarf wrapped around the head', 'name': 'turban'}, {'frequency': 'c', 'synset': 'turkey.n.04', 'synonyms': ['turkey_(food)'], 'id': 1128, 'def': 'flesh of large domesticated fowl usually roasted', 'name': 'turkey_(food)'}, {'frequency': 'r', 'synset': 'turnip.n.01', 'synonyms': ['turnip'], 'id': 1129, 'def': 'widely cultivated plant having a large fleshy edible white or yellow root', 'name': 'turnip'}, {'frequency': 'c', 'synset': 'turtle.n.02', 'synonyms': ['turtle'], 'id': 1130, 'def': 'any of various aquatic and land reptiles having a bony shell and flipper-like limbs for swimming', 'name': 'turtle'}, {'frequency': 'c', 'synset': 'turtleneck.n.01', 'synonyms': ['turtleneck_(clothing)', 'polo-neck'], 'id': 1131, 'def': 'a sweater or jersey with a high close-fitting collar', 'name': 'turtleneck_(clothing)'}, {'frequency': 'c', 'synset': 'typewriter.n.01', 'synonyms': ['typewriter'], 'id': 1132, 'def': 'hand-operated character printer for printing written messages one character at a time', 'name': 'typewriter'}, {'frequency': 'f', 'synset': 'umbrella.n.01', 'synonyms': ['umbrella'], 'id': 1133, 'def': 'a lightweight handheld collapsible canopy', 'name': 'umbrella'}, {'frequency': 'f', 'synset': 'underwear.n.01', 'synonyms': ['underwear', 'underclothes', 'underclothing', 'underpants'], 'id': 1134, 'def': 'undergarment worn next to the skin and under the outer garments', 'name': 'underwear'}, {'frequency': 'r', 'synset': 'unicycle.n.01', 'synonyms': ['unicycle'], 'id': 1135, 'def': 'a vehicle with a single wheel that is driven by pedals', 'name': 'unicycle'}, {'frequency': 'f', 'synset': 'urinal.n.01', 'synonyms': ['urinal'], 'id': 1136, 'def': 'a plumbing fixture (usually attached to the wall) used by men to urinate', 'name': 'urinal'}, {'frequency': 'c', 'synset': 'urn.n.01', 'synonyms': ['urn'], 'id': 1137, 'def': 'a large vase that usually has a pedestal or feet', 'name': 'urn'}, {'frequency': 'c', 'synset': 'vacuum.n.04', 'synonyms': ['vacuum_cleaner'], 'id': 1138, 'def': 'an electrical home appliance that cleans by suction', 'name': 'vacuum_cleaner'}, {'frequency': 'f', 'synset': 'vase.n.01', 'synonyms': ['vase'], 'id': 1139, 'def': 'an open jar of glass or porcelain used as an ornament or to hold flowers', 'name': 'vase'}, {'frequency': 'c', 'synset': 'vending_machine.n.01', 'synonyms': ['vending_machine'], 'id': 1140, 'def': 'a slot machine for selling goods', 'name': 'vending_machine'}, {'frequency': 'f', 'synset': 'vent.n.01', 'synonyms': ['vent', 'blowhole', 'air_vent'], 'id': 1141, 'def': 'a hole for the escape of gas or air', 'name': 'vent'}, {'frequency': 'f', 'synset': 'vest.n.01', 'synonyms': ['vest', 'waistcoat'], 'id': 1142, 'def': "a man's sleeveless garment worn underneath a coat", 'name': 'vest'}, {'frequency': 'c', 'synset': 'videotape.n.01', 'synonyms': ['videotape'], 'id': 1143, 'def': 'a video recording made on magnetic tape', 'name': 'videotape'}, {'frequency': 'r', 'synset': 'vinegar.n.01', 'synonyms': ['vinegar'], 'id': 1144, 'def': 'sour-tasting liquid produced usually by oxidation of the alcohol in wine or cider and used as a condiment or food preservative', 'name': 'vinegar'}, {'frequency': 'r', 'synset': 'violin.n.01', 'synonyms': ['violin', 'fiddle'], 'id': 1145, 'def': 'bowed stringed instrument that is the highest member of the violin family', 'name': 'violin'}, {'frequency': 'r', 'synset': 'vodka.n.01', 'synonyms': ['vodka'], 'id': 1146, 'def': 'unaged colorless liquor originating in Russia', 'name': 'vodka'}, {'frequency': 'c', 'synset': 'volleyball.n.02', 'synonyms': ['volleyball'], 'id': 1147, 'def': 'an inflated ball used in playing volleyball', 'name': 'volleyball'}, {'frequency': 'r', 'synset': 'vulture.n.01', 'synonyms': ['vulture'], 'id': 1148, 'def': 'any of various large birds of prey having naked heads and weak claws and feeding chiefly on carrion', 'name': 'vulture'}, {'frequency': 'c', 'synset': 'waffle.n.01', 'synonyms': ['waffle'], 'id': 1149, 'def': 'pancake batter baked in a waffle iron', 'name': 'waffle'}, {'frequency': 'r', 'synset': 'waffle_iron.n.01', 'synonyms': ['waffle_iron'], 'id': 1150, 'def': 'a kitchen appliance for baking waffles', 'name': 'waffle_iron'}, {'frequency': 'c', 'synset': 'wagon.n.01', 'synonyms': ['wagon'], 'id': 1151, 'def': 'any of various kinds of wheeled vehicles drawn by an animal or a tractor', 'name': 'wagon'}, {'frequency': 'c', 'synset': 'wagon_wheel.n.01', 'synonyms': ['wagon_wheel'], 'id': 1152, 'def': 'a wheel of a wagon', 'name': 'wagon_wheel'}, {'frequency': 'c', 'synset': 'walking_stick.n.01', 'synonyms': ['walking_stick'], 'id': 1153, 'def': 'a stick carried in the hand for support in walking', 'name': 'walking_stick'}, {'frequency': 'c', 'synset': 'wall_clock.n.01', 'synonyms': ['wall_clock'], 'id': 1154, 'def': 'a clock mounted on a wall', 'name': 'wall_clock'}, {'frequency': 'f', 'synset': 'wall_socket.n.01', 'synonyms': ['wall_socket', 'wall_plug', 'electric_outlet', 'electrical_outlet', 'outlet', 'electric_receptacle'], 'id': 1155, 'def': 'receptacle providing a place in a wiring system where current can be taken to run electrical devices', 'name': 'wall_socket'}, {'frequency': 'f', 'synset': 'wallet.n.01', 'synonyms': ['wallet', 'billfold'], 'id': 1156, 'def': 'a pocket-size case for holding papers and paper money', 'name': 'wallet'}, {'frequency': 'r', 'synset': 'walrus.n.01', 'synonyms': ['walrus'], 'id': 1157, 'def': 'either of two large northern marine mammals having ivory tusks and tough hide over thick blubber', 'name': 'walrus'}, {'frequency': 'r', 'synset': 'wardrobe.n.01', 'synonyms': ['wardrobe'], 'id': 1158, 'def': 'a tall piece of furniture that provides storage space for clothes; has a door and rails or hooks for hanging clothes', 'name': 'wardrobe'}, {'frequency': 'r', 'synset': 'washbasin.n.01', 'synonyms': ['washbasin', 'basin_(for_washing)', 'washbowl', 'washstand', 'handbasin'], 'id': 1159, 'def': 'a bathroom sink that is permanently installed and connected to a water supply and drainpipe; where you can wash your hands and face', 'name': 'washbasin'}, {'frequency': 'c', 'synset': 'washer.n.03', 'synonyms': ['automatic_washer', 'washing_machine'], 'id': 1160, 'def': 'a home appliance for washing clothes and linens automatically', 'name': 'automatic_washer'}, {'frequency': 'f', 'synset': 'watch.n.01', 'synonyms': ['watch', 'wristwatch'], 'id': 1161, 'def': 'a small, portable timepiece', 'name': 'watch'}, {'frequency': 'f', 'synset': 'water_bottle.n.01', 'synonyms': ['water_bottle'], 'id': 1162, 'def': 'a bottle for holding water', 'name': 'water_bottle'}, {'frequency': 'c', 'synset': 'water_cooler.n.01', 'synonyms': ['water_cooler'], 'id': 1163, 'def': 'a device for cooling and dispensing drinking water', 'name': 'water_cooler'}, {'frequency': 'c', 'synset': 'water_faucet.n.01', 'synonyms': ['water_faucet', 'water_tap', 'tap_(water_faucet)'], 'id': 1164, 'def': 'a faucet for drawing water from a pipe or cask', 'name': 'water_faucet'}, {'frequency': 'r', 'synset': 'water_heater.n.01', 'synonyms': ['water_heater', 'hot-water_heater'], 'id': 1165, 'def': 'a heater and storage tank to supply heated water', 'name': 'water_heater'}, {'frequency': 'c', 'synset': 'water_jug.n.01', 'synonyms': ['water_jug'], 'id': 1166, 'def': 'a jug that holds water', 'name': 'water_jug'}, {'frequency': 'r', 'synset': 'water_pistol.n.01', 'synonyms': ['water_gun', 'squirt_gun'], 'id': 1167, 'def': 'plaything consisting of a toy pistol that squirts water', 'name': 'water_gun'}, {'frequency': 'c', 'synset': 'water_scooter.n.01', 'synonyms': ['water_scooter', 'sea_scooter', 'jet_ski'], 'id': 1168, 'def': 'a motorboat resembling a motor scooter (NOT A SURFBOARD OR WATER SKI)', 'name': 'water_scooter'}, {'frequency': 'c', 'synset': 'water_ski.n.01', 'synonyms': ['water_ski'], 'id': 1169, 'def': 'broad ski for skimming over water towed by a speedboat (DO NOT MARK WATER)', 'name': 'water_ski'}, {'frequency': 'c', 'synset': 'water_tower.n.01', 'synonyms': ['water_tower'], 'id': 1170, 'def': 'a large reservoir for water', 'name': 'water_tower'}, {'frequency': 'c', 'synset': 'watering_can.n.01', 'synonyms': ['watering_can'], 'id': 1171, 'def': 'a container with a handle and a spout with a perforated nozzle; used to sprinkle water over plants', 'name': 'watering_can'}, {'frequency': 'f', 'synset': 'watermelon.n.02', 'synonyms': ['watermelon'], 'id': 1172, 'def': 'large oblong or roundish melon with a hard green rind and sweet watery red or occasionally yellowish pulp', 'name': 'watermelon'}, {'frequency': 'f', 'synset': 'weathervane.n.01', 'synonyms': ['weathervane', 'vane_(weathervane)', 'wind_vane'], 'id': 1173, 'def': 'mechanical device attached to an elevated structure; rotates freely to show the direction of the wind', 'name': 'weathervane'}, {'frequency': 'c', 'synset': 'webcam.n.01', 'synonyms': ['webcam'], 'id': 1174, 'def': 'a digital camera designed to take digital photographs and transmit them over the internet', 'name': 'webcam'}, {'frequency': 'c', 'synset': 'wedding_cake.n.01', 'synonyms': ['wedding_cake', 'bridecake'], 'id': 1175, 'def': 'a rich cake with two or more tiers and covered with frosting and decorations; served at a wedding reception', 'name': 'wedding_cake'}, {'frequency': 'c', 'synset': 'wedding_ring.n.01', 'synonyms': ['wedding_ring', 'wedding_band'], 'id': 1176, 'def': 'a ring given to the bride and/or groom at the wedding', 'name': 'wedding_ring'}, {'frequency': 'f', 'synset': 'wet_suit.n.01', 'synonyms': ['wet_suit'], 'id': 1177, 'def': 'a close-fitting garment made of a permeable material; worn in cold water to retain body heat', 'name': 'wet_suit'}, {'frequency': 'f', 'synset': 'wheel.n.01', 'synonyms': ['wheel'], 'id': 1178, 'def': 'a circular frame with spokes (or a solid disc) that can rotate on a shaft or axle', 'name': 'wheel'}, {'frequency': 'c', 'synset': 'wheelchair.n.01', 'synonyms': ['wheelchair'], 'id': 1179, 'def': 'a movable chair mounted on large wheels', 'name': 'wheelchair'}, {'frequency': 'c', 'synset': 'whipped_cream.n.01', 'synonyms': ['whipped_cream'], 'id': 1180, 'def': 'cream that has been beaten until light and fluffy', 'name': 'whipped_cream'}, {'frequency': 'c', 'synset': 'whistle.n.03', 'synonyms': ['whistle'], 'id': 1181, 'def': 'a small wind instrument that produces a whistling sound by blowing into it', 'name': 'whistle'}, {'frequency': 'c', 'synset': 'wig.n.01', 'synonyms': ['wig'], 'id': 1182, 'def': 'hairpiece covering the head and made of real or synthetic hair', 'name': 'wig'}, {'frequency': 'c', 'synset': 'wind_chime.n.01', 'synonyms': ['wind_chime'], 'id': 1183, 'def': 'a decorative arrangement of pieces of metal or glass or pottery that hang together loosely so the wind can cause them to tinkle', 'name': 'wind_chime'}, {'frequency': 'c', 'synset': 'windmill.n.01', 'synonyms': ['windmill'], 'id': 1184, 'def': 'A mill or turbine that is powered by wind', 'name': 'windmill'}, {'frequency': 'c', 'synset': 'window_box.n.01', 'synonyms': ['window_box_(for_plants)'], 'id': 1185, 'def': 'a container for growing plants on a windowsill', 'name': 'window_box_(for_plants)'}, {'frequency': 'f', 'synset': 'windshield_wiper.n.01', 'synonyms': ['windshield_wiper', 'windscreen_wiper', 'wiper_(for_windshield/screen)'], 'id': 1186, 'def': 'a mechanical device that cleans the windshield', 'name': 'windshield_wiper'}, {'frequency': 'c', 'synset': 'windsock.n.01', 'synonyms': ['windsock', 'air_sock', 'air-sleeve', 'wind_sleeve', 'wind_cone'], 'id': 1187, 'def': 'a truncated cloth cone mounted on a mast/pole; shows wind direction', 'name': 'windsock'}, {'frequency': 'f', 'synset': 'wine_bottle.n.01', 'synonyms': ['wine_bottle'], 'id': 1188, 'def': 'a bottle for holding wine', 'name': 'wine_bottle'}, {'frequency': 'c', 'synset': 'wine_bucket.n.01', 'synonyms': ['wine_bucket', 'wine_cooler'], 'id': 1189, 'def': 'a bucket of ice used to chill a bottle of wine', 'name': 'wine_bucket'}, {'frequency': 'f', 'synset': 'wineglass.n.01', 'synonyms': ['wineglass'], 'id': 1190, 'def': 'a glass that has a stem and in which wine is served', 'name': 'wineglass'}, {'frequency': 'f', 'synset': 'winker.n.02', 'synonyms': ['blinder_(for_horses)'], 'id': 1191, 'def': 'blinds that prevent a horse from seeing something on either side', 'name': 'blinder_(for_horses)'}, {'frequency': 'c', 'synset': 'wok.n.01', 'synonyms': ['wok'], 'id': 1192, 'def': 'pan with a convex bottom; used for frying in Chinese cooking', 'name': 'wok'}, {'frequency': 'r', 'synset': 'wolf.n.01', 'synonyms': ['wolf'], 'id': 1193, 'def': 'a wild carnivorous mammal of the dog family, living and hunting in packs', 'name': 'wolf'}, {'frequency': 'c', 'synset': 'wooden_spoon.n.02', 'synonyms': ['wooden_spoon'], 'id': 1194, 'def': 'a spoon made of wood', 'name': 'wooden_spoon'}, {'frequency': 'c', 'synset': 'wreath.n.01', 'synonyms': ['wreath'], 'id': 1195, 'def': 'an arrangement of flowers, leaves, or stems fastened in a ring', 'name': 'wreath'}, {'frequency': 'c', 'synset': 'wrench.n.03', 'synonyms': ['wrench', 'spanner'], 'id': 1196, 'def': 'a hand tool that is used to hold or twist a nut or bolt', 'name': 'wrench'}, {'frequency': 'f', 'synset': 'wristband.n.01', 'synonyms': ['wristband'], 'id': 1197, 'def': 'band consisting of a part of a sleeve that covers the wrist', 'name': 'wristband'}, {'frequency': 'f', 'synset': 'wristlet.n.01', 'synonyms': ['wristlet', 'wrist_band'], 'id': 1198, 'def': 'a band or bracelet worn around the wrist', 'name': 'wristlet'}, {'frequency': 'c', 'synset': 'yacht.n.01', 'synonyms': ['yacht'], 'id': 1199, 'def': 'an expensive vessel propelled by sail or power and used for cruising or racing', 'name': 'yacht'}, {'frequency': 'c', 'synset': 'yogurt.n.01', 'synonyms': ['yogurt', 'yoghurt', 'yoghourt'], 'id': 1200, 'def': 'a custard-like food made from curdled milk', 'name': 'yogurt'}, {'frequency': 'c', 'synset': 'yoke.n.07', 'synonyms': ['yoke_(animal_equipment)'], 'id': 1201, 'def': 'gear joining two animals at the neck; NOT egg yolk', 'name': 'yoke_(animal_equipment)'}, {'frequency': 'f', 'synset': 'zebra.n.01', 'synonyms': ['zebra'], 'id': 1202, 'def': 'any of several fleet black-and-white striped African equines', 'name': 'zebra'}, {'frequency': 'c', 'synset': 'zucchini.n.02', 'synonyms': ['zucchini', 'courgette'], 'id': 1203, 'def': 'small cucumber-shaped vegetable marrow; typically dark green', 'name': 'zucchini'}] # noqa +# fmt: on diff --git a/detectron2/data/datasets/pascal_voc.py b/detectron2/data/datasets/pascal_voc.py new file mode 100644 index 0000000000000000000000000000000000000000..dbbf82cb96442bfa0cf05ed0f4dddf3645434b7e --- /dev/null +++ b/detectron2/data/datasets/pascal_voc.py @@ -0,0 +1,82 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import numpy as np +import os +import xml.etree.ElementTree as ET +from typing import List, Tuple, Union + +from detectron2.data import DatasetCatalog, MetadataCatalog +from detectron2.structures import BoxMode +from detectron2.utils.file_io import PathManager + +__all__ = ["load_voc_instances", "register_pascal_voc"] + + +# fmt: off +CLASS_NAMES = ( + "aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", + "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", + "pottedplant", "sheep", "sofa", "train", "tvmonitor" +) +# fmt: on + + +def load_voc_instances(dirname: str, split: str, class_names: Union[List[str], Tuple[str, ...]]): + """ + Load Pascal VOC detection annotations to Detectron2 format. + + Args: + dirname: Contain "Annotations", "ImageSets", "JPEGImages" + split (str): one of "train", "test", "val", "trainval" + class_names: list or tuple of class names + """ + with PathManager.open(os.path.join(dirname, "ImageSets", "Main", split + ".txt")) as f: + fileids = np.loadtxt(f, dtype=np.str) + + # Needs to read many small annotation files. Makes sense at local + annotation_dirname = PathManager.get_local_path(os.path.join(dirname, "Annotations/")) + dicts = [] + for fileid in fileids: + anno_file = os.path.join(annotation_dirname, fileid + ".xml") + jpeg_file = os.path.join(dirname, "JPEGImages", fileid + ".jpg") + + with PathManager.open(anno_file) as f: + tree = ET.parse(f) + + r = { + "file_name": jpeg_file, + "image_id": fileid, + "height": int(tree.findall("./size/height")[0].text), + "width": int(tree.findall("./size/width")[0].text), + } + instances = [] + + for obj in tree.findall("object"): + cls = obj.find("name").text + # We include "difficult" samples in training. + # Based on limited experiments, they don't hurt accuracy. + # difficult = int(obj.find("difficult").text) + # if difficult == 1: + # continue + bbox = obj.find("bndbox") + bbox = [float(bbox.find(x).text) for x in ["xmin", "ymin", "xmax", "ymax"]] + # Original annotations are integers in the range [1, W or H] + # Assuming they mean 1-based pixel indices (inclusive), + # a box with annotation (xmin=1, xmax=W) covers the whole image. + # In coordinate space this is represented by (xmin=0, xmax=W) + bbox[0] -= 1.0 + bbox[1] -= 1.0 + instances.append( + {"category_id": class_names.index(cls), "bbox": bbox, "bbox_mode": BoxMode.XYXY_ABS} + ) + r["annotations"] = instances + dicts.append(r) + return dicts + + +def register_pascal_voc(name, dirname, split, year, class_names=CLASS_NAMES): + DatasetCatalog.register(name, lambda: load_voc_instances(dirname, split, class_names)) + MetadataCatalog.get(name).set( + thing_classes=list(class_names), dirname=dirname, year=year, split=split + ) diff --git a/detectron2/data/datasets/register_coco.py b/detectron2/data/datasets/register_coco.py new file mode 100644 index 0000000000000000000000000000000000000000..e564438d5bf016bcdbb65b4bbdc215d79f579f8a --- /dev/null +++ b/detectron2/data/datasets/register_coco.py @@ -0,0 +1,3 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .coco import register_coco_instances # noqa +from .coco_panoptic import register_coco_panoptic_separated # noqa diff --git a/detectron2/data/detection_utils.py b/detectron2/data/detection_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..cffcab4856aec210f54b994d05802f8aaf6040eb --- /dev/null +++ b/detectron2/data/detection_utils.py @@ -0,0 +1,615 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +""" +Common data processing utilities that are used in a +typical object detection data pipeline. +""" +import logging +import numpy as np +from typing import List, Union +import pycocotools.mask as mask_util +import torch +from PIL import Image + +from detectron2.structures import ( + BitMasks, + Boxes, + BoxMode, + Instances, + Keypoints, + PolygonMasks, + RotatedBoxes, + polygons_to_bitmask, +) +from detectron2.utils.file_io import PathManager + +from . import transforms as T +from .catalog import MetadataCatalog + +__all__ = [ + "SizeMismatchError", + "convert_image_to_rgb", + "check_image_size", + "transform_proposals", + "transform_instance_annotations", + "annotations_to_instances", + "annotations_to_instances_rotated", + "build_augmentation", + "build_transform_gen", + "create_keypoint_hflip_indices", + "filter_empty_instances", + "read_image", +] + + +class SizeMismatchError(ValueError): + """ + When loaded image has difference width/height compared with annotation. + """ + + +# https://en.wikipedia.org/wiki/YUV#SDTV_with_BT.601 +_M_RGB2YUV = [[0.299, 0.587, 0.114], [-0.14713, -0.28886, 0.436], [0.615, -0.51499, -0.10001]] +_M_YUV2RGB = [[1.0, 0.0, 1.13983], [1.0, -0.39465, -0.58060], [1.0, 2.03211, 0.0]] + +# https://www.exiv2.org/tags.html +_EXIF_ORIENT = 274 # exif 'Orientation' tag + + +def convert_PIL_to_numpy(image, format): + """ + Convert PIL image to numpy array of target format. + + Args: + image (PIL.Image): a PIL image + format (str): the format of output image + + Returns: + (np.ndarray): also see `read_image` + """ + if format is not None: + # PIL only supports RGB, so convert to RGB and flip channels over below + conversion_format = format + if format in ["BGR", "YUV-BT.601"]: + conversion_format = "RGB" + image = image.convert(conversion_format) + image = np.asarray(image) + # PIL squeezes out the channel dimension for "L", so make it HWC + if format == "L": + image = np.expand_dims(image, -1) + + # handle formats not supported by PIL + elif format == "BGR": + # flip channels if needed + image = image[:, :, ::-1] + elif format == "YUV-BT.601": + image = image / 255.0 + image = np.dot(image, np.array(_M_RGB2YUV).T) + + return image + + +def convert_image_to_rgb(image, format): + """ + Convert an image from given format to RGB. + + Args: + image (np.ndarray or Tensor): an HWC image + format (str): the format of input image, also see `read_image` + + Returns: + (np.ndarray): (H,W,3) RGB image in 0-255 range, can be either float or uint8 + """ + if isinstance(image, torch.Tensor): + image = image.cpu().numpy() + if format == "BGR": + image = image[:, :, [2, 1, 0]] + elif format == "YUV-BT.601": + image = np.dot(image, np.array(_M_YUV2RGB).T) + image = image * 255.0 + else: + if format == "L": + image = image[:, :, 0] + image = image.astype(np.uint8) + image = np.asarray(Image.fromarray(image, mode=format).convert("RGB")) + return image + + +def _apply_exif_orientation(image): + """ + Applies the exif orientation correctly. + + This code exists per the bug: + https://github.com/python-pillow/Pillow/issues/3973 + with the function `ImageOps.exif_transpose`. The Pillow source raises errors with + various methods, especially `tobytes` + + Function based on: + https://github.com/wkentaro/labelme/blob/v4.5.4/labelme/utils/image.py#L59 + https://github.com/python-pillow/Pillow/blob/7.1.2/src/PIL/ImageOps.py#L527 + + Args: + image (PIL.Image): a PIL image + + Returns: + (PIL.Image): the PIL image with exif orientation applied, if applicable + """ + if not hasattr(image, "getexif"): + return image + + try: + exif = image.getexif() + except Exception: # https://github.com/facebookresearch/detectron2/issues/1885 + exif = None + + if exif is None: + return image + + orientation = exif.get(_EXIF_ORIENT) + + method = { + 2: Image.FLIP_LEFT_RIGHT, + 3: Image.ROTATE_180, + 4: Image.FLIP_TOP_BOTTOM, + 5: Image.TRANSPOSE, + 6: Image.ROTATE_270, + 7: Image.TRANSVERSE, + 8: Image.ROTATE_90, + }.get(orientation) + + if method is not None: + return image.transpose(method) + return image + + +def read_image(file_name, format=None): + """ + Read an image into the given format. + Will apply rotation and flipping if the image has such exif information. + + Args: + file_name (str): image file path + format (str): one of the supported image modes in PIL, or "BGR" or "YUV-BT.601". + + Returns: + image (np.ndarray): + an HWC image in the given format, which is 0-255, uint8 for + supported image modes in PIL or "BGR"; float (0-1 for Y) for YUV-BT.601. + """ + with PathManager.open(file_name, "rb") as f: + image = Image.open(f) + + # work around this bug: https://github.com/python-pillow/Pillow/issues/3973 + image = _apply_exif_orientation(image) + return convert_PIL_to_numpy(image, format) + + +def check_image_size(dataset_dict, image): + """ + Raise an error if the image does not match the size specified in the dict. + """ + if "width" in dataset_dict or "height" in dataset_dict: + image_wh = (image.shape[1], image.shape[0]) + expected_wh = (dataset_dict["width"], dataset_dict["height"]) + if not image_wh == expected_wh: + raise SizeMismatchError( + "Mismatched image shape{}, got {}, expect {}.".format( + " for image " + dataset_dict["file_name"] + if "file_name" in dataset_dict + else "", + image_wh, + expected_wh, + ) + + " Please check the width/height in your annotation." + ) + + # To ensure bbox always remap to original image size + if "width" not in dataset_dict: + dataset_dict["width"] = image.shape[1] + if "height" not in dataset_dict: + dataset_dict["height"] = image.shape[0] + + +def transform_proposals(dataset_dict, image_shape, transforms, *, proposal_topk, min_box_size=0): + """ + Apply transformations to the proposals in dataset_dict, if any. + + Args: + dataset_dict (dict): a dict read from the dataset, possibly + contains fields "proposal_boxes", "proposal_objectness_logits", "proposal_bbox_mode" + image_shape (tuple): height, width + transforms (TransformList): + proposal_topk (int): only keep top-K scoring proposals + min_box_size (int): proposals with either side smaller than this + threshold are removed + + The input dict is modified in-place, with abovementioned keys removed. A new + key "proposals" will be added. Its value is an `Instances` + object which contains the transformed proposals in its field + "proposal_boxes" and "objectness_logits". + """ + if "proposal_boxes" in dataset_dict: + # Transform proposal boxes + boxes = transforms.apply_box( + BoxMode.convert( + dataset_dict.pop("proposal_boxes"), + dataset_dict.pop("proposal_bbox_mode"), + BoxMode.XYXY_ABS, + ) + ) + boxes = Boxes(boxes) + objectness_logits = torch.as_tensor( + dataset_dict.pop("proposal_objectness_logits").astype("float32") + ) + + boxes.clip(image_shape) + keep = boxes.nonempty(threshold=min_box_size) + boxes = boxes[keep] + objectness_logits = objectness_logits[keep] + + proposals = Instances(image_shape) + proposals.proposal_boxes = boxes[:proposal_topk] + proposals.objectness_logits = objectness_logits[:proposal_topk] + dataset_dict["proposals"] = proposals + + +def transform_instance_annotations( + annotation, transforms, image_size, *, keypoint_hflip_indices=None +): + """ + Apply transforms to box, segmentation and keypoints annotations of a single instance. + + It will use `transforms.apply_box` for the box, and + `transforms.apply_coords` for segmentation polygons & keypoints. + If you need anything more specially designed for each data structure, + you'll need to implement your own version of this function or the transforms. + + Args: + annotation (dict): dict of instance annotations for a single instance. + It will be modified in-place. + transforms (TransformList or list[Transform]): + image_size (tuple): the height, width of the transformed image + keypoint_hflip_indices (ndarray[int]): see `create_keypoint_hflip_indices`. + + Returns: + dict: + the same input dict with fields "bbox", "segmentation", "keypoints" + transformed according to `transforms`. + The "bbox_mode" field will be set to XYXY_ABS. + """ + # if isinstance(transforms, (tuple, list)): + # transforms = T.TransformList(transforms) + # # bbox is 1d (per-instance bounding box) + # bbox = BoxMode.convert(annotation["bbox"], annotation["bbox_mode"], BoxMode.XYXY_ABS) + # # clip transformed bbox to image size + # #bbox = transforms.apply_box(np.array([bbox]))[0].clip(min=0) + # annotation["bbox"] = np.minimum(bbox, list(image_size + image_size)[::-1]) + # annotation["bbox_mode"] = BoxMode.XYXY_ABS + # return annotation + + if isinstance(transforms, (tuple, list)): + transforms = T.TransformList(transforms) + # bbox is 1d (per-instance bounding box) + bbox = BoxMode.convert(annotation["bbox"], annotation["bbox_mode"], BoxMode.XYXY_ABS) + # clip transformed bbox to image size + bbox = transforms.apply_box(np.array([bbox]))[0].clip(min=0) + annotation["bbox"] = np.minimum(bbox, list(image_size + image_size)[::-1]) + annotation["bbox_mode"] = BoxMode.XYXY_ABS + + if "segmentation" in annotation: + # each instance contains 1 or more polygons + segm = annotation["segmentation"] + if isinstance(segm, list): + # polygons + polygons = [np.asarray(p).reshape(-1, 2) for p in segm] + annotation["segmentation"] = [ + p.reshape(-1) for p in transforms.apply_polygons(polygons) + ] + elif isinstance(segm, dict): + # RLE + mask = mask_util.decode(segm) + mask = transforms.apply_segmentation(mask) + assert tuple(mask.shape[:2]) == image_size + annotation["segmentation"] = mask + else: + raise ValueError( + "Cannot transform segmentation of type '{}'!" + "Supported types are: polygons as list[list[float] or ndarray]," + " COCO-style RLE as a dict.".format(type(segm)) + ) + + if "keypoints" in annotation: + keypoints = transform_keypoint_annotations( + annotation["keypoints"], transforms, image_size, keypoint_hflip_indices + ) + annotation["keypoints"] = keypoints + + return annotation + + +def transform_keypoint_annotations(keypoints, transforms, image_size, keypoint_hflip_indices=None): + """ + Transform keypoint annotations of an image. + If a keypoint is transformed out of image boundary, it will be marked "unlabeled" (visibility=0) + + Args: + keypoints (list[float]): Nx3 float in Detectron2's Dataset format. + Each point is represented by (x, y, visibility). + transforms (TransformList): + image_size (tuple): the height, width of the transformed image + keypoint_hflip_indices (ndarray[int]): see `create_keypoint_hflip_indices`. + When `transforms` includes horizontal flip, will use the index + mapping to flip keypoints. + """ + # (N*3,) -> (N, 3) + keypoints = np.asarray(keypoints, dtype="float64").reshape(-1, 3) + keypoints_xy = transforms.apply_coords(keypoints[:, :2]) + + # Set all out-of-boundary points to "unlabeled" + inside = (keypoints_xy >= np.array([0, 0])) & (keypoints_xy <= np.array(image_size[::-1])) + inside = inside.all(axis=1) + keypoints[:, :2] = keypoints_xy + keypoints[:, 2][~inside] = 0 + + # This assumes that HorizFlipTransform is the only one that does flip + do_hflip = sum(isinstance(t, T.HFlipTransform) for t in transforms.transforms) % 2 == 1 + + # Alternative way: check if probe points was horizontally flipped. + # probe = np.asarray([[0.0, 0.0], [image_width, 0.0]]) + # probe_aug = transforms.apply_coords(probe.copy()) + # do_hflip = np.sign(probe[1][0] - probe[0][0]) != np.sign(probe_aug[1][0] - probe_aug[0][0]) # noqa + + # If flipped, swap each keypoint with its opposite-handed equivalent + if do_hflip: + assert keypoint_hflip_indices is not None + keypoints = keypoints[np.asarray(keypoint_hflip_indices, dtype=np.int32), :] + + # Maintain COCO convention that if visibility == 0 (unlabeled), then x, y = 0 + keypoints[keypoints[:, 2] == 0] = 0 + return keypoints + + +def annotations_to_instances(annos, image_size, mask_format="polygon"): + """ + Create an :class:`Instances` object used by the models, + from instance annotations in the dataset dict. + + Args: + annos (list[dict]): a list of instance annotations in one image, each + element for one instance. + image_size (tuple): height, width + + Returns: + Instances: + It will contain fields "gt_boxes", "gt_classes", + "gt_masks", "gt_keypoints", if they can be obtained from `annos`. + This is the format that builtin models expect. + """ + boxes = [BoxMode.convert(obj["bbox"], obj["bbox_mode"], BoxMode.XYXY_ABS) for obj in annos] + target = Instances(image_size) + target.gt_boxes = Boxes(boxes) + + classes = [int(obj["category_id"]) for obj in annos] + classes = torch.tensor(classes, dtype=torch.int64) + target.gt_classes = classes + + if len(annos) and "segmentation" in annos[0]: + segms = [obj["segmentation"] for obj in annos] + if mask_format == "polygon": + try: + masks = PolygonMasks(segms) + except ValueError as e: + raise ValueError( + "Failed to use mask_format=='polygon' from the given annotations!" + ) from e + else: + assert mask_format == "bitmask", mask_format + masks = [] + for segm in segms: + if isinstance(segm, list): + # polygon + masks.append(polygons_to_bitmask(segm, *image_size)) + elif isinstance(segm, dict): + # COCO RLE + masks.append(mask_util.decode(segm)) + elif isinstance(segm, np.ndarray): + assert segm.ndim == 2, "Expect segmentation of 2 dimensions, got {}.".format( + segm.ndim + ) + # mask array + masks.append(segm) + else: + raise ValueError( + "Cannot convert segmentation of type '{}' to BitMasks!" + "Supported types are: polygons as list[list[float] or ndarray]," + " COCO-style RLE as a dict, or a binary segmentation mask " + " in a 2D numpy array of shape HxW.".format(type(segm)) + ) + # torch.from_numpy does not support array with negative stride. + masks = BitMasks( + torch.stack([torch.from_numpy(np.ascontiguousarray(x)) for x in masks]) + ) + target.gt_masks = masks + + if len(annos) and "keypoints" in annos[0]: + kpts = [obj.get("keypoints", []) for obj in annos] + target.gt_keypoints = Keypoints(kpts) + + return target + + +def annotations_to_instances_rotated(annos, image_size): + """ + Create an :class:`Instances` object used by the models, + from instance annotations in the dataset dict. + Compared to `annotations_to_instances`, this function is for rotated boxes only + + Args: + annos (list[dict]): a list of instance annotations in one image, each + element for one instance. + image_size (tuple): height, width + + Returns: + Instances: + Containing fields "gt_boxes", "gt_classes", + if they can be obtained from `annos`. + This is the format that builtin models expect. + """ + boxes = [obj["bbox"] for obj in annos] + target = Instances(image_size) + boxes = target.gt_boxes = RotatedBoxes(boxes) + boxes.clip(image_size) + + classes = [obj["category_id"] for obj in annos] + classes = torch.tensor(classes, dtype=torch.int64) + target.gt_classes = classes + + return target + + +def filter_empty_instances(instances, by_box=True, by_mask=True, box_threshold=1e-5): + """ + Filter out empty instances in an `Instances` object. + + Args: + instances (Instances): + by_box (bool): whether to filter out instances with empty boxes + by_mask (bool): whether to filter out instances with empty masks + box_threshold (float): minimum width and height to be considered non-empty + + Returns: + Instances: the filtered instances. + """ + assert by_box or by_mask + r = [] + if by_box: + r.append(instances.gt_boxes.nonempty(threshold=box_threshold)) + if instances.has("gt_masks") and by_mask: + r.append(instances.gt_masks.nonempty()) + + # TODO: can also filter visible keypoints + + if not r: + return instances + m = r[0] + for x in r[1:]: + m = m & x + return instances[m] + + +def create_keypoint_hflip_indices(dataset_names: Union[str, List[str]]) -> List[int]: + """ + Args: + dataset_names: list of dataset names + + Returns: + list[int]: a list of size=#keypoints, storing the + horizontally-flipped keypoint indices. + """ + if isinstance(dataset_names, str): + dataset_names = [dataset_names] + + check_metadata_consistency("keypoint_names", dataset_names) + check_metadata_consistency("keypoint_flip_map", dataset_names) + + meta = MetadataCatalog.get(dataset_names[0]) + names = meta.keypoint_names + # TODO flip -> hflip + flip_map = dict(meta.keypoint_flip_map) + flip_map.update({v: k for k, v in flip_map.items()}) + flipped_names = [i if i not in flip_map else flip_map[i] for i in names] + flip_indices = [names.index(i) for i in flipped_names] + return flip_indices + + +def gen_crop_transform_with_instance(crop_size, image_size, instance): + """ + Generate a CropTransform so that the cropping region contains + the center of the given instance. + + Args: + crop_size (tuple): h, w in pixels + image_size (tuple): h, w + instance (dict): an annotation dict of one instance, in Detectron2's + dataset format. + """ + crop_size = np.asarray(crop_size, dtype=np.int32) + bbox = BoxMode.convert(instance["bbox"], instance["bbox_mode"], BoxMode.XYXY_ABS) + center_yx = (bbox[1] + bbox[3]) * 0.5, (bbox[0] + bbox[2]) * 0.5 + assert ( + image_size[0] >= center_yx[0] and image_size[1] >= center_yx[1] + ), "The annotation bounding box is outside of the image!" + assert ( + image_size[0] >= crop_size[0] and image_size[1] >= crop_size[1] + ), "Crop size is larger than image size!" + + min_yx = np.maximum(np.floor(center_yx).astype(np.int32) - crop_size, 0) + max_yx = np.maximum(np.asarray(image_size, dtype=np.int32) - crop_size, 0) + max_yx = np.minimum(max_yx, np.ceil(center_yx).astype(np.int32)) + + y0 = np.random.randint(min_yx[0], max_yx[0] + 1) + x0 = np.random.randint(min_yx[1], max_yx[1] + 1) + return T.CropTransform(x0, y0, crop_size[1], crop_size[0]) + + +def check_metadata_consistency(key, dataset_names): + """ + Check that the datasets have consistent metadata. + + Args: + key (str): a metadata key + dataset_names (list[str]): a list of dataset names + + Raises: + AttributeError: if the key does not exist in the metadata + ValueError: if the given datasets do not have the same metadata values defined by key + """ + if len(dataset_names) == 0: + return + logger = logging.getLogger(__name__) + entries_per_dataset = [getattr(MetadataCatalog.get(d), key) for d in dataset_names] + for idx, entry in enumerate(entries_per_dataset): + if entry != entries_per_dataset[0]: + logger.error( + "Metadata '{}' for dataset '{}' is '{}'".format(key, dataset_names[idx], str(entry)) + ) + logger.error( + "Metadata '{}' for dataset '{}' is '{}'".format( + key, dataset_names[0], str(entries_per_dataset[0]) + ) + ) + raise ValueError("Datasets have different metadata '{}'!".format(key)) + + +def build_augmentation(cfg, is_train): + """ + Create a list of default :class:`Augmentation` from config. + Now it includes resizing and flipping. + + Returns: + list[Augmentation] + """ + if is_train: + min_size = cfg.INPUT.MIN_SIZE_TRAIN + max_size = cfg.INPUT.MAX_SIZE_TRAIN + sample_style = cfg.INPUT.MIN_SIZE_TRAIN_SAMPLING + else: + min_size = cfg.INPUT.MIN_SIZE_TEST + max_size = cfg.INPUT.MAX_SIZE_TEST + sample_style = "choice" + augmentation = [T.ResizeShortestEdge(min_size, max_size, sample_style)] + if is_train and cfg.INPUT.RANDOM_FLIP != "none": + augmentation.append( + T.RandomFlip( + horizontal=cfg.INPUT.RANDOM_FLIP == "horizontal", + vertical=cfg.INPUT.RANDOM_FLIP == "vertical", + ) + ) + return augmentation + + +build_transform_gen = build_augmentation +""" +Alias for backward-compatibility. +""" diff --git a/detectron2/data/samplers/__init__.py b/detectron2/data/samplers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..4bacd895756cedbc9b37fe24af6dbcd8a054246b --- /dev/null +++ b/detectron2/data/samplers/__init__.py @@ -0,0 +1,10 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .distributed_sampler import InferenceSampler, RepeatFactorTrainingSampler, TrainingSampler +from .grouped_batch_sampler import GroupedBatchSampler + +__all__ = [ + "GroupedBatchSampler", + "TrainingSampler", + "InferenceSampler", + "RepeatFactorTrainingSampler", +] diff --git a/detectron2/data/samplers/__pycache__/__init__.cpython-39.pyc b/detectron2/data/samplers/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ef7c161cece8659118e1ea1e506ba0c137502568 Binary files /dev/null and b/detectron2/data/samplers/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/data/samplers/__pycache__/distributed_sampler.cpython-39.pyc b/detectron2/data/samplers/__pycache__/distributed_sampler.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a79b0257434eebeff4346510cd5f29c87abf5630 Binary files /dev/null and b/detectron2/data/samplers/__pycache__/distributed_sampler.cpython-39.pyc differ diff --git a/detectron2/data/samplers/__pycache__/grouped_batch_sampler.cpython-39.pyc b/detectron2/data/samplers/__pycache__/grouped_batch_sampler.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..805d90b9c3100443974deeecbc378299c5e22a71 Binary files /dev/null and b/detectron2/data/samplers/__pycache__/grouped_batch_sampler.cpython-39.pyc differ diff --git a/detectron2/data/samplers/distributed_sampler.py b/detectron2/data/samplers/distributed_sampler.py new file mode 100644 index 0000000000000000000000000000000000000000..f0e8da28822ec071a04caaac2069e6148bc90622 --- /dev/null +++ b/detectron2/data/samplers/distributed_sampler.py @@ -0,0 +1,200 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import itertools +import math +from collections import defaultdict +from typing import Optional +import torch +from torch.utils.data.sampler import Sampler + +from detectron2.utils import comm + + +class TrainingSampler(Sampler): + """ + In training, we only care about the "infinite stream" of training data. + So this sampler produces an infinite stream of indices and + all workers cooperate to correctly shuffle the indices and sample different indices. + + The samplers in each worker effectively produces `indices[worker_id::num_workers]` + where `indices` is an infinite stream of indices consisting of + `shuffle(range(size)) + shuffle(range(size)) + ...` (if shuffle is True) + or `range(size) + range(size) + ...` (if shuffle is False) + """ + + def __init__(self, size: int, shuffle: bool = True, seed: Optional[int] = None): + """ + Args: + size (int): the total number of data of the underlying dataset to sample from + shuffle (bool): whether to shuffle the indices or not + seed (int): the initial seed of the shuffle. Must be the same + across all workers. If None, will use a random seed shared + among workers (require synchronization among all workers). + """ + self._size = size + assert size > 0 + self._shuffle = shuffle + if seed is None: + seed = comm.shared_random_seed() + self._seed = int(seed) + + self._rank = comm.get_rank() + self._world_size = comm.get_world_size() + + def __iter__(self): + start = self._rank + yield from itertools.islice(self._infinite_indices(), start, None, self._world_size) + + def _infinite_indices(self): + g = torch.Generator() + g.manual_seed(self._seed) + while True: + if self._shuffle: + yield from torch.randperm(self._size, generator=g).tolist() + else: + yield from torch.arange(self._size).tolist() + + +class RepeatFactorTrainingSampler(Sampler): + """ + Similar to TrainingSampler, but a sample may appear more times than others based + on its "repeat factor". This is suitable for training on class imbalanced datasets like LVIS. + """ + + def __init__(self, repeat_factors, *, shuffle=True, seed=None): + """ + Args: + repeat_factors (Tensor): a float vector, the repeat factor for each indice. When it's + full of ones, it is equivalent to ``TrainingSampler(len(repeat_factors), ...)``. + shuffle (bool): whether to shuffle the indices or not + seed (int): the initial seed of the shuffle. Must be the same + across all workers. If None, will use a random seed shared + among workers (require synchronization among all workers). + """ + self._shuffle = shuffle + if seed is None: + seed = comm.shared_random_seed() + self._seed = int(seed) + + self._rank = comm.get_rank() + self._world_size = comm.get_world_size() + + # Split into whole number (_int_part) and fractional (_frac_part) parts. + self._int_part = torch.trunc(repeat_factors) + self._frac_part = repeat_factors - self._int_part + + @staticmethod + def repeat_factors_from_category_frequency(dataset_dicts, repeat_thresh): + """ + Compute (fractional) per-image repeat factors based on category frequency. + The repeat factor for an image is a function of the frequency of the rarest + category labeled in that image. The "frequency of category c" in [0, 1] is defined + as the fraction of images in the training set (without repeats) in which category c + appears. + See :paper:`lvis` (>= v2) Appendix B.2. + + Args: + dataset_dicts (list[dict]): annotations in Detectron2 dataset format. + repeat_thresh (float): frequency threshold below which data is repeated. + If the frequency is half of `repeat_thresh`, the image will be + repeated twice. + + Returns: + torch.Tensor: + the i-th element is the repeat factor for the dataset image at index i. + """ + # 1. For each category c, compute the fraction of images that contain it: f(c) + category_freq = defaultdict(int) + for dataset_dict in dataset_dicts: # For each image (without repeats) + cat_ids = {ann["category_id"] for ann in dataset_dict["annotations"]} + for cat_id in cat_ids: + category_freq[cat_id] += 1 + num_images = len(dataset_dicts) + for k, v in category_freq.items(): + category_freq[k] = v / num_images + + # 2. For each category c, compute the category-level repeat factor: + # r(c) = max(1, sqrt(t / f(c))) + category_rep = { + cat_id: max(1.0, math.sqrt(repeat_thresh / cat_freq)) + for cat_id, cat_freq in category_freq.items() + } + + # 3. For each image I, compute the image-level repeat factor: + # r(I) = max_{c in I} r(c) + rep_factors = [] + for dataset_dict in dataset_dicts: + cat_ids = {ann["category_id"] for ann in dataset_dict["annotations"]} + rep_factor = max({category_rep[cat_id] for cat_id in cat_ids}, default=1.0) + rep_factors.append(rep_factor) + + return torch.tensor(rep_factors, dtype=torch.float32) + + def _get_epoch_indices(self, generator): + """ + Create a list of dataset indices (with repeats) to use for one epoch. + + Args: + generator (torch.Generator): pseudo random number generator used for + stochastic rounding. + + Returns: + torch.Tensor: list of dataset indices to use in one epoch. Each index + is repeated based on its calculated repeat factor. + """ + # Since repeat factors are fractional, we use stochastic rounding so + # that the target repeat factor is achieved in expectation over the + # course of training + rands = torch.rand(len(self._frac_part), generator=generator) + rep_factors = self._int_part + (rands < self._frac_part).float() + # Construct a list of indices in which we repeat images as specified + indices = [] + for dataset_index, rep_factor in enumerate(rep_factors): + indices.extend([dataset_index] * int(rep_factor.item())) + return torch.tensor(indices, dtype=torch.int64) + + def __iter__(self): + start = self._rank + yield from itertools.islice(self._infinite_indices(), start, None, self._world_size) + + def _infinite_indices(self): + g = torch.Generator() + g.manual_seed(self._seed) + while True: + # Sample indices with repeats determined by stochastic rounding; each + # "epoch" may have a slightly different size due to the rounding. + indices = self._get_epoch_indices(g) + if self._shuffle: + randperm = torch.randperm(len(indices), generator=g) + yield from indices[randperm].tolist() + else: + yield from indices.tolist() + + +class InferenceSampler(Sampler): + """ + Produce indices for inference across all workers. + Inference needs to run on the __exact__ set of samples, + therefore when the total number of samples is not divisible by the number of workers, + this sampler produces different number of samples on different workers. + """ + + def __init__(self, size: int): + """ + Args: + size (int): the total number of data of the underlying dataset to sample from + """ + self._size = size + assert size > 0 + self._rank = comm.get_rank() + self._world_size = comm.get_world_size() + + shard_size = (self._size - 1) // self._world_size + 1 + begin = shard_size * self._rank + end = min(shard_size * (self._rank + 1), self._size) + self._local_indices = range(begin, end) + + def __iter__(self): + yield from self._local_indices + + def __len__(self): + return len(self._local_indices) diff --git a/detectron2/data/samplers/grouped_batch_sampler.py b/detectron2/data/samplers/grouped_batch_sampler.py new file mode 100644 index 0000000000000000000000000000000000000000..5b247730aacd04dd0c752664acde3257c4eddd71 --- /dev/null +++ b/detectron2/data/samplers/grouped_batch_sampler.py @@ -0,0 +1,47 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import numpy as np +from torch.utils.data.sampler import BatchSampler, Sampler + + +class GroupedBatchSampler(BatchSampler): + """ + Wraps another sampler to yield a mini-batch of indices. + It enforces that the batch only contain elements from the same group. + It also tries to provide mini-batches which follows an ordering which is + as close as possible to the ordering from the original sampler. + """ + + def __init__(self, sampler, group_ids, batch_size): + """ + Args: + sampler (Sampler): Base sampler. + group_ids (list[int]): If the sampler produces indices in range [0, N), + `group_ids` must be a list of `N` ints which contains the group id of each sample. + The group ids must be a set of integers in the range [0, num_groups). + batch_size (int): Size of mini-batch. + """ + if not isinstance(sampler, Sampler): + raise ValueError( + "sampler should be an instance of " + "torch.utils.data.Sampler, but got sampler={}".format(sampler) + ) + self.sampler = sampler + self.group_ids = np.asarray(group_ids) + assert self.group_ids.ndim == 1 + self.batch_size = batch_size + groups = np.unique(self.group_ids).tolist() + + # buffer the indices of each group until batch size is reached + self.buffer_per_group = {k: [] for k in groups} + + def __iter__(self): + for idx in self.sampler: + group_id = self.group_ids[idx] + group_buffer = self.buffer_per_group[group_id] + group_buffer.append(idx) + if len(group_buffer) == self.batch_size: + yield group_buffer[:] # yield a copy of the list + del group_buffer[:] + + def __len__(self): + raise NotImplementedError("len() of GroupedBatchSampler is not well-defined.") diff --git a/detectron2/data/transforms/__init__.py b/detectron2/data/transforms/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ab3c63b5b456a7fb878757e25768a3634f76ae5b --- /dev/null +++ b/detectron2/data/transforms/__init__.py @@ -0,0 +1,14 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from fvcore.transforms.transform import Transform, TransformList # order them first +from fvcore.transforms.transform import * +from .transform import * +from .augmentation import * +from .augmentation_impl import * + +__all__ = [k for k in globals().keys() if not k.startswith("_")] + + +from detectron2.utils.env import fixup_module_metadata + +fixup_module_metadata(__name__, globals(), __all__) +del fixup_module_metadata diff --git a/detectron2/data/transforms/__pycache__/__init__.cpython-39.pyc b/detectron2/data/transforms/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2c03f17b1e17062d42be2fc7466140fd9475cd96 Binary files /dev/null and b/detectron2/data/transforms/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/data/transforms/__pycache__/augmentation.cpython-39.pyc b/detectron2/data/transforms/__pycache__/augmentation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d44f4edcce9e74e9dfd5c8116dab6bc827383c96 Binary files /dev/null and b/detectron2/data/transforms/__pycache__/augmentation.cpython-39.pyc differ diff --git a/detectron2/data/transforms/__pycache__/augmentation_impl.cpython-39.pyc b/detectron2/data/transforms/__pycache__/augmentation_impl.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6fb91145d6ae74cd35925b0a15bbc42117b3f54c Binary files /dev/null and b/detectron2/data/transforms/__pycache__/augmentation_impl.cpython-39.pyc differ diff --git a/detectron2/data/transforms/__pycache__/build.cpython-39.pyc b/detectron2/data/transforms/__pycache__/build.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..216f7f527df24f33c1ffdbcbcd58fad048c54a8c Binary files /dev/null and b/detectron2/data/transforms/__pycache__/build.cpython-39.pyc differ diff --git a/detectron2/data/transforms/__pycache__/transform.cpython-39.pyc b/detectron2/data/transforms/__pycache__/transform.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2a05885dec30e0e2027e11b0d9b19d3a4fd93f76 Binary files /dev/null and b/detectron2/data/transforms/__pycache__/transform.cpython-39.pyc differ diff --git a/detectron2/data/transforms/augmentation.py b/detectron2/data/transforms/augmentation.py new file mode 100644 index 0000000000000000000000000000000000000000..917290bf634b4bbd21fd70fbc14dabd6705fea33 --- /dev/null +++ b/detectron2/data/transforms/augmentation.py @@ -0,0 +1,377 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import inspect +import numpy as np +import pprint +from typing import Any, List, Optional, Tuple, Union +from fvcore.transforms.transform import Transform, TransformList + +""" +See "Data Augmentation" tutorial for an overview of the system: +https://detectron2.readthedocs.io/tutorials/augmentation.html +""" + + +__all__ = [ + "Augmentation", + "AugmentationList", + "AugInput", + "TransformGen", + "apply_transform_gens", + "StandardAugInput", + "apply_augmentations", +] + + +def _check_img_dtype(img): + assert isinstance(img, np.ndarray), "[Augmentation] Needs an numpy array, but got a {}!".format( + type(img) + ) + assert not isinstance(img.dtype, np.integer) or ( + img.dtype == np.uint8 + ), "[Augmentation] Got image of type {}, use uint8 or floating points instead!".format( + img.dtype + ) + assert img.ndim in [2, 3], img.ndim + + +def _get_aug_input_args(aug, aug_input) -> List[Any]: + """ + Get the arguments to be passed to ``aug.get_transform`` from the input ``aug_input``. + """ + if aug.input_args is None: + # Decide what attributes are needed automatically + prms = list(inspect.signature(aug.get_transform).parameters.items()) + # The default behavior is: if there is one parameter, then its "image" + # (work automatically for majority of use cases, and also avoid BC breaking), + # Otherwise, use the argument names. + if len(prms) == 1: + names = ("image",) + else: + names = [] + for name, prm in prms: + if prm.kind in (inspect.Parameter.VAR_POSITIONAL, inspect.Parameter.VAR_KEYWORD): + raise TypeError( + f""" \ +The default implementation of `{type(aug)}.__call__` does not allow \ +`{type(aug)}.get_transform` to use variable-length arguments (*args, **kwargs)! \ +If arguments are unknown, reimplement `__call__` instead. \ +""" + ) + names.append(name) + aug.input_args = tuple(names) + + args = [] + for f in aug.input_args: + try: + args.append(getattr(aug_input, f)) + except AttributeError as e: + raise AttributeError( + f"{type(aug)}.get_transform needs input attribute '{f}', " + f"but it is not an attribute of {type(aug_input)}!" + ) from e + return args + + +class Augmentation: + """ + Augmentation defines (often random) policies/strategies to generate :class:`Transform` + from data. It is often used for pre-processing of input data. + + A "policy" that generates a :class:`Transform` may, in the most general case, + need arbitrary information from input data in order to determine what transforms + to apply. Therefore, each :class:`Augmentation` instance defines the arguments + needed by its :meth:`get_transform` method. When called with the positional arguments, + the :meth:`get_transform` method executes the policy. + + Note that :class:`Augmentation` defines the policies to create a :class:`Transform`, + but not how to execute the actual transform operations to those data. + Its :meth:`__call__` method will use :meth:`AugInput.transform` to execute the transform. + + The returned `Transform` object is meant to describe deterministic transformation, which means + it can be re-applied on associated data, e.g. the geometry of an image and its segmentation + masks need to be transformed together. + (If such re-application is not needed, then determinism is not a crucial requirement.) + """ + + input_args: Optional[Tuple[str]] = None + """ + Stores the attribute names needed by :meth:`get_transform`, e.g. ``("image", "sem_seg")``. + By default, it is just a tuple of argument names in :meth:`self.get_transform`, which often only + contain "image". As long as the argument name convention is followed, there is no need for + users to touch this attribute. + """ + + def _init(self, params=None): + if params: + for k, v in params.items(): + if k != "self" and not k.startswith("_"): + setattr(self, k, v) + + def get_transform(self, *args) -> Transform: + """ + Execute the policy based on input data, and decide what transform to apply to inputs. + + Args: + args: Any fixed-length positional arguments. By default, the name of the arguments + should exist in the :class:`AugInput` to be used. + + Returns: + Transform: Returns the deterministic transform to apply to the input. + + Examples: + :: + class MyAug: + # if a policy needs to know both image and semantic segmentation + def get_transform(image, sem_seg) -> T.Transform: + pass + tfm: Transform = MyAug().get_transform(image, sem_seg) + new_image = tfm.apply_image(image) + + Notes: + Users can freely use arbitrary new argument names in custom + :meth:`get_transform` method, as long as they are available in the + input data. In detectron2 we use the following convention: + + * image: (H,W) or (H,W,C) ndarray of type uint8 in range [0, 255], or + floating point in range [0, 1] or [0, 255]. + * boxes: (N,4) ndarray of float32. It represents the instance bounding boxes + of N instances. Each is in XYXY format in unit of absolute coordinates. + * sem_seg: (H,W) ndarray of type uint8. Each element is an integer label of pixel. + + We do not specify convention for other types and do not include builtin + :class:`Augmentation` that uses other types in detectron2. + """ + raise NotImplementedError + + def __call__(self, aug_input) -> Transform: + """ + Augment the given `aug_input` **in-place**, and return the transform that's used. + + This method will be called to apply the augmentation. In most augmentation, it + is enough to use the default implementation, which calls :meth:`get_transform` + using the inputs. But a subclass can overwrite it to have more complicated logic. + + Args: + aug_input (AugInput): an object that has attributes needed by this augmentation + (defined by ``self.get_transform``). Its ``transform`` method will be called + to in-place transform it. + + Returns: + Transform: the transform that is applied on the input. + """ + args = _get_aug_input_args(self, aug_input) + tfm = self.get_transform(*args) + assert isinstance(tfm, (Transform, TransformList)), ( + f"{type(self)}.get_transform must return an instance of Transform! " + "Got {type(tfm)} instead." + ) + aug_input.transform(tfm) + return tfm + + def _rand_range(self, low=1.0, high=None, size=None): + """ + Uniform float random number between low and high. + """ + if high is None: + low, high = 0, low + if size is None: + size = [] + return np.random.uniform(low, high, size) + + def __repr__(self): + """ + Produce something like: + "MyAugmentation(field1={self.field1}, field2={self.field2})" + """ + try: + sig = inspect.signature(self.__init__) + classname = type(self).__name__ + argstr = [] + for name, param in sig.parameters.items(): + assert ( + param.kind != param.VAR_POSITIONAL and param.kind != param.VAR_KEYWORD + ), "The default __repr__ doesn't support *args or **kwargs" + assert hasattr(self, name), ( + "Attribute {} not found! " + "Default __repr__ only works if attributes match the constructor.".format(name) + ) + attr = getattr(self, name) + default = param.default + if default is attr: + continue + attr_str = pprint.pformat(attr) + if "\n" in attr_str: + # don't show it if pformat decides to use >1 lines + attr_str = "..." + argstr.append("{}={}".format(name, attr_str)) + return "{}({})".format(classname, ", ".join(argstr)) + except AssertionError: + return super().__repr__() + + __str__ = __repr__ + + +def _transform_to_aug(tfm_or_aug): + """ + Wrap Transform into Augmentation. + Private, used internally to implement augmentations. + """ + assert isinstance(tfm_or_aug, (Transform, Augmentation)), tfm_or_aug + if isinstance(tfm_or_aug, Augmentation): + return tfm_or_aug + else: + + class _TransformToAug(Augmentation): + def __init__(self, tfm: Transform): + self.tfm = tfm + + def get_transform(self, *args): + return self.tfm + + def __repr__(self): + return repr(self.tfm) + + __str__ = __repr__ + + return _TransformToAug(tfm_or_aug) + + +class AugmentationList(Augmentation): + """ + Apply a sequence of augmentations. + + It has ``__call__`` method to apply the augmentations. + + Note that :meth:`get_transform` method is impossible (will throw error if called) + for :class:`AugmentationList`, because in order to apply a sequence of augmentations, + the kth augmentation must be applied first, to provide inputs needed by the (k+1)th + augmentation. + """ + + def __init__(self, augs): + """ + Args: + augs (list[Augmentation or Transform]): + """ + super().__init__() + self.augs = [_transform_to_aug(x) for x in augs] + + def __call__(self, aug_input) -> Transform: + tfms = [] + for x in self.augs: + tfm = x(aug_input) + tfms.append(tfm) + return TransformList(tfms) + + def __repr__(self): + msgs = [str(x) for x in self.augs] + return "AugmentationList[{}]".format(", ".join(msgs)) + + __str__ = __repr__ + + +class AugInput: + """ + Input that can be used with :meth:`Augmentation.__call__`. + This is a standard implementation for the majority of use cases. + This class provides the standard attributes **"image", "boxes", "sem_seg"** + defined in :meth:`__init__` and they may be needed by different augmentations. + Most augmentation policies do not need attributes beyond these three. + + After applying augmentations to these attributes (using :meth:`AugInput.transform`), + the returned transforms can then be used to transform other data structures that users have. + + Examples: + :: + input = AugInput(image, boxes=boxes) + tfms = augmentation(input) + transformed_image = input.image + transformed_boxes = input.boxes + transformed_other_data = tfms.apply_other(other_data) + + An extended project that works with new data types may implement augmentation policies + that need other inputs. An algorithm may need to transform inputs in a way different + from the standard approach defined in this class. In those rare situations, users can + implement a class similar to this class, that satify the following condition: + + * The input must provide access to these data in the form of attribute access + (``getattr``). For example, if an :class:`Augmentation` to be applied needs "image" + and "sem_seg" arguments, its input must have the attribute "image" and "sem_seg". + * The input must have a ``transform(tfm: Transform) -> None`` method which + in-place transforms all its attributes. + """ + + # TODO maybe should support more builtin data types here + def __init__( + self, + image: np.ndarray, + *, + boxes: Optional[np.ndarray] = None, + sem_seg: Optional[np.ndarray] = None, + ): + """ + Args: + image (ndarray): (H,W) or (H,W,C) ndarray of type uint8 in range [0, 255], or + floating point in range [0, 1] or [0, 255]. The meaning of C is up + to users. + boxes (ndarray or None): Nx4 float32 boxes in XYXY_ABS mode + sem_seg (ndarray or None): HxW uint8 semantic segmentation mask. Each element + is an integer label of pixel. + """ + _check_img_dtype(image) + self.image = image + self.boxes = boxes + self.sem_seg = sem_seg + + def transform(self, tfm: Transform) -> None: + """ + In-place transform all attributes of this class. + + By "in-place", it means after calling this method, accessing an attribute such + as ``self.image`` will return transformed data. + """ + self.image = tfm.apply_image(self.image) + if self.boxes is not None: + self.boxes = tfm.apply_box(self.boxes) + if self.sem_seg is not None: + self.sem_seg = tfm.apply_segmentation(self.sem_seg) + + def apply_augmentations( + self, augmentations: List[Union[Augmentation, Transform]] + ) -> TransformList: + """ + Equivalent of ``AugmentationList(augmentations)(self)`` + """ + return AugmentationList(augmentations)(self) + + +def apply_augmentations(augmentations: List[Union[Transform, Augmentation]], inputs): + """ + Use ``T.AugmentationList(augmentations)(inputs)`` instead. + """ + if isinstance(inputs, np.ndarray): + # handle the common case of image-only Augmentation, also for backward compatibility + image_only = True + inputs = AugInput(inputs) + else: + image_only = False + tfms = inputs.apply_augmentations(augmentations) + return inputs.image if image_only else inputs, tfms + + +apply_transform_gens = apply_augmentations +""" +Alias for backward-compatibility. +""" + +TransformGen = Augmentation +""" +Alias for Augmentation, since it is something that generates :class:`Transform`s +""" + +StandardAugInput = AugInput +""" +Alias for compatibility. It's not worth the complexity to have two classes. +""" diff --git a/detectron2/data/transforms/augmentation_impl.py b/detectron2/data/transforms/augmentation_impl.py new file mode 100644 index 0000000000000000000000000000000000000000..db727cd246d145128d7e06ca0cd8bd776e4084e3 --- /dev/null +++ b/detectron2/data/transforms/augmentation_impl.py @@ -0,0 +1,579 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. +""" +Implement many useful :class:`Augmentation`. +""" +import numpy as np +import sys +from typing import Tuple +from fvcore.transforms.transform import ( + BlendTransform, + CropTransform, + HFlipTransform, + NoOpTransform, + PadTransform, + Transform, + TransformList, + VFlipTransform, +) +from PIL import Image + +from .augmentation import Augmentation, _transform_to_aug +from .transform import ExtentTransform, ResizeTransform, RotationTransform + +__all__ = [ + "FixedSizeCrop", + "RandomApply", + "RandomBrightness", + "RandomContrast", + "RandomCrop", + "RandomExtent", + "RandomFlip", + "RandomSaturation", + "RandomLighting", + "RandomRotation", + "Resize", + "ResizeScale", + "ResizeShortestEdge", + "RandomCrop_CategoryAreaConstraint", +] + + +class RandomApply(Augmentation): + """ + Randomly apply an augmentation with a given probability. + """ + + def __init__(self, tfm_or_aug, prob=0.5): + """ + Args: + tfm_or_aug (Transform, Augmentation): the transform or augmentation + to be applied. It can either be a `Transform` or `Augmentation` + instance. + prob (float): probability between 0.0 and 1.0 that + the wrapper transformation is applied + """ + super().__init__() + self.aug = _transform_to_aug(tfm_or_aug) + assert 0.0 <= prob <= 1.0, f"Probablity must be between 0.0 and 1.0 (given: {prob})" + self.prob = prob + + def get_transform(self, *args): + do = self._rand_range() < self.prob + if do: + return self.aug.get_transform(*args) + else: + return NoOpTransform() + + def __call__(self, aug_input): + do = self._rand_range() < self.prob + if do: + return self.aug(aug_input) + else: + return NoOpTransform() + + +class RandomFlip(Augmentation): + """ + Flip the image horizontally or vertically with the given probability. + """ + + def __init__(self, prob=0.5, *, horizontal=True, vertical=False): + """ + Args: + prob (float): probability of flip. + horizontal (boolean): whether to apply horizontal flipping + vertical (boolean): whether to apply vertical flipping + """ + super().__init__() + + if horizontal and vertical: + raise ValueError("Cannot do both horiz and vert. Please use two Flip instead.") + if not horizontal and not vertical: + raise ValueError("At least one of horiz or vert has to be True!") + self._init(locals()) + + def get_transform(self, image): + h, w = image.shape[:2] + do = self._rand_range() < self.prob + if do: + if self.horizontal: + return HFlipTransform(w) + elif self.vertical: + return VFlipTransform(h) + else: + return NoOpTransform() + + +class Resize(Augmentation): + """Resize image to a fixed target size""" + + def __init__(self, shape, interp=Image.BILINEAR): + """ + Args: + shape: (h, w) tuple or a int + interp: PIL interpolation method + """ + if isinstance(shape, int): + shape = (shape, shape) + shape = tuple(shape) + self._init(locals()) + + def get_transform(self, image): + return ResizeTransform( + image.shape[0], image.shape[1], self.shape[0], self.shape[1], self.interp + ) + + +class ResizeShortestEdge(Augmentation): + """ + Scale the shorter edge to the given size, with a limit of `max_size` on the longer edge. + If `max_size` is reached, then downscale so that the longer edge does not exceed max_size. + """ + + def __init__( + self, short_edge_length, max_size=sys.maxsize, sample_style="range", interp=Image.BILINEAR + ): + """ + Args: + short_edge_length (list[int]): If ``sample_style=="range"``, + a [min, max] interval from which to sample the shortest edge length. + If ``sample_style=="choice"``, a list of shortest edge lengths to sample from. + max_size (int): maximum allowed longest edge length. + sample_style (str): either "range" or "choice". + """ + super().__init__() + assert sample_style in ["range", "choice"], sample_style + + self.is_range = sample_style == "range" + if isinstance(short_edge_length, int): + short_edge_length = (short_edge_length, short_edge_length) + if self.is_range: + assert len(short_edge_length) == 2, ( + "short_edge_length must be two values using 'range' sample style." + f" Got {short_edge_length}!" + ) + self._init(locals()) + + def get_transform(self, image): + h, w = image.shape[:2] + if self.is_range: + size = np.random.randint(self.short_edge_length[0], self.short_edge_length[1] + 1) + else: + size = np.random.choice(self.short_edge_length) + if size == 0: + return NoOpTransform() + + scale = size * 1.0 / min(h, w) + if h < w: + newh, neww = size, scale * w + else: + newh, neww = scale * h, size + if max(newh, neww) > self.max_size: + scale = self.max_size * 1.0 / max(newh, neww) + newh = newh * scale + neww = neww * scale + neww = int(neww + 0.5) + newh = int(newh + 0.5) + return ResizeTransform(h, w, newh, neww, self.interp) + + +class ResizeScale(Augmentation): + """ + Takes target size as input and randomly scales the given target size between `min_scale` + and `max_scale`. It then scales the input image such that it fits inside the scaled target + box, keeping the aspect ratio constant. + This implements the resize part of the Google's 'resize_and_crop' data augmentation: + https://github.com/tensorflow/tpu/blob/master/models/official/detection/utils/input_utils.py#L127 + """ + + def __init__( + self, + min_scale: float, + max_scale: float, + target_height: int, + target_width: int, + interp: int = Image.BILINEAR, + ): + """ + Args: + min_scale: minimum image scale range. + max_scale: maximum image scale range. + target_height: target image height. + target_width: target image width. + interp: image interpolation method. + """ + super().__init__() + self._init(locals()) + + def get_transform(self, image: np.ndarray) -> Transform: + # Compute the image scale and scaled size. + input_size = image.shape[:2] + output_size = (self.target_height, self.target_width) + random_scale = np.random.uniform(self.min_scale, self.max_scale) + random_scale_size = np.multiply(output_size, random_scale) + scale = np.minimum( + random_scale_size[0] / input_size[0], random_scale_size[1] / input_size[1] + ) + scaled_size = np.round(np.multiply(input_size, scale)).astype(int) + return ResizeTransform( + input_size[0], input_size[1], scaled_size[0], scaled_size[1], self.interp + ) + + +class RandomRotation(Augmentation): + """ + This method returns a copy of this image, rotated the given + number of degrees counter clockwise around the given center. + """ + + def __init__(self, angle, expand=True, center=None, sample_style="range", interp=None): + """ + Args: + angle (list[float]): If ``sample_style=="range"``, + a [min, max] interval from which to sample the angle (in degrees). + If ``sample_style=="choice"``, a list of angles to sample from + expand (bool): choose if the image should be resized to fit the whole + rotated image (default), or simply cropped + center (list[[float, float]]): If ``sample_style=="range"``, + a [[minx, miny], [maxx, maxy]] relative interval from which to sample the center, + [0, 0] being the top left of the image and [1, 1] the bottom right. + If ``sample_style=="choice"``, a list of centers to sample from + Default: None, which means that the center of rotation is the center of the image + center has no effect if expand=True because it only affects shifting + """ + super().__init__() + assert sample_style in ["range", "choice"], sample_style + self.is_range = sample_style == "range" + if isinstance(angle, (float, int)): + angle = (angle, angle) + if center is not None and isinstance(center[0], (float, int)): + center = (center, center) + self._init(locals()) + + def get_transform(self, image): + h, w = image.shape[:2] + center = None + if self.is_range: + angle = np.random.uniform(self.angle[0], self.angle[1]) + if self.center is not None: + center = ( + np.random.uniform(self.center[0][0], self.center[1][0]), + np.random.uniform(self.center[0][1], self.center[1][1]), + ) + else: + angle = np.random.choice(self.angle) + if self.center is not None: + center = np.random.choice(self.center) + + if center is not None: + center = (w * center[0], h * center[1]) # Convert to absolute coordinates + + if angle % 360 == 0: + return NoOpTransform() + + return RotationTransform(h, w, angle, expand=self.expand, center=center, interp=self.interp) + + +class FixedSizeCrop(Augmentation): + """ + If `crop_size` is smaller than the input image size, then it uses a random crop of + the crop size. If `crop_size` is larger than the input image size, then it pads + the right and the bottom of the image to the crop size. + """ + + def __init__(self, crop_size: Tuple[int], pad_value: float = 128.0): + """ + Args: + crop_size: target image (height, width). + pad_value: the padding value. + """ + super().__init__() + self._init(locals()) + + def get_transform(self, image: np.ndarray) -> TransformList: + # Compute the image scale and scaled size. + input_size = image.shape[:2] + output_size = self.crop_size + + # Add random crop if the image is scaled up. + max_offset = np.subtract(input_size, output_size) + max_offset = np.maximum(max_offset, 0) + offset = np.multiply(max_offset, np.random.uniform(0.0, 1.0)) + offset = np.round(offset).astype(int) + crop_transform = CropTransform( + offset[1], offset[0], output_size[1], output_size[0], input_size[1], input_size[0] + ) + + # Add padding if the image is scaled down. + pad_size = np.subtract(output_size, input_size) + pad_size = np.maximum(pad_size, 0) + original_size = np.minimum(input_size, output_size) + pad_transform = PadTransform( + 0, 0, pad_size[1], pad_size[0], original_size[1], original_size[0], self.pad_value + ) + + return TransformList([crop_transform, pad_transform]) + + +class RandomCrop(Augmentation): + """ + Randomly crop a rectangle region out of an image. + """ + + def __init__(self, crop_type: str, crop_size): + """ + Args: + crop_type (str): one of "relative_range", "relative", "absolute", "absolute_range". + crop_size (tuple[float, float]): two floats, explained below. + + - "relative": crop a (H * crop_size[0], W * crop_size[1]) region from an input image of + size (H, W). crop size should be in (0, 1] + - "relative_range": uniformly sample two values from [crop_size[0], 1] + and [crop_size[1]], 1], and use them as in "relative" crop type. + - "absolute" crop a (crop_size[0], crop_size[1]) region from input image. + crop_size must be smaller than the input image size. + - "absolute_range", for an input of size (H, W), uniformly sample H_crop in + [crop_size[0], min(H, crop_size[1])] and W_crop in [crop_size[0], min(W, crop_size[1])]. + Then crop a region (H_crop, W_crop). + """ + # TODO style of relative_range and absolute_range are not consistent: + # one takes (h, w) but another takes (min, max) + super().__init__() + assert crop_type in ["relative_range", "relative", "absolute", "absolute_range"] + self._init(locals()) + + def get_transform(self, image): + h, w = image.shape[:2] + croph, cropw = self.get_crop_size((h, w)) + assert h >= croph and w >= cropw, "Shape computation in {} has bugs.".format(self) + h0 = np.random.randint(h - croph + 1) + w0 = np.random.randint(w - cropw + 1) + return CropTransform(w0, h0, cropw, croph) + + def get_crop_size(self, image_size): + """ + Args: + image_size (tuple): height, width + + Returns: + crop_size (tuple): height, width in absolute pixels + """ + h, w = image_size + if self.crop_type == "relative": + ch, cw = self.crop_size + return int(h * ch + 0.5), int(w * cw + 0.5) + elif self.crop_type == "relative_range": + crop_size = np.asarray(self.crop_size, dtype=np.float32) + ch, cw = crop_size + np.random.rand(2) * (1 - crop_size) + return int(h * ch + 0.5), int(w * cw + 0.5) + elif self.crop_type == "absolute": + return (min(self.crop_size[0], h), min(self.crop_size[1], w)) + elif self.crop_type == "absolute_range": + assert self.crop_size[0] <= self.crop_size[1] + ch = np.random.randint(min(h, self.crop_size[0]), min(h, self.crop_size[1]) + 1) + cw = np.random.randint(min(w, self.crop_size[0]), min(w, self.crop_size[1]) + 1) + return ch, cw + else: + NotImplementedError("Unknown crop type {}".format(self.crop_type)) + + +class RandomCrop_CategoryAreaConstraint(Augmentation): + """ + Similar to :class:`RandomCrop`, but find a cropping window such that no single category + occupies a ratio of more than `single_category_max_area` in semantic segmentation ground + truth, which can cause unstability in training. The function attempts to find such a valid + cropping window for at most 10 times. + """ + + def __init__( + self, + crop_type: str, + crop_size, + single_category_max_area: float = 1.0, + ignored_category: int = None, + ): + """ + Args: + crop_type, crop_size: same as in :class:`RandomCrop` + single_category_max_area: the maximum allowed area ratio of a + category. Set to 1.0 to disable + ignored_category: allow this category in the semantic segmentation + ground truth to exceed the area ratio. Usually set to the category + that's ignored in training. + """ + self.crop_aug = RandomCrop(crop_type, crop_size) + self._init(locals()) + + def get_transform(self, image, sem_seg): + if self.single_category_max_area >= 1.0: + return self.crop_aug.get_transform(image) + else: + h, w = sem_seg.shape + for _ in range(10): + crop_size = self.crop_aug.get_crop_size((h, w)) + y0 = np.random.randint(h - crop_size[0] + 1) + x0 = np.random.randint(w - crop_size[1] + 1) + sem_seg_temp = sem_seg[y0 : y0 + crop_size[0], x0 : x0 + crop_size[1]] + labels, cnt = np.unique(sem_seg_temp, return_counts=True) + if self.ignored_category is not None: + cnt = cnt[labels != self.ignored_category] + if len(cnt) > 1 and np.max(cnt) < np.sum(cnt) * self.single_category_max_area: + break + crop_tfm = CropTransform(x0, y0, crop_size[1], crop_size[0]) + return crop_tfm + + +class RandomExtent(Augmentation): + """ + Outputs an image by cropping a random "subrect" of the source image. + + The subrect can be parameterized to include pixels outside the source image, + in which case they will be set to zeros (i.e. black). The size of the output + image will vary with the size of the random subrect. + """ + + def __init__(self, scale_range, shift_range): + """ + Args: + output_size (h, w): Dimensions of output image + scale_range (l, h): Range of input-to-output size scaling factor + shift_range (x, y): Range of shifts of the cropped subrect. The rect + is shifted by [w / 2 * Uniform(-x, x), h / 2 * Uniform(-y, y)], + where (w, h) is the (width, height) of the input image. Set each + component to zero to crop at the image's center. + """ + super().__init__() + self._init(locals()) + + def get_transform(self, image): + img_h, img_w = image.shape[:2] + + # Initialize src_rect to fit the input image. + src_rect = np.array([-0.5 * img_w, -0.5 * img_h, 0.5 * img_w, 0.5 * img_h]) + + # Apply a random scaling to the src_rect. + src_rect *= np.random.uniform(self.scale_range[0], self.scale_range[1]) + + # Apply a random shift to the coordinates origin. + src_rect[0::2] += self.shift_range[0] * img_w * (np.random.rand() - 0.5) + src_rect[1::2] += self.shift_range[1] * img_h * (np.random.rand() - 0.5) + + # Map src_rect coordinates into image coordinates (center at corner). + src_rect[0::2] += 0.5 * img_w + src_rect[1::2] += 0.5 * img_h + + return ExtentTransform( + src_rect=(src_rect[0], src_rect[1], src_rect[2], src_rect[3]), + output_size=(int(src_rect[3] - src_rect[1]), int(src_rect[2] - src_rect[0])), + ) + + +class RandomContrast(Augmentation): + """ + Randomly transforms image contrast. + + Contrast intensity is uniformly sampled in (intensity_min, intensity_max). + - intensity < 1 will reduce contrast + - intensity = 1 will preserve the input image + - intensity > 1 will increase contrast + + See: https://pillow.readthedocs.io/en/3.0.x/reference/ImageEnhance.html + """ + + def __init__(self, intensity_min, intensity_max): + """ + Args: + intensity_min (float): Minimum augmentation + intensity_max (float): Maximum augmentation + """ + super().__init__() + self._init(locals()) + + def get_transform(self, image): + w = np.random.uniform(self.intensity_min, self.intensity_max) + return BlendTransform(src_image=image.mean(), src_weight=1 - w, dst_weight=w) + + +class RandomBrightness(Augmentation): + """ + Randomly transforms image brightness. + + Brightness intensity is uniformly sampled in (intensity_min, intensity_max). + - intensity < 1 will reduce brightness + - intensity = 1 will preserve the input image + - intensity > 1 will increase brightness + + See: https://pillow.readthedocs.io/en/3.0.x/reference/ImageEnhance.html + """ + + def __init__(self, intensity_min, intensity_max): + """ + Args: + intensity_min (float): Minimum augmentation + intensity_max (float): Maximum augmentation + """ + super().__init__() + self._init(locals()) + + def get_transform(self, image): + w = np.random.uniform(self.intensity_min, self.intensity_max) + return BlendTransform(src_image=0, src_weight=1 - w, dst_weight=w) + + +class RandomSaturation(Augmentation): + """ + Randomly transforms saturation of an RGB image. + Input images are assumed to have 'RGB' channel order. + + Saturation intensity is uniformly sampled in (intensity_min, intensity_max). + - intensity < 1 will reduce saturation (make the image more grayscale) + - intensity = 1 will preserve the input image + - intensity > 1 will increase saturation + + See: https://pillow.readthedocs.io/en/3.0.x/reference/ImageEnhance.html + """ + + def __init__(self, intensity_min, intensity_max): + """ + Args: + intensity_min (float): Minimum augmentation (1 preserves input). + intensity_max (float): Maximum augmentation (1 preserves input). + """ + super().__init__() + self._init(locals()) + + def get_transform(self, image): + assert image.shape[-1] == 3, "RandomSaturation only works on RGB images" + w = np.random.uniform(self.intensity_min, self.intensity_max) + grayscale = image.dot([0.299, 0.587, 0.114])[:, :, np.newaxis] + return BlendTransform(src_image=grayscale, src_weight=1 - w, dst_weight=w) + + +class RandomLighting(Augmentation): + """ + The "lighting" augmentation described in AlexNet, using fixed PCA over ImageNet. + Input images are assumed to have 'RGB' channel order. + + The degree of color jittering is randomly sampled via a normal distribution, + with standard deviation given by the scale parameter. + """ + + def __init__(self, scale): + """ + Args: + scale (float): Standard deviation of principal component weighting. + """ + super().__init__() + self._init(locals()) + self.eigen_vecs = np.array( + [[-0.5675, 0.7192, 0.4009], [-0.5808, -0.0045, -0.8140], [-0.5836, -0.6948, 0.4203]] + ) + self.eigen_vals = np.array([0.2175, 0.0188, 0.0045]) + + def get_transform(self, image): + assert image.shape[-1] == 3, "RandomLighting only works on RGB images" + weights = np.random.normal(scale=self.scale, size=3) + return BlendTransform( + src_image=self.eigen_vecs.dot(weights * self.eigen_vals), src_weight=1.0, dst_weight=1.0 + ) diff --git a/detectron2/data/transforms/build.py b/detectron2/data/transforms/build.py new file mode 100644 index 0000000000000000000000000000000000000000..0732481a39d3086ffc6708ea4408808be8656f06 --- /dev/null +++ b/detectron2/data/transforms/build.py @@ -0,0 +1,89 @@ +# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. +# from . import transforms as T +import torchvision.transforms as T +from PIL import Image +from timm.data import create_transform +from .torchvision_transforms.transforms import Resize as New_Resize + +def build_clip_transforms(cfg, is_train=True): + if cfg.AUG.USE_TIMM and is_train: + print('=> use timm transform for training') + timm_cfg = cfg.AUG.TIMM_AUG + transforms = create_transform( + input_size=cfg.TRAIN.IMAGE_SIZE[0], + is_training=True, + use_prefetcher=False, + no_aug=False, + re_prob=timm_cfg.RE_PROB, + re_mode=timm_cfg.RE_MODE, + re_count=timm_cfg.RE_COUNT, + scale=cfg.AUG.SCALE, + ratio=cfg.AUG.RATIO, + hflip=timm_cfg.HFLIP, + vflip=timm_cfg.VFLIP, + color_jitter=timm_cfg.COLOR_JITTER, + auto_augment=timm_cfg.AUTO_AUGMENT, + interpolation=timm_cfg.INTERPOLATION, + mean=cfg.MODEL.PIXEL_MEAN, + std=cfg.MODEL.PIXEL_STD, + ) + + return transforms + + # normalize_transform = T.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)) + # assert isinstance(cfg.DATASET.OUTPUT_SIZE, (list, tuple)), 'DATASET.OUTPUT_SIZE should be list or tuple' + # NOTE: normalization is applied in rcnn.py, to keep consistent as Detectron2 + # normalize = T.Normalize(mean=cfg.MODEL.PIXEL_MEAN, std=cfg.MODEL.PIXEL_STD) # T.Normalize(mean=cfg.INPUT.PIXEL_MEAN, std=cfg.INPUT.PIXEL_STD) + + transforms = None + if is_train: + aug = cfg.AUG + scale = aug.SCALE + ratio = aug.RATIO + if len(cfg.AUG.TRAIN.IMAGE_SIZE) == 2: # Data Augmentation from MSR-CLIP + ts = [ + T.RandomResizedCrop( + cfg.AUG.TRAIN.IMAGE_SIZE[0], scale=scale, ratio=ratio, + interpolation=cfg.AUG.INTERPOLATION + ), + T.RandomHorizontalFlip(), + ] + elif len(cfg.AUG.TRAIN.IMAGE_SIZE) == 1 and cfg.AUG.TRAIN.MAX_SIZE is not None: # designed for pretraining fastrcnn + ts = [ + New_Resize( + cfg.AUG.TRAIN.IMAGE_SIZE[0], max_size=cfg.AUG.TRAIN.MAX_SIZE, + interpolation=cfg.AUG.INTERPOLATION + ), + T.RandomHorizontalFlip(), + ] + + cj = aug.COLOR_JITTER + if cj[-1] > 0.0: + ts.append(T.RandomApply([T.ColorJitter(*cj[:-1])], p=cj[-1])) + + gs = aug.GRAY_SCALE + if gs > 0.0: + ts.append(T.RandomGrayscale(gs)) + + gb = aug.GAUSSIAN_BLUR + if gb > 0.0: + ts.append(T.RandomApply([GaussianBlur([.1, 2.])], p=gb)) + + ts.append(T.ToTensor()) + # NOTE: normalization is applied in rcnn.py, to keep consistent as Detectron2 + #ts.append(normalize) + + transforms = T.Compose(ts) + else: + # for zeroshot inference of grounding evaluation + transforms = T.Compose([ + T.Resize( + cfg.AUG.TEST.IMAGE_SIZE[0], + interpolation=cfg.AUG.TEST.INTERPOLATION + ), + T.ToTensor(), + ]) + return transforms + + return transforms + diff --git a/detectron2/data/transforms/torchvision_transforms/__init__.py b/detectron2/data/transforms/torchvision_transforms/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/detectron2/data/transforms/torchvision_transforms/__pycache__/__init__.cpython-39.pyc b/detectron2/data/transforms/torchvision_transforms/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..331b9b254d8b7060ff47b425625236d8338c8657 Binary files /dev/null and b/detectron2/data/transforms/torchvision_transforms/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/data/transforms/torchvision_transforms/__pycache__/functional.cpython-39.pyc b/detectron2/data/transforms/torchvision_transforms/__pycache__/functional.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..16e53675d42d334dfb54cdb0aa762dcca67d9e43 Binary files /dev/null and b/detectron2/data/transforms/torchvision_transforms/__pycache__/functional.cpython-39.pyc differ diff --git a/detectron2/data/transforms/torchvision_transforms/__pycache__/functional_pil.cpython-39.pyc b/detectron2/data/transforms/torchvision_transforms/__pycache__/functional_pil.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1a83c27ce17f34fdb1886f5bafca855266dd16e8 Binary files /dev/null and b/detectron2/data/transforms/torchvision_transforms/__pycache__/functional_pil.cpython-39.pyc differ diff --git a/detectron2/data/transforms/torchvision_transforms/__pycache__/functional_tensor.cpython-39.pyc b/detectron2/data/transforms/torchvision_transforms/__pycache__/functional_tensor.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7a7e1dac753202898048108904150d991d4b52ee Binary files /dev/null and b/detectron2/data/transforms/torchvision_transforms/__pycache__/functional_tensor.cpython-39.pyc differ diff --git a/detectron2/data/transforms/torchvision_transforms/__pycache__/transforms.cpython-39.pyc b/detectron2/data/transforms/torchvision_transforms/__pycache__/transforms.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..435182f4ce2875ca9f8e45058b3ebe62612327de Binary files /dev/null and b/detectron2/data/transforms/torchvision_transforms/__pycache__/transforms.cpython-39.pyc differ diff --git a/detectron2/data/transforms/torchvision_transforms/functional.py b/detectron2/data/transforms/torchvision_transforms/functional.py new file mode 100644 index 0000000000000000000000000000000000000000..dc3a9f8f68b0ad75f36c26696be6a9903358c619 --- /dev/null +++ b/detectron2/data/transforms/torchvision_transforms/functional.py @@ -0,0 +1,1365 @@ +import math +import numbers +import warnings +from enum import Enum + +import numpy as np +from PIL import Image + +import torch +from torch import Tensor +from typing import List, Tuple, Any, Optional + +try: + import accimage +except ImportError: + accimage = None + +from . import functional_pil as F_pil +from . import functional_tensor as F_t + + +class InterpolationMode(Enum): + """Interpolation modes + Available interpolation methods are ``nearest``, ``bilinear``, ``bicubic``, ``box``, ``hamming``, and ``lanczos``. + """ + NEAREST = "nearest" + BILINEAR = "bilinear" + BICUBIC = "bicubic" + # For PIL compatibility + BOX = "box" + HAMMING = "hamming" + LANCZOS = "lanczos" + + +# TODO: Once torchscript supports Enums with staticmethod +# this can be put into InterpolationMode as staticmethod +def _interpolation_modes_from_int(i: int) -> InterpolationMode: + inverse_modes_mapping = { + 0: InterpolationMode.NEAREST, + 2: InterpolationMode.BILINEAR, + 3: InterpolationMode.BICUBIC, + 4: InterpolationMode.BOX, + 5: InterpolationMode.HAMMING, + 1: InterpolationMode.LANCZOS, + } + return inverse_modes_mapping[i] + + +pil_modes_mapping = { + InterpolationMode.NEAREST: 0, + InterpolationMode.BILINEAR: 2, + InterpolationMode.BICUBIC: 3, + InterpolationMode.BOX: 4, + InterpolationMode.HAMMING: 5, + InterpolationMode.LANCZOS: 1, +} + +_is_pil_image = F_pil._is_pil_image + + +def _get_image_size(img: Tensor) -> List[int]: + """Returns image size as [w, h] + """ + if isinstance(img, torch.Tensor): + return F_t._get_image_size(img) + + return F_pil._get_image_size(img) + + +def _get_image_num_channels(img: Tensor) -> int: + """Returns number of image channels + """ + if isinstance(img, torch.Tensor): + return F_t._get_image_num_channels(img) + + return F_pil._get_image_num_channels(img) + + +@torch.jit.unused +def _is_numpy(img: Any) -> bool: + return isinstance(img, np.ndarray) + + +@torch.jit.unused +def _is_numpy_image(img: Any) -> bool: + return img.ndim in {2, 3} + + +def to_tensor(pic): + """Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor. + This function does not support torchscript. + + See :class:`~torchvision.transforms.ToTensor` for more details. + + Args: + pic (PIL Image or numpy.ndarray): Image to be converted to tensor. + + Returns: + Tensor: Converted image. + """ + if not(F_pil._is_pil_image(pic) or _is_numpy(pic)): + raise TypeError('pic should be PIL Image or ndarray. Got {}'.format(type(pic))) + + if _is_numpy(pic) and not _is_numpy_image(pic): + raise ValueError('pic should be 2/3 dimensional. Got {} dimensions.'.format(pic.ndim)) + + default_float_dtype = torch.get_default_dtype() + + if isinstance(pic, np.ndarray): + # handle numpy array + if pic.ndim == 2: + pic = pic[:, :, None] + + img = torch.from_numpy(pic.transpose((2, 0, 1))).contiguous() + # backward compatibility + if isinstance(img, torch.ByteTensor): + return img.to(dtype=default_float_dtype).div(255) + else: + return img + + if accimage is not None and isinstance(pic, accimage.Image): + nppic = np.zeros([pic.channels, pic.height, pic.width], dtype=np.float32) + pic.copyto(nppic) + return torch.from_numpy(nppic).to(dtype=default_float_dtype) + + # handle PIL Image + mode_to_nptype = {'I': np.int32, 'I;16': np.int16, 'F': np.float32} + img = torch.from_numpy( + np.array(pic, mode_to_nptype.get(pic.mode, np.uint8), copy=True) + ) + + if pic.mode == '1': + img = 255 * img + img = img.view(pic.size[1], pic.size[0], len(pic.getbands())) + # put it from HWC to CHW format + img = img.permute((2, 0, 1)).contiguous() + if isinstance(img, torch.ByteTensor): + return img.to(dtype=default_float_dtype).div(255) + else: + return img + + +def pil_to_tensor(pic): + """Convert a ``PIL Image`` to a tensor of the same type. + This function does not support torchscript. + + See :class:`~torchvision.transforms.PILToTensor` for more details. + + Args: + pic (PIL Image): Image to be converted to tensor. + + Returns: + Tensor: Converted image. + """ + if not F_pil._is_pil_image(pic): + raise TypeError('pic should be PIL Image. Got {}'.format(type(pic))) + + if accimage is not None and isinstance(pic, accimage.Image): + # accimage format is always uint8 internally, so always return uint8 here + nppic = np.zeros([pic.channels, pic.height, pic.width], dtype=np.uint8) + pic.copyto(nppic) + return torch.as_tensor(nppic) + + # handle PIL Image + img = torch.as_tensor(np.asarray(pic)) + img = img.view(pic.size[1], pic.size[0], len(pic.getbands())) + # put it from HWC to CHW format + img = img.permute((2, 0, 1)) + return img + + +def convert_image_dtype(image: torch.Tensor, dtype: torch.dtype = torch.float) -> torch.Tensor: + """Convert a tensor image to the given ``dtype`` and scale the values accordingly + This function does not support PIL Image. + + Args: + image (torch.Tensor): Image to be converted + dtype (torch.dtype): Desired data type of the output + + Returns: + Tensor: Converted image + + .. note:: + + When converting from a smaller to a larger integer ``dtype`` the maximum values are **not** mapped exactly. + If converted back and forth, this mismatch has no effect. + + Raises: + RuntimeError: When trying to cast :class:`torch.float32` to :class:`torch.int32` or :class:`torch.int64` as + well as for trying to cast :class:`torch.float64` to :class:`torch.int64`. These conversions might lead to + overflow errors since the floating point ``dtype`` cannot store consecutive integers over the whole range + of the integer ``dtype``. + """ + if not isinstance(image, torch.Tensor): + raise TypeError('Input img should be Tensor Image') + + return F_t.convert_image_dtype(image, dtype) + + +def to_pil_image(pic, mode=None): + """Convert a tensor or an ndarray to PIL Image. This function does not support torchscript. + + See :class:`~torchvision.transforms.ToPILImage` for more details. + + Args: + pic (Tensor or numpy.ndarray): Image to be converted to PIL Image. + mode (`PIL.Image mode`_): color space and pixel depth of input data (optional). + + .. _PIL.Image mode: https://pillow.readthedocs.io/en/latest/handbook/concepts.html#concept-modes + + Returns: + PIL Image: Image converted to PIL Image. + """ + if not(isinstance(pic, torch.Tensor) or isinstance(pic, np.ndarray)): + raise TypeError('pic should be Tensor or ndarray. Got {}.'.format(type(pic))) + + elif isinstance(pic, torch.Tensor): + if pic.ndimension() not in {2, 3}: + raise ValueError('pic should be 2/3 dimensional. Got {} dimensions.'.format(pic.ndimension())) + + elif pic.ndimension() == 2: + # if 2D image, add channel dimension (CHW) + pic = pic.unsqueeze(0) + + # check number of channels + if pic.shape[-3] > 4: + raise ValueError('pic should not have > 4 channels. Got {} channels.'.format(pic.shape[-3])) + + elif isinstance(pic, np.ndarray): + if pic.ndim not in {2, 3}: + raise ValueError('pic should be 2/3 dimensional. Got {} dimensions.'.format(pic.ndim)) + + elif pic.ndim == 2: + # if 2D image, add channel dimension (HWC) + pic = np.expand_dims(pic, 2) + + # check number of channels + if pic.shape[-1] > 4: + raise ValueError('pic should not have > 4 channels. Got {} channels.'.format(pic.shape[-1])) + + npimg = pic + if isinstance(pic, torch.Tensor): + if pic.is_floating_point() and mode != 'F': + pic = pic.mul(255).byte() + npimg = np.transpose(pic.cpu().numpy(), (1, 2, 0)) + + if not isinstance(npimg, np.ndarray): + raise TypeError('Input pic must be a torch.Tensor or NumPy ndarray, ' + + 'not {}'.format(type(npimg))) + + if npimg.shape[2] == 1: + expected_mode = None + npimg = npimg[:, :, 0] + if npimg.dtype == np.uint8: + expected_mode = 'L' + elif npimg.dtype == np.int16: + expected_mode = 'I;16' + elif npimg.dtype == np.int32: + expected_mode = 'I' + elif npimg.dtype == np.float32: + expected_mode = 'F' + if mode is not None and mode != expected_mode: + raise ValueError("Incorrect mode ({}) supplied for input type {}. Should be {}" + .format(mode, np.dtype, expected_mode)) + mode = expected_mode + + elif npimg.shape[2] == 2: + permitted_2_channel_modes = ['LA'] + if mode is not None and mode not in permitted_2_channel_modes: + raise ValueError("Only modes {} are supported for 2D inputs".format(permitted_2_channel_modes)) + + if mode is None and npimg.dtype == np.uint8: + mode = 'LA' + + elif npimg.shape[2] == 4: + permitted_4_channel_modes = ['RGBA', 'CMYK', 'RGBX'] + if mode is not None and mode not in permitted_4_channel_modes: + raise ValueError("Only modes {} are supported for 4D inputs".format(permitted_4_channel_modes)) + + if mode is None and npimg.dtype == np.uint8: + mode = 'RGBA' + else: + permitted_3_channel_modes = ['RGB', 'YCbCr', 'HSV'] + if mode is not None and mode not in permitted_3_channel_modes: + raise ValueError("Only modes {} are supported for 3D inputs".format(permitted_3_channel_modes)) + if mode is None and npimg.dtype == np.uint8: + mode = 'RGB' + + if mode is None: + raise TypeError('Input type {} is not supported'.format(npimg.dtype)) + + return Image.fromarray(npimg, mode=mode) + + +def normalize(tensor: Tensor, mean: List[float], std: List[float], inplace: bool = False) -> Tensor: + """Normalize a float tensor image with mean and standard deviation. + This transform does not support PIL Image. + + .. note:: + This transform acts out of place by default, i.e., it does not mutates the input tensor. + + See :class:`~torchvision.transforms.Normalize` for more details. + + Args: + tensor (Tensor): Float tensor image of size (C, H, W) or (B, C, H, W) to be normalized. + mean (sequence): Sequence of means for each channel. + std (sequence): Sequence of standard deviations for each channel. + inplace(bool,optional): Bool to make this operation inplace. + + Returns: + Tensor: Normalized Tensor image. + """ + if not isinstance(tensor, torch.Tensor): + raise TypeError('Input tensor should be a torch tensor. Got {}.'.format(type(tensor))) + + if not tensor.is_floating_point(): + raise TypeError('Input tensor should be a float tensor. Got {}.'.format(tensor.dtype)) + + if tensor.ndim < 3: + raise ValueError('Expected tensor to be a tensor image of size (..., C, H, W). Got tensor.size() = ' + '{}.'.format(tensor.size())) + + if not inplace: + tensor = tensor.clone() + + dtype = tensor.dtype + mean = torch.as_tensor(mean, dtype=dtype, device=tensor.device) + std = torch.as_tensor(std, dtype=dtype, device=tensor.device) + if (std == 0).any(): + raise ValueError('std evaluated to zero after conversion to {}, leading to division by zero.'.format(dtype)) + if mean.ndim == 1: + mean = mean.view(-1, 1, 1) + if std.ndim == 1: + std = std.view(-1, 1, 1) + tensor.sub_(mean).div_(std) + return tensor + + +def resize(img: Tensor, size: List[int], interpolation: InterpolationMode = InterpolationMode.BILINEAR, + max_size: Optional[int] = None, antialias: Optional[bool] = None) -> Tensor: + r"""Resize the input image to the given size. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions + + .. warning:: + The output image might be different depending on its type: when downsampling, the interpolation of PIL images + and tensors is slightly different, because PIL applies antialiasing. This may lead to significant differences + in the performance of a network. Therefore, it is preferable to train and serve a model with the same input + types. See also below the ``antialias`` parameter, which can help making the output of PIL images and tensors + closer. + + Args: + img (PIL Image or Tensor): Image to be resized. + size (sequence or int): Desired output size. If size is a sequence like + (h, w), the output size will be matched to this. If size is an int, + the smaller edge of the image will be matched to this number maintaining + the aspect ratio. i.e, if height > width, then image will be rescaled to + :math:`\left(\text{size} \times \frac{\text{height}}{\text{width}}, \text{size}\right)`. + + .. note:: + In torchscript mode size as single int is not supported, use a sequence of length 1: ``[size, ]``. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. + Default is ``InterpolationMode.BILINEAR``. If input is Tensor, only ``InterpolationMode.NEAREST``, + ``InterpolationMode.BILINEAR`` and ``InterpolationMode.BICUBIC`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + max_size (int, optional): The maximum allowed for the longer edge of + the resized image: if the longer edge of the image is greater + than ``max_size`` after being resized according to ``size``, then + the image is resized again so that the longer edge is equal to + ``max_size``. As a result, ``size`` might be overruled, i.e the + smaller edge may be shorter than ``size``. This is only supported + if ``size`` is an int (or a sequence of length 1 in torchscript + mode). + antialias (bool, optional): antialias flag. If ``img`` is PIL Image, the flag is ignored and anti-alias + is always used. If ``img`` is Tensor, the flag is False by default and can be set to True for + ``InterpolationMode.BILINEAR`` only mode. This can help making the output for PIL images and tensors + closer. + + .. warning:: + There is no autodiff support for ``antialias=True`` option with input ``img`` as Tensor. + + Returns: + PIL Image or Tensor: Resized image. + """ + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + if not isinstance(interpolation, InterpolationMode): + raise TypeError("Argument interpolation should be a InterpolationMode") + + if not isinstance(img, torch.Tensor): + if antialias is not None and not antialias: + warnings.warn( + "Anti-alias option is always applied for PIL Image input. Argument antialias is ignored." + ) + pil_interpolation = pil_modes_mapping[interpolation] + return F_pil.resize(img, size=size, interpolation=pil_interpolation, max_size=max_size) + + return F_t.resize(img, size=size, interpolation=interpolation.value, max_size=max_size, antialias=antialias) + + +def scale(*args, **kwargs): + warnings.warn("The use of the transforms.Scale transform is deprecated, " + + "please use transforms.Resize instead.") + return resize(*args, **kwargs) + + +def pad(img: Tensor, padding: List[int], fill: int = 0, padding_mode: str = "constant") -> Tensor: + r"""Pad the given image on all sides with the given "pad" value. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means at most 2 leading dimensions for mode reflect and symmetric, + at most 3 leading dimensions for mode edge, + and an arbitrary number of leading dimensions for mode constant + + Args: + img (PIL Image or Tensor): Image to be padded. + padding (int or sequence): Padding on each border. If a single int is provided this + is used to pad all borders. If sequence of length 2 is provided this is the padding + on left/right and top/bottom respectively. If a sequence of length 4 is provided + this is the padding for the left, top, right and bottom borders respectively. + + .. note:: + In torchscript mode padding as single int is not supported, use a sequence of + length 1: ``[padding, ]``. + fill (number or str or tuple): Pixel fill value for constant fill. Default is 0. + If a tuple of length 3, it is used to fill R, G, B channels respectively. + This value is only used when the padding_mode is constant. + Only number is supported for torch Tensor. + Only int or str or tuple value is supported for PIL Image. + padding_mode (str): Type of padding. Should be: constant, edge, reflect or symmetric. + Default is constant. + + - constant: pads with a constant value, this value is specified with fill + + - edge: pads with the last value at the edge of the image. + If input a 5D torch Tensor, the last 3 dimensions will be padded instead of the last 2 + + - reflect: pads with reflection of image without repeating the last value on the edge. + For example, padding [1, 2, 3, 4] with 2 elements on both sides in reflect mode + will result in [3, 2, 1, 2, 3, 4, 3, 2] + + - symmetric: pads with reflection of image repeating the last value on the edge. + For example, padding [1, 2, 3, 4] with 2 elements on both sides in symmetric mode + will result in [2, 1, 1, 2, 3, 4, 4, 3] + + Returns: + PIL Image or Tensor: Padded image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.pad(img, padding=padding, fill=fill, padding_mode=padding_mode) + + return F_t.pad(img, padding=padding, fill=fill, padding_mode=padding_mode) + + +def crop(img: Tensor, top: int, left: int, height: int, width: int) -> Tensor: + """Crop the given image at specified location and output size. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + If image size is smaller than output size along any edge, image is padded with 0 and then cropped. + + Args: + img (PIL Image or Tensor): Image to be cropped. (0,0) denotes the top left corner of the image. + top (int): Vertical component of the top left corner of the crop box. + left (int): Horizontal component of the top left corner of the crop box. + height (int): Height of the crop box. + width (int): Width of the crop box. + + Returns: + PIL Image or Tensor: Cropped image. + """ + + if not isinstance(img, torch.Tensor): + return F_pil.crop(img, top, left, height, width) + + return F_t.crop(img, top, left, height, width) + + +def center_crop(img: Tensor, output_size: List[int]) -> Tensor: + """Crops the given image at the center. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + If image size is smaller than output size along any edge, image is padded with 0 and then center cropped. + + Args: + img (PIL Image or Tensor): Image to be cropped. + output_size (sequence or int): (height, width) of the crop box. If int or sequence with single int, + it is used for both directions. + + Returns: + PIL Image or Tensor: Cropped image. + """ + if isinstance(output_size, numbers.Number): + output_size = (int(output_size), int(output_size)) + elif isinstance(output_size, (tuple, list)) and len(output_size) == 1: + output_size = (output_size[0], output_size[0]) + + image_width, image_height = _get_image_size(img) + crop_height, crop_width = output_size + + if crop_width > image_width or crop_height > image_height: + padding_ltrb = [ + (crop_width - image_width) // 2 if crop_width > image_width else 0, + (crop_height - image_height) // 2 if crop_height > image_height else 0, + (crop_width - image_width + 1) // 2 if crop_width > image_width else 0, + (crop_height - image_height + 1) // 2 if crop_height > image_height else 0, + ] + img = pad(img, padding_ltrb, fill=0) # PIL uses fill value 0 + image_width, image_height = _get_image_size(img) + if crop_width == image_width and crop_height == image_height: + return img + + crop_top = int(round((image_height - crop_height) / 2.)) + crop_left = int(round((image_width - crop_width) / 2.)) + return crop(img, crop_top, crop_left, crop_height, crop_width) + + +def resized_crop( + img: Tensor, top: int, left: int, height: int, width: int, size: List[int], + interpolation: InterpolationMode = InterpolationMode.BILINEAR +) -> Tensor: + """Crop the given image and resize it to desired size. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions + + Notably used in :class:`~torchvision.transforms.RandomResizedCrop`. + + Args: + img (PIL Image or Tensor): Image to be cropped. (0,0) denotes the top left corner of the image. + top (int): Vertical component of the top left corner of the crop box. + left (int): Horizontal component of the top left corner of the crop box. + height (int): Height of the crop box. + width (int): Width of the crop box. + size (sequence or int): Desired output size. Same semantics as ``resize``. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. + Default is ``InterpolationMode.BILINEAR``. If input is Tensor, only ``InterpolationMode.NEAREST``, + ``InterpolationMode.BILINEAR`` and ``InterpolationMode.BICUBIC`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + + Returns: + PIL Image or Tensor: Cropped image. + """ + img = crop(img, top, left, height, width) + img = resize(img, size, interpolation) + return img + + +def hflip(img: Tensor) -> Tensor: + """Horizontally flip the given image. + + Args: + img (PIL Image or Tensor): Image to be flipped. If img + is a Tensor, it is expected to be in [..., H, W] format, + where ... means it can have an arbitrary number of leading + dimensions. + + Returns: + PIL Image or Tensor: Horizontally flipped image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.hflip(img) + + return F_t.hflip(img) + + +def _get_perspective_coeffs( + startpoints: List[List[int]], endpoints: List[List[int]] +) -> List[float]: + """Helper function to get the coefficients (a, b, c, d, e, f, g, h) for the perspective transforms. + + In Perspective Transform each pixel (x, y) in the original image gets transformed as, + (x, y) -> ( (ax + by + c) / (gx + hy + 1), (dx + ey + f) / (gx + hy + 1) ) + + Args: + startpoints (list of list of ints): List containing four lists of two integers corresponding to four corners + ``[top-left, top-right, bottom-right, bottom-left]`` of the original image. + endpoints (list of list of ints): List containing four lists of two integers corresponding to four corners + ``[top-left, top-right, bottom-right, bottom-left]`` of the transformed image. + + Returns: + octuple (a, b, c, d, e, f, g, h) for transforming each pixel. + """ + a_matrix = torch.zeros(2 * len(startpoints), 8, dtype=torch.float) + + for i, (p1, p2) in enumerate(zip(endpoints, startpoints)): + a_matrix[2 * i, :] = torch.tensor([p1[0], p1[1], 1, 0, 0, 0, -p2[0] * p1[0], -p2[0] * p1[1]]) + a_matrix[2 * i + 1, :] = torch.tensor([0, 0, 0, p1[0], p1[1], 1, -p2[1] * p1[0], -p2[1] * p1[1]]) + + b_matrix = torch.tensor(startpoints, dtype=torch.float).view(8) + res = torch.linalg.lstsq(a_matrix, b_matrix, driver='gels').solution + + output: List[float] = res.tolist() + return output + + +def perspective( + img: Tensor, + startpoints: List[List[int]], + endpoints: List[List[int]], + interpolation: InterpolationMode = InterpolationMode.BILINEAR, + fill: Optional[List[float]] = None +) -> Tensor: + """Perform perspective transform of the given image. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + img (PIL Image or Tensor): Image to be transformed. + startpoints (list of list of ints): List containing four lists of two integers corresponding to four corners + ``[top-left, top-right, bottom-right, bottom-left]`` of the original image. + endpoints (list of list of ints): List containing four lists of two integers corresponding to four corners + ``[top-left, top-right, bottom-right, bottom-left]`` of the transformed image. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.BILINEAR``. + If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + fill (sequence or number, optional): Pixel fill value for the area outside the transformed + image. If given a number, the value is used for all bands respectively. + + .. note:: + In torchscript mode single int/float value is not supported, please use a sequence + of length 1: ``[value, ]``. + + Returns: + PIL Image or Tensor: transformed Image. + """ + + coeffs = _get_perspective_coeffs(startpoints, endpoints) + + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + if not isinstance(interpolation, InterpolationMode): + raise TypeError("Argument interpolation should be a InterpolationMode") + + if not isinstance(img, torch.Tensor): + pil_interpolation = pil_modes_mapping[interpolation] + return F_pil.perspective(img, coeffs, interpolation=pil_interpolation, fill=fill) + + return F_t.perspective(img, coeffs, interpolation=interpolation.value, fill=fill) + + +def vflip(img: Tensor) -> Tensor: + """Vertically flip the given image. + + Args: + img (PIL Image or Tensor): Image to be flipped. If img + is a Tensor, it is expected to be in [..., H, W] format, + where ... means it can have an arbitrary number of leading + dimensions. + + Returns: + PIL Image or Tensor: Vertically flipped image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.vflip(img) + + return F_t.vflip(img) + + +def five_crop(img: Tensor, size: List[int]) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor]: + """Crop the given image into four corners and the central crop. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions + + .. Note:: + This transform returns a tuple of images and there may be a + mismatch in the number of inputs and targets your ``Dataset`` returns. + + Args: + img (PIL Image or Tensor): Image to be cropped. + size (sequence or int): Desired output size of the crop. If size is an + int instead of sequence like (h, w), a square crop (size, size) is + made. If provided a sequence of length 1, it will be interpreted as (size[0], size[0]). + + Returns: + tuple: tuple (tl, tr, bl, br, center) + Corresponding top left, top right, bottom left, bottom right and center crop. + """ + if isinstance(size, numbers.Number): + size = (int(size), int(size)) + elif isinstance(size, (tuple, list)) and len(size) == 1: + size = (size[0], size[0]) + + if len(size) != 2: + raise ValueError("Please provide only two dimensions (h, w) for size.") + + image_width, image_height = _get_image_size(img) + crop_height, crop_width = size + if crop_width > image_width or crop_height > image_height: + msg = "Requested crop size {} is bigger than input size {}" + raise ValueError(msg.format(size, (image_height, image_width))) + + tl = crop(img, 0, 0, crop_height, crop_width) + tr = crop(img, 0, image_width - crop_width, crop_height, crop_width) + bl = crop(img, image_height - crop_height, 0, crop_height, crop_width) + br = crop(img, image_height - crop_height, image_width - crop_width, crop_height, crop_width) + + center = center_crop(img, [crop_height, crop_width]) + + return tl, tr, bl, br, center + + +def ten_crop(img: Tensor, size: List[int], vertical_flip: bool = False) -> List[Tensor]: + """Generate ten cropped images from the given image. + Crop the given image into four corners and the central crop plus the + flipped version of these (horizontal flipping is used by default). + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions + + .. Note:: + This transform returns a tuple of images and there may be a + mismatch in the number of inputs and targets your ``Dataset`` returns. + + Args: + img (PIL Image or Tensor): Image to be cropped. + size (sequence or int): Desired output size of the crop. If size is an + int instead of sequence like (h, w), a square crop (size, size) is + made. If provided a sequence of length 1, it will be interpreted as (size[0], size[0]). + vertical_flip (bool): Use vertical flipping instead of horizontal + + Returns: + tuple: tuple (tl, tr, bl, br, center, tl_flip, tr_flip, bl_flip, br_flip, center_flip) + Corresponding top left, top right, bottom left, bottom right and + center crop and same for the flipped image. + """ + if isinstance(size, numbers.Number): + size = (int(size), int(size)) + elif isinstance(size, (tuple, list)) and len(size) == 1: + size = (size[0], size[0]) + + if len(size) != 2: + raise ValueError("Please provide only two dimensions (h, w) for size.") + + first_five = five_crop(img, size) + + if vertical_flip: + img = vflip(img) + else: + img = hflip(img) + + second_five = five_crop(img, size) + return first_five + second_five + + +def adjust_brightness(img: Tensor, brightness_factor: float) -> Tensor: + """Adjust brightness of an image. + + Args: + img (PIL Image or Tensor): Image to be adjusted. + If img is torch Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + brightness_factor (float): How much to adjust the brightness. Can be + any non negative number. 0 gives a black image, 1 gives the + original image while 2 increases the brightness by a factor of 2. + + Returns: + PIL Image or Tensor: Brightness adjusted image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.adjust_brightness(img, brightness_factor) + + return F_t.adjust_brightness(img, brightness_factor) + + +def adjust_contrast(img: Tensor, contrast_factor: float) -> Tensor: + """Adjust contrast of an image. + + Args: + img (PIL Image or Tensor): Image to be adjusted. + If img is torch Tensor, it is expected to be in [..., 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + contrast_factor (float): How much to adjust the contrast. Can be any + non negative number. 0 gives a solid gray image, 1 gives the + original image while 2 increases the contrast by a factor of 2. + + Returns: + PIL Image or Tensor: Contrast adjusted image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.adjust_contrast(img, contrast_factor) + + return F_t.adjust_contrast(img, contrast_factor) + + +def adjust_saturation(img: Tensor, saturation_factor: float) -> Tensor: + """Adjust color saturation of an image. + + Args: + img (PIL Image or Tensor): Image to be adjusted. + If img is torch Tensor, it is expected to be in [..., 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + saturation_factor (float): How much to adjust the saturation. 0 will + give a black and white image, 1 will give the original image while + 2 will enhance the saturation by a factor of 2. + + Returns: + PIL Image or Tensor: Saturation adjusted image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.adjust_saturation(img, saturation_factor) + + return F_t.adjust_saturation(img, saturation_factor) + + +def adjust_hue(img: Tensor, hue_factor: float) -> Tensor: + """Adjust hue of an image. + + The image hue is adjusted by converting the image to HSV and + cyclically shifting the intensities in the hue channel (H). + The image is then converted back to original image mode. + + `hue_factor` is the amount of shift in H channel and must be in the + interval `[-0.5, 0.5]`. + + See `Hue`_ for more details. + + .. _Hue: https://en.wikipedia.org/wiki/Hue + + Args: + img (PIL Image or Tensor): Image to be adjusted. + If img is torch Tensor, it is expected to be in [..., 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + If img is PIL Image mode "1", "L", "I", "F" and modes with transparency (alpha channel) are not supported. + hue_factor (float): How much to shift the hue channel. Should be in + [-0.5, 0.5]. 0.5 and -0.5 give complete reversal of hue channel in + HSV space in positive and negative direction respectively. + 0 means no shift. Therefore, both -0.5 and 0.5 will give an image + with complementary colors while 0 gives the original image. + + Returns: + PIL Image or Tensor: Hue adjusted image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.adjust_hue(img, hue_factor) + + return F_t.adjust_hue(img, hue_factor) + + +def adjust_gamma(img: Tensor, gamma: float, gain: float = 1) -> Tensor: + r"""Perform gamma correction on an image. + + Also known as Power Law Transform. Intensities in RGB mode are adjusted + based on the following equation: + + .. math:: + I_{\text{out}} = 255 \times \text{gain} \times \left(\frac{I_{\text{in}}}{255}\right)^{\gamma} + + See `Gamma Correction`_ for more details. + + .. _Gamma Correction: https://en.wikipedia.org/wiki/Gamma_correction + + Args: + img (PIL Image or Tensor): PIL Image to be adjusted. + If img is torch Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + If img is PIL Image, modes with transparency (alpha channel) are not supported. + gamma (float): Non negative real number, same as :math:`\gamma` in the equation. + gamma larger than 1 make the shadows darker, + while gamma smaller than 1 make dark regions lighter. + gain (float): The constant multiplier. + Returns: + PIL Image or Tensor: Gamma correction adjusted image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.adjust_gamma(img, gamma, gain) + + return F_t.adjust_gamma(img, gamma, gain) + + +def _get_inverse_affine_matrix( + center: List[float], angle: float, translate: List[float], scale: float, shear: List[float] +) -> List[float]: + # Helper method to compute inverse matrix for affine transformation + + # As it is explained in PIL.Image.rotate + # We need compute INVERSE of affine transformation matrix: M = T * C * RSS * C^-1 + # where T is translation matrix: [1, 0, tx | 0, 1, ty | 0, 0, 1] + # C is translation matrix to keep center: [1, 0, cx | 0, 1, cy | 0, 0, 1] + # RSS is rotation with scale and shear matrix + # RSS(a, s, (sx, sy)) = + # = R(a) * S(s) * SHy(sy) * SHx(sx) + # = [ s*cos(a - sy)/cos(sy), s*(-cos(a - sy)*tan(x)/cos(y) - sin(a)), 0 ] + # [ s*sin(a + sy)/cos(sy), s*(-sin(a - sy)*tan(x)/cos(y) + cos(a)), 0 ] + # [ 0 , 0 , 1 ] + # + # where R is a rotation matrix, S is a scaling matrix, and SHx and SHy are the shears: + # SHx(s) = [1, -tan(s)] and SHy(s) = [1 , 0] + # [0, 1 ] [-tan(s), 1] + # + # Thus, the inverse is M^-1 = C * RSS^-1 * C^-1 * T^-1 + + rot = math.radians(angle) + sx, sy = [math.radians(s) for s in shear] + + cx, cy = center + tx, ty = translate + + # RSS without scaling + a = math.cos(rot - sy) / math.cos(sy) + b = -math.cos(rot - sy) * math.tan(sx) / math.cos(sy) - math.sin(rot) + c = math.sin(rot - sy) / math.cos(sy) + d = -math.sin(rot - sy) * math.tan(sx) / math.cos(sy) + math.cos(rot) + + # Inverted rotation matrix with scale and shear + # det([[a, b], [c, d]]) == 1, since det(rotation) = 1 and det(shear) = 1 + matrix = [d, -b, 0.0, -c, a, 0.0] + matrix = [x / scale for x in matrix] + + # Apply inverse of translation and of center translation: RSS^-1 * C^-1 * T^-1 + matrix[2] += matrix[0] * (-cx - tx) + matrix[1] * (-cy - ty) + matrix[5] += matrix[3] * (-cx - tx) + matrix[4] * (-cy - ty) + + # Apply center translation: C * RSS^-1 * C^-1 * T^-1 + matrix[2] += cx + matrix[5] += cy + + return matrix + + +def rotate( + img: Tensor, angle: float, interpolation: InterpolationMode = InterpolationMode.NEAREST, + expand: bool = False, center: Optional[List[int]] = None, + fill: Optional[List[float]] = None, resample: Optional[int] = None +) -> Tensor: + """Rotate the image by angle. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + img (PIL Image or Tensor): image to be rotated. + angle (number): rotation angle value in degrees, counter-clockwise. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. + If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + expand (bool, optional): Optional expansion flag. + If true, expands the output image to make it large enough to hold the entire rotated image. + If false or omitted, make the output image the same size as the input image. + Note that the expand flag assumes rotation around the center and no translation. + center (sequence, optional): Optional center of rotation. Origin is the upper left corner. + Default is the center of the image. + fill (sequence or number, optional): Pixel fill value for the area outside the transformed + image. If given a number, the value is used for all bands respectively. + + .. note:: + In torchscript mode single int/float value is not supported, please use a sequence + of length 1: ``[value, ]``. + + Returns: + PIL Image or Tensor: Rotated image. + + .. _filters: https://pillow.readthedocs.io/en/latest/handbook/concepts.html#filters + + """ + if resample is not None: + warnings.warn( + "Argument resample is deprecated and will be removed since v0.10.0. Please, use interpolation instead" + ) + interpolation = _interpolation_modes_from_int(resample) + + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + if not isinstance(angle, (int, float)): + raise TypeError("Argument angle should be int or float") + + if center is not None and not isinstance(center, (list, tuple)): + raise TypeError("Argument center should be a sequence") + + if not isinstance(interpolation, InterpolationMode): + raise TypeError("Argument interpolation should be a InterpolationMode") + + if not isinstance(img, torch.Tensor): + pil_interpolation = pil_modes_mapping[interpolation] + return F_pil.rotate(img, angle=angle, interpolation=pil_interpolation, expand=expand, center=center, fill=fill) + + center_f = [0.0, 0.0] + if center is not None: + img_size = _get_image_size(img) + # Center values should be in pixel coordinates but translated such that (0, 0) corresponds to image center. + center_f = [1.0 * (c - s * 0.5) for c, s in zip(center, img_size)] + + # due to current incoherence of rotation angle direction between affine and rotate implementations + # we need to set -angle. + matrix = _get_inverse_affine_matrix(center_f, -angle, [0.0, 0.0], 1.0, [0.0, 0.0]) + return F_t.rotate(img, matrix=matrix, interpolation=interpolation.value, expand=expand, fill=fill) + + +def affine( + img: Tensor, angle: float, translate: List[int], scale: float, shear: List[float], + interpolation: InterpolationMode = InterpolationMode.NEAREST, fill: Optional[List[float]] = None, + resample: Optional[int] = None, fillcolor: Optional[List[float]] = None +) -> Tensor: + """Apply affine transformation on the image keeping image center invariant. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + img (PIL Image or Tensor): image to transform. + angle (number): rotation angle in degrees between -180 and 180, clockwise direction. + translate (sequence of integers): horizontal and vertical translations (post-rotation translation) + scale (float): overall scale + shear (float or sequence): shear angle value in degrees between -180 to 180, clockwise direction. + If a sequence is specified, the first value corresponds to a shear parallel to the x axis, while + the second value corresponds to a shear parallel to the y axis. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. + If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + fill (sequence or number, optional): Pixel fill value for the area outside the transformed + image. If given a number, the value is used for all bands respectively. + + .. note:: + In torchscript mode single int/float value is not supported, please use a sequence + of length 1: ``[value, ]``. + fillcolor (sequence, int, float): deprecated argument and will be removed since v0.10.0. + Please use the ``fill`` parameter instead. + resample (int, optional): deprecated argument and will be removed since v0.10.0. + Please use the ``interpolation`` parameter instead. + + Returns: + PIL Image or Tensor: Transformed image. + """ + if resample is not None: + warnings.warn( + "Argument resample is deprecated and will be removed since v0.10.0. Please, use interpolation instead" + ) + interpolation = _interpolation_modes_from_int(resample) + + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + if fillcolor is not None: + warnings.warn( + "Argument fillcolor is deprecated and will be removed since v0.10.0. Please, use fill instead" + ) + fill = fillcolor + + if not isinstance(angle, (int, float)): + raise TypeError("Argument angle should be int or float") + + if not isinstance(translate, (list, tuple)): + raise TypeError("Argument translate should be a sequence") + + if len(translate) != 2: + raise ValueError("Argument translate should be a sequence of length 2") + + if scale <= 0.0: + raise ValueError("Argument scale should be positive") + + if not isinstance(shear, (numbers.Number, (list, tuple))): + raise TypeError("Shear should be either a single value or a sequence of two values") + + if not isinstance(interpolation, InterpolationMode): + raise TypeError("Argument interpolation should be a InterpolationMode") + + if isinstance(angle, int): + angle = float(angle) + + if isinstance(translate, tuple): + translate = list(translate) + + if isinstance(shear, numbers.Number): + shear = [shear, 0.0] + + if isinstance(shear, tuple): + shear = list(shear) + + if len(shear) == 1: + shear = [shear[0], shear[0]] + + if len(shear) != 2: + raise ValueError("Shear should be a sequence containing two values. Got {}".format(shear)) + + img_size = _get_image_size(img) + if not isinstance(img, torch.Tensor): + # center = (img_size[0] * 0.5 + 0.5, img_size[1] * 0.5 + 0.5) + # it is visually better to estimate the center without 0.5 offset + # otherwise image rotated by 90 degrees is shifted vs output image of torch.rot90 or F_t.affine + center = [img_size[0] * 0.5, img_size[1] * 0.5] + matrix = _get_inverse_affine_matrix(center, angle, translate, scale, shear) + pil_interpolation = pil_modes_mapping[interpolation] + return F_pil.affine(img, matrix=matrix, interpolation=pil_interpolation, fill=fill) + + translate_f = [1.0 * t for t in translate] + matrix = _get_inverse_affine_matrix([0.0, 0.0], angle, translate_f, scale, shear) + return F_t.affine(img, matrix=matrix, interpolation=interpolation.value, fill=fill) + + +@torch.jit.unused +def to_grayscale(img, num_output_channels=1): + """Convert PIL image of any mode (RGB, HSV, LAB, etc) to grayscale version of image. + This transform does not support torch Tensor. + + Args: + img (PIL Image): PIL Image to be converted to grayscale. + num_output_channels (int): number of channels of the output image. Value can be 1 or 3. Default is 1. + + Returns: + PIL Image: Grayscale version of the image. + + - if num_output_channels = 1 : returned image is single channel + - if num_output_channels = 3 : returned image is 3 channel with r = g = b + """ + if isinstance(img, Image.Image): + return F_pil.to_grayscale(img, num_output_channels) + + raise TypeError("Input should be PIL Image") + + +def rgb_to_grayscale(img: Tensor, num_output_channels: int = 1) -> Tensor: + """Convert RGB image to grayscale version of image. + If the image is torch Tensor, it is expected + to have [..., 3, H, W] shape, where ... means an arbitrary number of leading dimensions + + Note: + Please, note that this method supports only RGB images as input. For inputs in other color spaces, + please, consider using meth:`~torchvision.transforms.functional.to_grayscale` with PIL Image. + + Args: + img (PIL Image or Tensor): RGB Image to be converted to grayscale. + num_output_channels (int): number of channels of the output image. Value can be 1 or 3. Default, 1. + + Returns: + PIL Image or Tensor: Grayscale version of the image. + + - if num_output_channels = 1 : returned image is single channel + - if num_output_channels = 3 : returned image is 3 channel with r = g = b + """ + if not isinstance(img, torch.Tensor): + return F_pil.to_grayscale(img, num_output_channels) + + return F_t.rgb_to_grayscale(img, num_output_channels) + + +def erase(img: Tensor, i: int, j: int, h: int, w: int, v: Tensor, inplace: bool = False) -> Tensor: + """ Erase the input Tensor Image with given value. + This transform does not support PIL Image. + + Args: + img (Tensor Image): Tensor image of size (C, H, W) to be erased + i (int): i in (i,j) i.e coordinates of the upper left corner. + j (int): j in (i,j) i.e coordinates of the upper left corner. + h (int): Height of the erased region. + w (int): Width of the erased region. + v: Erasing value. + inplace(bool, optional): For in-place operations. By default is set False. + + Returns: + Tensor Image: Erased image. + """ + if not isinstance(img, torch.Tensor): + raise TypeError('img should be Tensor Image. Got {}'.format(type(img))) + + if not inplace: + img = img.clone() + + img[..., i:i + h, j:j + w] = v + return img + + +def gaussian_blur(img: Tensor, kernel_size: List[int], sigma: Optional[List[float]] = None) -> Tensor: + """Performs Gaussian blurring on the image by given kernel. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + img (PIL Image or Tensor): Image to be blurred + kernel_size (sequence of ints or int): Gaussian kernel size. Can be a sequence of integers + like ``(kx, ky)`` or a single integer for square kernels. + + .. note:: + In torchscript mode kernel_size as single int is not supported, use a sequence of + length 1: ``[ksize, ]``. + sigma (sequence of floats or float, optional): Gaussian kernel standard deviation. Can be a + sequence of floats like ``(sigma_x, sigma_y)`` or a single float to define the + same sigma in both X/Y directions. If None, then it is computed using + ``kernel_size`` as ``sigma = 0.3 * ((kernel_size - 1) * 0.5 - 1) + 0.8``. + Default, None. + + .. note:: + In torchscript mode sigma as single float is + not supported, use a sequence of length 1: ``[sigma, ]``. + + Returns: + PIL Image or Tensor: Gaussian Blurred version of the image. + """ + if not isinstance(kernel_size, (int, list, tuple)): + raise TypeError('kernel_size should be int or a sequence of integers. Got {}'.format(type(kernel_size))) + if isinstance(kernel_size, int): + kernel_size = [kernel_size, kernel_size] + if len(kernel_size) != 2: + raise ValueError('If kernel_size is a sequence its length should be 2. Got {}'.format(len(kernel_size))) + for ksize in kernel_size: + if ksize % 2 == 0 or ksize < 0: + raise ValueError('kernel_size should have odd and positive integers. Got {}'.format(kernel_size)) + + if sigma is None: + sigma = [ksize * 0.15 + 0.35 for ksize in kernel_size] + + if sigma is not None and not isinstance(sigma, (int, float, list, tuple)): + raise TypeError('sigma should be either float or sequence of floats. Got {}'.format(type(sigma))) + if isinstance(sigma, (int, float)): + sigma = [float(sigma), float(sigma)] + if isinstance(sigma, (list, tuple)) and len(sigma) == 1: + sigma = [sigma[0], sigma[0]] + if len(sigma) != 2: + raise ValueError('If sigma is a sequence, its length should be 2. Got {}'.format(len(sigma))) + for s in sigma: + if s <= 0.: + raise ValueError('sigma should have positive values. Got {}'.format(sigma)) + + t_img = img + if not isinstance(img, torch.Tensor): + if not F_pil._is_pil_image(img): + raise TypeError('img should be PIL Image or Tensor. Got {}'.format(type(img))) + + t_img = to_tensor(img) + + output = F_t.gaussian_blur(t_img, kernel_size, sigma) + + if not isinstance(img, torch.Tensor): + output = to_pil_image(output) + return output + + +def invert(img: Tensor) -> Tensor: + """Invert the colors of an RGB/grayscale image. + + Args: + img (PIL Image or Tensor): Image to have its colors inverted. + If img is torch Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "L" or "RGB". + + Returns: + PIL Image or Tensor: Color inverted image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.invert(img) + + return F_t.invert(img) + + +def posterize(img: Tensor, bits: int) -> Tensor: + """Posterize an image by reducing the number of bits for each color channel. + + Args: + img (PIL Image or Tensor): Image to have its colors posterized. + If img is torch Tensor, it should be of type torch.uint8 and + it is expected to be in [..., 1 or 3, H, W] format, where ... means + it can have an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "L" or "RGB". + bits (int): The number of bits to keep for each channel (0-8). + Returns: + PIL Image or Tensor: Posterized image. + """ + if not (0 <= bits <= 8): + raise ValueError('The number if bits should be between 0 and 8. Got {}'.format(bits)) + + if not isinstance(img, torch.Tensor): + return F_pil.posterize(img, bits) + + return F_t.posterize(img, bits) + + +def solarize(img: Tensor, threshold: float) -> Tensor: + """Solarize an RGB/grayscale image by inverting all pixel values above a threshold. + + Args: + img (PIL Image or Tensor): Image to have its colors inverted. + If img is torch Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "L" or "RGB". + threshold (float): All pixels equal or above this value are inverted. + Returns: + PIL Image or Tensor: Solarized image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.solarize(img, threshold) + + return F_t.solarize(img, threshold) + + +def adjust_sharpness(img: Tensor, sharpness_factor: float) -> Tensor: + """Adjust the sharpness of an image. + + Args: + img (PIL Image or Tensor): Image to be adjusted. + If img is torch Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + sharpness_factor (float): How much to adjust the sharpness. Can be + any non negative number. 0 gives a blurred image, 1 gives the + original image while 2 increases the sharpness by a factor of 2. + + Returns: + PIL Image or Tensor: Sharpness adjusted image. + """ + if not isinstance(img, torch.Tensor): + return F_pil.adjust_sharpness(img, sharpness_factor) + + return F_t.adjust_sharpness(img, sharpness_factor) + + +def autocontrast(img: Tensor) -> Tensor: + """Maximize contrast of an image by remapping its + pixels per channel so that the lowest becomes black and the lightest + becomes white. + + Args: + img (PIL Image or Tensor): Image on which autocontrast is applied. + If img is torch Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "L" or "RGB". + + Returns: + PIL Image or Tensor: An image that was autocontrasted. + """ + if not isinstance(img, torch.Tensor): + return F_pil.autocontrast(img) + + return F_t.autocontrast(img) + + +def equalize(img: Tensor) -> Tensor: + """Equalize the histogram of an image by applying + a non-linear mapping to the input in order to create a uniform + distribution of grayscale values in the output. + + Args: + img (PIL Image or Tensor): Image on which equalize is applied. + If img is torch Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + The tensor dtype must be ``torch.uint8`` and values are expected to be in ``[0, 255]``. + If img is PIL Image, it is expected to be in mode "P", "L" or "RGB". + + Returns: + PIL Image or Tensor: An image that was equalized. + """ + if not isinstance(img, torch.Tensor): + return F_pil.equalize(img) + + return F_t.equalize(img) diff --git a/detectron2/data/transforms/torchvision_transforms/functional_pil.py b/detectron2/data/transforms/torchvision_transforms/functional_pil.py new file mode 100644 index 0000000000000000000000000000000000000000..3829637fdb723a9d75ac63e99f7d2a1be8b754c2 --- /dev/null +++ b/detectron2/data/transforms/torchvision_transforms/functional_pil.py @@ -0,0 +1,352 @@ +import numbers +from typing import Any, List, Sequence + +import numpy as np +import torch +from PIL import Image, ImageOps, ImageEnhance + +try: + import accimage +except ImportError: + accimage = None + + +@torch.jit.unused +def _is_pil_image(img: Any) -> bool: + if accimage is not None: + return isinstance(img, (Image.Image, accimage.Image)) + else: + return isinstance(img, Image.Image) + + +@torch.jit.unused +def _get_image_size(img: Any) -> List[int]: + if _is_pil_image(img): + return img.size + raise TypeError("Unexpected type {}".format(type(img))) + + +@torch.jit.unused +def _get_image_num_channels(img: Any) -> int: + if _is_pil_image(img): + return 1 if img.mode == 'L' else 3 + raise TypeError("Unexpected type {}".format(type(img))) + + +@torch.jit.unused +def hflip(img): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + return img.transpose(Image.FLIP_LEFT_RIGHT) + + +@torch.jit.unused +def vflip(img): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + return img.transpose(Image.FLIP_TOP_BOTTOM) + + +@torch.jit.unused +def adjust_brightness(img, brightness_factor): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + enhancer = ImageEnhance.Brightness(img) + img = enhancer.enhance(brightness_factor) + return img + + +@torch.jit.unused +def adjust_contrast(img, contrast_factor): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + enhancer = ImageEnhance.Contrast(img) + img = enhancer.enhance(contrast_factor) + return img + + +@torch.jit.unused +def adjust_saturation(img, saturation_factor): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + enhancer = ImageEnhance.Color(img) + img = enhancer.enhance(saturation_factor) + return img + + +@torch.jit.unused +def adjust_hue(img, hue_factor): + if not(-0.5 <= hue_factor <= 0.5): + raise ValueError('hue_factor ({}) is not in [-0.5, 0.5].'.format(hue_factor)) + + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + input_mode = img.mode + if input_mode in {'L', '1', 'I', 'F'}: + return img + + h, s, v = img.convert('HSV').split() + + np_h = np.array(h, dtype=np.uint8) + # uint8 addition take cares of rotation across boundaries + with np.errstate(over='ignore'): + np_h += np.uint8(hue_factor * 255) + h = Image.fromarray(np_h, 'L') + + img = Image.merge('HSV', (h, s, v)).convert(input_mode) + return img + + +@torch.jit.unused +def adjust_gamma(img, gamma, gain=1): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + if gamma < 0: + raise ValueError('Gamma should be a non-negative real number') + + input_mode = img.mode + img = img.convert('RGB') + gamma_map = [(255 + 1 - 1e-3) * gain * pow(ele / 255., gamma) for ele in range(256)] * 3 + img = img.point(gamma_map) # use PIL's point-function to accelerate this part + + img = img.convert(input_mode) + return img + + +@torch.jit.unused +def pad(img, padding, fill=0, padding_mode="constant"): + if not _is_pil_image(img): + raise TypeError("img should be PIL Image. Got {}".format(type(img))) + + if not isinstance(padding, (numbers.Number, tuple, list)): + raise TypeError("Got inappropriate padding arg") + if not isinstance(fill, (numbers.Number, str, tuple)): + raise TypeError("Got inappropriate fill arg") + if not isinstance(padding_mode, str): + raise TypeError("Got inappropriate padding_mode arg") + + if isinstance(padding, list): + padding = tuple(padding) + + if isinstance(padding, tuple) and len(padding) not in [1, 2, 4]: + raise ValueError("Padding must be an int or a 1, 2, or 4 element tuple, not a " + + "{} element tuple".format(len(padding))) + + if isinstance(padding, tuple) and len(padding) == 1: + # Compatibility with `functional_tensor.pad` + padding = padding[0] + + if padding_mode not in ["constant", "edge", "reflect", "symmetric"]: + raise ValueError("Padding mode should be either constant, edge, reflect or symmetric") + + if padding_mode == "constant": + opts = _parse_fill(fill, img, name="fill") + if img.mode == "P": + palette = img.getpalette() + image = ImageOps.expand(img, border=padding, **opts) + image.putpalette(palette) + return image + + return ImageOps.expand(img, border=padding, **opts) + else: + if isinstance(padding, int): + pad_left = pad_right = pad_top = pad_bottom = padding + if isinstance(padding, tuple) and len(padding) == 2: + pad_left = pad_right = padding[0] + pad_top = pad_bottom = padding[1] + if isinstance(padding, tuple) and len(padding) == 4: + pad_left = padding[0] + pad_top = padding[1] + pad_right = padding[2] + pad_bottom = padding[3] + + p = [pad_left, pad_top, pad_right, pad_bottom] + cropping = -np.minimum(p, 0) + + if cropping.any(): + crop_left, crop_top, crop_right, crop_bottom = cropping + img = img.crop((crop_left, crop_top, img.width - crop_right, img.height - crop_bottom)) + + pad_left, pad_top, pad_right, pad_bottom = np.maximum(p, 0) + + if img.mode == 'P': + palette = img.getpalette() + img = np.asarray(img) + img = np.pad(img, ((pad_top, pad_bottom), (pad_left, pad_right)), padding_mode) + img = Image.fromarray(img) + img.putpalette(palette) + return img + + img = np.asarray(img) + # RGB image + if len(img.shape) == 3: + img = np.pad(img, ((pad_top, pad_bottom), (pad_left, pad_right), (0, 0)), padding_mode) + # Grayscale image + if len(img.shape) == 2: + img = np.pad(img, ((pad_top, pad_bottom), (pad_left, pad_right)), padding_mode) + + return Image.fromarray(img) + + +@torch.jit.unused +def crop(img: Image.Image, top: int, left: int, height: int, width: int) -> Image.Image: + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + return img.crop((left, top, left + width, top + height)) + + +@torch.jit.unused +def resize(img, size, interpolation=Image.BILINEAR, max_size=None): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + if not (isinstance(size, int) or (isinstance(size, Sequence) and len(size) in (1, 2))): + raise TypeError('Got inappropriate size arg: {}'.format(size)) + + if isinstance(size, Sequence) and len(size) == 1: + size = size[0] + if isinstance(size, int): + w, h = img.size + + short, long = (w, h) if w <= h else (h, w) + if short == size: + return img + + new_short, new_long = size, int(size * long / short) + + if max_size is not None: + if max_size <= size: + raise ValueError( + f"max_size = {max_size} must be strictly greater than the requested " + f"size for the smaller edge size = {size}" + ) + if new_long > max_size: + new_short, new_long = int(max_size * new_short / new_long), max_size + + new_w, new_h = (new_short, new_long) if w <= h else (new_long, new_short) + return img.resize((new_w, new_h), interpolation) + else: + if max_size is not None: + raise ValueError( + "max_size should only be passed if size specifies the length of the smaller edge, " + "i.e. size should be an int or a sequence of length 1 in torchscript mode." + ) + return img.resize(size[::-1], interpolation) + + +@torch.jit.unused +def _parse_fill(fill, img, name="fillcolor"): + # Process fill color for affine transforms + num_bands = len(img.getbands()) + if fill is None: + fill = 0 + if isinstance(fill, (int, float)) and num_bands > 1: + fill = tuple([fill] * num_bands) + if isinstance(fill, (list, tuple)): + if len(fill) != num_bands: + msg = ("The number of elements in 'fill' does not match the number of " + "bands of the image ({} != {})") + raise ValueError(msg.format(len(fill), num_bands)) + + fill = tuple(fill) + + return {name: fill} + + +@torch.jit.unused +def affine(img, matrix, interpolation=0, fill=None): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + output_size = img.size + opts = _parse_fill(fill, img) + return img.transform(output_size, Image.AFFINE, matrix, interpolation, **opts) + + +@torch.jit.unused +def rotate(img, angle, interpolation=0, expand=False, center=None, fill=None): + if not _is_pil_image(img): + raise TypeError("img should be PIL Image. Got {}".format(type(img))) + + opts = _parse_fill(fill, img) + return img.rotate(angle, interpolation, expand, center, **opts) + + +@torch.jit.unused +def perspective(img, perspective_coeffs, interpolation=Image.BICUBIC, fill=None): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + opts = _parse_fill(fill, img) + + return img.transform(img.size, Image.PERSPECTIVE, perspective_coeffs, interpolation, **opts) + + +@torch.jit.unused +def to_grayscale(img, num_output_channels): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + if num_output_channels == 1: + img = img.convert('L') + elif num_output_channels == 3: + img = img.convert('L') + np_img = np.array(img, dtype=np.uint8) + np_img = np.dstack([np_img, np_img, np_img]) + img = Image.fromarray(np_img, 'RGB') + else: + raise ValueError('num_output_channels should be either 1 or 3') + + return img + + +@torch.jit.unused +def invert(img): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + return ImageOps.invert(img) + + +@torch.jit.unused +def posterize(img, bits): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + return ImageOps.posterize(img, bits) + + +@torch.jit.unused +def solarize(img, threshold): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + return ImageOps.solarize(img, threshold) + + +@torch.jit.unused +def adjust_sharpness(img, sharpness_factor): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + + enhancer = ImageEnhance.Sharpness(img) + img = enhancer.enhance(sharpness_factor) + return img + + +@torch.jit.unused +def autocontrast(img): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + return ImageOps.autocontrast(img) + + +@torch.jit.unused +def equalize(img): + if not _is_pil_image(img): + raise TypeError('img should be PIL Image. Got {}'.format(type(img))) + return ImageOps.equalize(img) diff --git a/detectron2/data/transforms/torchvision_transforms/functional_tensor.py b/detectron2/data/transforms/torchvision_transforms/functional_tensor.py new file mode 100644 index 0000000000000000000000000000000000000000..5a13bd5d3925c1e2206a69b0a3e7ab72e3e955f2 --- /dev/null +++ b/detectron2/data/transforms/torchvision_transforms/functional_tensor.py @@ -0,0 +1,966 @@ +import warnings + +import torch +from torch import Tensor +from torch.nn.functional import grid_sample, conv2d, interpolate, pad as torch_pad +from torch.jit.annotations import BroadcastingList2 +from typing import Optional, Tuple, List + + +def _is_tensor_a_torch_image(x: Tensor) -> bool: + return x.ndim >= 2 + + +def _assert_image_tensor(img): + if not _is_tensor_a_torch_image(img): + raise TypeError("Tensor is not a torch image.") + + +def _get_image_size(img: Tensor) -> List[int]: + # Returns (w, h) of tensor image + _assert_image_tensor(img) + return [img.shape[-1], img.shape[-2]] + + +def _get_image_num_channels(img: Tensor) -> int: + if img.ndim == 2: + return 1 + elif img.ndim > 2: + return img.shape[-3] + + raise TypeError("Input ndim should be 2 or more. Got {}".format(img.ndim)) + + +def _max_value(dtype: torch.dtype) -> float: + # TODO: replace this method with torch.iinfo when it gets torchscript support. + # https://github.com/pytorch/pytorch/issues/41492 + + a = torch.tensor(2, dtype=dtype) + signed = 1 if torch.tensor(0, dtype=dtype).is_signed() else 0 + bits = 1 + max_value = torch.tensor(-signed, dtype=torch.long) + while True: + next_value = a.pow(bits - signed).sub(1) + if next_value > max_value: + max_value = next_value + bits *= 2 + else: + break + return max_value.item() + + +def _assert_channels(img: Tensor, permitted: List[int]) -> None: + c = _get_image_num_channels(img) + if c not in permitted: + raise TypeError("Input image tensor permitted channel values are {}, but found {}".format(permitted, c)) + + +def convert_image_dtype(image: torch.Tensor, dtype: torch.dtype = torch.float) -> torch.Tensor: + if image.dtype == dtype: + return image + + if image.is_floating_point(): + + # TODO: replace with dtype.is_floating_point when torchscript supports it + if torch.tensor(0, dtype=dtype).is_floating_point(): + return image.to(dtype) + + # float to int + if (image.dtype == torch.float32 and dtype in (torch.int32, torch.int64)) or ( + image.dtype == torch.float64 and dtype == torch.int64 + ): + msg = f"The cast from {image.dtype} to {dtype} cannot be performed safely." + raise RuntimeError(msg) + + # https://github.com/pytorch/vision/pull/2078#issuecomment-612045321 + # For data in the range 0-1, (float * 255).to(uint) is only 255 + # when float is exactly 1.0. + # `max + 1 - epsilon` provides more evenly distributed mapping of + # ranges of floats to ints. + eps = 1e-3 + max_val = _max_value(dtype) + result = image.mul(max_val + 1.0 - eps) + return result.to(dtype) + else: + input_max = _max_value(image.dtype) + + # int to float + # TODO: replace with dtype.is_floating_point when torchscript supports it + if torch.tensor(0, dtype=dtype).is_floating_point(): + image = image.to(dtype) + return image / input_max + + output_max = _max_value(dtype) + + # int to int + if input_max > output_max: + # factor should be forced to int for torch jit script + # otherwise factor is a float and image // factor can produce different results + factor = int((input_max + 1) // (output_max + 1)) + image = torch.div(image, factor, rounding_mode='floor') + return image.to(dtype) + else: + # factor should be forced to int for torch jit script + # otherwise factor is a float and image * factor can produce different results + factor = int((output_max + 1) // (input_max + 1)) + image = image.to(dtype) + return image * factor + + +def vflip(img: Tensor) -> Tensor: + _assert_image_tensor(img) + + return img.flip(-2) + + +def hflip(img: Tensor) -> Tensor: + _assert_image_tensor(img) + + return img.flip(-1) + + +def crop(img: Tensor, top: int, left: int, height: int, width: int) -> Tensor: + _assert_image_tensor(img) + + w, h = _get_image_size(img) + right = left + width + bottom = top + height + + if left < 0 or top < 0 or right > w or bottom > h: + padding_ltrb = [max(-left, 0), max(-top, 0), max(right - w, 0), max(bottom - h, 0)] + return pad(img[..., max(top, 0):bottom, max(left, 0):right], padding_ltrb, fill=0) + return img[..., top:bottom, left:right] + + +def rgb_to_grayscale(img: Tensor, num_output_channels: int = 1) -> Tensor: + if img.ndim < 3: + raise TypeError("Input image tensor should have at least 3 dimensions, but found {}".format(img.ndim)) + _assert_channels(img, [3]) + + if num_output_channels not in (1, 3): + raise ValueError('num_output_channels should be either 1 or 3') + + r, g, b = img.unbind(dim=-3) + # This implementation closely follows the TF one: + # https://github.com/tensorflow/tensorflow/blob/v2.3.0/tensorflow/python/ops/image_ops_impl.py#L2105-L2138 + l_img = (0.2989 * r + 0.587 * g + 0.114 * b).to(img.dtype) + l_img = l_img.unsqueeze(dim=-3) + + if num_output_channels == 3: + return l_img.expand(img.shape) + + return l_img + + +def adjust_brightness(img: Tensor, brightness_factor: float) -> Tensor: + if brightness_factor < 0: + raise ValueError('brightness_factor ({}) is not non-negative.'.format(brightness_factor)) + + _assert_image_tensor(img) + + _assert_channels(img, [1, 3]) + + return _blend(img, torch.zeros_like(img), brightness_factor) + + +def adjust_contrast(img: Tensor, contrast_factor: float) -> Tensor: + if contrast_factor < 0: + raise ValueError('contrast_factor ({}) is not non-negative.'.format(contrast_factor)) + + _assert_image_tensor(img) + + _assert_channels(img, [3]) + + dtype = img.dtype if torch.is_floating_point(img) else torch.float32 + mean = torch.mean(rgb_to_grayscale(img).to(dtype), dim=(-3, -2, -1), keepdim=True) + + return _blend(img, mean, contrast_factor) + + +def adjust_hue(img: Tensor, hue_factor: float) -> Tensor: + if not (-0.5 <= hue_factor <= 0.5): + raise ValueError('hue_factor ({}) is not in [-0.5, 0.5].'.format(hue_factor)) + + if not (isinstance(img, torch.Tensor)): + raise TypeError('Input img should be Tensor image') + + _assert_image_tensor(img) + + _assert_channels(img, [1, 3]) + if _get_image_num_channels(img) == 1: # Match PIL behaviour + return img + + orig_dtype = img.dtype + if img.dtype == torch.uint8: + img = img.to(dtype=torch.float32) / 255.0 + + img = _rgb2hsv(img) + h, s, v = img.unbind(dim=-3) + h = (h + hue_factor) % 1.0 + img = torch.stack((h, s, v), dim=-3) + img_hue_adj = _hsv2rgb(img) + + if orig_dtype == torch.uint8: + img_hue_adj = (img_hue_adj * 255.0).to(dtype=orig_dtype) + + return img_hue_adj + + +def adjust_saturation(img: Tensor, saturation_factor: float) -> Tensor: + if saturation_factor < 0: + raise ValueError('saturation_factor ({}) is not non-negative.'.format(saturation_factor)) + + _assert_image_tensor(img) + + _assert_channels(img, [3]) + + return _blend(img, rgb_to_grayscale(img), saturation_factor) + + +def adjust_gamma(img: Tensor, gamma: float, gain: float = 1) -> Tensor: + if not isinstance(img, torch.Tensor): + raise TypeError('Input img should be a Tensor.') + + _assert_channels(img, [1, 3]) + + if gamma < 0: + raise ValueError('Gamma should be a non-negative real number') + + result = img + dtype = img.dtype + if not torch.is_floating_point(img): + result = convert_image_dtype(result, torch.float32) + + result = (gain * result ** gamma).clamp(0, 1) + + result = convert_image_dtype(result, dtype) + return result + + +def center_crop(img: Tensor, output_size: BroadcastingList2[int]) -> Tensor: + """DEPRECATED + """ + warnings.warn( + "This method is deprecated and will be removed in future releases. " + "Please, use ``F.center_crop`` instead." + ) + + _assert_image_tensor(img) + + _, image_width, image_height = img.size() + crop_height, crop_width = output_size + # crop_top = int(round((image_height - crop_height) / 2.)) + # Result can be different between python func and scripted func + # Temporary workaround: + crop_top = int((image_height - crop_height + 1) * 0.5) + # crop_left = int(round((image_width - crop_width) / 2.)) + # Result can be different between python func and scripted func + # Temporary workaround: + crop_left = int((image_width - crop_width + 1) * 0.5) + + return crop(img, crop_top, crop_left, crop_height, crop_width) + + +def five_crop(img: Tensor, size: BroadcastingList2[int]) -> List[Tensor]: + """DEPRECATED + """ + warnings.warn( + "This method is deprecated and will be removed in future releases. " + "Please, use ``F.five_crop`` instead." + ) + + _assert_image_tensor(img) + + assert len(size) == 2, "Please provide only two dimensions (h, w) for size." + + _, image_width, image_height = img.size() + crop_height, crop_width = size + if crop_width > image_width or crop_height > image_height: + msg = "Requested crop size {} is bigger than input size {}" + raise ValueError(msg.format(size, (image_height, image_width))) + + tl = crop(img, 0, 0, crop_width, crop_height) + tr = crop(img, image_width - crop_width, 0, image_width, crop_height) + bl = crop(img, 0, image_height - crop_height, crop_width, image_height) + br = crop(img, image_width - crop_width, image_height - crop_height, image_width, image_height) + center = center_crop(img, (crop_height, crop_width)) + + return [tl, tr, bl, br, center] + + +def ten_crop(img: Tensor, size: BroadcastingList2[int], vertical_flip: bool = False) -> List[Tensor]: + """DEPRECATED + """ + warnings.warn( + "This method is deprecated and will be removed in future releases. " + "Please, use ``F.ten_crop`` instead." + ) + + _assert_image_tensor(img) + + assert len(size) == 2, "Please provide only two dimensions (h, w) for size." + first_five = five_crop(img, size) + + if vertical_flip: + img = vflip(img) + else: + img = hflip(img) + + second_five = five_crop(img, size) + + return first_five + second_five + + +def _blend(img1: Tensor, img2: Tensor, ratio: float) -> Tensor: + ratio = float(ratio) + bound = 1.0 if img1.is_floating_point() else 255.0 + return (ratio * img1 + (1.0 - ratio) * img2).clamp(0, bound).to(img1.dtype) + + +def _rgb2hsv(img): + r, g, b = img.unbind(dim=-3) + + # Implementation is based on https://github.com/python-pillow/Pillow/blob/4174d4267616897df3746d315d5a2d0f82c656ee/ + # src/libImaging/Convert.c#L330 + maxc = torch.max(img, dim=-3).values + minc = torch.min(img, dim=-3).values + + # The algorithm erases S and H channel where `maxc = minc`. This avoids NaN + # from happening in the results, because + # + S channel has division by `maxc`, which is zero only if `maxc = minc` + # + H channel has division by `(maxc - minc)`. + # + # Instead of overwriting NaN afterwards, we just prevent it from occuring so + # we don't need to deal with it in case we save the NaN in a buffer in + # backprop, if it is ever supported, but it doesn't hurt to do so. + eqc = maxc == minc + + cr = maxc - minc + # Since `eqc => cr = 0`, replacing denominator with 1 when `eqc` is fine. + ones = torch.ones_like(maxc) + s = cr / torch.where(eqc, ones, maxc) + # Note that `eqc => maxc = minc = r = g = b`. So the following calculation + # of `h` would reduce to `bc - gc + 2 + rc - bc + 4 + rc - bc = 6` so it + # would not matter what values `rc`, `gc`, and `bc` have here, and thus + # replacing denominator with 1 when `eqc` is fine. + cr_divisor = torch.where(eqc, ones, cr) + rc = (maxc - r) / cr_divisor + gc = (maxc - g) / cr_divisor + bc = (maxc - b) / cr_divisor + + hr = (maxc == r) * (bc - gc) + hg = ((maxc == g) & (maxc != r)) * (2.0 + rc - bc) + hb = ((maxc != g) & (maxc != r)) * (4.0 + gc - rc) + h = (hr + hg + hb) + h = torch.fmod((h / 6.0 + 1.0), 1.0) + return torch.stack((h, s, maxc), dim=-3) + + +def _hsv2rgb(img): + h, s, v = img.unbind(dim=-3) + i = torch.floor(h * 6.0) + f = (h * 6.0) - i + i = i.to(dtype=torch.int32) + + p = torch.clamp((v * (1.0 - s)), 0.0, 1.0) + q = torch.clamp((v * (1.0 - s * f)), 0.0, 1.0) + t = torch.clamp((v * (1.0 - s * (1.0 - f))), 0.0, 1.0) + i = i % 6 + + mask = i.unsqueeze(dim=-3) == torch.arange(6, device=i.device).view(-1, 1, 1) + + a1 = torch.stack((v, q, p, p, t, v), dim=-3) + a2 = torch.stack((t, v, v, q, p, p), dim=-3) + a3 = torch.stack((p, p, t, v, v, q), dim=-3) + a4 = torch.stack((a1, a2, a3), dim=-4) + + return torch.einsum("...ijk, ...xijk -> ...xjk", mask.to(dtype=img.dtype), a4) + + +def _pad_symmetric(img: Tensor, padding: List[int]) -> Tensor: + # padding is left, right, top, bottom + + # crop if needed + if padding[0] < 0 or padding[1] < 0 or padding[2] < 0 or padding[3] < 0: + crop_left, crop_right, crop_top, crop_bottom = [-min(x, 0) for x in padding] + img = img[..., crop_top:img.shape[-2] - crop_bottom, crop_left:img.shape[-1] - crop_right] + padding = [max(x, 0) for x in padding] + + in_sizes = img.size() + + x_indices = [i for i in range(in_sizes[-1])] # [0, 1, 2, 3, ...] + left_indices = [i for i in range(padding[0] - 1, -1, -1)] # e.g. [3, 2, 1, 0] + right_indices = [-(i + 1) for i in range(padding[1])] # e.g. [-1, -2, -3] + x_indices = torch.tensor(left_indices + x_indices + right_indices, device=img.device) + + y_indices = [i for i in range(in_sizes[-2])] + top_indices = [i for i in range(padding[2] - 1, -1, -1)] + bottom_indices = [-(i + 1) for i in range(padding[3])] + y_indices = torch.tensor(top_indices + y_indices + bottom_indices, device=img.device) + + ndim = img.ndim + if ndim == 3: + return img[:, y_indices[:, None], x_indices[None, :]] + elif ndim == 4: + return img[:, :, y_indices[:, None], x_indices[None, :]] + else: + raise RuntimeError("Symmetric padding of N-D tensors are not supported yet") + + +def pad(img: Tensor, padding: List[int], fill: int = 0, padding_mode: str = "constant") -> Tensor: + _assert_image_tensor(img) + + if not isinstance(padding, (int, tuple, list)): + raise TypeError("Got inappropriate padding arg") + if not isinstance(fill, (int, float)): + raise TypeError("Got inappropriate fill arg") + if not isinstance(padding_mode, str): + raise TypeError("Got inappropriate padding_mode arg") + + if isinstance(padding, tuple): + padding = list(padding) + + if isinstance(padding, list) and len(padding) not in [1, 2, 4]: + raise ValueError("Padding must be an int or a 1, 2, or 4 element tuple, not a " + + "{} element tuple".format(len(padding))) + + if padding_mode not in ["constant", "edge", "reflect", "symmetric"]: + raise ValueError("Padding mode should be either constant, edge, reflect or symmetric") + + if isinstance(padding, int): + if torch.jit.is_scripting(): + # This maybe unreachable + raise ValueError("padding can't be an int while torchscripting, set it as a list [value, ]") + pad_left = pad_right = pad_top = pad_bottom = padding + elif len(padding) == 1: + pad_left = pad_right = pad_top = pad_bottom = padding[0] + elif len(padding) == 2: + pad_left = pad_right = padding[0] + pad_top = pad_bottom = padding[1] + else: + pad_left = padding[0] + pad_top = padding[1] + pad_right = padding[2] + pad_bottom = padding[3] + + p = [pad_left, pad_right, pad_top, pad_bottom] + + if padding_mode == "edge": + # remap padding_mode str + padding_mode = "replicate" + elif padding_mode == "symmetric": + # route to another implementation + return _pad_symmetric(img, p) + + need_squeeze = False + if img.ndim < 4: + img = img.unsqueeze(dim=0) + need_squeeze = True + + out_dtype = img.dtype + need_cast = False + if (padding_mode != "constant") and img.dtype not in (torch.float32, torch.float64): + # Here we temporary cast input tensor to float + # until pytorch issue is resolved : + # https://github.com/pytorch/pytorch/issues/40763 + need_cast = True + img = img.to(torch.float32) + + img = torch_pad(img, p, mode=padding_mode, value=float(fill)) + + if need_squeeze: + img = img.squeeze(dim=0) + + if need_cast: + img = img.to(out_dtype) + + return img + + +def resize( + img: Tensor, + size: List[int], + interpolation: str = "bilinear", + max_size: Optional[int] = None, + antialias: Optional[bool] = None +) -> Tensor: + _assert_image_tensor(img) + + if not isinstance(size, (int, tuple, list)): + raise TypeError("Got inappropriate size arg") + if not isinstance(interpolation, str): + raise TypeError("Got inappropriate interpolation arg") + + if interpolation not in ["nearest", "bilinear", "bicubic"]: + raise ValueError("This interpolation mode is unsupported with Tensor input") + + if isinstance(size, tuple): + size = list(size) + + if isinstance(size, list): + if len(size) not in [1, 2]: + raise ValueError("Size must be an int or a 1 or 2 element tuple/list, not a " + "{} element tuple/list".format(len(size))) + if max_size is not None and len(size) != 1: + raise ValueError( + "max_size should only be passed if size specifies the length of the smaller edge, " + "i.e. size should be an int or a sequence of length 1 in torchscript mode." + ) + + if antialias is None: + antialias = False + + if antialias and interpolation not in ["bilinear", "bicubic"]: + raise ValueError("Antialias option is supported for bilinear and bicubic interpolation modes only") + + w, h = _get_image_size(img) + + if isinstance(size, int) or len(size) == 1: # specified size only for the smallest edge + short, long = (w, h) if w <= h else (h, w) + requested_new_short = size if isinstance(size, int) else size[0] + + if short == requested_new_short: + return img + + new_short, new_long = requested_new_short, int(requested_new_short * long / short) + + if max_size is not None: + if max_size <= requested_new_short: + raise ValueError( + f"max_size = {max_size} must be strictly greater than the requested " + f"size for the smaller edge size = {size}" + ) + if new_long > max_size: + new_short, new_long = int(max_size * new_short / new_long), max_size + + new_w, new_h = (new_short, new_long) if w <= h else (new_long, new_short) + + else: # specified both h and w + new_w, new_h = size[1], size[0] + + img, need_cast, need_squeeze, out_dtype = _cast_squeeze_in(img, [torch.float32, torch.float64]) + + # Define align_corners to avoid warnings + align_corners = False if interpolation in ["bilinear", "bicubic"] else None + + if antialias: + if interpolation == "bilinear": + img = torch.ops.torchvision._interpolate_bilinear2d_aa(img, [new_h, new_w], align_corners=False) + elif interpolation == "bicubic": + img = torch.ops.torchvision._interpolate_bicubic2d_aa(img, [new_h, new_w], align_corners=False) + else: + img = interpolate(img, size=[new_h, new_w], mode=interpolation, align_corners=align_corners) + + if interpolation == "bicubic" and out_dtype == torch.uint8: + img = img.clamp(min=0, max=255) + + img = _cast_squeeze_out(img, need_cast=need_cast, need_squeeze=need_squeeze, out_dtype=out_dtype) + + return img + + +def _assert_grid_transform_inputs( + img: Tensor, + matrix: Optional[List[float]], + interpolation: str, + fill: Optional[List[float]], + supported_interpolation_modes: List[str], + coeffs: Optional[List[float]] = None, +): + + if not (isinstance(img, torch.Tensor)): + raise TypeError("Input img should be Tensor") + + _assert_image_tensor(img) + + if matrix is not None and not isinstance(matrix, list): + raise TypeError("Argument matrix should be a list") + + if matrix is not None and len(matrix) != 6: + raise ValueError("Argument matrix should have 6 float values") + + if coeffs is not None and len(coeffs) != 8: + raise ValueError("Argument coeffs should have 8 float values") + + if fill is not None and not isinstance(fill, (int, float, tuple, list)): + warnings.warn("Argument fill should be either int, float, tuple or list") + + # Check fill + num_channels = _get_image_num_channels(img) + if isinstance(fill, (tuple, list)) and (len(fill) > 1 and len(fill) != num_channels): + msg = ("The number of elements in 'fill' cannot broadcast to match the number of " + "channels of the image ({} != {})") + raise ValueError(msg.format(len(fill), num_channels)) + + if interpolation not in supported_interpolation_modes: + raise ValueError("Interpolation mode '{}' is unsupported with Tensor input".format(interpolation)) + + +def _cast_squeeze_in(img: Tensor, req_dtypes: List[torch.dtype]) -> Tuple[Tensor, bool, bool, torch.dtype]: + need_squeeze = False + # make image NCHW + if img.ndim < 4: + img = img.unsqueeze(dim=0) + need_squeeze = True + + out_dtype = img.dtype + need_cast = False + if out_dtype not in req_dtypes: + need_cast = True + req_dtype = req_dtypes[0] + img = img.to(req_dtype) + return img, need_cast, need_squeeze, out_dtype + + +def _cast_squeeze_out(img: Tensor, need_cast: bool, need_squeeze: bool, out_dtype: torch.dtype): + if need_squeeze: + img = img.squeeze(dim=0) + + if need_cast: + if out_dtype in (torch.uint8, torch.int8, torch.int16, torch.int32, torch.int64): + # it is better to round before cast + img = torch.round(img) + img = img.to(out_dtype) + + return img + + +def _apply_grid_transform(img: Tensor, grid: Tensor, mode: str, fill: Optional[List[float]]) -> Tensor: + + img, need_cast, need_squeeze, out_dtype = _cast_squeeze_in(img, [grid.dtype, ]) + + if img.shape[0] > 1: + # Apply same grid to a batch of images + grid = grid.expand(img.shape[0], grid.shape[1], grid.shape[2], grid.shape[3]) + + # Append a dummy mask for customized fill colors, should be faster than grid_sample() twice + if fill is not None: + dummy = torch.ones((img.shape[0], 1, img.shape[2], img.shape[3]), dtype=img.dtype, device=img.device) + img = torch.cat((img, dummy), dim=1) + + img = grid_sample(img, grid, mode=mode, padding_mode="zeros", align_corners=False) + + # Fill with required color + if fill is not None: + mask = img[:, -1:, :, :] # N * 1 * H * W + img = img[:, :-1, :, :] # N * C * H * W + mask = mask.expand_as(img) + len_fill = len(fill) if isinstance(fill, (tuple, list)) else 1 + fill_img = torch.tensor(fill, dtype=img.dtype, device=img.device).view(1, len_fill, 1, 1).expand_as(img) + if mode == 'nearest': + mask = mask < 0.5 + img[mask] = fill_img[mask] + else: # 'bilinear' + img = img * mask + (1.0 - mask) * fill_img + + img = _cast_squeeze_out(img, need_cast, need_squeeze, out_dtype) + return img + + +def _gen_affine_grid( + theta: Tensor, w: int, h: int, ow: int, oh: int, +) -> Tensor: + # https://github.com/pytorch/pytorch/blob/74b65c32be68b15dc7c9e8bb62459efbfbde33d8/aten/src/ATen/native/ + # AffineGridGenerator.cpp#L18 + # Difference with AffineGridGenerator is that: + # 1) we normalize grid values after applying theta + # 2) we can normalize by other image size, such that it covers "extend" option like in PIL.Image.rotate + + d = 0.5 + base_grid = torch.empty(1, oh, ow, 3, dtype=theta.dtype, device=theta.device) + x_grid = torch.linspace(-ow * 0.5 + d, ow * 0.5 + d - 1, steps=ow, device=theta.device) + base_grid[..., 0].copy_(x_grid) + y_grid = torch.linspace(-oh * 0.5 + d, oh * 0.5 + d - 1, steps=oh, device=theta.device).unsqueeze_(-1) + base_grid[..., 1].copy_(y_grid) + base_grid[..., 2].fill_(1) + + rescaled_theta = theta.transpose(1, 2) / torch.tensor([0.5 * w, 0.5 * h], dtype=theta.dtype, device=theta.device) + output_grid = base_grid.view(1, oh * ow, 3).bmm(rescaled_theta) + return output_grid.view(1, oh, ow, 2) + + +def affine( + img: Tensor, matrix: List[float], interpolation: str = "nearest", fill: Optional[List[float]] = None +) -> Tensor: + _assert_grid_transform_inputs(img, matrix, interpolation, fill, ["nearest", "bilinear"]) + + dtype = img.dtype if torch.is_floating_point(img) else torch.float32 + theta = torch.tensor(matrix, dtype=dtype, device=img.device).reshape(1, 2, 3) + shape = img.shape + # grid will be generated on the same device as theta and img + grid = _gen_affine_grid(theta, w=shape[-1], h=shape[-2], ow=shape[-1], oh=shape[-2]) + return _apply_grid_transform(img, grid, interpolation, fill=fill) + + +def _compute_output_size(matrix: List[float], w: int, h: int) -> Tuple[int, int]: + + # Inspired of PIL implementation: + # https://github.com/python-pillow/Pillow/blob/11de3318867e4398057373ee9f12dcb33db7335c/src/PIL/Image.py#L2054 + + # pts are Top-Left, Top-Right, Bottom-Left, Bottom-Right points. + pts = torch.tensor([ + [-0.5 * w, -0.5 * h, 1.0], + [-0.5 * w, 0.5 * h, 1.0], + [0.5 * w, 0.5 * h, 1.0], + [0.5 * w, -0.5 * h, 1.0], + ]) + theta = torch.tensor(matrix, dtype=torch.float).reshape(1, 2, 3) + new_pts = pts.view(1, 4, 3).bmm(theta.transpose(1, 2)).view(4, 2) + min_vals, _ = new_pts.min(dim=0) + max_vals, _ = new_pts.max(dim=0) + + # Truncate precision to 1e-4 to avoid ceil of Xe-15 to 1.0 + tol = 1e-4 + cmax = torch.ceil((max_vals / tol).trunc_() * tol) + cmin = torch.floor((min_vals / tol).trunc_() * tol) + size = cmax - cmin + return int(size[0]), int(size[1]) + + +def rotate( + img: Tensor, matrix: List[float], interpolation: str = "nearest", + expand: bool = False, fill: Optional[List[float]] = None +) -> Tensor: + _assert_grid_transform_inputs(img, matrix, interpolation, fill, ["nearest", "bilinear"]) + w, h = img.shape[-1], img.shape[-2] + ow, oh = _compute_output_size(matrix, w, h) if expand else (w, h) + dtype = img.dtype if torch.is_floating_point(img) else torch.float32 + theta = torch.tensor(matrix, dtype=dtype, device=img.device).reshape(1, 2, 3) + # grid will be generated on the same device as theta and img + grid = _gen_affine_grid(theta, w=w, h=h, ow=ow, oh=oh) + + return _apply_grid_transform(img, grid, interpolation, fill=fill) + + +def _perspective_grid(coeffs: List[float], ow: int, oh: int, dtype: torch.dtype, device: torch.device): + # https://github.com/python-pillow/Pillow/blob/4634eafe3c695a014267eefdce830b4a825beed7/ + # src/libImaging/Geometry.c#L394 + + # + # x_out = (coeffs[0] * x + coeffs[1] * y + coeffs[2]) / (coeffs[6] * x + coeffs[7] * y + 1) + # y_out = (coeffs[3] * x + coeffs[4] * y + coeffs[5]) / (coeffs[6] * x + coeffs[7] * y + 1) + # + theta1 = torch.tensor([[ + [coeffs[0], coeffs[1], coeffs[2]], + [coeffs[3], coeffs[4], coeffs[5]] + ]], dtype=dtype, device=device) + theta2 = torch.tensor([[ + [coeffs[6], coeffs[7], 1.0], + [coeffs[6], coeffs[7], 1.0] + ]], dtype=dtype, device=device) + + d = 0.5 + base_grid = torch.empty(1, oh, ow, 3, dtype=dtype, device=device) + x_grid = torch.linspace(d, ow * 1.0 + d - 1.0, steps=ow, device=device) + base_grid[..., 0].copy_(x_grid) + y_grid = torch.linspace(d, oh * 1.0 + d - 1.0, steps=oh, device=device).unsqueeze_(-1) + base_grid[..., 1].copy_(y_grid) + base_grid[..., 2].fill_(1) + + rescaled_theta1 = theta1.transpose(1, 2) / torch.tensor([0.5 * ow, 0.5 * oh], dtype=dtype, device=device) + output_grid1 = base_grid.view(1, oh * ow, 3).bmm(rescaled_theta1) + output_grid2 = base_grid.view(1, oh * ow, 3).bmm(theta2.transpose(1, 2)) + + output_grid = output_grid1 / output_grid2 - 1.0 + return output_grid.view(1, oh, ow, 2) + + +def perspective( + img: Tensor, perspective_coeffs: List[float], interpolation: str = "bilinear", fill: Optional[List[float]] = None +) -> Tensor: + if not (isinstance(img, torch.Tensor)): + raise TypeError('Input img should be Tensor.') + + _assert_image_tensor(img) + + _assert_grid_transform_inputs( + img, + matrix=None, + interpolation=interpolation, + fill=fill, + supported_interpolation_modes=["nearest", "bilinear"], + coeffs=perspective_coeffs + ) + + ow, oh = img.shape[-1], img.shape[-2] + dtype = img.dtype if torch.is_floating_point(img) else torch.float32 + grid = _perspective_grid(perspective_coeffs, ow=ow, oh=oh, dtype=dtype, device=img.device) + return _apply_grid_transform(img, grid, interpolation, fill=fill) + + +def _get_gaussian_kernel1d(kernel_size: int, sigma: float) -> Tensor: + ksize_half = (kernel_size - 1) * 0.5 + + x = torch.linspace(-ksize_half, ksize_half, steps=kernel_size) + pdf = torch.exp(-0.5 * (x / sigma).pow(2)) + kernel1d = pdf / pdf.sum() + + return kernel1d + + +def _get_gaussian_kernel2d( + kernel_size: List[int], sigma: List[float], dtype: torch.dtype, device: torch.device +) -> Tensor: + kernel1d_x = _get_gaussian_kernel1d(kernel_size[0], sigma[0]).to(device, dtype=dtype) + kernel1d_y = _get_gaussian_kernel1d(kernel_size[1], sigma[1]).to(device, dtype=dtype) + kernel2d = torch.mm(kernel1d_y[:, None], kernel1d_x[None, :]) + return kernel2d + + +def gaussian_blur(img: Tensor, kernel_size: List[int], sigma: List[float]) -> Tensor: + if not (isinstance(img, torch.Tensor)): + raise TypeError('img should be Tensor. Got {}'.format(type(img))) + + _assert_image_tensor(img) + + dtype = img.dtype if torch.is_floating_point(img) else torch.float32 + kernel = _get_gaussian_kernel2d(kernel_size, sigma, dtype=dtype, device=img.device) + kernel = kernel.expand(img.shape[-3], 1, kernel.shape[0], kernel.shape[1]) + + img, need_cast, need_squeeze, out_dtype = _cast_squeeze_in(img, [kernel.dtype, ]) + + # padding = (left, right, top, bottom) + padding = [kernel_size[0] // 2, kernel_size[0] // 2, kernel_size[1] // 2, kernel_size[1] // 2] + img = torch_pad(img, padding, mode="reflect") + img = conv2d(img, kernel, groups=img.shape[-3]) + + img = _cast_squeeze_out(img, need_cast, need_squeeze, out_dtype) + return img + + +def invert(img: Tensor) -> Tensor: + + _assert_image_tensor(img) + + if img.ndim < 3: + raise TypeError("Input image tensor should have at least 3 dimensions, but found {}".format(img.ndim)) + + _assert_channels(img, [1, 3]) + + bound = torch.tensor(1 if img.is_floating_point() else 255, dtype=img.dtype, device=img.device) + return bound - img + + +def posterize(img: Tensor, bits: int) -> Tensor: + + _assert_image_tensor(img) + + if img.ndim < 3: + raise TypeError("Input image tensor should have at least 3 dimensions, but found {}".format(img.ndim)) + if img.dtype != torch.uint8: + raise TypeError("Only torch.uint8 image tensors are supported, but found {}".format(img.dtype)) + + _assert_channels(img, [1, 3]) + mask = -int(2**(8 - bits)) # JIT-friendly for: ~(2 ** (8 - bits) - 1) + return img & mask + + +def solarize(img: Tensor, threshold: float) -> Tensor: + + _assert_image_tensor(img) + + if img.ndim < 3: + raise TypeError("Input image tensor should have at least 3 dimensions, but found {}".format(img.ndim)) + + _assert_channels(img, [1, 3]) + + inverted_img = invert(img) + return torch.where(img >= threshold, inverted_img, img) + + +def _blurred_degenerate_image(img: Tensor) -> Tensor: + dtype = img.dtype if torch.is_floating_point(img) else torch.float32 + + kernel = torch.ones((3, 3), dtype=dtype, device=img.device) + kernel[1, 1] = 5.0 + kernel /= kernel.sum() + kernel = kernel.expand(img.shape[-3], 1, kernel.shape[0], kernel.shape[1]) + + result_tmp, need_cast, need_squeeze, out_dtype = _cast_squeeze_in(img, [kernel.dtype, ]) + result_tmp = conv2d(result_tmp, kernel, groups=result_tmp.shape[-3]) + result_tmp = _cast_squeeze_out(result_tmp, need_cast, need_squeeze, out_dtype) + + result = img.clone() + result[..., 1:-1, 1:-1] = result_tmp + + return result + + +def adjust_sharpness(img: Tensor, sharpness_factor: float) -> Tensor: + if sharpness_factor < 0: + raise ValueError('sharpness_factor ({}) is not non-negative.'.format(sharpness_factor)) + + _assert_image_tensor(img) + + _assert_channels(img, [1, 3]) + + if img.size(-1) <= 2 or img.size(-2) <= 2: + return img + + return _blend(img, _blurred_degenerate_image(img), sharpness_factor) + + +def autocontrast(img: Tensor) -> Tensor: + + _assert_image_tensor(img) + + if img.ndim < 3: + raise TypeError("Input image tensor should have at least 3 dimensions, but found {}".format(img.ndim)) + + _assert_channels(img, [1, 3]) + + bound = 1.0 if img.is_floating_point() else 255.0 + dtype = img.dtype if torch.is_floating_point(img) else torch.float32 + + minimum = img.amin(dim=(-2, -1), keepdim=True).to(dtype) + maximum = img.amax(dim=(-2, -1), keepdim=True).to(dtype) + eq_idxs = torch.where(minimum == maximum)[0] + minimum[eq_idxs] = 0 + maximum[eq_idxs] = bound + scale = bound / (maximum - minimum) + + return ((img - minimum) * scale).clamp(0, bound).to(img.dtype) + + +def _scale_channel(img_chan): + # TODO: we should expect bincount to always be faster than histc, but this + # isn't always the case. Once + # https://github.com/pytorch/pytorch/issues/53194 is fixed, remove the if + # block and only use bincount. + if img_chan.is_cuda: + hist = torch.histc(img_chan.to(torch.float32), bins=256, min=0, max=255) + else: + hist = torch.bincount(img_chan.view(-1), minlength=256) + + nonzero_hist = hist[hist != 0] + step = torch.div(nonzero_hist[:-1].sum(), 255, rounding_mode='floor') + if step == 0: + return img_chan + + lut = torch.div( + torch.cumsum(hist, 0) + torch.div(step, 2, rounding_mode='floor'), + step, rounding_mode='floor') + lut = torch.nn.functional.pad(lut, [1, 0])[:-1].clamp(0, 255) + + return lut[img_chan.to(torch.int64)].to(torch.uint8) + + +def _equalize_single_image(img: Tensor) -> Tensor: + return torch.stack([_scale_channel(img[c]) for c in range(img.size(0))]) + + +def equalize(img: Tensor) -> Tensor: + + _assert_image_tensor(img) + + if not (3 <= img.ndim <= 4): + raise TypeError("Input image tensor should have 3 or 4 dimensions, but found {}".format(img.ndim)) + if img.dtype != torch.uint8: + raise TypeError("Only torch.uint8 image tensors are supported, but found {}".format(img.dtype)) + + _assert_channels(img, [1, 3]) + + if img.ndim == 3: + return _equalize_single_image(img) + + return torch.stack([_equalize_single_image(x) for x in img]) diff --git a/detectron2/data/transforms/torchvision_transforms/transforms.py b/detectron2/data/transforms/torchvision_transforms/transforms.py new file mode 100644 index 0000000000000000000000000000000000000000..954d5f5f06490309eeace247bc14ce101095ae9f --- /dev/null +++ b/detectron2/data/transforms/torchvision_transforms/transforms.py @@ -0,0 +1,1955 @@ +import math +import numbers +import random +import warnings +from collections.abc import Sequence +from typing import Tuple, List, Optional + +import torch +from torch import Tensor + +try: + import accimage +except ImportError: + accimage = None + +from . import functional as F +from .functional import InterpolationMode, _interpolation_modes_from_int + + +__all__ = ["Compose", "ToTensor", "PILToTensor", "ConvertImageDtype", "ToPILImage", "Normalize", "Resize", "Scale", + "CenterCrop", "Pad", "Lambda", "RandomApply", "RandomChoice", "RandomOrder", "RandomCrop", + "RandomHorizontalFlip", "RandomVerticalFlip", "RandomResizedCrop", "RandomSizedCrop", "FiveCrop", "TenCrop", + "LinearTransformation", "ColorJitter", "RandomRotation", "RandomAffine", "Grayscale", "RandomGrayscale", + "RandomPerspective", "RandomErasing", "GaussianBlur", "InterpolationMode", "RandomInvert", "RandomPosterize", + "RandomSolarize", "RandomAdjustSharpness", "RandomAutocontrast", "RandomEqualize"] + + +class Compose: + """Composes several transforms together. This transform does not support torchscript. + Please, see the note below. + + Args: + transforms (list of ``Transform`` objects): list of transforms to compose. + + Example: + >>> transforms.Compose([ + >>> transforms.CenterCrop(10), + >>> transforms.ToTensor(), + >>> ]) + + .. note:: + In order to script the transformations, please use ``torch.nn.Sequential`` as below. + + >>> transforms = torch.nn.Sequential( + >>> transforms.CenterCrop(10), + >>> transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), + >>> ) + >>> scripted_transforms = torch.jit.script(transforms) + + Make sure to use only scriptable transformations, i.e. that work with ``torch.Tensor``, does not require + `lambda` functions or ``PIL.Image``. + + """ + + def __init__(self, transforms): + self.transforms = transforms + + def __call__(self, img): + for t in self.transforms: + img = t(img) + return img + + def __repr__(self): + format_string = self.__class__.__name__ + '(' + for t in self.transforms: + format_string += '\n' + format_string += ' {0}'.format(t) + format_string += '\n)' + return format_string + + +class ToTensor: + """Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor. This transform does not support torchscript. + + Converts a PIL Image or numpy.ndarray (H x W x C) in the range + [0, 255] to a torch.FloatTensor of shape (C x H x W) in the range [0.0, 1.0] + if the PIL Image belongs to one of the modes (L, LA, P, I, F, RGB, YCbCr, RGBA, CMYK, 1) + or if the numpy.ndarray has dtype = np.uint8 + + In the other cases, tensors are returned without scaling. + + .. note:: + Because the input image is scaled to [0.0, 1.0], this transformation should not be used when + transforming target image masks. See the `references`_ for implementing the transforms for image masks. + + .. _references: https://github.com/pytorch/vision/tree/master/references/segmentation + """ + + def __call__(self, pic): + """ + Args: + pic (PIL Image or numpy.ndarray): Image to be converted to tensor. + + Returns: + Tensor: Converted image. + """ + return F.to_tensor(pic) + + def __repr__(self): + return self.__class__.__name__ + '()' + + +class PILToTensor: + """Convert a ``PIL Image`` to a tensor of the same type. This transform does not support torchscript. + + Converts a PIL Image (H x W x C) to a Tensor of shape (C x H x W). + """ + + def __call__(self, pic): + """ + Args: + pic (PIL Image): Image to be converted to tensor. + + Returns: + Tensor: Converted image. + """ + return F.pil_to_tensor(pic) + + def __repr__(self): + return self.__class__.__name__ + '()' + + +class ConvertImageDtype(torch.nn.Module): + """Convert a tensor image to the given ``dtype`` and scale the values accordingly + This function does not support PIL Image. + + Args: + dtype (torch.dtype): Desired data type of the output + + .. note:: + + When converting from a smaller to a larger integer ``dtype`` the maximum values are **not** mapped exactly. + If converted back and forth, this mismatch has no effect. + + Raises: + RuntimeError: When trying to cast :class:`torch.float32` to :class:`torch.int32` or :class:`torch.int64` as + well as for trying to cast :class:`torch.float64` to :class:`torch.int64`. These conversions might lead to + overflow errors since the floating point ``dtype`` cannot store consecutive integers over the whole range + of the integer ``dtype``. + """ + + def __init__(self, dtype: torch.dtype) -> None: + super().__init__() + self.dtype = dtype + + def forward(self, image): + return F.convert_image_dtype(image, self.dtype) + + +class ToPILImage: + """Convert a tensor or an ndarray to PIL Image. This transform does not support torchscript. + + Converts a torch.*Tensor of shape C x H x W or a numpy ndarray of shape + H x W x C to a PIL Image while preserving the value range. + + Args: + mode (`PIL.Image mode`_): color space and pixel depth of input data (optional). + If ``mode`` is ``None`` (default) there are some assumptions made about the input data: + - If the input has 4 channels, the ``mode`` is assumed to be ``RGBA``. + - If the input has 3 channels, the ``mode`` is assumed to be ``RGB``. + - If the input has 2 channels, the ``mode`` is assumed to be ``LA``. + - If the input has 1 channel, the ``mode`` is determined by the data type (i.e ``int``, ``float``, + ``short``). + + .. _PIL.Image mode: https://pillow.readthedocs.io/en/latest/handbook/concepts.html#concept-modes + """ + def __init__(self, mode=None): + self.mode = mode + + def __call__(self, pic): + """ + Args: + pic (Tensor or numpy.ndarray): Image to be converted to PIL Image. + + Returns: + PIL Image: Image converted to PIL Image. + + """ + return F.to_pil_image(pic, self.mode) + + def __repr__(self): + format_string = self.__class__.__name__ + '(' + if self.mode is not None: + format_string += 'mode={0}'.format(self.mode) + format_string += ')' + return format_string + + +class Normalize(torch.nn.Module): + """Normalize a tensor image with mean and standard deviation. + This transform does not support PIL Image. + Given mean: ``(mean[1],...,mean[n])`` and std: ``(std[1],..,std[n])`` for ``n`` + channels, this transform will normalize each channel of the input + ``torch.*Tensor`` i.e., + ``output[channel] = (input[channel] - mean[channel]) / std[channel]`` + + .. note:: + This transform acts out of place, i.e., it does not mutate the input tensor. + + Args: + mean (sequence): Sequence of means for each channel. + std (sequence): Sequence of standard deviations for each channel. + inplace(bool,optional): Bool to make this operation in-place. + + """ + + def __init__(self, mean, std, inplace=False): + super().__init__() + self.mean = mean + self.std = std + self.inplace = inplace + + def forward(self, tensor: Tensor) -> Tensor: + """ + Args: + tensor (Tensor): Tensor image to be normalized. + + Returns: + Tensor: Normalized Tensor image. + """ + return F.normalize(tensor, self.mean, self.std, self.inplace) + + def __repr__(self): + return self.__class__.__name__ + '(mean={0}, std={1})'.format(self.mean, self.std) + + +class Resize(torch.nn.Module): + """Resize the input image to the given size. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions + + .. warning:: + The output image might be different depending on its type: when downsampling, the interpolation of PIL images + and tensors is slightly different, because PIL applies antialiasing. This may lead to significant differences + in the performance of a network. Therefore, it is preferable to train and serve a model with the same input + types. See also below the ``antialias`` parameter, which can help making the output of PIL images and tensors + closer. + + Args: + size (sequence or int): Desired output size. If size is a sequence like + (h, w), output size will be matched to this. If size is an int, + smaller edge of the image will be matched to this number. + i.e, if height > width, then image will be rescaled to + (size * height / width, size). + + .. note:: + In torchscript mode size as single int is not supported, use a sequence of length 1: ``[size, ]``. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.BILINEAR``. + If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` and + ``InterpolationMode.BICUBIC`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + max_size (int, optional): The maximum allowed for the longer edge of + the resized image: if the longer edge of the image is greater + than ``max_size`` after being resized according to ``size``, then + the image is resized again so that the longer edge is equal to + ``max_size``. As a result, ``size`` might be overruled, i.e the + smaller edge may be shorter than ``size``. This is only supported + if ``size`` is an int (or a sequence of length 1 in torchscript + mode). + antialias (bool, optional): antialias flag. If ``img`` is PIL Image, the flag is ignored and anti-alias + is always used. If ``img`` is Tensor, the flag is False by default and can be set to True for + ``InterpolationMode.BILINEAR`` only mode. This can help making the output for PIL images and tensors + closer. + + .. warning:: + There is no autodiff support for ``antialias=True`` option with input ``img`` as Tensor. + + """ + + def __init__(self, size, interpolation=InterpolationMode.BILINEAR, max_size=None, antialias=None): + super().__init__() + if not isinstance(size, (int, Sequence)): + raise TypeError("Size should be int or sequence. Got {}".format(type(size))) + if isinstance(size, Sequence) and len(size) not in (1, 2): + raise ValueError("If size is a sequence, it should have 1 or 2 values") + self.size = size + self.max_size = max_size + + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + self.interpolation = interpolation + self.antialias = antialias + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be scaled. + + Returns: + PIL Image or Tensor: Rescaled image. + """ + return F.resize(img, self.size, self.interpolation, self.max_size, self.antialias) + + def __repr__(self): + interpolate_str = self.interpolation.value + return self.__class__.__name__ + '(size={0}, interpolation={1}, max_size={2}, antialias={3})'.format( + self.size, interpolate_str, self.max_size, self.antialias) + + +class Scale(Resize): + """ + Note: This transform is deprecated in favor of Resize. + """ + def __init__(self, *args, **kwargs): + warnings.warn("The use of the transforms.Scale transform is deprecated, " + + "please use transforms.Resize instead.") + super(Scale, self).__init__(*args, **kwargs) + + +class CenterCrop(torch.nn.Module): + """Crops the given image at the center. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + If image size is smaller than output size along any edge, image is padded with 0 and then center cropped. + + Args: + size (sequence or int): Desired output size of the crop. If size is an + int instead of sequence like (h, w), a square crop (size, size) is + made. If provided a sequence of length 1, it will be interpreted as (size[0], size[0]). + """ + + def __init__(self, size): + super().__init__() + self.size = _setup_size(size, error_msg="Please provide only two dimensions (h, w) for size.") + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be cropped. + + Returns: + PIL Image or Tensor: Cropped image. + """ + return F.center_crop(img, self.size) + + def __repr__(self): + return self.__class__.__name__ + '(size={0})'.format(self.size) + + +class Pad(torch.nn.Module): + """Pad the given image on all sides with the given "pad" value. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means at most 2 leading dimensions for mode reflect and symmetric, + at most 3 leading dimensions for mode edge, + and an arbitrary number of leading dimensions for mode constant + + Args: + padding (int or sequence): Padding on each border. If a single int is provided this + is used to pad all borders. If sequence of length 2 is provided this is the padding + on left/right and top/bottom respectively. If a sequence of length 4 is provided + this is the padding for the left, top, right and bottom borders respectively. + + .. note:: + In torchscript mode padding as single int is not supported, use a sequence of + length 1: ``[padding, ]``. + fill (number or str or tuple): Pixel fill value for constant fill. Default is 0. If a tuple of + length 3, it is used to fill R, G, B channels respectively. + This value is only used when the padding_mode is constant. + Only number is supported for torch Tensor. + Only int or str or tuple value is supported for PIL Image. + padding_mode (str): Type of padding. Should be: constant, edge, reflect or symmetric. + Default is constant. + + - constant: pads with a constant value, this value is specified with fill + + - edge: pads with the last value at the edge of the image. + If input a 5D torch Tensor, the last 3 dimensions will be padded instead of the last 2 + + - reflect: pads with reflection of image without repeating the last value on the edge. + For example, padding [1, 2, 3, 4] with 2 elements on both sides in reflect mode + will result in [3, 2, 1, 2, 3, 4, 3, 2] + + - symmetric: pads with reflection of image repeating the last value on the edge. + For example, padding [1, 2, 3, 4] with 2 elements on both sides in symmetric mode + will result in [2, 1, 1, 2, 3, 4, 4, 3] + """ + + def __init__(self, padding, fill=0, padding_mode="constant"): + super().__init__() + if not isinstance(padding, (numbers.Number, tuple, list)): + raise TypeError("Got inappropriate padding arg") + + if not isinstance(fill, (numbers.Number, str, tuple)): + raise TypeError("Got inappropriate fill arg") + + if padding_mode not in ["constant", "edge", "reflect", "symmetric"]: + raise ValueError("Padding mode should be either constant, edge, reflect or symmetric") + + if isinstance(padding, Sequence) and len(padding) not in [1, 2, 4]: + raise ValueError("Padding must be an int or a 1, 2, or 4 element tuple, not a " + + "{} element tuple".format(len(padding))) + + self.padding = padding + self.fill = fill + self.padding_mode = padding_mode + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be padded. + + Returns: + PIL Image or Tensor: Padded image. + """ + return F.pad(img, self.padding, self.fill, self.padding_mode) + + def __repr__(self): + return self.__class__.__name__ + '(padding={0}, fill={1}, padding_mode={2})'.\ + format(self.padding, self.fill, self.padding_mode) + + +class Lambda: + """Apply a user-defined lambda as a transform. This transform does not support torchscript. + + Args: + lambd (function): Lambda/function to be used for transform. + """ + + def __init__(self, lambd): + if not callable(lambd): + raise TypeError("Argument lambd should be callable, got {}".format(repr(type(lambd).__name__))) + self.lambd = lambd + + def __call__(self, img): + return self.lambd(img) + + def __repr__(self): + return self.__class__.__name__ + '()' + + +class RandomTransforms: + """Base class for a list of transformations with randomness + + Args: + transforms (sequence): list of transformations + """ + + def __init__(self, transforms): + if not isinstance(transforms, Sequence): + raise TypeError("Argument transforms should be a sequence") + self.transforms = transforms + + def __call__(self, *args, **kwargs): + raise NotImplementedError() + + def __repr__(self): + format_string = self.__class__.__name__ + '(' + for t in self.transforms: + format_string += '\n' + format_string += ' {0}'.format(t) + format_string += '\n)' + return format_string + + +class RandomApply(torch.nn.Module): + """Apply randomly a list of transformations with a given probability. + + .. note:: + In order to script the transformation, please use ``torch.nn.ModuleList`` as input instead of list/tuple of + transforms as shown below: + + >>> transforms = transforms.RandomApply(torch.nn.ModuleList([ + >>> transforms.ColorJitter(), + >>> ]), p=0.3) + >>> scripted_transforms = torch.jit.script(transforms) + + Make sure to use only scriptable transformations, i.e. that work with ``torch.Tensor``, does not require + `lambda` functions or ``PIL.Image``. + + Args: + transforms (sequence or torch.nn.Module): list of transformations + p (float): probability + """ + + def __init__(self, transforms, p=0.5): + super().__init__() + self.transforms = transforms + self.p = p + + def forward(self, img): + if self.p < torch.rand(1): + return img + for t in self.transforms: + img = t(img) + return img + + def __repr__(self): + format_string = self.__class__.__name__ + '(' + format_string += '\n p={}'.format(self.p) + for t in self.transforms: + format_string += '\n' + format_string += ' {0}'.format(t) + format_string += '\n)' + return format_string + + +class RandomOrder(RandomTransforms): + """Apply a list of transformations in a random order. This transform does not support torchscript. + """ + def __call__(self, img): + order = list(range(len(self.transforms))) + random.shuffle(order) + for i in order: + img = self.transforms[i](img) + return img + + +class RandomChoice(RandomTransforms): + """Apply single transformation randomly picked from a list. This transform does not support torchscript. + """ + def __call__(self, img): + t = random.choice(self.transforms) + return t(img) + + +class RandomCrop(torch.nn.Module): + """Crop the given image at a random location. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions, + but if non-constant padding is used, the input is expected to have at most 2 leading dimensions + + Args: + size (sequence or int): Desired output size of the crop. If size is an + int instead of sequence like (h, w), a square crop (size, size) is + made. If provided a sequence of length 1, it will be interpreted as (size[0], size[0]). + padding (int or sequence, optional): Optional padding on each border + of the image. Default is None. If a single int is provided this + is used to pad all borders. If sequence of length 2 is provided this is the padding + on left/right and top/bottom respectively. If a sequence of length 4 is provided + this is the padding for the left, top, right and bottom borders respectively. + + .. note:: + In torchscript mode padding as single int is not supported, use a sequence of + length 1: ``[padding, ]``. + pad_if_needed (boolean): It will pad the image if smaller than the + desired size to avoid raising an exception. Since cropping is done + after padding, the padding seems to be done at a random offset. + fill (number or str or tuple): Pixel fill value for constant fill. Default is 0. If a tuple of + length 3, it is used to fill R, G, B channels respectively. + This value is only used when the padding_mode is constant. + Only number is supported for torch Tensor. + Only int or str or tuple value is supported for PIL Image. + padding_mode (str): Type of padding. Should be: constant, edge, reflect or symmetric. + Default is constant. + + - constant: pads with a constant value, this value is specified with fill + + - edge: pads with the last value at the edge of the image. + If input a 5D torch Tensor, the last 3 dimensions will be padded instead of the last 2 + + - reflect: pads with reflection of image without repeating the last value on the edge. + For example, padding [1, 2, 3, 4] with 2 elements on both sides in reflect mode + will result in [3, 2, 1, 2, 3, 4, 3, 2] + + - symmetric: pads with reflection of image repeating the last value on the edge. + For example, padding [1, 2, 3, 4] with 2 elements on both sides in symmetric mode + will result in [2, 1, 1, 2, 3, 4, 4, 3] + """ + + @staticmethod + def get_params(img: Tensor, output_size: Tuple[int, int]) -> Tuple[int, int, int, int]: + """Get parameters for ``crop`` for a random crop. + + Args: + img (PIL Image or Tensor): Image to be cropped. + output_size (tuple): Expected output size of the crop. + + Returns: + tuple: params (i, j, h, w) to be passed to ``crop`` for random crop. + """ + w, h = F._get_image_size(img) + th, tw = output_size + + if h + 1 < th or w + 1 < tw: + raise ValueError( + "Required crop size {} is larger then input image size {}".format((th, tw), (h, w)) + ) + + if w == tw and h == th: + return 0, 0, h, w + + i = torch.randint(0, h - th + 1, size=(1, )).item() + j = torch.randint(0, w - tw + 1, size=(1, )).item() + return i, j, th, tw + + def __init__(self, size, padding=None, pad_if_needed=False, fill=0, padding_mode="constant"): + super().__init__() + + self.size = tuple(_setup_size( + size, error_msg="Please provide only two dimensions (h, w) for size." + )) + + self.padding = padding + self.pad_if_needed = pad_if_needed + self.fill = fill + self.padding_mode = padding_mode + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be cropped. + + Returns: + PIL Image or Tensor: Cropped image. + """ + if self.padding is not None: + img = F.pad(img, self.padding, self.fill, self.padding_mode) + + width, height = F._get_image_size(img) + # pad the width if needed + if self.pad_if_needed and width < self.size[1]: + padding = [self.size[1] - width, 0] + img = F.pad(img, padding, self.fill, self.padding_mode) + # pad the height if needed + if self.pad_if_needed and height < self.size[0]: + padding = [0, self.size[0] - height] + img = F.pad(img, padding, self.fill, self.padding_mode) + + i, j, h, w = self.get_params(img, self.size) + + return F.crop(img, i, j, h, w) + + def __repr__(self): + return self.__class__.__name__ + "(size={0}, padding={1})".format(self.size, self.padding) + + +class RandomHorizontalFlip(torch.nn.Module): + """Horizontally flip the given image randomly with a given probability. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading + dimensions + + Args: + p (float): probability of the image being flipped. Default value is 0.5 + """ + + def __init__(self, p=0.5): + super().__init__() + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be flipped. + + Returns: + PIL Image or Tensor: Randomly flipped image. + """ + if torch.rand(1) < self.p: + return F.hflip(img) + return img + + def __repr__(self): + return self.__class__.__name__ + '(p={})'.format(self.p) + + +class RandomVerticalFlip(torch.nn.Module): + """Vertically flip the given image randomly with a given probability. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading + dimensions + + Args: + p (float): probability of the image being flipped. Default value is 0.5 + """ + + def __init__(self, p=0.5): + super().__init__() + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be flipped. + + Returns: + PIL Image or Tensor: Randomly flipped image. + """ + if torch.rand(1) < self.p: + return F.vflip(img) + return img + + def __repr__(self): + return self.__class__.__name__ + '(p={})'.format(self.p) + + +class RandomPerspective(torch.nn.Module): + """Performs a random perspective transformation of the given image with a given probability. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + distortion_scale (float): argument to control the degree of distortion and ranges from 0 to 1. + Default is 0.5. + p (float): probability of the image being transformed. Default is 0.5. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.BILINEAR``. + If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + fill (sequence or number): Pixel fill value for the area outside the transformed + image. Default is ``0``. If given a number, the value is used for all bands respectively. + """ + + def __init__(self, distortion_scale=0.5, p=0.5, interpolation=InterpolationMode.BILINEAR, fill=0): + super().__init__() + self.p = p + + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + self.interpolation = interpolation + self.distortion_scale = distortion_scale + + if fill is None: + fill = 0 + elif not isinstance(fill, (Sequence, numbers.Number)): + raise TypeError("Fill should be either a sequence or a number.") + + self.fill = fill + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be Perspectively transformed. + + Returns: + PIL Image or Tensor: Randomly transformed image. + """ + + fill = self.fill + if isinstance(img, Tensor): + if isinstance(fill, (int, float)): + fill = [float(fill)] * F._get_image_num_channels(img) + else: + fill = [float(f) for f in fill] + + if torch.rand(1) < self.p: + width, height = F._get_image_size(img) + startpoints, endpoints = self.get_params(width, height, self.distortion_scale) + return F.perspective(img, startpoints, endpoints, self.interpolation, fill) + return img + + @staticmethod + def get_params(width: int, height: int, distortion_scale: float) -> Tuple[List[List[int]], List[List[int]]]: + """Get parameters for ``perspective`` for a random perspective transform. + + Args: + width (int): width of the image. + height (int): height of the image. + distortion_scale (float): argument to control the degree of distortion and ranges from 0 to 1. + + Returns: + List containing [top-left, top-right, bottom-right, bottom-left] of the original image, + List containing [top-left, top-right, bottom-right, bottom-left] of the transformed image. + """ + half_height = height // 2 + half_width = width // 2 + topleft = [ + int(torch.randint(0, int(distortion_scale * half_width) + 1, size=(1, )).item()), + int(torch.randint(0, int(distortion_scale * half_height) + 1, size=(1, )).item()) + ] + topright = [ + int(torch.randint(width - int(distortion_scale * half_width) - 1, width, size=(1, )).item()), + int(torch.randint(0, int(distortion_scale * half_height) + 1, size=(1, )).item()) + ] + botright = [ + int(torch.randint(width - int(distortion_scale * half_width) - 1, width, size=(1, )).item()), + int(torch.randint(height - int(distortion_scale * half_height) - 1, height, size=(1, )).item()) + ] + botleft = [ + int(torch.randint(0, int(distortion_scale * half_width) + 1, size=(1, )).item()), + int(torch.randint(height - int(distortion_scale * half_height) - 1, height, size=(1, )).item()) + ] + startpoints = [[0, 0], [width - 1, 0], [width - 1, height - 1], [0, height - 1]] + endpoints = [topleft, topright, botright, botleft] + return startpoints, endpoints + + def __repr__(self): + return self.__class__.__name__ + '(p={})'.format(self.p) + + +class RandomResizedCrop(torch.nn.Module): + """Crop a random portion of image and resize it to a given size. + + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions + + A crop of the original image is made: the crop has a random area (H * W) + and a random aspect ratio. This crop is finally resized to the given + size. This is popularly used to train the Inception networks. + + Args: + size (int or sequence): expected output size of the crop, for each edge. If size is an + int instead of sequence like (h, w), a square output size ``(size, size)`` is + made. If provided a sequence of length 1, it will be interpreted as (size[0], size[0]). + + .. note:: + In torchscript mode size as single int is not supported, use a sequence of length 1: ``[size, ]``. + scale (tuple of float): Specifies the lower and upper bounds for the random area of the crop, + before resizing. The scale is defined with respect to the area of the original image. + ratio (tuple of float): lower and upper bounds for the random aspect ratio of the crop, before + resizing. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.BILINEAR``. + If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` and + ``InterpolationMode.BICUBIC`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + + """ + + def __init__(self, size, scale=(0.08, 1.0), ratio=(3. / 4., 4. / 3.), interpolation=InterpolationMode.BILINEAR): + super().__init__() + self.size = _setup_size(size, error_msg="Please provide only two dimensions (h, w) for size.") + + if not isinstance(scale, Sequence): + raise TypeError("Scale should be a sequence") + if not isinstance(ratio, Sequence): + raise TypeError("Ratio should be a sequence") + if (scale[0] > scale[1]) or (ratio[0] > ratio[1]): + warnings.warn("Scale and ratio should be of kind (min, max)") + + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + self.interpolation = interpolation + self.scale = scale + self.ratio = ratio + + @staticmethod + def get_params( + img: Tensor, scale: List[float], ratio: List[float] + ) -> Tuple[int, int, int, int]: + """Get parameters for ``crop`` for a random sized crop. + + Args: + img (PIL Image or Tensor): Input image. + scale (list): range of scale of the origin size cropped + ratio (list): range of aspect ratio of the origin aspect ratio cropped + + Returns: + tuple: params (i, j, h, w) to be passed to ``crop`` for a random + sized crop. + """ + width, height = F._get_image_size(img) + area = height * width + + log_ratio = torch.log(torch.tensor(ratio)) + for _ in range(10): + target_area = area * torch.empty(1).uniform_(scale[0], scale[1]).item() + aspect_ratio = torch.exp( + torch.empty(1).uniform_(log_ratio[0], log_ratio[1]) + ).item() + + w = int(round(math.sqrt(target_area * aspect_ratio))) + h = int(round(math.sqrt(target_area / aspect_ratio))) + + if 0 < w <= width and 0 < h <= height: + i = torch.randint(0, height - h + 1, size=(1,)).item() + j = torch.randint(0, width - w + 1, size=(1,)).item() + return i, j, h, w + + # Fallback to central crop + in_ratio = float(width) / float(height) + if in_ratio < min(ratio): + w = width + h = int(round(w / min(ratio))) + elif in_ratio > max(ratio): + h = height + w = int(round(h * max(ratio))) + else: # whole image + w = width + h = height + i = (height - h) // 2 + j = (width - w) // 2 + return i, j, h, w + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be cropped and resized. + + Returns: + PIL Image or Tensor: Randomly cropped and resized image. + """ + i, j, h, w = self.get_params(img, self.scale, self.ratio) + return F.resized_crop(img, i, j, h, w, self.size, self.interpolation) + + def __repr__(self): + interpolate_str = self.interpolation.value + format_string = self.__class__.__name__ + '(size={0}'.format(self.size) + format_string += ', scale={0}'.format(tuple(round(s, 4) for s in self.scale)) + format_string += ', ratio={0}'.format(tuple(round(r, 4) for r in self.ratio)) + format_string += ', interpolation={0})'.format(interpolate_str) + return format_string + + +class RandomSizedCrop(RandomResizedCrop): + """ + Note: This transform is deprecated in favor of RandomResizedCrop. + """ + def __init__(self, *args, **kwargs): + warnings.warn("The use of the transforms.RandomSizedCrop transform is deprecated, " + + "please use transforms.RandomResizedCrop instead.") + super(RandomSizedCrop, self).__init__(*args, **kwargs) + + +class FiveCrop(torch.nn.Module): + """Crop the given image into four corners and the central crop. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading + dimensions + + .. Note:: + This transform returns a tuple of images and there may be a mismatch in the number of + inputs and targets your Dataset returns. See below for an example of how to deal with + this. + + Args: + size (sequence or int): Desired output size of the crop. If size is an ``int`` + instead of sequence like (h, w), a square crop of size (size, size) is made. + If provided a sequence of length 1, it will be interpreted as (size[0], size[0]). + + Example: + >>> transform = Compose([ + >>> FiveCrop(size), # this is a list of PIL Images + >>> Lambda(lambda crops: torch.stack([ToTensor()(crop) for crop in crops])) # returns a 4D tensor + >>> ]) + >>> #In your test loop you can do the following: + >>> input, target = batch # input is a 5d tensor, target is 2d + >>> bs, ncrops, c, h, w = input.size() + >>> result = model(input.view(-1, c, h, w)) # fuse batch size and ncrops + >>> result_avg = result.view(bs, ncrops, -1).mean(1) # avg over crops + """ + + def __init__(self, size): + super().__init__() + self.size = _setup_size(size, error_msg="Please provide only two dimensions (h, w) for size.") + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be cropped. + + Returns: + tuple of 5 images. Image can be PIL Image or Tensor + """ + return F.five_crop(img, self.size) + + def __repr__(self): + return self.__class__.__name__ + '(size={0})'.format(self.size) + + +class TenCrop(torch.nn.Module): + """Crop the given image into four corners and the central crop plus the flipped version of + these (horizontal flipping is used by default). + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading + dimensions + + .. Note:: + This transform returns a tuple of images and there may be a mismatch in the number of + inputs and targets your Dataset returns. See below for an example of how to deal with + this. + + Args: + size (sequence or int): Desired output size of the crop. If size is an + int instead of sequence like (h, w), a square crop (size, size) is + made. If provided a sequence of length 1, it will be interpreted as (size[0], size[0]). + vertical_flip (bool): Use vertical flipping instead of horizontal + + Example: + >>> transform = Compose([ + >>> TenCrop(size), # this is a list of PIL Images + >>> Lambda(lambda crops: torch.stack([ToTensor()(crop) for crop in crops])) # returns a 4D tensor + >>> ]) + >>> #In your test loop you can do the following: + >>> input, target = batch # input is a 5d tensor, target is 2d + >>> bs, ncrops, c, h, w = input.size() + >>> result = model(input.view(-1, c, h, w)) # fuse batch size and ncrops + >>> result_avg = result.view(bs, ncrops, -1).mean(1) # avg over crops + """ + + def __init__(self, size, vertical_flip=False): + super().__init__() + self.size = _setup_size(size, error_msg="Please provide only two dimensions (h, w) for size.") + self.vertical_flip = vertical_flip + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be cropped. + + Returns: + tuple of 10 images. Image can be PIL Image or Tensor + """ + return F.ten_crop(img, self.size, self.vertical_flip) + + def __repr__(self): + return self.__class__.__name__ + '(size={0}, vertical_flip={1})'.format(self.size, self.vertical_flip) + + +class LinearTransformation(torch.nn.Module): + """Transform a tensor image with a square transformation matrix and a mean_vector computed + offline. + This transform does not support PIL Image. + Given transformation_matrix and mean_vector, will flatten the torch.*Tensor and + subtract mean_vector from it which is then followed by computing the dot + product with the transformation matrix and then reshaping the tensor to its + original shape. + + Applications: + whitening transformation: Suppose X is a column vector zero-centered data. + Then compute the data covariance matrix [D x D] with torch.mm(X.t(), X), + perform SVD on this matrix and pass it as transformation_matrix. + + Args: + transformation_matrix (Tensor): tensor [D x D], D = C x H x W + mean_vector (Tensor): tensor [D], D = C x H x W + """ + + def __init__(self, transformation_matrix, mean_vector): + super().__init__() + if transformation_matrix.size(0) != transformation_matrix.size(1): + raise ValueError("transformation_matrix should be square. Got " + + "[{} x {}] rectangular matrix.".format(*transformation_matrix.size())) + + if mean_vector.size(0) != transformation_matrix.size(0): + raise ValueError("mean_vector should have the same length {}".format(mean_vector.size(0)) + + " as any one of the dimensions of the transformation_matrix [{}]" + .format(tuple(transformation_matrix.size()))) + + if transformation_matrix.device != mean_vector.device: + raise ValueError("Input tensors should be on the same device. Got {} and {}" + .format(transformation_matrix.device, mean_vector.device)) + + self.transformation_matrix = transformation_matrix + self.mean_vector = mean_vector + + def forward(self, tensor: Tensor) -> Tensor: + """ + Args: + tensor (Tensor): Tensor image to be whitened. + + Returns: + Tensor: Transformed image. + """ + shape = tensor.shape + n = shape[-3] * shape[-2] * shape[-1] + if n != self.transformation_matrix.shape[0]: + raise ValueError("Input tensor and transformation matrix have incompatible shape." + + "[{} x {} x {}] != ".format(shape[-3], shape[-2], shape[-1]) + + "{}".format(self.transformation_matrix.shape[0])) + + if tensor.device.type != self.mean_vector.device.type: + raise ValueError("Input tensor should be on the same device as transformation matrix and mean vector. " + "Got {} vs {}".format(tensor.device, self.mean_vector.device)) + + flat_tensor = tensor.view(-1, n) - self.mean_vector + transformed_tensor = torch.mm(flat_tensor, self.transformation_matrix) + tensor = transformed_tensor.view(shape) + return tensor + + def __repr__(self): + format_string = self.__class__.__name__ + '(transformation_matrix=' + format_string += (str(self.transformation_matrix.tolist()) + ')') + format_string += (", (mean_vector=" + str(self.mean_vector.tolist()) + ')') + return format_string + + +class ColorJitter(torch.nn.Module): + """Randomly change the brightness, contrast, saturation and hue of an image. + If the image is torch Tensor, it is expected + to have [..., 3, H, W] shape, where ... means an arbitrary number of leading dimensions. + If img is PIL Image, mode "1", "L", "I", "F" and modes with transparency (alpha channel) are not supported. + + Args: + brightness (float or tuple of float (min, max)): How much to jitter brightness. + brightness_factor is chosen uniformly from [max(0, 1 - brightness), 1 + brightness] + or the given [min, max]. Should be non negative numbers. + contrast (float or tuple of float (min, max)): How much to jitter contrast. + contrast_factor is chosen uniformly from [max(0, 1 - contrast), 1 + contrast] + or the given [min, max]. Should be non negative numbers. + saturation (float or tuple of float (min, max)): How much to jitter saturation. + saturation_factor is chosen uniformly from [max(0, 1 - saturation), 1 + saturation] + or the given [min, max]. Should be non negative numbers. + hue (float or tuple of float (min, max)): How much to jitter hue. + hue_factor is chosen uniformly from [-hue, hue] or the given [min, max]. + Should have 0<= hue <= 0.5 or -0.5 <= min <= max <= 0.5. + """ + + def __init__(self, brightness=0, contrast=0, saturation=0, hue=0): + super().__init__() + self.brightness = self._check_input(brightness, 'brightness') + self.contrast = self._check_input(contrast, 'contrast') + self.saturation = self._check_input(saturation, 'saturation') + self.hue = self._check_input(hue, 'hue', center=0, bound=(-0.5, 0.5), + clip_first_on_zero=False) + + @torch.jit.unused + def _check_input(self, value, name, center=1, bound=(0, float('inf')), clip_first_on_zero=True): + if isinstance(value, numbers.Number): + if value < 0: + raise ValueError("If {} is a single number, it must be non negative.".format(name)) + value = [center - float(value), center + float(value)] + if clip_first_on_zero: + value[0] = max(value[0], 0.0) + elif isinstance(value, (tuple, list)) and len(value) == 2: + if not bound[0] <= value[0] <= value[1] <= bound[1]: + raise ValueError("{} values should be between {}".format(name, bound)) + else: + raise TypeError("{} should be a single number or a list/tuple with length 2.".format(name)) + + # if value is 0 or (1., 1.) for brightness/contrast/saturation + # or (0., 0.) for hue, do nothing + if value[0] == value[1] == center: + value = None + return value + + @staticmethod + def get_params(brightness: Optional[List[float]], + contrast: Optional[List[float]], + saturation: Optional[List[float]], + hue: Optional[List[float]] + ) -> Tuple[Tensor, Optional[float], Optional[float], Optional[float], Optional[float]]: + """Get the parameters for the randomized transform to be applied on image. + + Args: + brightness (tuple of float (min, max), optional): The range from which the brightness_factor is chosen + uniformly. Pass None to turn off the transformation. + contrast (tuple of float (min, max), optional): The range from which the contrast_factor is chosen + uniformly. Pass None to turn off the transformation. + saturation (tuple of float (min, max), optional): The range from which the saturation_factor is chosen + uniformly. Pass None to turn off the transformation. + hue (tuple of float (min, max), optional): The range from which the hue_factor is chosen uniformly. + Pass None to turn off the transformation. + + Returns: + tuple: The parameters used to apply the randomized transform + along with their random order. + """ + fn_idx = torch.randperm(4) + + b = None if brightness is None else float(torch.empty(1).uniform_(brightness[0], brightness[1])) + c = None if contrast is None else float(torch.empty(1).uniform_(contrast[0], contrast[1])) + s = None if saturation is None else float(torch.empty(1).uniform_(saturation[0], saturation[1])) + h = None if hue is None else float(torch.empty(1).uniform_(hue[0], hue[1])) + + return fn_idx, b, c, s, h + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Input image. + + Returns: + PIL Image or Tensor: Color jittered image. + """ + fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = \ + self.get_params(self.brightness, self.contrast, self.saturation, self.hue) + + for fn_id in fn_idx: + if fn_id == 0 and brightness_factor is not None: + img = F.adjust_brightness(img, brightness_factor) + elif fn_id == 1 and contrast_factor is not None: + img = F.adjust_contrast(img, contrast_factor) + elif fn_id == 2 and saturation_factor is not None: + img = F.adjust_saturation(img, saturation_factor) + elif fn_id == 3 and hue_factor is not None: + img = F.adjust_hue(img, hue_factor) + + return img + + def __repr__(self): + format_string = self.__class__.__name__ + '(' + format_string += 'brightness={0}'.format(self.brightness) + format_string += ', contrast={0}'.format(self.contrast) + format_string += ', saturation={0}'.format(self.saturation) + format_string += ', hue={0})'.format(self.hue) + return format_string + + +class RandomRotation(torch.nn.Module): + """Rotate the image by angle. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + degrees (sequence or number): Range of degrees to select from. + If degrees is a number instead of sequence like (min, max), the range of degrees + will be (-degrees, +degrees). + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. + If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + expand (bool, optional): Optional expansion flag. + If true, expands the output to make it large enough to hold the entire rotated image. + If false or omitted, make the output image the same size as the input image. + Note that the expand flag assumes rotation around the center and no translation. + center (sequence, optional): Optional center of rotation, (x, y). Origin is the upper left corner. + Default is the center of the image. + fill (sequence or number): Pixel fill value for the area outside the rotated + image. Default is ``0``. If given a number, the value is used for all bands respectively. + resample (int, optional): deprecated argument and will be removed since v0.10.0. + Please use the ``interpolation`` parameter instead. + + .. _filters: https://pillow.readthedocs.io/en/latest/handbook/concepts.html#filters + + """ + + def __init__( + self, degrees, interpolation=InterpolationMode.NEAREST, expand=False, center=None, fill=0, resample=None + ): + super().__init__() + if resample is not None: + warnings.warn( + "Argument resample is deprecated and will be removed since v0.10.0. Please, use interpolation instead" + ) + interpolation = _interpolation_modes_from_int(resample) + + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + self.degrees = _setup_angle(degrees, name="degrees", req_sizes=(2, )) + + if center is not None: + _check_sequence_input(center, "center", req_sizes=(2, )) + + self.center = center + + self.resample = self.interpolation = interpolation + self.expand = expand + + if fill is None: + fill = 0 + elif not isinstance(fill, (Sequence, numbers.Number)): + raise TypeError("Fill should be either a sequence or a number.") + + self.fill = fill + + @staticmethod + def get_params(degrees: List[float]) -> float: + """Get parameters for ``rotate`` for a random rotation. + + Returns: + float: angle parameter to be passed to ``rotate`` for random rotation. + """ + angle = float(torch.empty(1).uniform_(float(degrees[0]), float(degrees[1])).item()) + return angle + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be rotated. + + Returns: + PIL Image or Tensor: Rotated image. + """ + fill = self.fill + if isinstance(img, Tensor): + if isinstance(fill, (int, float)): + fill = [float(fill)] * F._get_image_num_channels(img) + else: + fill = [float(f) for f in fill] + angle = self.get_params(self.degrees) + + return F.rotate(img, angle, self.resample, self.expand, self.center, fill) + + def __repr__(self): + interpolate_str = self.interpolation.value + format_string = self.__class__.__name__ + '(degrees={0}'.format(self.degrees) + format_string += ', interpolation={0}'.format(interpolate_str) + format_string += ', expand={0}'.format(self.expand) + if self.center is not None: + format_string += ', center={0}'.format(self.center) + if self.fill is not None: + format_string += ', fill={0}'.format(self.fill) + format_string += ')' + return format_string + + +class RandomAffine(torch.nn.Module): + """Random affine transformation of the image keeping center invariant. + If the image is torch Tensor, it is expected + to have [..., H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + degrees (sequence or number): Range of degrees to select from. + If degrees is a number instead of sequence like (min, max), the range of degrees + will be (-degrees, +degrees). Set to 0 to deactivate rotations. + translate (tuple, optional): tuple of maximum absolute fraction for horizontal + and vertical translations. For example translate=(a, b), then horizontal shift + is randomly sampled in the range -img_width * a < dx < img_width * a and vertical shift is + randomly sampled in the range -img_height * b < dy < img_height * b. Will not translate by default. + scale (tuple, optional): scaling factor interval, e.g (a, b), then scale is + randomly sampled from the range a <= scale <= b. Will keep original scale by default. + shear (sequence or number, optional): Range of degrees to select from. + If shear is a number, a shear parallel to the x axis in the range (-shear, +shear) + will be applied. Else if shear is a sequence of 2 values a shear parallel to the x axis in the + range (shear[0], shear[1]) will be applied. Else if shear is a sequence of 4 values, + a x-axis shear in (shear[0], shear[1]) and y-axis shear in (shear[2], shear[3]) will be applied. + Will not apply shear by default. + interpolation (InterpolationMode): Desired interpolation enum defined by + :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. + If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. + For backward compatibility integer values (e.g. ``PIL.Image.NEAREST``) are still acceptable. + fill (sequence or number): Pixel fill value for the area outside the transformed + image. Default is ``0``. If given a number, the value is used for all bands respectively. + fillcolor (sequence or number, optional): deprecated argument and will be removed since v0.10.0. + Please use the ``fill`` parameter instead. + resample (int, optional): deprecated argument and will be removed since v0.10.0. + Please use the ``interpolation`` parameter instead. + + .. _filters: https://pillow.readthedocs.io/en/latest/handbook/concepts.html#filters + + """ + + def __init__( + self, degrees, translate=None, scale=None, shear=None, interpolation=InterpolationMode.NEAREST, fill=0, + fillcolor=None, resample=None + ): + super().__init__() + if resample is not None: + warnings.warn( + "Argument resample is deprecated and will be removed since v0.10.0. Please, use interpolation instead" + ) + interpolation = _interpolation_modes_from_int(resample) + + # Backward compatibility with integer value + if isinstance(interpolation, int): + warnings.warn( + "Argument interpolation should be of type InterpolationMode instead of int. " + "Please, use InterpolationMode enum." + ) + interpolation = _interpolation_modes_from_int(interpolation) + + if fillcolor is not None: + warnings.warn( + "Argument fillcolor is deprecated and will be removed since v0.10.0. Please, use fill instead" + ) + fill = fillcolor + + self.degrees = _setup_angle(degrees, name="degrees", req_sizes=(2, )) + + if translate is not None: + _check_sequence_input(translate, "translate", req_sizes=(2, )) + for t in translate: + if not (0.0 <= t <= 1.0): + raise ValueError("translation values should be between 0 and 1") + self.translate = translate + + if scale is not None: + _check_sequence_input(scale, "scale", req_sizes=(2, )) + for s in scale: + if s <= 0: + raise ValueError("scale values should be positive") + self.scale = scale + + if shear is not None: + self.shear = _setup_angle(shear, name="shear", req_sizes=(2, 4)) + else: + self.shear = shear + + self.resample = self.interpolation = interpolation + + if fill is None: + fill = 0 + elif not isinstance(fill, (Sequence, numbers.Number)): + raise TypeError("Fill should be either a sequence or a number.") + + self.fillcolor = self.fill = fill + + @staticmethod + def get_params( + degrees: List[float], + translate: Optional[List[float]], + scale_ranges: Optional[List[float]], + shears: Optional[List[float]], + img_size: List[int] + ) -> Tuple[float, Tuple[int, int], float, Tuple[float, float]]: + """Get parameters for affine transformation + + Returns: + params to be passed to the affine transformation + """ + angle = float(torch.empty(1).uniform_(float(degrees[0]), float(degrees[1])).item()) + if translate is not None: + max_dx = float(translate[0] * img_size[0]) + max_dy = float(translate[1] * img_size[1]) + tx = int(round(torch.empty(1).uniform_(-max_dx, max_dx).item())) + ty = int(round(torch.empty(1).uniform_(-max_dy, max_dy).item())) + translations = (tx, ty) + else: + translations = (0, 0) + + if scale_ranges is not None: + scale = float(torch.empty(1).uniform_(scale_ranges[0], scale_ranges[1]).item()) + else: + scale = 1.0 + + shear_x = shear_y = 0.0 + if shears is not None: + shear_x = float(torch.empty(1).uniform_(shears[0], shears[1]).item()) + if len(shears) == 4: + shear_y = float(torch.empty(1).uniform_(shears[2], shears[3]).item()) + + shear = (shear_x, shear_y) + + return angle, translations, scale, shear + + def forward(self, img): + """ + img (PIL Image or Tensor): Image to be transformed. + + Returns: + PIL Image or Tensor: Affine transformed image. + """ + fill = self.fill + if isinstance(img, Tensor): + if isinstance(fill, (int, float)): + fill = [float(fill)] * F._get_image_num_channels(img) + else: + fill = [float(f) for f in fill] + + img_size = F._get_image_size(img) + + ret = self.get_params(self.degrees, self.translate, self.scale, self.shear, img_size) + + return F.affine(img, *ret, interpolation=self.interpolation, fill=fill) + + def __repr__(self): + s = '{name}(degrees={degrees}' + if self.translate is not None: + s += ', translate={translate}' + if self.scale is not None: + s += ', scale={scale}' + if self.shear is not None: + s += ', shear={shear}' + if self.interpolation != InterpolationMode.NEAREST: + s += ', interpolation={interpolation}' + if self.fill != 0: + s += ', fill={fill}' + s += ')' + d = dict(self.__dict__) + d['interpolation'] = self.interpolation.value + return s.format(name=self.__class__.__name__, **d) + + +class Grayscale(torch.nn.Module): + """Convert image to grayscale. + If the image is torch Tensor, it is expected + to have [..., 3, H, W] shape, where ... means an arbitrary number of leading dimensions + + Args: + num_output_channels (int): (1 or 3) number of channels desired for output image + + Returns: + PIL Image: Grayscale version of the input. + + - If ``num_output_channels == 1`` : returned image is single channel + - If ``num_output_channels == 3`` : returned image is 3 channel with r == g == b + + """ + + def __init__(self, num_output_channels=1): + super().__init__() + self.num_output_channels = num_output_channels + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be converted to grayscale. + + Returns: + PIL Image or Tensor: Grayscaled image. + """ + return F.rgb_to_grayscale(img, num_output_channels=self.num_output_channels) + + def __repr__(self): + return self.__class__.__name__ + '(num_output_channels={0})'.format(self.num_output_channels) + + +class RandomGrayscale(torch.nn.Module): + """Randomly convert image to grayscale with a probability of p (default 0.1). + If the image is torch Tensor, it is expected + to have [..., 3, H, W] shape, where ... means an arbitrary number of leading dimensions + + Args: + p (float): probability that image should be converted to grayscale. + + Returns: + PIL Image or Tensor: Grayscale version of the input image with probability p and unchanged + with probability (1-p). + - If input image is 1 channel: grayscale version is 1 channel + - If input image is 3 channel: grayscale version is 3 channel with r == g == b + + """ + + def __init__(self, p=0.1): + super().__init__() + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be converted to grayscale. + + Returns: + PIL Image or Tensor: Randomly grayscaled image. + """ + num_output_channels = F._get_image_num_channels(img) + if torch.rand(1) < self.p: + return F.rgb_to_grayscale(img, num_output_channels=num_output_channels) + return img + + def __repr__(self): + return self.__class__.__name__ + '(p={0})'.format(self.p) + + +class RandomErasing(torch.nn.Module): + """ Randomly selects a rectangle region in an torch Tensor image and erases its pixels. + This transform does not support PIL Image. + 'Random Erasing Data Augmentation' by Zhong et al. See https://arxiv.org/abs/1708.04896 + + Args: + p: probability that the random erasing operation will be performed. + scale: range of proportion of erased area against input image. + ratio: range of aspect ratio of erased area. + value: erasing value. Default is 0. If a single int, it is used to + erase all pixels. If a tuple of length 3, it is used to erase + R, G, B channels respectively. + If a str of 'random', erasing each pixel with random values. + inplace: boolean to make this transform inplace. Default set to False. + + Returns: + Erased Image. + + Example: + >>> transform = transforms.Compose([ + >>> transforms.RandomHorizontalFlip(), + >>> transforms.ToTensor(), + >>> transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), + >>> transforms.RandomErasing(), + >>> ]) + """ + + def __init__(self, p=0.5, scale=(0.02, 0.33), ratio=(0.3, 3.3), value=0, inplace=False): + super().__init__() + if not isinstance(value, (numbers.Number, str, tuple, list)): + raise TypeError("Argument value should be either a number or str or a sequence") + if isinstance(value, str) and value != "random": + raise ValueError("If value is str, it should be 'random'") + if not isinstance(scale, (tuple, list)): + raise TypeError("Scale should be a sequence") + if not isinstance(ratio, (tuple, list)): + raise TypeError("Ratio should be a sequence") + if (scale[0] > scale[1]) or (ratio[0] > ratio[1]): + warnings.warn("Scale and ratio should be of kind (min, max)") + if scale[0] < 0 or scale[1] > 1: + raise ValueError("Scale should be between 0 and 1") + if p < 0 or p > 1: + raise ValueError("Random erasing probability should be between 0 and 1") + + self.p = p + self.scale = scale + self.ratio = ratio + self.value = value + self.inplace = inplace + + @staticmethod + def get_params( + img: Tensor, scale: Tuple[float, float], ratio: Tuple[float, float], value: Optional[List[float]] = None + ) -> Tuple[int, int, int, int, Tensor]: + """Get parameters for ``erase`` for a random erasing. + + Args: + img (Tensor): Tensor image to be erased. + scale (sequence): range of proportion of erased area against input image. + ratio (sequence): range of aspect ratio of erased area. + value (list, optional): erasing value. If None, it is interpreted as "random" + (erasing each pixel with random values). If ``len(value)`` is 1, it is interpreted as a number, + i.e. ``value[0]``. + + Returns: + tuple: params (i, j, h, w, v) to be passed to ``erase`` for random erasing. + """ + img_c, img_h, img_w = img.shape[-3], img.shape[-2], img.shape[-1] + area = img_h * img_w + + log_ratio = torch.log(torch.tensor(ratio)) + for _ in range(10): + erase_area = area * torch.empty(1).uniform_(scale[0], scale[1]).item() + aspect_ratio = torch.exp( + torch.empty(1).uniform_(log_ratio[0], log_ratio[1]) + ).item() + + h = int(round(math.sqrt(erase_area * aspect_ratio))) + w = int(round(math.sqrt(erase_area / aspect_ratio))) + if not (h < img_h and w < img_w): + continue + + if value is None: + v = torch.empty([img_c, h, w], dtype=torch.float32).normal_() + else: + v = torch.tensor(value)[:, None, None] + + i = torch.randint(0, img_h - h + 1, size=(1, )).item() + j = torch.randint(0, img_w - w + 1, size=(1, )).item() + return i, j, h, w, v + + # Return original image + return 0, 0, img_h, img_w, img + + def forward(self, img): + """ + Args: + img (Tensor): Tensor image to be erased. + + Returns: + img (Tensor): Erased Tensor image. + """ + if torch.rand(1) < self.p: + + # cast self.value to script acceptable type + if isinstance(self.value, (int, float)): + value = [self.value, ] + elif isinstance(self.value, str): + value = None + elif isinstance(self.value, tuple): + value = list(self.value) + else: + value = self.value + + if value is not None and not (len(value) in (1, img.shape[-3])): + raise ValueError( + "If value is a sequence, it should have either a single value or " + "{} (number of input channels)".format(img.shape[-3]) + ) + + x, y, h, w, v = self.get_params(img, scale=self.scale, ratio=self.ratio, value=value) + return F.erase(img, x, y, h, w, v, self.inplace) + return img + + def __repr__(self): + s = '(p={}, '.format(self.p) + s += 'scale={}, '.format(self.scale) + s += 'ratio={}, '.format(self.ratio) + s += 'value={}, '.format(self.value) + s += 'inplace={})'.format(self.inplace) + return self.__class__.__name__ + s + + +class GaussianBlur(torch.nn.Module): + """Blurs image with randomly chosen Gaussian blur. + If the image is torch Tensor, it is expected + to have [..., C, H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + kernel_size (int or sequence): Size of the Gaussian kernel. + sigma (float or tuple of float (min, max)): Standard deviation to be used for + creating kernel to perform blurring. If float, sigma is fixed. If it is tuple + of float (min, max), sigma is chosen uniformly at random to lie in the + given range. + + Returns: + PIL Image or Tensor: Gaussian blurred version of the input image. + + """ + + def __init__(self, kernel_size, sigma=(0.1, 2.0)): + super().__init__() + self.kernel_size = _setup_size(kernel_size, "Kernel size should be a tuple/list of two integers") + for ks in self.kernel_size: + if ks <= 0 or ks % 2 == 0: + raise ValueError("Kernel size value should be an odd and positive number.") + + if isinstance(sigma, numbers.Number): + if sigma <= 0: + raise ValueError("If sigma is a single number, it must be positive.") + sigma = (sigma, sigma) + elif isinstance(sigma, Sequence) and len(sigma) == 2: + if not 0. < sigma[0] <= sigma[1]: + raise ValueError("sigma values should be positive and of the form (min, max).") + else: + raise ValueError("sigma should be a single number or a list/tuple with length 2.") + + self.sigma = sigma + + @staticmethod + def get_params(sigma_min: float, sigma_max: float) -> float: + """Choose sigma for random gaussian blurring. + + Args: + sigma_min (float): Minimum standard deviation that can be chosen for blurring kernel. + sigma_max (float): Maximum standard deviation that can be chosen for blurring kernel. + + Returns: + float: Standard deviation to be passed to calculate kernel for gaussian blurring. + """ + return torch.empty(1).uniform_(sigma_min, sigma_max).item() + + def forward(self, img: Tensor) -> Tensor: + """ + Args: + img (PIL Image or Tensor): image to be blurred. + + Returns: + PIL Image or Tensor: Gaussian blurred image + """ + sigma = self.get_params(self.sigma[0], self.sigma[1]) + return F.gaussian_blur(img, self.kernel_size, [sigma, sigma]) + + def __repr__(self): + s = '(kernel_size={}, '.format(self.kernel_size) + s += 'sigma={})'.format(self.sigma) + return self.__class__.__name__ + s + + +def _setup_size(size, error_msg): + if isinstance(size, numbers.Number): + return int(size), int(size) + + if isinstance(size, Sequence) and len(size) == 1: + return size[0], size[0] + + if len(size) != 2: + raise ValueError(error_msg) + + return size + + +def _check_sequence_input(x, name, req_sizes): + msg = req_sizes[0] if len(req_sizes) < 2 else " or ".join([str(s) for s in req_sizes]) + if not isinstance(x, Sequence): + raise TypeError("{} should be a sequence of length {}.".format(name, msg)) + if len(x) not in req_sizes: + raise ValueError("{} should be sequence of length {}.".format(name, msg)) + + +def _setup_angle(x, name, req_sizes=(2, )): + if isinstance(x, numbers.Number): + if x < 0: + raise ValueError("If {} is a single number, it must be positive.".format(name)) + x = [-x, x] + else: + _check_sequence_input(x, name, req_sizes) + + return [float(d) for d in x] + + +class RandomInvert(torch.nn.Module): + """Inverts the colors of the given image randomly with a given probability. + If img is a Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "L" or "RGB". + + Args: + p (float): probability of the image being color inverted. Default value is 0.5 + """ + + def __init__(self, p=0.5): + super().__init__() + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be inverted. + + Returns: + PIL Image or Tensor: Randomly color inverted image. + """ + if torch.rand(1).item() < self.p: + return F.invert(img) + return img + + def __repr__(self): + return self.__class__.__name__ + '(p={})'.format(self.p) + + +class RandomPosterize(torch.nn.Module): + """Posterize the image randomly with a given probability by reducing the + number of bits for each color channel. If the image is torch Tensor, it should be of type torch.uint8, + and it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "L" or "RGB". + + Args: + bits (int): number of bits to keep for each channel (0-8) + p (float): probability of the image being color inverted. Default value is 0.5 + """ + + def __init__(self, bits, p=0.5): + super().__init__() + self.bits = bits + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be posterized. + + Returns: + PIL Image or Tensor: Randomly posterized image. + """ + if torch.rand(1).item() < self.p: + return F.posterize(img, self.bits) + return img + + def __repr__(self): + return self.__class__.__name__ + '(bits={},p={})'.format(self.bits, self.p) + + +class RandomSolarize(torch.nn.Module): + """Solarize the image randomly with a given probability by inverting all pixel + values above a threshold. If img is a Tensor, it is expected to be in [..., 1 or 3, H, W] format, + where ... means it can have an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "L" or "RGB". + + Args: + threshold (float): all pixels equal or above this value are inverted. + p (float): probability of the image being color inverted. Default value is 0.5 + """ + + def __init__(self, threshold, p=0.5): + super().__init__() + self.threshold = threshold + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be solarized. + + Returns: + PIL Image or Tensor: Randomly solarized image. + """ + if torch.rand(1).item() < self.p: + return F.solarize(img, self.threshold) + return img + + def __repr__(self): + return self.__class__.__name__ + '(threshold={},p={})'.format(self.threshold, self.p) + + +class RandomAdjustSharpness(torch.nn.Module): + """Adjust the sharpness of the image randomly with a given probability. If the image is torch Tensor, + it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. + + Args: + sharpness_factor (float): How much to adjust the sharpness. Can be + any non negative number. 0 gives a blurred image, 1 gives the + original image while 2 increases the sharpness by a factor of 2. + p (float): probability of the image being color inverted. Default value is 0.5 + """ + + def __init__(self, sharpness_factor, p=0.5): + super().__init__() + self.sharpness_factor = sharpness_factor + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be sharpened. + + Returns: + PIL Image or Tensor: Randomly sharpened image. + """ + if torch.rand(1).item() < self.p: + return F.adjust_sharpness(img, self.sharpness_factor) + return img + + def __repr__(self): + return self.__class__.__name__ + '(sharpness_factor={},p={})'.format(self.sharpness_factor, self.p) + + +class RandomAutocontrast(torch.nn.Module): + """Autocontrast the pixels of the given image randomly with a given probability. + If the image is torch Tensor, it is expected + to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "L" or "RGB". + + Args: + p (float): probability of the image being autocontrasted. Default value is 0.5 + """ + + def __init__(self, p=0.5): + super().__init__() + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be autocontrasted. + + Returns: + PIL Image or Tensor: Randomly autocontrasted image. + """ + if torch.rand(1).item() < self.p: + return F.autocontrast(img) + return img + + def __repr__(self): + return self.__class__.__name__ + '(p={})'.format(self.p) + + +class RandomEqualize(torch.nn.Module): + """Equalize the histogram of the given image randomly with a given probability. + If the image is torch Tensor, it is expected + to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. + If img is PIL Image, it is expected to be in mode "P", "L" or "RGB". + + Args: + p (float): probability of the image being equalized. Default value is 0.5 + """ + + def __init__(self, p=0.5): + super().__init__() + self.p = p + + def forward(self, img): + """ + Args: + img (PIL Image or Tensor): Image to be equalized. + + Returns: + PIL Image or Tensor: Randomly equalized image. + """ + if torch.rand(1).item() < self.p: + return F.equalize(img) + return img + + def __repr__(self): + return self.__class__.__name__ + '(p={})'.format(self.p) diff --git a/detectron2/data/transforms/transform.py b/detectron2/data/transforms/transform.py new file mode 100644 index 0000000000000000000000000000000000000000..de44b991d7ab0d920ffb769e1402f08e358d37f7 --- /dev/null +++ b/detectron2/data/transforms/transform.py @@ -0,0 +1,351 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +""" +See "Data Augmentation" tutorial for an overview of the system: +https://detectron2.readthedocs.io/tutorials/augmentation.html +""" + +import numpy as np +import torch +import torch.nn.functional as F +from fvcore.transforms.transform import ( + CropTransform, + HFlipTransform, + NoOpTransform, + Transform, + TransformList, +) +from PIL import Image + +try: + import cv2 # noqa +except ImportError: + # OpenCV is an optional dependency at the moment + pass + +__all__ = [ + "ExtentTransform", + "ResizeTransform", + "RotationTransform", + "ColorTransform", + "PILColorTransform", +] + + +class ExtentTransform(Transform): + """ + Extracts a subregion from the source image and scales it to the output size. + + The fill color is used to map pixels from the source rect that fall outside + the source image. + + See: https://pillow.readthedocs.io/en/latest/PIL.html#PIL.ImageTransform.ExtentTransform + """ + + def __init__(self, src_rect, output_size, interp=Image.LINEAR, fill=0): + """ + Args: + src_rect (x0, y0, x1, y1): src coordinates + output_size (h, w): dst image size + interp: PIL interpolation methods + fill: Fill color used when src_rect extends outside image + """ + super().__init__() + self._set_attributes(locals()) + + def apply_image(self, img, interp=None): + h, w = self.output_size + if len(img.shape) > 2 and img.shape[2] == 1: + pil_image = Image.fromarray(img[:, :, 0], mode="L") + else: + pil_image = Image.fromarray(img) + pil_image = pil_image.transform( + size=(w, h), + method=Image.EXTENT, + data=self.src_rect, + resample=interp if interp else self.interp, + fill=self.fill, + ) + ret = np.asarray(pil_image) + if len(img.shape) > 2 and img.shape[2] == 1: + ret = np.expand_dims(ret, -1) + return ret + + def apply_coords(self, coords): + # Transform image center from source coordinates into output coordinates + # and then map the new origin to the corner of the output image. + h, w = self.output_size + x0, y0, x1, y1 = self.src_rect + new_coords = coords.astype(np.float32) + new_coords[:, 0] -= 0.5 * (x0 + x1) + new_coords[:, 1] -= 0.5 * (y0 + y1) + new_coords[:, 0] *= w / (x1 - x0) + new_coords[:, 1] *= h / (y1 - y0) + new_coords[:, 0] += 0.5 * w + new_coords[:, 1] += 0.5 * h + return new_coords + + def apply_segmentation(self, segmentation): + segmentation = self.apply_image(segmentation, interp=Image.NEAREST) + return segmentation + + +class ResizeTransform(Transform): + """ + Resize the image to a target size. + """ + + def __init__(self, h, w, new_h, new_w, interp=None): + """ + Args: + h, w (int): original image size + new_h, new_w (int): new image size + interp: PIL interpolation methods, defaults to bilinear. + """ + # TODO decide on PIL vs opencv + super().__init__() + if interp is None: + interp = Image.BILINEAR + self._set_attributes(locals()) + + def apply_image(self, img, interp=None): + assert img.shape[:2] == (self.h, self.w) + assert len(img.shape) <= 4 + interp_method = interp if interp is not None else self.interp + + if img.dtype == np.uint8: + if len(img.shape) > 2 and img.shape[2] == 1: + pil_image = Image.fromarray(img[:, :, 0], mode="L") + else: + pil_image = Image.fromarray(img) + pil_image = pil_image.resize((self.new_w, self.new_h), interp_method) + ret = np.asarray(pil_image) + if len(img.shape) > 2 and img.shape[2] == 1: + ret = np.expand_dims(ret, -1) + else: + # PIL only supports uint8 + if any(x < 0 for x in img.strides): + img = np.ascontiguousarray(img) + img = torch.from_numpy(img) + shape = list(img.shape) + shape_4d = shape[:2] + [1] * (4 - len(shape)) + shape[2:] + img = img.view(shape_4d).permute(2, 3, 0, 1) # hw(c) -> nchw + _PIL_RESIZE_TO_INTERPOLATE_MODE = { + Image.NEAREST: "nearest", + Image.BILINEAR: "bilinear", + Image.BICUBIC: "bicubic", + } + mode = _PIL_RESIZE_TO_INTERPOLATE_MODE[interp_method] + align_corners = None if mode == "nearest" else False + img = F.interpolate( + img, (self.new_h, self.new_w), mode=mode, align_corners=align_corners + ) + shape[:2] = (self.new_h, self.new_w) + ret = img.permute(2, 3, 0, 1).view(shape).numpy() # nchw -> hw(c) + + return ret + + def apply_coords(self, coords): + coords[:, 0] = coords[:, 0] * (self.new_w * 1.0 / self.w) + coords[:, 1] = coords[:, 1] * (self.new_h * 1.0 / self.h) + return coords + + def apply_segmentation(self, segmentation): + segmentation = self.apply_image(segmentation, interp=Image.NEAREST) + return segmentation + + def inverse(self): + return ResizeTransform(self.new_h, self.new_w, self.h, self.w, self.interp) + + +class RotationTransform(Transform): + """ + This method returns a copy of this image, rotated the given + number of degrees counter clockwise around its center. + """ + + def __init__(self, h, w, angle, expand=True, center=None, interp=None): + """ + Args: + h, w (int): original image size + angle (float): degrees for rotation + expand (bool): choose if the image should be resized to fit the whole + rotated image (default), or simply cropped + center (tuple (width, height)): coordinates of the rotation center + if left to None, the center will be fit to the center of each image + center has no effect if expand=True because it only affects shifting + interp: cv2 interpolation method, default cv2.INTER_LINEAR + """ + super().__init__() + image_center = np.array((w / 2, h / 2)) + if center is None: + center = image_center + if interp is None: + interp = cv2.INTER_LINEAR + abs_cos, abs_sin = (abs(np.cos(np.deg2rad(angle))), abs(np.sin(np.deg2rad(angle)))) + if expand: + # find the new width and height bounds + bound_w, bound_h = np.rint( + [h * abs_sin + w * abs_cos, h * abs_cos + w * abs_sin] + ).astype(int) + else: + bound_w, bound_h = w, h + + self._set_attributes(locals()) + self.rm_coords = self.create_rotation_matrix() + # Needed because of this problem https://github.com/opencv/opencv/issues/11784 + self.rm_image = self.create_rotation_matrix(offset=-0.5) + + def apply_image(self, img, interp=None): + """ + img should be a numpy array, formatted as Height * Width * Nchannels + """ + if len(img) == 0 or self.angle % 360 == 0: + return img + assert img.shape[:2] == (self.h, self.w) + interp = interp if interp is not None else self.interp + return cv2.warpAffine(img, self.rm_image, (self.bound_w, self.bound_h), flags=interp) + + def apply_coords(self, coords): + """ + coords should be a N * 2 array-like, containing N couples of (x, y) points + """ + coords = np.asarray(coords, dtype=float) + if len(coords) == 0 or self.angle % 360 == 0: + return coords + return cv2.transform(coords[:, np.newaxis, :], self.rm_coords)[:, 0, :] + + def apply_segmentation(self, segmentation): + segmentation = self.apply_image(segmentation, interp=cv2.INTER_NEAREST) + return segmentation + + def create_rotation_matrix(self, offset=0): + center = (self.center[0] + offset, self.center[1] + offset) + rm = cv2.getRotationMatrix2D(tuple(center), self.angle, 1) + if self.expand: + # Find the coordinates of the center of rotation in the new image + # The only point for which we know the future coordinates is the center of the image + rot_im_center = cv2.transform(self.image_center[None, None, :] + offset, rm)[0, 0, :] + new_center = np.array([self.bound_w / 2, self.bound_h / 2]) + offset - rot_im_center + # shift the rotation center to the new coordinates + rm[:, 2] += new_center + return rm + + def inverse(self): + """ + The inverse is to rotate it back with expand, and crop to get the original shape. + """ + if not self.expand: # Not possible to inverse if a part of the image is lost + raise NotImplementedError() + rotation = RotationTransform( + self.bound_h, self.bound_w, -self.angle, True, None, self.interp + ) + crop = CropTransform( + (rotation.bound_w - self.w) // 2, (rotation.bound_h - self.h) // 2, self.w, self.h + ) + return TransformList([rotation, crop]) + + +class ColorTransform(Transform): + """ + Generic wrapper for any photometric transforms. + These transformations should only affect the color space and + not the coordinate space of the image (e.g. annotation + coordinates such as bounding boxes should not be changed) + """ + + def __init__(self, op): + """ + Args: + op (Callable): operation to be applied to the image, + which takes in an ndarray and returns an ndarray. + """ + if not callable(op): + raise ValueError("op parameter should be callable") + super().__init__() + self._set_attributes(locals()) + + def apply_image(self, img): + return self.op(img) + + def apply_coords(self, coords): + return coords + + def inverse(self): + return NoOpTransform() + + def apply_segmentation(self, segmentation): + return segmentation + + +class PILColorTransform(ColorTransform): + """ + Generic wrapper for PIL Photometric image transforms, + which affect the color space and not the coordinate + space of the image + """ + + def __init__(self, op): + """ + Args: + op (Callable): operation to be applied to the image, + which takes in a PIL Image and returns a transformed + PIL Image. + For reference on possible operations see: + - https://pillow.readthedocs.io/en/stable/ + """ + if not callable(op): + raise ValueError("op parameter should be callable") + super().__init__(op) + + def apply_image(self, img): + img = Image.fromarray(img) + return np.asarray(super().apply_image(img)) + + +def HFlip_rotated_box(transform, rotated_boxes): + """ + Apply the horizontal flip transform on rotated boxes. + + Args: + rotated_boxes (ndarray): Nx5 floating point array of + (x_center, y_center, width, height, angle_degrees) format + in absolute coordinates. + """ + # Transform x_center + rotated_boxes[:, 0] = transform.width - rotated_boxes[:, 0] + # Transform angle + rotated_boxes[:, 4] = -rotated_boxes[:, 4] + return rotated_boxes + + +def Resize_rotated_box(transform, rotated_boxes): + """ + Apply the resizing transform on rotated boxes. For details of how these (approximation) + formulas are derived, please refer to :meth:`RotatedBoxes.scale`. + + Args: + rotated_boxes (ndarray): Nx5 floating point array of + (x_center, y_center, width, height, angle_degrees) format + in absolute coordinates. + """ + scale_factor_x = transform.new_w * 1.0 / transform.w + scale_factor_y = transform.new_h * 1.0 / transform.h + rotated_boxes[:, 0] *= scale_factor_x + rotated_boxes[:, 1] *= scale_factor_y + theta = rotated_boxes[:, 4] * np.pi / 180.0 + c = np.cos(theta) + s = np.sin(theta) + rotated_boxes[:, 2] *= np.sqrt(np.square(scale_factor_x * c) + np.square(scale_factor_y * s)) + rotated_boxes[:, 3] *= np.sqrt(np.square(scale_factor_x * s) + np.square(scale_factor_y * c)) + rotated_boxes[:, 4] = np.arctan2(scale_factor_x * s, scale_factor_y * c) * 180 / np.pi + + return rotated_boxes + + +HFlipTransform.register_type("rotated_box", HFlip_rotated_box) +ResizeTransform.register_type("rotated_box", Resize_rotated_box) + +# not necessary any more with latest fvcore +NoOpTransform.register_type("rotated_box", lambda t, x: x) diff --git a/detectron2/engine/__init__.py b/detectron2/engine/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..08a61572b4c7d09c8d400e903a96cbf5b2cc4763 --- /dev/null +++ b/detectron2/engine/__init__.py @@ -0,0 +1,12 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +from .launch import * +from .train_loop import * + +__all__ = [k for k in globals().keys() if not k.startswith("_")] + + +# prefer to let hooks and defaults live in separate namespaces (therefore not in __all__) +# but still make them available here +from .hooks import * +from .defaults import * diff --git a/detectron2/engine/__pycache__/__init__.cpython-39.pyc b/detectron2/engine/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b0f69f0e86f226b36c1d3420251eda6b4707946b Binary files /dev/null and b/detectron2/engine/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/engine/__pycache__/defaults.cpython-39.pyc b/detectron2/engine/__pycache__/defaults.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..92b9cc2b4d81b8cf6eeaf94893e917ba23443d0b Binary files /dev/null and b/detectron2/engine/__pycache__/defaults.cpython-39.pyc differ diff --git a/detectron2/engine/__pycache__/hooks.cpython-39.pyc b/detectron2/engine/__pycache__/hooks.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ee5682cfcd904e3ef677cc079c7e66505ba16f66 Binary files /dev/null and b/detectron2/engine/__pycache__/hooks.cpython-39.pyc differ diff --git a/detectron2/engine/__pycache__/launch.cpython-39.pyc b/detectron2/engine/__pycache__/launch.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..74ae113db757ab782d50701eb6e5c0c7269cc1e1 Binary files /dev/null and b/detectron2/engine/__pycache__/launch.cpython-39.pyc differ diff --git a/detectron2/engine/__pycache__/train_loop.cpython-39.pyc b/detectron2/engine/__pycache__/train_loop.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..821b22ee5996f06524d1ed83bd88294cc9178d6f Binary files /dev/null and b/detectron2/engine/__pycache__/train_loop.cpython-39.pyc differ diff --git a/detectron2/engine/defaults.py b/detectron2/engine/defaults.py new file mode 100644 index 0000000000000000000000000000000000000000..2360b3fe7971b15dd5a951c310d8ec522f5f630c --- /dev/null +++ b/detectron2/engine/defaults.py @@ -0,0 +1,705 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +""" +This file contains components with some default boilerplate logic user may need +in training / testing. They will not work for everyone, but many users may find them useful. + +The behavior of functions/classes in this file is subject to change, +since they are meant to represent the "common default behavior" people need in their projects. +""" + +import argparse +import logging +import os +import sys +import weakref +from collections import OrderedDict +from typing import Optional +import torch +from fvcore.nn.precise_bn import get_bn_modules +from omegaconf import OmegaConf +from torch.nn.parallel import DistributedDataParallel + +import detectron2.data.transforms as T +from detectron2.checkpoint import DetectionCheckpointer +from detectron2.config import CfgNode, LazyConfig +from detectron2.data import ( + MetadataCatalog, + build_detection_test_loader, + build_detection_train_loader, +) +from detectron2.evaluation import ( + DatasetEvaluator, + inference_on_dataset, + print_csv_format, + verify_results, +) +from detectron2.modeling import build_model +from detectron2.solver import build_lr_scheduler, build_optimizer +from detectron2.utils import comm +from detectron2.utils.collect_env import collect_env_info +from detectron2.utils.env import seed_all_rng +from detectron2.utils.events import CommonMetricPrinter, JSONWriter, TensorboardXWriter +from detectron2.utils.file_io import PathManager +from detectron2.utils.logger import setup_logger + +from . import hooks +from .train_loop import AMPTrainer, SimpleTrainer, TrainerBase + +__all__ = [ + "create_ddp_model", + "default_argument_parser", + "default_setup", + "default_writers", + "DefaultPredictor", + "DefaultTrainer", +] + + +def create_ddp_model(model, *, fp16_compression=False, **kwargs): + """ + Create a DistributedDataParallel model if there are >1 processes. + + Args: + model: a torch.nn.Module + fp16_compression: add fp16 compression hooks to the ddp object. + See more at https://pytorch.org/docs/stable/ddp_comm_hooks.html#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_hook + kwargs: other arguments of :module:`torch.nn.parallel.DistributedDataParallel`. + """ # noqa + if comm.get_world_size() == 1: + return model + if "device_ids" not in kwargs: + kwargs["device_ids"] = [comm.get_local_rank()] + ddp = DistributedDataParallel(model, **kwargs) + if fp16_compression: + from torch.distributed.algorithms.ddp_comm_hooks import default as comm_hooks + + ddp.register_comm_hook(state=None, hook=comm_hooks.fp16_compress_hook) + return ddp + + +def default_argument_parser(epilog=None): + """ + Create a parser with some common arguments used by detectron2 users. + + Args: + epilog (str): epilog passed to ArgumentParser describing the usage. + + Returns: + argparse.ArgumentParser: + """ + parser = argparse.ArgumentParser( + epilog=epilog + or f""" +Examples: + +Run on single machine: + $ {sys.argv[0]} --num-gpus 8 --config-file cfg.yaml + +Change some config options: + $ {sys.argv[0]} --config-file cfg.yaml MODEL.WEIGHTS /path/to/weight.pth SOLVER.BASE_LR 0.001 + +Run on multiple machines: + (machine0)$ {sys.argv[0]} --machine-rank 0 --num-machines 2 --dist-url [--other-flags] + (machine1)$ {sys.argv[0]} --machine-rank 1 --num-machines 2 --dist-url [--other-flags] +""", + formatter_class=argparse.RawDescriptionHelpFormatter, + ) + parser.add_argument("--config-file", default="", metavar="FILE", help="path to config file") + parser.add_argument( + "--resume", + action="store_true", + help="Whether to attempt to resume from the checkpoint directory. " + "See documentation of `DefaultTrainer.resume_or_load()` for what it means.", + ) + parser.add_argument("--eval-only", action="store_true", help="perform evaluation only") + parser.add_argument("--num-gpus", type=int, default=1, help="number of gpus *per machine*") + parser.add_argument("--num-machines", type=int, default=1, help="total number of machines") + parser.add_argument( + "--machine-rank", type=int, default=0, help="the rank of this machine (unique per machine)" + ) + + # PyTorch still may leave orphan processes in multi-gpu training. + # Therefore we use a deterministic way to obtain port, + # so that users are aware of orphan processes by seeing the port occupied. + port = 2 ** 15 + 2 ** 14 + hash(os.getuid() if sys.platform != "win32" else 1) % 2 ** 14 + parser.add_argument( + "--dist-url", + default="tcp://127.0.0.1:{}".format(port), + help="initialization URL for pytorch distributed backend. See " + "https://pytorch.org/docs/stable/distributed.html for details.", + ) + parser.add_argument( + "opts", + help="Modify config options by adding 'KEY VALUE' pairs at the end of the command. " + "See config references at " + "https://detectron2.readthedocs.io/modules/config.html#config-references", + default=None, + nargs=argparse.REMAINDER, + ) + return parser + + +def _try_get_key(cfg, *keys, default=None): + """ + Try select keys from cfg until the first key that exists. Otherwise return default. + """ + if isinstance(cfg, CfgNode): + cfg = OmegaConf.create(cfg.dump()) + for k in keys: + parts = k.split(".") + # https://github.com/omry/omegaconf/issues/674 + for p in parts: + if p not in cfg: + break + cfg = OmegaConf.select(cfg, p) + else: + return cfg + return default + + +def _highlight(code, filename): + try: + import pygments + except ImportError: + return code + + from pygments.lexers import Python3Lexer, YamlLexer + from pygments.formatters import Terminal256Formatter + + lexer = Python3Lexer() if filename.endswith(".py") else YamlLexer() + code = pygments.highlight(code, lexer, Terminal256Formatter(style="monokai")) + return code + + +def default_setup(cfg, args): + """ + Perform some basic common setups at the beginning of a job, including: + + 1. Set up the detectron2 logger + 2. Log basic information about environment, cmdline arguments, and config + 3. Backup the config to the output directory + + Args: + cfg (CfgNode or omegaconf.DictConfig): the full config to be used + args (argparse.NameSpace): the command line arguments to be logged + """ + output_dir = _try_get_key(cfg, "OUTPUT_DIR", "output_dir", "train.output_dir") + if comm.is_main_process() and output_dir: + PathManager.mkdirs(output_dir) + + rank = comm.get_rank() + setup_logger(output_dir, distributed_rank=rank, name="fvcore") + logger = setup_logger(output_dir, distributed_rank=rank) + + logger.info("Rank of current process: {}. World size: {}".format(rank, comm.get_world_size())) + logger.info("Environment info:\n" + collect_env_info()) + + logger.info("Command line arguments: " + str(args)) + if hasattr(args, "config_file") and args.config_file != "": + logger.info( + "Contents of args.config_file={}:\n{}".format( + args.config_file, + _highlight(PathManager.open(args.config_file, "r").read(), args.config_file), + ) + ) + + if comm.is_main_process() and output_dir: + # Note: some of our scripts may expect the existence of + # config.yaml in output directory + path = os.path.join(output_dir, "config.yaml") + if isinstance(cfg, CfgNode): + logger.info("Running with full config:\n{}".format(_highlight(cfg.dump(), ".yaml"))) + with PathManager.open(path, "w") as f: + f.write(cfg.dump()) + else: + LazyConfig.save(cfg, path) + logger.info("Full config saved to {}".format(path)) + + # make sure each worker has a different, yet deterministic seed if specified + seed = _try_get_key(cfg, "SEED", "train.seed", default=-1) + seed_all_rng(None if seed < 0 else seed + rank) + + # cudnn benchmark has large overhead. It shouldn't be used considering the small size of + # typical validation set. + if not (hasattr(args, "eval_only") and args.eval_only): + torch.backends.cudnn.benchmark = _try_get_key( + cfg, "CUDNN_BENCHMARK", "train.cudnn_benchmark", default=False + ) + + +def default_writers(output_dir: str, max_iter: Optional[int] = None): + """ + Build a list of :class:`EventWriter` to be used. + It now consists of a :class:`CommonMetricPrinter`, + :class:`TensorboardXWriter` and :class:`JSONWriter`. + + Args: + output_dir: directory to store JSON metrics and tensorboard events + max_iter: the total number of iterations + + Returns: + list[EventWriter]: a list of :class:`EventWriter` objects. + """ + return [ + # It may not always print what you want to see, since it prints "common" metrics only. + CommonMetricPrinter(max_iter), + JSONWriter(os.path.join(output_dir, "metrics.json")), + TensorboardXWriter(output_dir), + ] + + +class DefaultPredictor: + """ + Create a simple end-to-end predictor with the given config that runs on + single device for a single input image. + + Compared to using the model directly, this class does the following additions: + + 1. Load checkpoint from `cfg.MODEL.WEIGHTS`. + 2. Always take BGR image as the input and apply conversion defined by `cfg.INPUT.FORMAT`. + 3. Apply resizing defined by `cfg.INPUT.{MIN,MAX}_SIZE_TEST`. + 4. Take one input image and produce a single output, instead of a batch. + + This is meant for simple demo purposes, so it does the above steps automatically. + This is not meant for benchmarks or running complicated inference logic. + If you'd like to do anything more fancy, please refer to its source code as examples + to build and use the model manually. + + Attributes: + metadata (Metadata): the metadata of the underlying dataset, obtained from + cfg.DATASETS.TEST. + + Examples: + :: + pred = DefaultPredictor(cfg) + inputs = cv2.imread("input.jpg") + outputs = pred(inputs) + """ + + def __init__(self, cfg): + self.cfg = cfg.clone() # cfg can be modified by model + self.model = build_model(self.cfg) + self.model.eval() + if len(cfg.DATASETS.TEST): + self.metadata = MetadataCatalog.get(cfg.DATASETS.TEST[0]) + + checkpointer = DetectionCheckpointer(self.model) + checkpointer.load(cfg.MODEL.WEIGHTS) + + self.aug = T.ResizeShortestEdge( + [cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MIN_SIZE_TEST], cfg.INPUT.MAX_SIZE_TEST + ) + + self.input_format = cfg.INPUT.FORMAT + assert self.input_format in ["RGB", "BGR"], self.input_format + + def __call__(self, original_image): + """ + Args: + original_image (np.ndarray): an image of shape (H, W, C) (in BGR order). + + Returns: + predictions (dict): + the output of the model for one image only. + See :doc:`/tutorials/models` for details about the format. + """ + with torch.no_grad(): # https://github.com/sphinx-doc/sphinx/issues/4258 + # Apply pre-processing to image. + if self.input_format == "RGB": + # whether the model expects BGR inputs or RGB + original_image = original_image[:, :, ::-1] + height, width = original_image.shape[:2] + image = self.aug.get_transform(original_image).apply_image(original_image) + image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1)) + + inputs = {"image": image, "height": height, "width": width} + predictions = self.model([inputs])[0] + return predictions + + +class DefaultTrainer(TrainerBase): + """ + A trainer with default training logic. It does the following: + + 1. Create a :class:`SimpleTrainer` using model, optimizer, dataloader + defined by the given config. Create a LR scheduler defined by the config. + 2. Load the last checkpoint or `cfg.MODEL.WEIGHTS`, if exists, when + `resume_or_load` is called. + 3. Register a few common hooks defined by the config. + + It is created to simplify the **standard model training workflow** and reduce code boilerplate + for users who only need the standard training workflow, with standard features. + It means this class makes *many assumptions* about your training logic that + may easily become invalid in a new research. In fact, any assumptions beyond those made in the + :class:`SimpleTrainer` are too much for research. + + The code of this class has been annotated about restrictive assumptions it makes. + When they do not work for you, you're encouraged to: + + 1. Overwrite methods of this class, OR: + 2. Use :class:`SimpleTrainer`, which only does minimal SGD training and + nothing else. You can then add your own hooks if needed. OR: + 3. Write your own training loop similar to `tools/plain_train_net.py`. + + See the :doc:`/tutorials/training` tutorials for more details. + + Note that the behavior of this class, like other functions/classes in + this file, is not stable, since it is meant to represent the "common default behavior". + It is only guaranteed to work well with the standard models and training workflow in detectron2. + To obtain more stable behavior, write your own training logic with other public APIs. + + Examples: + :: + trainer = DefaultTrainer(cfg) + trainer.resume_or_load() # load last checkpoint or MODEL.WEIGHTS + trainer.train() + + Attributes: + scheduler: + checkpointer (DetectionCheckpointer): + cfg (CfgNode): + """ + + def __init__(self, cfg): + """ + Args: + cfg (CfgNode): + """ + super().__init__() + logger = logging.getLogger("detectron2") + if not logger.isEnabledFor(logging.INFO): # setup_logger is not called for d2 + setup_logger() + cfg = DefaultTrainer.auto_scale_workers(cfg, comm.get_world_size()) + + # Assume these objects must be constructed in this order. + model = self.build_model(cfg) + optimizer = self.build_optimizer(cfg, model) + data_loader = self.build_train_loader(cfg) + + model = create_ddp_model(model, broadcast_buffers=False) + self._trainer = (AMPTrainer if cfg.SOLVER.AMP.ENABLED else SimpleTrainer)( + model, data_loader, optimizer + ) + + self.scheduler = self.build_lr_scheduler(cfg, optimizer) + self.checkpointer = DetectionCheckpointer( + # Assume you want to save checkpoints together with logs/statistics + model, + cfg.OUTPUT_DIR, + trainer=weakref.proxy(self), + ) + # load 2nd pretrained model + if cfg.MODEL.META_ARCHITECTURE in ['CLIPRCNN', 'CLIPFastRCNN', 'PretrainFastRCNN'] \ + and cfg.MODEL.CLIP.BB_RPN_WEIGHTS is not None\ + and cfg.MODEL.CLIP.CROP_REGION_TYPE == 'RPN': # load 2nd pretrained model + self.second_checkpointer = DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR, bb_rpn_weights=True) + else: + self.second_checkpointer = None + self.start_iter = 0 + self.max_iter = cfg.SOLVER.MAX_ITER + self.cfg = cfg + + self.register_hooks(self.build_hooks()) + + def resume_or_load(self, resume=True): + """ + If `resume==True` and `cfg.OUTPUT_DIR` contains the last checkpoint (defined by + a `last_checkpoint` file), resume from the file. Resuming means loading all + available states (eg. optimizer and scheduler) and update iteration counter + from the checkpoint. ``cfg.MODEL.WEIGHTS`` will not be used. + + Otherwise, this is considered as an independent training. The method will load model + weights from the file `cfg.MODEL.WEIGHTS` (but will not load other states) and start + from iteration 0. + + Args: + resume (bool): whether to do resume or not + """ + self.checkpointer.resume_or_load(self.cfg.MODEL.WEIGHTS, resume=resume) + if self.second_checkpointer: + self.second_checkpointer.resume_or_load( + self.cfg.MODEL.CLIP.BB_RPN_WEIGHTS, resume=False + ) + if resume and self.checkpointer.has_checkpoint(): + # The checkpoint stores the training iteration that just finished, thus we start + # at the next iteration + self.start_iter = self.iter + 1 + + def build_hooks(self): + """ + Build a list of default hooks, including timing, evaluation, + checkpointing, lr scheduling, precise BN, writing events. + + Returns: + list[HookBase]: + """ + cfg = self.cfg.clone() + cfg.defrost() + cfg.DATALOADER.NUM_WORKERS = 0 # save some memory and time for PreciseBN + + ret = [ + hooks.IterationTimer(), + hooks.LRScheduler(), + hooks.PreciseBN( + # Run at the same freq as (but before) evaluation. + cfg.TEST.EVAL_PERIOD, + self.model, + # Build a new data loader to not affect training + self.build_train_loader(cfg), + cfg.TEST.PRECISE_BN.NUM_ITER, + ) + if cfg.TEST.PRECISE_BN.ENABLED and get_bn_modules(self.model) + else None, + ] + + # Do PreciseBN before checkpointer, because it updates the model and need to + # be saved by checkpointer. + # This is not always the best: if checkpointing has a different frequency, + # some checkpoints may have more precise statistics than others. + if comm.is_main_process(): + ret.append(hooks.PeriodicCheckpointer(self.checkpointer, cfg.SOLVER.CHECKPOINT_PERIOD)) + + def test_and_save_results(): + self._last_eval_results = self.test(self.cfg, self.model) + return self._last_eval_results + + # Do evaluation after checkpointer, because then if it fails, + # we can use the saved checkpoint to debug. + ret.append(hooks.EvalHook(cfg.TEST.EVAL_PERIOD, test_and_save_results)) + + if comm.is_main_process(): + # Here the default print/log frequency of each writer is used. + # run writers in the end, so that evaluation metrics are written + ret.append(hooks.PeriodicWriter(self.build_writers(), period=20)) + return ret + + def build_writers(self): + """ + Build a list of writers to be used using :func:`default_writers()`. + If you'd like a different list of writers, you can overwrite it in + your trainer. + + Returns: + list[EventWriter]: a list of :class:`EventWriter` objects. + """ + return default_writers(self.cfg.OUTPUT_DIR, self.max_iter) + + def train(self): + """ + Run training. + + Returns: + OrderedDict of results, if evaluation is enabled. Otherwise None. + """ + super().train(self.start_iter, self.max_iter) + if len(self.cfg.TEST.EXPECTED_RESULTS) and comm.is_main_process(): + assert hasattr( + self, "_last_eval_results" + ), "No evaluation results obtained during training!" + verify_results(self.cfg, self._last_eval_results) + return self._last_eval_results + + def run_step(self): + self._trainer.iter = self.iter + self._trainer.run_step() + + @classmethod + def build_model(cls, cfg): + """ + Returns: + torch.nn.Module: + + It now calls :func:`detectron2.modeling.build_model`. + Overwrite it if you'd like a different model. + """ + model = build_model(cfg) + logger = logging.getLogger(__name__) + logger.info("Model:\n{}".format(model)) + return model + + @classmethod + def build_optimizer(cls, cfg, model): + """ + Returns: + torch.optim.Optimizer: + + It now calls :func:`detectron2.solver.build_optimizer`. + Overwrite it if you'd like a different optimizer. + """ + return build_optimizer(cfg, model) + + @classmethod + def build_lr_scheduler(cls, cfg, optimizer): + """ + It now calls :func:`detectron2.solver.build_lr_scheduler`. + Overwrite it if you'd like a different scheduler. + """ + return build_lr_scheduler(cfg, optimizer) + + @classmethod + def build_train_loader(cls, cfg): + """ + Returns: + iterable + + It now calls :func:`detectron2.data.build_detection_train_loader`. + Overwrite it if you'd like a different data loader. + """ + return build_detection_train_loader(cfg) + + @classmethod + def build_test_loader(cls, cfg, dataset_name): + """ + Returns: + iterable + + It now calls :func:`detectron2.data.build_detection_test_loader`. + Overwrite it if you'd like a different data loader. + """ + return build_detection_test_loader(cfg, dataset_name) + + @classmethod + def build_evaluator(cls, cfg, dataset_name): + """ + Returns: + DatasetEvaluator or None + + It is not implemented by default. + """ + raise NotImplementedError( + """ +If you want DefaultTrainer to automatically run evaluation, +please implement `build_evaluator()` in subclasses (see train_net.py for example). +Alternatively, you can call evaluation functions yourself (see Colab balloon tutorial for example). +""" + ) + + @classmethod + def test(cls, cfg, model, queries, evaluators=None): + """ + Args: + cfg (CfgNode): + model (nn.Module): + evaluators (list[DatasetEvaluator] or None): if None, will call + :meth:`build_evaluator`. Otherwise, must have the same length as + ``cfg.DATASETS.TEST``. + + Returns: + dict: a dict of result metrics + """ + logger = logging.getLogger(__name__) + if isinstance(evaluators, DatasetEvaluator): + evaluators = [evaluators] + if evaluators is not None: + assert len(cfg.DATASETS.TEST) == len(evaluators), "{} != {}".format( + len(cfg.DATASETS.TEST), len(evaluators) + ) + + results = OrderedDict() + for idx, dataset_name in enumerate(cfg.DATASETS.TEST): + data_loader = cls.build_test_loader(cfg, dataset_name) + # When evaluators are passed in as arguments, + # implicitly assume that evaluators can be created before data_loader. + if evaluators is not None: + evaluator = evaluators[idx] + results_i = inference_on_dataset(model, data_loader, queries, None) + results[dataset_name] = results_i + if comm.is_main_process(): + assert isinstance( + results_i, dict + ), "Evaluator must return a dict on the main process. Got {} instead.".format( + results_i + ) + logger.info("Evaluation results for {} in csv format:".format(dataset_name)) + print_csv_format(results_i) + + if len(results) == 1: + results = list(results.values())[0] + return results + + @staticmethod + def auto_scale_workers(cfg, num_workers: int): + """ + When the config is defined for certain number of workers (according to + ``cfg.SOLVER.REFERENCE_WORLD_SIZE``) that's different from the number of + workers currently in use, returns a new cfg where the total batch size + is scaled so that the per-GPU batch size stays the same as the + original ``IMS_PER_BATCH // REFERENCE_WORLD_SIZE``. + + Other config options are also scaled accordingly: + * training steps and warmup steps are scaled inverse proportionally. + * learning rate are scaled proportionally, following :paper:`ImageNet in 1h`. + + For example, with the original config like the following: + + .. code-block:: yaml + + IMS_PER_BATCH: 16 + BASE_LR: 0.1 + REFERENCE_WORLD_SIZE: 8 + MAX_ITER: 5000 + STEPS: (4000,) + CHECKPOINT_PERIOD: 1000 + + When this config is used on 16 GPUs instead of the reference number 8, + calling this method will return a new config with: + + .. code-block:: yaml + + IMS_PER_BATCH: 32 + BASE_LR: 0.2 + REFERENCE_WORLD_SIZE: 16 + MAX_ITER: 2500 + STEPS: (2000,) + CHECKPOINT_PERIOD: 500 + + Note that both the original config and this new config can be trained on 16 GPUs. + It's up to user whether to enable this feature (by setting ``REFERENCE_WORLD_SIZE``). + + Returns: + CfgNode: a new config. Same as original if ``cfg.SOLVER.REFERENCE_WORLD_SIZE==0``. + """ + old_world_size = cfg.SOLVER.REFERENCE_WORLD_SIZE + if old_world_size == 0 or old_world_size == num_workers: + return cfg + cfg = cfg.clone() + frozen = cfg.is_frozen() + cfg.defrost() + + assert ( + cfg.SOLVER.IMS_PER_BATCH % old_world_size == 0 + ), "Invalid REFERENCE_WORLD_SIZE in config!" + scale = num_workers / old_world_size + bs = cfg.SOLVER.IMS_PER_BATCH = int(round(cfg.SOLVER.IMS_PER_BATCH * scale)) + lr = cfg.SOLVER.BASE_LR = cfg.SOLVER.BASE_LR * scale + max_iter = cfg.SOLVER.MAX_ITER = int(round(cfg.SOLVER.MAX_ITER / scale)) + warmup_iter = cfg.SOLVER.WARMUP_ITERS = int(round(cfg.SOLVER.WARMUP_ITERS / scale)) + cfg.SOLVER.STEPS = tuple(int(round(s / scale)) for s in cfg.SOLVER.STEPS) + cfg.TEST.EVAL_PERIOD = int(round(cfg.TEST.EVAL_PERIOD / scale)) + cfg.SOLVER.CHECKPOINT_PERIOD = int(round(cfg.SOLVER.CHECKPOINT_PERIOD / scale)) + cfg.SOLVER.REFERENCE_WORLD_SIZE = num_workers # maintain invariant + logger = logging.getLogger(__name__) + logger.info( + f"Auto-scaling the config to batch_size={bs}, learning_rate={lr}, " + f"max_iter={max_iter}, warmup={warmup_iter}." + ) + + if frozen: + cfg.freeze() + return cfg + + +# Access basic attributes from the underlying trainer +for _attr in ["model", "data_loader", "optimizer"]: + setattr( + DefaultTrainer, + _attr, + property( + # getter + lambda self, x=_attr: getattr(self._trainer, x), + # setter + lambda self, value, x=_attr: setattr(self._trainer, x, value), + ), + ) diff --git a/detectron2/engine/hooks.py b/detectron2/engine/hooks.py new file mode 100644 index 0000000000000000000000000000000000000000..79e1b164a54ef14889c3c7082534b6c38d85413a --- /dev/null +++ b/detectron2/engine/hooks.py @@ -0,0 +1,466 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import datetime +import itertools +import logging +import os +import tempfile +import time +from collections import Counter +import torch +from fvcore.common.checkpoint import PeriodicCheckpointer as _PeriodicCheckpointer +from fvcore.common.param_scheduler import ParamScheduler +from fvcore.common.timer import Timer +from fvcore.nn.precise_bn import get_bn_modules, update_bn_stats + +import detectron2.utils.comm as comm +from detectron2.evaluation.testing import flatten_results_dict +from detectron2.solver import LRMultiplier +from detectron2.utils.events import EventStorage, EventWriter +from detectron2.utils.file_io import PathManager + +from .train_loop import HookBase + +__all__ = [ + "CallbackHook", + "IterationTimer", + "PeriodicWriter", + "PeriodicCheckpointer", + "LRScheduler", + "AutogradProfiler", + "EvalHook", + "PreciseBN", +] + + +""" +Implement some common hooks. +""" + + +class CallbackHook(HookBase): + """ + Create a hook using callback functions provided by the user. + """ + + def __init__(self, *, before_train=None, after_train=None, before_step=None, after_step=None): + """ + Each argument is a function that takes one argument: the trainer. + """ + self._before_train = before_train + self._before_step = before_step + self._after_step = after_step + self._after_train = after_train + + def before_train(self): + if self._before_train: + self._before_train(self.trainer) + + def after_train(self): + if self._after_train: + self._after_train(self.trainer) + # The functions may be closures that hold reference to the trainer + # Therefore, delete them to avoid circular reference. + del self._before_train, self._after_train + del self._before_step, self._after_step + + def before_step(self): + if self._before_step: + self._before_step(self.trainer) + + def after_step(self): + if self._after_step: + self._after_step(self.trainer) + + +class IterationTimer(HookBase): + """ + Track the time spent for each iteration (each run_step call in the trainer). + Print a summary in the end of training. + + This hook uses the time between the call to its :meth:`before_step` + and :meth:`after_step` methods. + Under the convention that :meth:`before_step` of all hooks should only + take negligible amount of time, the :class:`IterationTimer` hook should be + placed at the beginning of the list of hooks to obtain accurate timing. + """ + + def __init__(self, warmup_iter=3): + """ + Args: + warmup_iter (int): the number of iterations at the beginning to exclude + from timing. + """ + self._warmup_iter = warmup_iter + self._step_timer = Timer() + self._start_time = time.perf_counter() + self._total_timer = Timer() + + def before_train(self): + self._start_time = time.perf_counter() + self._total_timer.reset() + self._total_timer.pause() + + def after_train(self): + logger = logging.getLogger(__name__) + total_time = time.perf_counter() - self._start_time + total_time_minus_hooks = self._total_timer.seconds() + hook_time = total_time - total_time_minus_hooks + + num_iter = self.trainer.iter + 1 - self.trainer.start_iter - self._warmup_iter + + if num_iter > 0 and total_time_minus_hooks > 0: + # Speed is meaningful only after warmup + # NOTE this format is parsed by grep in some scripts + logger.info( + "Overall training speed: {} iterations in {} ({:.4f} s / it)".format( + num_iter, + str(datetime.timedelta(seconds=int(total_time_minus_hooks))), + total_time_minus_hooks / num_iter, + ) + ) + + logger.info( + "Total training time: {} ({} on hooks)".format( + str(datetime.timedelta(seconds=int(total_time))), + str(datetime.timedelta(seconds=int(hook_time))), + ) + ) + + def before_step(self): + self._step_timer.reset() + self._total_timer.resume() + + def after_step(self): + # +1 because we're in after_step, the current step is done + # but not yet counted + iter_done = self.trainer.iter - self.trainer.start_iter + 1 + if iter_done >= self._warmup_iter: + sec = self._step_timer.seconds() + self.trainer.storage.put_scalars(time=sec) + else: + self._start_time = time.perf_counter() + self._total_timer.reset() + + self._total_timer.pause() + + +class PeriodicWriter(HookBase): + """ + Write events to EventStorage (by calling ``writer.write()``) periodically. + + It is executed every ``period`` iterations and after the last iteration. + Note that ``period`` does not affect how data is smoothed by each writer. + """ + + def __init__(self, writers, period=20): + """ + Args: + writers (list[EventWriter]): a list of EventWriter objects + period (int): + """ + self._writers = writers + for w in writers: + assert isinstance(w, EventWriter), w + self._period = period + + def after_step(self): + if (self.trainer.iter + 1) % self._period == 0 or ( + self.trainer.iter == self.trainer.max_iter - 1 + ): + for writer in self._writers: + writer.write() + + def after_train(self): + for writer in self._writers: + # If any new data is found (e.g. produced by other after_train), + # write them before closing + writer.write() + writer.close() + + +class PeriodicCheckpointer(_PeriodicCheckpointer, HookBase): + """ + Same as :class:`detectron2.checkpoint.PeriodicCheckpointer`, but as a hook. + + Note that when used as a hook, + it is unable to save additional data other than what's defined + by the given `checkpointer`. + + It is executed every ``period`` iterations and after the last iteration. + """ + + def before_train(self): + self.max_iter = self.trainer.max_iter + + def after_step(self): + # No way to use **kwargs + self.step(self.trainer.iter) + + +class LRScheduler(HookBase): + """ + A hook which executes a torch builtin LR scheduler and summarizes the LR. + It is executed after every iteration. + """ + + def __init__(self, optimizer=None, scheduler=None): + """ + Args: + optimizer (torch.optim.Optimizer): + scheduler (torch.optim.LRScheduler or fvcore.common.param_scheduler.ParamScheduler): + if a :class:`ParamScheduler` object, it defines the multiplier over the base LR + in the optimizer. + + If any argument is not given, will try to obtain it from the trainer. + """ + self._optimizer = optimizer + self._scheduler = scheduler + + def before_train(self): + self._optimizer = self._optimizer or self.trainer.optimizer + if isinstance(self.scheduler, ParamScheduler): + self._scheduler = LRMultiplier( + self._optimizer, + self.scheduler, + self.trainer.max_iter, + last_iter=self.trainer.iter - 1, + ) + + # NOTE: some heuristics on what LR to summarize + # summarize the param group with most parameters + largest_group = max(len(g["params"]) for g in self._optimizer.param_groups) + + if largest_group == 1: + # If all groups have one parameter, + # then find the most common initial LR, and use it for summary + lr_count = Counter([g["lr"] for g in self._optimizer.param_groups]) + lr = lr_count.most_common()[0][0] + for i, g in enumerate(self._optimizer.param_groups): + if g["lr"] == lr: + self._best_param_group_id = i + break + else: + for i, g in enumerate(self._optimizer.param_groups): + if len(g["params"]) == largest_group: + self._best_param_group_id = i + break + + def after_step(self): + lr = self._optimizer.param_groups[self._best_param_group_id]["lr"] + self.trainer.storage.put_scalar("lr", lr, smoothing_hint=False) + self.scheduler.step() + + @property + def scheduler(self): + return self._scheduler or self.trainer.scheduler + + def state_dict(self): + if isinstance(self.scheduler, torch.optim.lr_scheduler._LRScheduler): + return self.scheduler.state_dict() + return {} + + def load_state_dict(self, state_dict): + if isinstance(self.scheduler, torch.optim.lr_scheduler._LRScheduler): + logger = logging.getLogger(__name__) + logger.info("Loading scheduler from state_dict ...") + self.scheduler.load_state_dict(state_dict) + + +class AutogradProfiler(HookBase): + """ + A hook which runs `torch.autograd.profiler.profile`. + + Examples: + :: + hooks.AutogradProfiler( + lambda trainer: trainer.iter > 10 and trainer.iter < 20, self.cfg.OUTPUT_DIR + ) + + The above example will run the profiler for iteration 10~20 and dump + results to ``OUTPUT_DIR``. We did not profile the first few iterations + because they are typically slower than the rest. + The result files can be loaded in the ``chrome://tracing`` page in chrome browser. + + Note: + When used together with NCCL on older version of GPUs, + autograd profiler may cause deadlock because it unnecessarily allocates + memory on every device it sees. The memory management calls, if + interleaved with NCCL calls, lead to deadlock on GPUs that do not + support ``cudaLaunchCooperativeKernelMultiDevice``. + """ + + def __init__(self, enable_predicate, output_dir, *, use_cuda=True): + """ + Args: + enable_predicate (callable[trainer -> bool]): a function which takes a trainer, + and returns whether to enable the profiler. + It will be called once every step, and can be used to select which steps to profile. + output_dir (str): the output directory to dump tracing files. + use_cuda (bool): same as in `torch.autograd.profiler.profile`. + """ + self._enable_predicate = enable_predicate + self._use_cuda = use_cuda + self._output_dir = output_dir + + def before_step(self): + if self._enable_predicate(self.trainer): + self._profiler = torch.autograd.profiler.profile(use_cuda=self._use_cuda) + self._profiler.__enter__() + else: + self._profiler = None + + def after_step(self): + if self._profiler is None: + return + self._profiler.__exit__(None, None, None) + PathManager.mkdirs(self._output_dir) + out_file = os.path.join( + self._output_dir, "profiler-trace-iter{}.json".format(self.trainer.iter) + ) + if "://" not in out_file: + self._profiler.export_chrome_trace(out_file) + else: + # Support non-posix filesystems + with tempfile.TemporaryDirectory(prefix="detectron2_profiler") as d: + tmp_file = os.path.join(d, "tmp.json") + self._profiler.export_chrome_trace(tmp_file) + with open(tmp_file) as f: + content = f.read() + with PathManager.open(out_file, "w") as f: + f.write(content) + + +class EvalHook(HookBase): + """ + Run an evaluation function periodically, and at the end of training. + + It is executed every ``eval_period`` iterations and after the last iteration. + """ + + def __init__(self, eval_period, eval_function): + """ + Args: + eval_period (int): the period to run `eval_function`. Set to 0 to + not evaluate periodically (but still after the last iteration). + eval_function (callable): a function which takes no arguments, and + returns a nested dict of evaluation metrics. + + Note: + This hook must be enabled in all or none workers. + If you would like only certain workers to perform evaluation, + give other workers a no-op function (`eval_function=lambda: None`). + """ + self._period = eval_period + self._func = eval_function + + def _do_eval(self): + results = self._func() + + if results: + assert isinstance( + results, dict + ), "Eval function must return a dict. Got {} instead.".format(results) + + flattened_results = flatten_results_dict(results) + for k, v in flattened_results.items(): + try: + v = float(v) + except Exception as e: + raise ValueError( + "[EvalHook] eval_function should return a nested dict of float. " + "Got '{}: {}' instead.".format(k, v) + ) from e + self.trainer.storage.put_scalars(**flattened_results, smoothing_hint=False) + + # Evaluation may take different time among workers. + # A barrier make them start the next iteration together. + comm.synchronize() + + def after_step(self): + next_iter = self.trainer.iter + 1 + if self._period > 0 and next_iter % self._period == 0: + # do the last eval in after_train + if next_iter != self.trainer.max_iter: + self._do_eval() + + def after_train(self): + # This condition is to prevent the eval from running after a failed training + if self.trainer.iter + 1 >= self.trainer.max_iter: + self._do_eval() + # func is likely a closure that holds reference to the trainer + # therefore we clean it to avoid circular reference in the end + del self._func + + +class PreciseBN(HookBase): + """ + The standard implementation of BatchNorm uses EMA in inference, which is + sometimes suboptimal. + This class computes the true average of statistics rather than the moving average, + and put true averages to every BN layer in the given model. + + It is executed every ``period`` iterations and after the last iteration. + """ + + def __init__(self, period, model, data_loader, num_iter): + """ + Args: + period (int): the period this hook is run, or 0 to not run during training. + The hook will always run in the end of training. + model (nn.Module): a module whose all BN layers in training mode will be + updated by precise BN. + Note that user is responsible for ensuring the BN layers to be + updated are in training mode when this hook is triggered. + data_loader (iterable): it will produce data to be run by `model(data)`. + num_iter (int): number of iterations used to compute the precise + statistics. + """ + self._logger = logging.getLogger(__name__) + if len(get_bn_modules(model)) == 0: + self._logger.info( + "PreciseBN is disabled because model does not contain BN layers in training mode." + ) + self._disabled = True + return + + self._model = model + self._data_loader = data_loader + self._num_iter = num_iter + self._period = period + self._disabled = False + + self._data_iter = None + + def after_step(self): + next_iter = self.trainer.iter + 1 + is_final = next_iter == self.trainer.max_iter + if is_final or (self._period > 0 and next_iter % self._period == 0): + self.update_stats() + + def update_stats(self): + """ + Update the model with precise statistics. Users can manually call this method. + """ + if self._disabled: + return + + if self._data_iter is None: + self._data_iter = iter(self._data_loader) + + def data_loader(): + for num_iter in itertools.count(1): + if num_iter % 100 == 0: + self._logger.info( + "Running precise-BN ... {}/{} iterations.".format(num_iter, self._num_iter) + ) + # This way we can reuse the same iterator + yield next(self._data_iter) + + with EventStorage(): # capture events in a new storage to discard them + self._logger.info( + "Running precise-BN for {} iterations... ".format(self._num_iter) + + "Note that this could produce different statistics every time." + ) + update_bn_stats(self._model, data_loader(), self._num_iter) diff --git a/detectron2/engine/launch.py b/detectron2/engine/launch.py new file mode 100644 index 0000000000000000000000000000000000000000..40dad262dd9929e6e4e9c60424b3fda1ab97318c --- /dev/null +++ b/detectron2/engine/launch.py @@ -0,0 +1,125 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +from datetime import timedelta +import torch +import torch.distributed as dist +import torch.multiprocessing as mp + +from detectron2.utils import comm + +__all__ = ["DEFAULT_TIMEOUT", "launch"] + +DEFAULT_TIMEOUT = timedelta(minutes=30) + + +def _find_free_port(): + import socket + + sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) + # Binding to port 0 will cause the OS to find an available port for us + sock.bind(("", 0)) + port = sock.getsockname()[1] + sock.close() + # NOTE: there is still a chance the port could be taken by other processes. + return port + + +def launch( + main_func, + num_gpus_per_machine, + num_machines=1, + machine_rank=0, + dist_url=None, + args=(), + timeout=DEFAULT_TIMEOUT, +): + """ + Launch multi-gpu or distributed training. + This function must be called on all machines involved in the training. + It will spawn child processes (defined by ``num_gpus_per_machine``) on each machine. + + Args: + main_func: a function that will be called by `main_func(*args)` + num_gpus_per_machine (int): number of GPUs per machine + num_machines (int): the total number of machines + machine_rank (int): the rank of this machine + dist_url (str): url to connect to for distributed jobs, including protocol + e.g. "tcp://127.0.0.1:8686". + Can be set to "auto" to automatically select a free port on localhost + timeout (timedelta): timeout of the distributed workers + args (tuple): arguments passed to main_func + """ + world_size = num_machines * num_gpus_per_machine + if world_size > 1: + # https://github.com/pytorch/pytorch/pull/14391 + # TODO prctl in spawned processes + + if dist_url == "auto": + assert num_machines == 1, "dist_url=auto not supported in multi-machine jobs." + port = _find_free_port() + dist_url = f"tcp://127.0.0.1:{port}" + if num_machines > 1 and dist_url.startswith("file://"): + logger = logging.getLogger(__name__) + logger.warning( + "file:// is not a reliable init_method in multi-machine jobs. Prefer tcp://" + ) + + mp.spawn( + _distributed_worker, + nprocs=num_gpus_per_machine, + args=( + main_func, + world_size, + num_gpus_per_machine, + machine_rank, + dist_url, + args, + timeout, + ), + daemon=False, + ) + else: + main_func(*args) + + +def _distributed_worker( + local_rank, + main_func, + world_size, + num_gpus_per_machine, + machine_rank, + dist_url, + args, + timeout=DEFAULT_TIMEOUT, +): + assert torch.cuda.is_available(), "cuda is not available. Please check your installation." + global_rank = machine_rank * num_gpus_per_machine + local_rank + try: + dist.init_process_group( + backend="NCCL", + init_method=dist_url, + world_size=world_size, + rank=global_rank, + timeout=timeout, + ) + except Exception as e: + logger = logging.getLogger(__name__) + logger.error("Process group URL: {}".format(dist_url)) + raise e + # synchronize is needed here to prevent a possible timeout after calling init_process_group + # See: https://github.com/facebookresearch/maskrcnn-benchmark/issues/172 + comm.synchronize() + + assert num_gpus_per_machine <= torch.cuda.device_count() + torch.cuda.set_device(local_rank) + + # Setup the local process group (which contains ranks within the same machine) + assert comm._LOCAL_PROCESS_GROUP is None + num_machines = world_size // num_gpus_per_machine + for i in range(num_machines): + ranks_on_i = list(range(i * num_gpus_per_machine, (i + 1) * num_gpus_per_machine)) + pg = dist.new_group(ranks_on_i) + if i == machine_rank: + comm._LOCAL_PROCESS_GROUP = pg + + main_func(*args) diff --git a/detectron2/engine/train_loop.py b/detectron2/engine/train_loop.py new file mode 100644 index 0000000000000000000000000000000000000000..1a4f8bfd09d033ec64bc5dbc7c59391aa2346961 --- /dev/null +++ b/detectron2/engine/train_loop.py @@ -0,0 +1,408 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import logging +import numpy as np +import time +import weakref +from typing import Dict, List, Optional +import torch +from torch.nn.parallel import DataParallel, DistributedDataParallel + +import detectron2.utils.comm as comm +from detectron2.utils.events import EventStorage, get_event_storage +from detectron2.utils.logger import _log_api_usage + +__all__ = ["HookBase", "TrainerBase", "SimpleTrainer", "AMPTrainer"] + + +class HookBase: + """ + Base class for hooks that can be registered with :class:`TrainerBase`. + + Each hook can implement 4 methods. The way they are called is demonstrated + in the following snippet: + :: + hook.before_train() + for iter in range(start_iter, max_iter): + hook.before_step() + trainer.run_step() + hook.after_step() + iter += 1 + hook.after_train() + + Notes: + 1. In the hook method, users can access ``self.trainer`` to access more + properties about the context (e.g., model, current iteration, or config + if using :class:`DefaultTrainer`). + + 2. A hook that does something in :meth:`before_step` can often be + implemented equivalently in :meth:`after_step`. + If the hook takes non-trivial time, it is strongly recommended to + implement the hook in :meth:`after_step` instead of :meth:`before_step`. + The convention is that :meth:`before_step` should only take negligible time. + + Following this convention will allow hooks that do care about the difference + between :meth:`before_step` and :meth:`after_step` (e.g., timer) to + function properly. + + """ + + trainer: "TrainerBase" = None + """ + A weak reference to the trainer object. Set by the trainer when the hook is registered. + """ + + def before_train(self): + """ + Called before the first iteration. + """ + pass + + def after_train(self): + """ + Called after the last iteration. + """ + pass + + def before_step(self): + """ + Called before each iteration. + """ + pass + + def after_step(self): + """ + Called after each iteration. + """ + pass + + def state_dict(self): + """ + Hooks are stateless by default, but can be made checkpointable by + implementing `state_dict` and `load_state_dict`. + """ + return {} + + +class TrainerBase: + """ + Base class for iterative trainer with hooks. + + The only assumption we made here is: the training runs in a loop. + A subclass can implement what the loop is. + We made no assumptions about the existence of dataloader, optimizer, model, etc. + + Attributes: + iter(int): the current iteration. + + start_iter(int): The iteration to start with. + By convention the minimum possible value is 0. + + max_iter(int): The iteration to end training. + + storage(EventStorage): An EventStorage that's opened during the course of training. + """ + + def __init__(self) -> None: + self._hooks: List[HookBase] = [] + self.iter: int = 0 + self.start_iter: int = 0 + self.max_iter: int + self.storage: EventStorage + _log_api_usage("trainer." + self.__class__.__name__) + + def register_hooks(self, hooks: List[Optional[HookBase]]) -> None: + """ + Register hooks to the trainer. The hooks are executed in the order + they are registered. + + Args: + hooks (list[Optional[HookBase]]): list of hooks + """ + hooks = [h for h in hooks if h is not None] + for h in hooks: + assert isinstance(h, HookBase) + # To avoid circular reference, hooks and trainer cannot own each other. + # This normally does not matter, but will cause memory leak if the + # involved objects contain __del__: + # See http://engineering.hearsaysocial.com/2013/06/16/circular-references-in-python/ + h.trainer = weakref.proxy(self) + self._hooks.extend(hooks) + + def train(self, start_iter: int, max_iter: int): + """ + Args: + start_iter, max_iter (int): See docs above + """ + logger = logging.getLogger(__name__) + logger.info("Starting training from iteration {}".format(start_iter)) + + self.iter = self.start_iter = start_iter + self.max_iter = max_iter + + with EventStorage(start_iter) as self.storage: + try: + self.before_train() + for self.iter in range(start_iter, max_iter): + self.before_step() + self.run_step() + self.after_step() + # self.iter == max_iter can be used by `after_train` to + # tell whether the training successfully finished or failed + # due to exceptions. + self.iter += 1 + except Exception: + logger.exception("Exception during training:") + raise + finally: + self.after_train() + + def before_train(self): + for h in self._hooks: + h.before_train() + + def after_train(self): + self.storage.iter = self.iter + for h in self._hooks: + h.after_train() + + def before_step(self): + # Maintain the invariant that storage.iter == trainer.iter + # for the entire execution of each step + self.storage.iter = self.iter + + for h in self._hooks: + h.before_step() + + def after_step(self): + for h in self._hooks: + h.after_step() + + def run_step(self): + raise NotImplementedError + + def state_dict(self): + ret = {"iteration": self.iter} + hooks_state = {} + for h in self._hooks: + sd = h.state_dict() + if sd: + name = type(h).__qualname__ + if name in hooks_state: + # TODO handle repetitive stateful hooks + continue + hooks_state[name] = sd + if hooks_state: + ret["hooks"] = hooks_state + return ret + + def load_state_dict(self, state_dict): + logger = logging.getLogger(__name__) + self.iter = state_dict["iteration"] + for key, value in state_dict.get("hooks", {}).items(): + for h in self._hooks: + try: + name = type(h).__qualname__ + except AttributeError: + continue + if name == key: + h.load_state_dict(value) + break + else: + logger.warning(f"Cannot find the hook '{key}', its state_dict is ignored.") + + +class SimpleTrainer(TrainerBase): + """ + A simple trainer for the most common type of task: + single-cost single-optimizer single-data-source iterative optimization, + optionally using data-parallelism. + It assumes that every step, you: + + 1. Compute the loss with a data from the data_loader. + 2. Compute the gradients with the above loss. + 3. Update the model with the optimizer. + + All other tasks during training (checkpointing, logging, evaluation, LR schedule) + are maintained by hooks, which can be registered by :meth:`TrainerBase.register_hooks`. + + If you want to do anything fancier than this, + either subclass TrainerBase and implement your own `run_step`, + or write your own training loop. + """ + + def __init__(self, model, data_loader, optimizer): + """ + Args: + model: a torch Module. Takes a data from data_loader and returns a + dict of losses. + data_loader: an iterable. Contains data to be used to call model. + optimizer: a torch optimizer. + """ + super().__init__() + + """ + We set the model to training mode in the trainer. + However it's valid to train a model that's in eval mode. + If you want your model (or a submodule of it) to behave + like evaluation during training, you can overwrite its train() method. + """ + model.train() + + self.model = model + self.data_loader = data_loader + self._data_loader_iter = iter(data_loader) + self.optimizer = optimizer + + def run_step(self): + """ + Implement the standard training logic described above. + """ + assert self.model.training, "[SimpleTrainer] model was changed to eval mode!" + start = time.perf_counter() + """ + If you want to do something with the data, you can wrap the dataloader. + """ + data = next(self._data_loader_iter) + data_time = time.perf_counter() - start + + """ + If you want to do something with the losses, you can wrap the model. + """ + loss_dict = self.model(data) + if isinstance(loss_dict, torch.Tensor): + losses = loss_dict + loss_dict = {"total_loss": loss_dict} + else: + losses = sum(loss_dict.values()) + + """ + If you need to accumulate gradients or do something similar, you can + wrap the optimizer with your custom `zero_grad()` method. + """ + self.optimizer.zero_grad() + losses.backward() + + self._write_metrics(loss_dict, data_time) + + """ + If you need gradient clipping/scaling or other processing, you can + wrap the optimizer with your custom `step()` method. But it is + suboptimal as explained in https://arxiv.org/abs/2006.15704 Sec 3.2.4 + """ + self.optimizer.step() + + def _write_metrics( + self, + loss_dict: Dict[str, torch.Tensor], + data_time: float, + prefix: str = "", + ): + """ + Args: + loss_dict (dict): dict of scalar losses + data_time (float): time taken by the dataloader iteration + """ + metrics_dict = {k: v.detach().cpu().item() for k, v in loss_dict.items()} + metrics_dict["data_time"] = data_time + + # Gather metrics among all workers for logging + # This assumes we do DDP-style training, which is currently the only + # supported method in detectron2. + all_metrics_dict = comm.gather(metrics_dict) + + if comm.is_main_process(): + storage = get_event_storage() + + # data_time among workers can have high variance. The actual latency + # caused by data_time is the maximum among workers. + data_time = np.max([x.pop("data_time") for x in all_metrics_dict]) + storage.put_scalar("data_time", data_time) + + # average the rest metrics + metrics_dict = { + k: np.mean([x[k] for x in all_metrics_dict]) for k in all_metrics_dict[0].keys() + } + total_losses_reduced = sum(metrics_dict.values()) + if not np.isfinite(total_losses_reduced): + raise FloatingPointError( + f"Loss became infinite or NaN at iteration={self.iter}!\n" + f"loss_dict = {metrics_dict}" + ) + + storage.put_scalar("{}total_loss".format(prefix), total_losses_reduced) + if len(metrics_dict) > 1: + storage.put_scalars(**metrics_dict) + + def state_dict(self): + ret = super().state_dict() + ret["optimizer"] = self.optimizer.state_dict() + return ret + + def load_state_dict(self, state_dict): + super().load_state_dict(state_dict) + self.optimizer.load_state_dict(state_dict["optimizer"]) + + +class AMPTrainer(SimpleTrainer): + """ + Like :class:`SimpleTrainer`, but uses PyTorch's native automatic mixed precision + in the training loop. + """ + + def __init__(self, model, data_loader, optimizer, grad_scaler=None): + """ + Args: + model, data_loader, optimizer: same as in :class:`SimpleTrainer`. + grad_scaler: torch GradScaler to automatically scale gradients. + """ + unsupported = "AMPTrainer does not support single-process multi-device training!" + if isinstance(model, DistributedDataParallel): + assert not (model.device_ids and len(model.device_ids) > 1), unsupported + assert not isinstance(model, DataParallel), unsupported + + super().__init__(model, data_loader, optimizer) + + if grad_scaler is None: + from torch.cuda.amp import GradScaler + + grad_scaler = GradScaler() + self.grad_scaler = grad_scaler + + def run_step(self): + """ + Implement the AMP training logic. + """ + assert self.model.training, "[AMPTrainer] model was changed to eval mode!" + assert torch.cuda.is_available(), "[AMPTrainer] CUDA is required for AMP training!" + from torch.cuda.amp import autocast + + start = time.perf_counter() + data = next(self._data_loader_iter) + data_time = time.perf_counter() - start + + with autocast(): + loss_dict = self.model(data) + if isinstance(loss_dict, torch.Tensor): + losses = loss_dict + loss_dict = {"total_loss": loss_dict} + else: + losses = sum(loss_dict.values()) + + self.optimizer.zero_grad() + self.grad_scaler.scale(losses).backward() + + self._write_metrics(loss_dict, data_time) + + self.grad_scaler.step(self.optimizer) + self.grad_scaler.update() + + def state_dict(self): + ret = super().state_dict() + ret["grad_scaler"] = self.grad_scaler.state_dict() + return ret + + def load_state_dict(self, state_dict): + super().load_state_dict(state_dict) + self.grad_scaler.load_state_dict(state_dict["grad_scaler"]) diff --git a/detectron2/evaluation/__init__.py b/detectron2/evaluation/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6be52b555f7fb49c7d8d2fa3792a51be81db0274 --- /dev/null +++ b/detectron2/evaluation/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .cityscapes_evaluation import CityscapesInstanceEvaluator, CityscapesSemSegEvaluator +from .coco_evaluation import COCOEvaluator +from .rotated_coco_evaluation import RotatedCOCOEvaluator +from .evaluator import DatasetEvaluator, DatasetEvaluators, inference_context, inference_on_dataset +from .lvis_evaluation import LVISEvaluator +from .panoptic_evaluation import COCOPanopticEvaluator +from .pascal_voc_evaluation import PascalVOCDetectionEvaluator +from .sem_seg_evaluation import SemSegEvaluator +from .testing import print_csv_format, verify_results +from .flickr30k_evaluation import FLICKR30KEvaluator + +__all__ = [k for k in globals().keys() if not k.startswith("_")] diff --git a/detectron2/evaluation/__pycache__/__init__.cpython-39.pyc b/detectron2/evaluation/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..469fa043337ace0b6592bd0242eb40b7d9ad3ebd Binary files /dev/null and b/detectron2/evaluation/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/cityscapes_evaluation.cpython-39.pyc b/detectron2/evaluation/__pycache__/cityscapes_evaluation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f54eef15b96988130188935caf047cc9434d766e Binary files /dev/null and b/detectron2/evaluation/__pycache__/cityscapes_evaluation.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/coco_evaluation.cpython-39.pyc b/detectron2/evaluation/__pycache__/coco_evaluation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3b80aae0b42a2729b456a150bbf5b6d5f432a2ae Binary files /dev/null and b/detectron2/evaluation/__pycache__/coco_evaluation.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/evaluator.cpython-39.pyc b/detectron2/evaluation/__pycache__/evaluator.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ef9d013106b68e37c7d94403c2ceac58bf375561 Binary files /dev/null and b/detectron2/evaluation/__pycache__/evaluator.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/fast_eval_api.cpython-39.pyc b/detectron2/evaluation/__pycache__/fast_eval_api.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..47022b910b3535a75b46391921b0f6e659d631f1 Binary files /dev/null and b/detectron2/evaluation/__pycache__/fast_eval_api.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/flickr30k_evaluation.cpython-39.pyc b/detectron2/evaluation/__pycache__/flickr30k_evaluation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..be69aeb652dad6085f8679234bb0ebb872a96487 Binary files /dev/null and b/detectron2/evaluation/__pycache__/flickr30k_evaluation.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/lvis_evaluation.cpython-39.pyc b/detectron2/evaluation/__pycache__/lvis_evaluation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1302cc24874bbbb239a9659562797b7c74239387 Binary files /dev/null and b/detectron2/evaluation/__pycache__/lvis_evaluation.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/panoptic_evaluation.cpython-39.pyc b/detectron2/evaluation/__pycache__/panoptic_evaluation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9c9e785df32f2dfe884e229c546e4eabb3f7d964 Binary files /dev/null and b/detectron2/evaluation/__pycache__/panoptic_evaluation.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/pascal_voc_evaluation.cpython-39.pyc b/detectron2/evaluation/__pycache__/pascal_voc_evaluation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7922ffca38e6b261680e3b6fd6d6662ed92fd2e7 Binary files /dev/null and b/detectron2/evaluation/__pycache__/pascal_voc_evaluation.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/rotated_coco_evaluation.cpython-39.pyc b/detectron2/evaluation/__pycache__/rotated_coco_evaluation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..42401dd86eac3e214a2f496611e345ad82104e7f Binary files /dev/null and b/detectron2/evaluation/__pycache__/rotated_coco_evaluation.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/sem_seg_evaluation.cpython-39.pyc b/detectron2/evaluation/__pycache__/sem_seg_evaluation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..93b4774b78032a314990a9f744eee51a5103c66d Binary files /dev/null and b/detectron2/evaluation/__pycache__/sem_seg_evaluation.cpython-39.pyc differ diff --git a/detectron2/evaluation/__pycache__/testing.cpython-39.pyc b/detectron2/evaluation/__pycache__/testing.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..906ce7323561653eced8f1898090192a7ec57dcc Binary files /dev/null and b/detectron2/evaluation/__pycache__/testing.cpython-39.pyc differ diff --git a/detectron2/evaluation/cityscapes_evaluation.py b/detectron2/evaluation/cityscapes_evaluation.py new file mode 100644 index 0000000000000000000000000000000000000000..3fb6c4cd5f752d639570d022cb23ce18491c370a --- /dev/null +++ b/detectron2/evaluation/cityscapes_evaluation.py @@ -0,0 +1,194 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import glob +import logging +import numpy as np +import os +import tempfile +from collections import OrderedDict +import torch +from PIL import Image + +from detectron2.data import MetadataCatalog +from detectron2.utils import comm +from detectron2.utils.file_io import PathManager + +from .evaluator import DatasetEvaluator + + +class CityscapesEvaluator(DatasetEvaluator): + """ + Base class for evaluation using cityscapes API. + """ + + def __init__(self, dataset_name): + """ + Args: + dataset_name (str): the name of the dataset. + It must have the following metadata associated with it: + "thing_classes", "gt_dir". + """ + self._metadata = MetadataCatalog.get(dataset_name) + self._cpu_device = torch.device("cpu") + self._logger = logging.getLogger(__name__) + + def reset(self): + self._working_dir = tempfile.TemporaryDirectory(prefix="cityscapes_eval_") + self._temp_dir = self._working_dir.name + # All workers will write to the same results directory + # TODO this does not work in distributed training + self._temp_dir = comm.all_gather(self._temp_dir)[0] + if self._temp_dir != self._working_dir.name: + self._working_dir.cleanup() + self._logger.info( + "Writing cityscapes results to temporary directory {} ...".format(self._temp_dir) + ) + + +class CityscapesInstanceEvaluator(CityscapesEvaluator): + """ + Evaluate instance segmentation results on cityscapes dataset using cityscapes API. + + Note: + * It does not work in multi-machine distributed training. + * It contains a synchronization, therefore has to be used on all ranks. + * Only the main process runs evaluation. + """ + + def process(self, inputs, outputs): + from cityscapesscripts.helpers.labels import name2label + + for input, output in zip(inputs, outputs): + file_name = input["file_name"] + basename = os.path.splitext(os.path.basename(file_name))[0] + pred_txt = os.path.join(self._temp_dir, basename + "_pred.txt") + + if "instances" in output: + output = output["instances"].to(self._cpu_device) + num_instances = len(output) + with open(pred_txt, "w") as fout: + for i in range(num_instances): + pred_class = output.pred_classes[i] + classes = self._metadata.thing_classes[pred_class] + class_id = name2label[classes].id + score = output.scores[i] + mask = output.pred_masks[i].numpy().astype("uint8") + png_filename = os.path.join( + self._temp_dir, basename + "_{}_{}.png".format(i, classes) + ) + + Image.fromarray(mask * 255).save(png_filename) + fout.write( + "{} {} {}\n".format(os.path.basename(png_filename), class_id, score) + ) + else: + # Cityscapes requires a prediction file for every ground truth image. + with open(pred_txt, "w") as fout: + pass + + def evaluate(self): + """ + Returns: + dict: has a key "segm", whose value is a dict of "AP" and "AP50". + """ + comm.synchronize() + if comm.get_rank() > 0: + return + import cityscapesscripts.evaluation.evalInstanceLevelSemanticLabeling as cityscapes_eval + + self._logger.info("Evaluating results under {} ...".format(self._temp_dir)) + + # set some global states in cityscapes evaluation API, before evaluating + cityscapes_eval.args.predictionPath = os.path.abspath(self._temp_dir) + cityscapes_eval.args.predictionWalk = None + cityscapes_eval.args.JSONOutput = False + cityscapes_eval.args.colorized = False + cityscapes_eval.args.gtInstancesFile = os.path.join(self._temp_dir, "gtInstances.json") + + # These lines are adopted from + # https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/evalInstanceLevelSemanticLabeling.py # noqa + gt_dir = PathManager.get_local_path(self._metadata.gt_dir) + groundTruthImgList = glob.glob(os.path.join(gt_dir, "*", "*_gtFine_instanceIds.png")) + assert len( + groundTruthImgList + ), "Cannot find any ground truth images to use for evaluation. Searched for: {}".format( + cityscapes_eval.args.groundTruthSearch + ) + predictionImgList = [] + for gt in groundTruthImgList: + predictionImgList.append(cityscapes_eval.getPrediction(gt, cityscapes_eval.args)) + results = cityscapes_eval.evaluateImgLists( + predictionImgList, groundTruthImgList, cityscapes_eval.args + )["averages"] + + ret = OrderedDict() + ret["segm"] = {"AP": results["allAp"] * 100, "AP50": results["allAp50%"] * 100} + self._working_dir.cleanup() + return ret + + +class CityscapesSemSegEvaluator(CityscapesEvaluator): + """ + Evaluate semantic segmentation results on cityscapes dataset using cityscapes API. + + Note: + * It does not work in multi-machine distributed training. + * It contains a synchronization, therefore has to be used on all ranks. + * Only the main process runs evaluation. + """ + + def process(self, inputs, outputs): + from cityscapesscripts.helpers.labels import trainId2label + + for input, output in zip(inputs, outputs): + file_name = input["file_name"] + basename = os.path.splitext(os.path.basename(file_name))[0] + pred_filename = os.path.join(self._temp_dir, basename + "_pred.png") + + output = output["sem_seg"].argmax(dim=0).to(self._cpu_device).numpy() + pred = 255 * np.ones(output.shape, dtype=np.uint8) + for train_id, label in trainId2label.items(): + if label.ignoreInEval: + continue + pred[output == train_id] = label.id + Image.fromarray(pred).save(pred_filename) + + def evaluate(self): + comm.synchronize() + if comm.get_rank() > 0: + return + # Load the Cityscapes eval script *after* setting the required env var, + # since the script reads CITYSCAPES_DATASET into global variables at load time. + import cityscapesscripts.evaluation.evalPixelLevelSemanticLabeling as cityscapes_eval + + self._logger.info("Evaluating results under {} ...".format(self._temp_dir)) + + # set some global states in cityscapes evaluation API, before evaluating + cityscapes_eval.args.predictionPath = os.path.abspath(self._temp_dir) + cityscapes_eval.args.predictionWalk = None + cityscapes_eval.args.JSONOutput = False + cityscapes_eval.args.colorized = False + + # These lines are adopted from + # https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/evaluation/evalPixelLevelSemanticLabeling.py # noqa + gt_dir = PathManager.get_local_path(self._metadata.gt_dir) + groundTruthImgList = glob.glob(os.path.join(gt_dir, "*", "*_gtFine_labelIds.png")) + assert len( + groundTruthImgList + ), "Cannot find any ground truth images to use for evaluation. Searched for: {}".format( + cityscapes_eval.args.groundTruthSearch + ) + predictionImgList = [] + for gt in groundTruthImgList: + predictionImgList.append(cityscapes_eval.getPrediction(cityscapes_eval.args, gt)) + results = cityscapes_eval.evaluateImgLists( + predictionImgList, groundTruthImgList, cityscapes_eval.args + ) + ret = OrderedDict() + ret["sem_seg"] = { + "IoU": 100.0 * results["averageScoreClasses"], + "iIoU": 100.0 * results["averageScoreInstClasses"], + "IoU_sup": 100.0 * results["averageScoreCategories"], + "iIoU_sup": 100.0 * results["averageScoreInstCategories"], + } + self._working_dir.cleanup() + return ret diff --git a/detectron2/evaluation/coco_evaluation.py b/detectron2/evaluation/coco_evaluation.py new file mode 100644 index 0000000000000000000000000000000000000000..9ed0a434c2fb4d351aedb9c84e76fc5dc3cc4e49 --- /dev/null +++ b/detectron2/evaluation/coco_evaluation.py @@ -0,0 +1,610 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import contextlib +import copy +import io +import itertools +import json +import logging +import numpy as np +import os +import pickle +from collections import OrderedDict +import pycocotools.mask as mask_util +import torch +from pycocotools.coco import COCO +from pycocotools.cocoeval import COCOeval +from tabulate import tabulate + +import detectron2.utils.comm as comm +from detectron2.config import CfgNode +from detectron2.data import MetadataCatalog +from detectron2.data.datasets.coco import convert_to_coco_json +from detectron2.data.datasets.coco_zeroshot_categories import COCO_UNSEEN_CLS, COCO_SEEN_CLS, COCO_OVD_ALL_CLS +from detectron2.evaluation.fast_eval_api import COCOeval_opt +from detectron2.structures import Boxes, BoxMode, pairwise_iou +from detectron2.utils.file_io import PathManager +from detectron2.utils.logger import create_small_table + +from .evaluator import DatasetEvaluator + + +class COCOEvaluator(DatasetEvaluator): + """ + Evaluate AR for object proposals, AP for instance detection/segmentation, AP + for keypoint detection outputs using COCO's metrics. + See http://cocodataset.org/#detection-eval and + http://cocodataset.org/#keypoints-eval to understand its metrics. + The metrics range from 0 to 100 (instead of 0 to 1), where a -1 or NaN means + the metric cannot be computed (e.g. due to no predictions made). + + In addition to COCO, this evaluator is able to support any bounding box detection, + instance segmentation, or keypoint detection dataset. + """ + + def __init__( + self, + dataset_name, + tasks=None, + distributed=True, + output_dir=None, + *, + use_fast_impl=True, + kpt_oks_sigmas=(), + ): + """ + Args: + dataset_name (str): name of the dataset to be evaluated. + It must have either the following corresponding metadata: + + "json_file": the path to the COCO format annotation + + Or it must be in detectron2's standard dataset format + so it can be converted to COCO format automatically. + tasks (tuple[str]): tasks that can be evaluated under the given + configuration. A task is one of "bbox", "segm", "keypoints". + By default, will infer this automatically from predictions. + distributed (True): if True, will collect results from all ranks and run evaluation + in the main process. + Otherwise, will only evaluate the results in the current process. + output_dir (str): optional, an output directory to dump all + results predicted on the dataset. The dump contains two files: + + 1. "instances_predictions.pth" a file that can be loaded with `torch.load` and + contains all the results in the format they are produced by the model. + 2. "coco_instances_results.json" a json file in COCO's result format. + use_fast_impl (bool): use a fast but **unofficial** implementation to compute AP. + Although the results should be very close to the official implementation in COCO + API, it is still recommended to compute results with the official API for use in + papers. The faster implementation also uses more RAM. + kpt_oks_sigmas (list[float]): The sigmas used to calculate keypoint OKS. + See http://cocodataset.org/#keypoints-eval + When empty, it will use the defaults in COCO. + Otherwise it should be the same length as ROI_KEYPOINT_HEAD.NUM_KEYPOINTS. + """ + self._logger = logging.getLogger(__name__) + self._distributed = distributed + self._output_dir = output_dir + self._use_fast_impl = use_fast_impl + + if tasks is not None and isinstance(tasks, CfgNode): + kpt_oks_sigmas = ( + tasks.TEST.KEYPOINT_OKS_SIGMAS if not kpt_oks_sigmas else kpt_oks_sigmas + ) + self._logger.warn( + "COCO Evaluator instantiated using config, this is deprecated behavior." + " Please pass in explicit arguments instead." + ) + self._tasks = None # Infering it from predictions should be better + else: + self._tasks = tasks + + self._cpu_device = torch.device("cpu") + + self._metadata = MetadataCatalog.get(dataset_name) + if not hasattr(self._metadata, "json_file"): + self._logger.info( + f"'{dataset_name}' is not registered by `register_coco_instances`." + " Therefore trying to convert it to COCO format ..." + ) + + cache_path = os.path.join(output_dir, f"{dataset_name}_coco_format.json") + self._metadata.json_file = cache_path + convert_to_coco_json(dataset_name, cache_path) + + json_file = PathManager.get_local_path(self._metadata.json_file) + with contextlib.redirect_stdout(io.StringIO()): + self._coco_api = COCO(json_file) + + # Test set json files do not contain annotations (evaluation must be + # performed using the COCO evaluation server). + self._do_evaluation = "annotations" in self._coco_api.dataset + if self._do_evaluation: + self._kpt_oks_sigmas = kpt_oks_sigmas + + def reset(self): + self._predictions = [] + + def process(self, inputs, outputs): + """ + Args: + inputs: the inputs to a COCO model (e.g., GeneralizedRCNN). + It is a list of dict. Each dict corresponds to an image and + contains keys like "height", "width", "file_name", "image_id". + outputs: the outputs of a COCO model. It is a list of dicts with key + "instances" that contains :class:`Instances`. + """ + for input, output in zip(inputs, outputs): + prediction = {"image_id": input["image_id"]} + + if "instances" in output: + instances = output["instances"].to(self._cpu_device) + prediction["instances"] = instances_to_coco_json(instances, input["image_id"]) + if "proposals" in output: + prediction["proposals"] = output["proposals"].to(self._cpu_device) + if len(prediction) > 1: + self._predictions.append(prediction) + + def evaluate(self, img_ids=None): + """ + Args: + img_ids: a list of image IDs to evaluate on. Default to None for the whole dataset + """ + if self._distributed: + comm.synchronize() + predictions = comm.gather(self._predictions, dst=0) + predictions = list(itertools.chain(*predictions)) + + if not comm.is_main_process(): + return {} + else: + predictions = self._predictions + + if len(predictions) == 0: + self._logger.warning("[COCOEvaluator] Did not receive valid predictions.") + return {} + + if self._output_dir: + PathManager.mkdirs(self._output_dir) + file_path = os.path.join(self._output_dir, "instances_predictions.pth") + with PathManager.open(file_path, "wb") as f: + torch.save(predictions, f) + + self._results = OrderedDict() + if "proposals" in predictions[0]: + self._eval_box_proposals(predictions) + if "instances" in predictions[0]: + self._eval_predictions(predictions, img_ids=img_ids) + # Copy so the caller can do whatever with results + return copy.deepcopy(self._results) + + def _tasks_from_predictions(self, predictions): + """ + Get COCO API "tasks" (i.e. iou_type) from COCO-format predictions. + """ + tasks = {"bbox"} + for pred in predictions: + if "segmentation" in pred: + tasks.add("segm") + if "keypoints" in pred: + tasks.add("keypoints") + return sorted(tasks) + + def _eval_predictions(self, predictions, img_ids=None): + """ + Evaluate predictions. Fill self._results with the metrics of the tasks. + """ + self._logger.info("Preparing results for COCO format ...") + coco_results = list(itertools.chain(*[x["instances"] for x in predictions])) + tasks = self._tasks or self._tasks_from_predictions(coco_results) + + # unmap the category ids for COCO + if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"): + dataset_id_to_contiguous_id = self._metadata.thing_dataset_id_to_contiguous_id + all_contiguous_ids = list(dataset_id_to_contiguous_id.values()) + num_classes = len(all_contiguous_ids) + assert min(all_contiguous_ids) == 0 and max(all_contiguous_ids) == num_classes - 1 + + reverse_id_mapping = {v: k for k, v in dataset_id_to_contiguous_id.items()} + for result in coco_results: + category_id = result["category_id"] + assert category_id < num_classes, ( + f"A prediction has class={category_id}, " + f"but the dataset only has {num_classes} classes and " + f"predicted class id should be in [0, {num_classes - 1}]." + ) + result["category_id"] = reverse_id_mapping[category_id] + + if self._output_dir: + file_path = os.path.join(self._output_dir, "coco_instances_results.json") + self._logger.info("Saving results to {}".format(file_path)) + with PathManager.open(file_path, "w") as f: + f.write(json.dumps(coco_results)) + f.flush() + + if not self._do_evaluation: + self._logger.info("Annotations are not available for evaluation.") + return + + self._logger.info( + "Evaluating predictions with {} COCO API...".format( + "unofficial" if self._use_fast_impl else "official" + ) + ) + for task in sorted(tasks): + assert task in {"bbox", "segm", "keypoints"}, f"Got unknown task: {task}!" + coco_eval = ( + _evaluate_predictions_on_coco( + self._coco_api, + coco_results, + task, + kpt_oks_sigmas=self._kpt_oks_sigmas, + use_fast_impl=self._use_fast_impl, + img_ids=img_ids, + ) + if len(coco_results) > 0 + else None # cocoapi does not handle empty results very well + ) + + res = self._derive_coco_results( + coco_eval, task, class_names=self._metadata.get("thing_classes") + ) + self._results[task] = res + + def _eval_box_proposals(self, predictions): + """ + Evaluate the box proposals in predictions. + Fill self._results with the metrics for "box_proposals" task. + """ + if self._output_dir: + # Saving generated box proposals to file. + # Predicted box_proposals are in XYXY_ABS mode. + bbox_mode = BoxMode.XYXY_ABS.value + ids, boxes, objectness_logits = [], [], [] + for prediction in predictions: + ids.append(prediction["image_id"]) + boxes.append(prediction["proposals"].proposal_boxes.tensor.numpy()) + objectness_logits.append(prediction["proposals"].objectness_logits.numpy()) + + proposal_data = { + "boxes": boxes, + "objectness_logits": objectness_logits, + "ids": ids, + "bbox_mode": bbox_mode, + } + with PathManager.open(os.path.join(self._output_dir, "box_proposals.pkl"), "wb") as f: + pickle.dump(proposal_data, f) + + if not self._do_evaluation: + self._logger.info("Annotations are not available for evaluation.") + return + + self._logger.info("Evaluating bbox proposals ...") + res = {} + areas = {"all": "", "small": "s", "medium": "m", "large": "l"} + for limit in [100, 1000]: + for area, suffix in areas.items(): + stats = _evaluate_box_proposals(predictions, self._coco_api, area=area, limit=limit) + key = "AR{}@{:d}".format(suffix, limit) + res[key] = float(stats["ar"].item() * 100) + self._logger.info("Proposal metrics: \n" + create_small_table(res)) + self._results["box_proposals"] = res + + def _derive_coco_results(self, coco_eval, iou_type, class_names=None): + """ + Derive the desired score numbers from summarized COCOeval. + + Args: + coco_eval (None or COCOEval): None represents no predictions from model. + iou_type (str): + class_names (None or list[str]): if provided, will use it to predict + per-category AP. + + Returns: + a dict of {metric name: score} + """ + + metrics = { + "bbox": ["AP", "AP50", "AP75", "APs", "APm", "APl"], + "segm": ["AP", "AP50", "AP75", "APs", "APm", "APl"], + "keypoints": ["AP", "AP50", "AP75", "APm", "APl"], + }[iou_type] + + if coco_eval is None: + self._logger.warn("No predictions from the model!") + return {metric: float("nan") for metric in metrics} + + # the standard metrics + results = { + metric: float(coco_eval.stats[idx] * 100 if coco_eval.stats[idx] >= 0 else "nan") + for idx, metric in enumerate(metrics) + } + self._logger.info( + "Evaluation results for {}: \n".format(iou_type) + create_small_table(results) + ) + if not np.isfinite(sum(results.values())): + self._logger.info("Some metrics cannot be computed and is shown as NaN.") + + if class_names is None or len(class_names) <= 1: + return results + # Compute per-category AP + # from https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L222-L252 # noqa + precisions = coco_eval.eval["precision"] + # precision has dims (iou, recall, cls, area range, max dets) + assert len(class_names) == precisions.shape[2] + + results_per_category = [] + for idx, name in enumerate(class_names): + # area range index 0: all area ranges + # max dets index -1: typically 100 per image + precision = precisions[:, :, idx, 0, -1] + precision = precision[precision > -1] + ap = np.mean(precision) if precision.size else float("nan") + results_per_category.append(("{}".format(name), float(ap * 100))) + + # Computing AP50 for (seen/unseen) split in generalized zeroshot setting (eg. all 65 categories) + # from https://github.com/alirezazareian/ovr-cnn/blob/master/maskrcnn_benchmark/data/datasets/evaluation/coco/coco_eval.py + if len(class_names) == 65: + p = coco_eval.params + maxDets = p.maxDets[2] + areaRng = 'all' + iouThr = 0.5 + aind = [i for i, aRng in enumerate(p.areaRngLbl) if aRng == areaRng] + mind = [i for i, mDet in enumerate(p.maxDets) if mDet == maxDets] + t = np.where(iouThr == p.iouThrs)[0] + s = coco_eval.eval['precision'] + s = s[t,:,:,aind,mind] + + unseen_cids = [p.catIds[i] for i, c in enumerate(class_names) if c in COCO_UNSEEN_CLS] + seen_cids = [p.catIds[i] for i, c in enumerate(class_names) if c in COCO_SEEN_CLS] + all_cids = [p.catIds[i] for i, c in enumerate(class_names) if c in COCO_OVD_ALL_CLS] + res = {} + for split, cid_list in [('target',unseen_cids), ('base',seen_cids), ('all',all_cids)]: + cinds = [] + for cid in cid_list: + cinds.extend([i for i, c in enumerate(p.catIds) if c == cid]) + s_split = s[:, :, cinds] + if len(s_split[s_split>-1])==0: + mean_s = -1 + else: + mean_s = np.mean(s_split[s_split>-1]) + res[f'AP50_split_{split}'] = mean_s + for res_item in res: + self._logger.info("{} AP: {}\n".format(res_item, res[res_item])) + + # tabulate it + N_COLS = min(6, len(results_per_category) * 2) + results_flatten = list(itertools.chain(*results_per_category)) + results_2d = itertools.zip_longest(*[results_flatten[i::N_COLS] for i in range(N_COLS)]) + table = tabulate( + results_2d, + tablefmt="pipe", + floatfmt=".3f", + headers=["category", "AP"] * (N_COLS // 2), + numalign="left", + ) + self._logger.info("Per-category {} AP: \n".format(iou_type) + table) + + results.update({"AP-" + name: ap for name, ap in results_per_category}) + return results + + +def instances_to_coco_json(instances, img_id): + """ + Dump an "Instances" object to a COCO-format json that's used for evaluation. + + Args: + instances (Instances): + img_id (int): the image id + + Returns: + list[dict]: list of json annotations in COCO format. + """ + num_instance = len(instances) + if num_instance == 0: + return [] + + boxes = instances.pred_boxes.tensor.numpy() + boxes = BoxMode.convert(boxes, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS) + boxes = boxes.tolist() + scores = instances.scores.tolist() + classes = instances.pred_classes.tolist() + + has_mask = instances.has("pred_masks") + if has_mask: + # use RLE to encode the masks, because they are too large and takes memory + # since this evaluator stores outputs of the entire dataset + rles = [ + mask_util.encode(np.array(mask[:, :, None], order="F", dtype="uint8"))[0] + for mask in instances.pred_masks + ] + for rle in rles: + # "counts" is an array encoded by mask_util as a byte-stream. Python3's + # json writer which always produces strings cannot serialize a bytestream + # unless you decode it. Thankfully, utf-8 works out (which is also what + # the pycocotools/_mask.pyx does). + rle["counts"] = rle["counts"].decode("utf-8") + + has_keypoints = instances.has("pred_keypoints") + if has_keypoints: + keypoints = instances.pred_keypoints + + results = [] + for k in range(num_instance): + result = { + "image_id": img_id, + "category_id": classes[k], + "bbox": boxes[k], + "score": scores[k], + } + if has_mask: + result["segmentation"] = rles[k] + if has_keypoints: + # In COCO annotations, + # keypoints coordinates are pixel indices. + # However our predictions are floating point coordinates. + # Therefore we subtract 0.5 to be consistent with the annotation format. + # This is the inverse of data loading logic in `datasets/coco.py`. + keypoints[k][:, :2] -= 0.5 + result["keypoints"] = keypoints[k].flatten().tolist() + results.append(result) + return results + + +# inspired from Detectron: +# https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L255 # noqa +def _evaluate_box_proposals(dataset_predictions, coco_api, thresholds=None, area="all", limit=None): + """ + Evaluate detection proposal recall metrics. This function is a much + faster alternative to the official COCO API recall evaluation code. However, + it produces slightly different results. + """ + # Record max overlap value for each gt box + # Return vector of overlap values + areas = { + "all": 0, + "small": 1, + "medium": 2, + "large": 3, + "96-128": 4, + "128-256": 5, + "256-512": 6, + "512-inf": 7, + } + area_ranges = [ + [0 ** 2, 1e5 ** 2], # all + [0 ** 2, 32 ** 2], # small + [32 ** 2, 96 ** 2], # medium + [96 ** 2, 1e5 ** 2], # large + [96 ** 2, 128 ** 2], # 96-128 + [128 ** 2, 256 ** 2], # 128-256 + [256 ** 2, 512 ** 2], # 256-512 + [512 ** 2, 1e5 ** 2], + ] # 512-inf + assert area in areas, "Unknown area range: {}".format(area) + area_range = area_ranges[areas[area]] + gt_overlaps = [] + num_pos = 0 + + for prediction_dict in dataset_predictions: + predictions = prediction_dict["proposals"] + + # sort predictions in descending order + # TODO maybe remove this and make it explicit in the documentation + inds = predictions.objectness_logits.sort(descending=True)[1] + predictions = predictions[inds] + + ann_ids = coco_api.getAnnIds(imgIds=prediction_dict["image_id"]) + anno = coco_api.loadAnns(ann_ids) + gt_boxes = [ + BoxMode.convert(obj["bbox"], BoxMode.XYWH_ABS, BoxMode.XYXY_ABS) + for obj in anno + if obj["iscrowd"] == 0 + ] + gt_boxes = torch.as_tensor(gt_boxes).reshape(-1, 4) # guard against no boxes + gt_boxes = Boxes(gt_boxes) + gt_areas = torch.as_tensor([obj["area"] for obj in anno if obj["iscrowd"] == 0]) + + if len(gt_boxes) == 0 or len(predictions) == 0: + continue + + valid_gt_inds = (gt_areas >= area_range[0]) & (gt_areas <= area_range[1]) + gt_boxes = gt_boxes[valid_gt_inds] + + num_pos += len(gt_boxes) + + if len(gt_boxes) == 0: + continue + + if limit is not None and len(predictions) > limit: + predictions = predictions[:limit] + + overlaps = pairwise_iou(predictions.proposal_boxes, gt_boxes) + + _gt_overlaps = torch.zeros(len(gt_boxes)) + for j in range(min(len(predictions), len(gt_boxes))): + # find which proposal box maximally covers each gt box + # and get the iou amount of coverage for each gt box + max_overlaps, argmax_overlaps = overlaps.max(dim=0) + + # find which gt box is 'best' covered (i.e. 'best' = most iou) + gt_ovr, gt_ind = max_overlaps.max(dim=0) + assert gt_ovr >= 0 + # find the proposal box that covers the best covered gt box + box_ind = argmax_overlaps[gt_ind] + # record the iou coverage of this gt box + _gt_overlaps[j] = overlaps[box_ind, gt_ind] + assert _gt_overlaps[j] == gt_ovr + # mark the proposal box and the gt box as used + overlaps[box_ind, :] = -1 + overlaps[:, gt_ind] = -1 + + # append recorded iou coverage level + gt_overlaps.append(_gt_overlaps) + gt_overlaps = ( + torch.cat(gt_overlaps, dim=0) if len(gt_overlaps) else torch.zeros(0, dtype=torch.float32) + ) + gt_overlaps, _ = torch.sort(gt_overlaps) + + if thresholds is None: + step = 0.05 + thresholds = torch.arange(0.5, 0.95 + 1e-5, step, dtype=torch.float32) + recalls = torch.zeros_like(thresholds) + # compute recall for each iou threshold + for i, t in enumerate(thresholds): + recalls[i] = (gt_overlaps >= t).float().sum() / float(num_pos) + # ar = 2 * np.trapz(recalls, thresholds) + ar = recalls.mean() + return { + "ar": ar, + "recalls": recalls, + "thresholds": thresholds, + "gt_overlaps": gt_overlaps, + "num_pos": num_pos, + } + + +def _evaluate_predictions_on_coco( + coco_gt, coco_results, iou_type, kpt_oks_sigmas=None, use_fast_impl=True, img_ids=None +): + """ + Evaluate the coco results using COCOEval API. + """ + assert len(coco_results) > 0 + + if iou_type == "segm": + coco_results = copy.deepcopy(coco_results) + # When evaluating mask AP, if the results contain bbox, cocoapi will + # use the box area as the area of the instance, instead of the mask area. + # This leads to a different definition of small/medium/large. + # We remove the bbox field to let mask AP use mask area. + for c in coco_results: + c.pop("bbox", None) + + coco_dt = coco_gt.loadRes(coco_results) + coco_eval = (COCOeval_opt if use_fast_impl else COCOeval)(coco_gt, coco_dt, iou_type) + if img_ids is not None: + coco_eval.params.imgIds = img_ids + + if iou_type == "keypoints": + # Use the COCO default keypoint OKS sigmas unless overrides are specified + if kpt_oks_sigmas: + assert hasattr(coco_eval.params, "kpt_oks_sigmas"), "pycocotools is too old!" + coco_eval.params.kpt_oks_sigmas = np.array(kpt_oks_sigmas) + # COCOAPI requires every detection and every gt to have keypoints, so + # we just take the first entry from both + num_keypoints_dt = len(coco_results[0]["keypoints"]) // 3 + num_keypoints_gt = len(next(iter(coco_gt.anns.values()))["keypoints"]) // 3 + num_keypoints_oks = len(coco_eval.params.kpt_oks_sigmas) + assert num_keypoints_oks == num_keypoints_dt == num_keypoints_gt, ( + f"[COCOEvaluator] Prediction contain {num_keypoints_dt} keypoints. " + f"Ground truth contains {num_keypoints_gt} keypoints. " + f"The length of cfg.TEST.KEYPOINT_OKS_SIGMAS is {num_keypoints_oks}. " + "They have to agree with each other. For meaning of OKS, please refer to " + "http://cocodataset.org/#keypoints-eval." + ) + + coco_eval.evaluate() + coco_eval.accumulate() + coco_eval.summarize() + + return coco_eval diff --git a/detectron2/evaluation/evaluator.py b/detectron2/evaluation/evaluator.py new file mode 100644 index 0000000000000000000000000000000000000000..0f27f01d19dd1ca7434a6dae1bb7f78614d23895 --- /dev/null +++ b/detectron2/evaluation/evaluator.py @@ -0,0 +1,226 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import datetime +import logging +import time +from collections import OrderedDict, abc +from contextlib import ExitStack, contextmanager +from typing import List, Union +import torch +from torch import nn + +from detectron2.utils.comm import get_world_size, is_main_process +from detectron2.utils.logger import log_every_n_seconds + + +class DatasetEvaluator: + """ + Base class for a dataset evaluator. + + The function :func:`inference_on_dataset` runs the model over + all samples in the dataset, and have a DatasetEvaluator to process the inputs/outputs. + + This class will accumulate information of the inputs/outputs (by :meth:`process`), + and produce evaluation results in the end (by :meth:`evaluate`). + """ + + def reset(self): + """ + Preparation for a new round of evaluation. + Should be called before starting a round of evaluation. + """ + pass + + def process(self, inputs, outputs): + """ + Process the pair of inputs and outputs. + If they contain batches, the pairs can be consumed one-by-one using `zip`: + + .. code-block:: python + + for input_, output in zip(inputs, outputs): + # do evaluation on single input/output pair + ... + + Args: + inputs (list): the inputs that's used to call the model. + outputs (list): the return value of `model(inputs)` + """ + pass + + def evaluate(self): + """ + Evaluate/summarize the performance, after processing all input/output pairs. + + Returns: + dict: + A new evaluator class can return a dict of arbitrary format + as long as the user can process the results. + In our train_net.py, we expect the following format: + + * key: the name of the task (e.g., bbox) + * value: a dict of {metric name: score}, e.g.: {"AP50": 80} + """ + pass + + +class DatasetEvaluators(DatasetEvaluator): + """ + Wrapper class to combine multiple :class:`DatasetEvaluator` instances. + + This class dispatches every evaluation call to + all of its :class:`DatasetEvaluator`. + """ + + def __init__(self, evaluators): + """ + Args: + evaluators (list): the evaluators to combine. + """ + super().__init__() + self._evaluators = evaluators + + def reset(self): + for evaluator in self._evaluators: + evaluator.reset() + + def process(self, inputs, outputs): + for evaluator in self._evaluators: + evaluator.process(inputs, outputs) + + def evaluate(self): + results = OrderedDict() + for evaluator in self._evaluators: + result = evaluator.evaluate() + if is_main_process() and result is not None: + for k, v in result.items(): + assert ( + k not in results + ), "Different evaluators produce results with the same key {}".format(k) + results[k] = v + return results + + +def inference_on_dataset( + model, data_loader, queries, evaluator: Union[DatasetEvaluator, List[DatasetEvaluator], None] +): + """ + Run model on the data_loader and evaluate the metrics with evaluator. + Also benchmark the inference speed of `model.__call__` accurately. + The model will be used in eval mode. + + Args: + model (callable): a callable which takes an object from + `data_loader` and returns some outputs. + + If it's an nn.Module, it will be temporarily set to `eval` mode. + If you wish to evaluate a model in `training` mode instead, you can + wrap the given model and override its behavior of `.eval()` and `.train()`. + data_loader: an iterable object with a length. + The elements it generates will be the inputs to the model. + evaluator: the evaluator(s) to run. Use `None` if you only want to benchmark, + but don't want to do any evaluation. + + Returns: + The return value of `evaluator.evaluate()` + """ + num_devices = get_world_size() + logger = logging.getLogger(__name__) + logger.info("Start inference on {} batches".format(len(data_loader))) + + total = len(data_loader) # inference data loader must have a fixed length + if evaluator is None: + # create a no-op evaluator + evaluator = DatasetEvaluators([]) + if isinstance(evaluator, abc.MutableSequence): + evaluator = DatasetEvaluators(evaluator) + evaluator.reset() + + num_warmup = min(5, total - 1) + start_time = time.perf_counter() + total_data_time = 0 + total_compute_time = 0 + total_eval_time = 0 + with ExitStack() as stack: + if isinstance(model, nn.Module): + stack.enter_context(inference_context(model)) + stack.enter_context(torch.no_grad()) + + start_data_time = time.perf_counter() + for idx, inputs in enumerate(data_loader): + total_data_time += time.perf_counter() - start_data_time + if idx == num_warmup: + start_time = time.perf_counter() + total_data_time = 0 + total_compute_time = 0 + total_eval_time = 0 + + start_compute_time = time.perf_counter() + + outputs = model(queries, inputs) + + if torch.cuda.is_available(): + torch.cuda.synchronize() + total_compute_time += time.perf_counter() - start_compute_time + + start_eval_time = time.perf_counter() + evaluator.process(inputs, outputs) + total_eval_time += time.perf_counter() - start_eval_time + + iters_after_start = idx + 1 - num_warmup * int(idx >= num_warmup) + data_seconds_per_iter = total_data_time / iters_after_start + compute_seconds_per_iter = total_compute_time / iters_after_start + eval_seconds_per_iter = total_eval_time / iters_after_start + total_seconds_per_iter = (time.perf_counter() - start_time) / iters_after_start + if idx >= num_warmup * 2 or compute_seconds_per_iter > 5: + eta = datetime.timedelta(seconds=int(total_seconds_per_iter * (total - idx - 1))) + log_every_n_seconds( + logging.INFO, + ( + f"Inference done {idx + 1}/{total}. " + f"Dataloading: {data_seconds_per_iter:.4f} s / iter. " + f"Inference: {compute_seconds_per_iter:.4f} s / iter. " + f"Eval: {eval_seconds_per_iter:.4f} s / iter. " + f"Total: {total_seconds_per_iter:.4f} s / iter. " + f"ETA={eta}" + ), + n=5, + ) + start_data_time = time.perf_counter() + + # Measure the time only for this worker (before the synchronization barrier) + total_time = time.perf_counter() - start_time + total_time_str = str(datetime.timedelta(seconds=total_time)) + # NOTE this format is parsed by grep + logger.info( + "Total inference time: {} ({:.6f} s / iter per device, on {} devices)".format( + total_time_str, total_time / (total - num_warmup), num_devices + ) + ) + total_compute_time_str = str(datetime.timedelta(seconds=int(total_compute_time))) + logger.info( + "Total inference pure compute time: {} ({:.6f} s / iter per device, on {} devices)".format( + total_compute_time_str, total_compute_time / (total - num_warmup), num_devices + ) + ) + + results = evaluator.evaluate() + # An evaluator may return None when not in main process. + # Replace it by an empty dict instead to make it easier for downstream code to handle + if results is None: + results = {} + return results + + +@contextmanager +def inference_context(model): + """ + A context where the model is temporarily changed to eval mode, + and restored to previous mode afterwards. + + Args: + model: a torch Module + """ + training_mode = model.training + model.eval() + yield + model.train(training_mode) diff --git a/detectron2/evaluation/fast_eval_api.py b/detectron2/evaluation/fast_eval_api.py new file mode 100644 index 0000000000000000000000000000000000000000..2eb202bd5efa3ec3d366027b1debffc269ae8b17 --- /dev/null +++ b/detectron2/evaluation/fast_eval_api.py @@ -0,0 +1,121 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import logging +import numpy as np +import time +from pycocotools.cocoeval import COCOeval + +from detectron2 import _C + +logger = logging.getLogger(__name__) + + +class COCOeval_opt(COCOeval): + """ + This is a slightly modified version of the original COCO API, where the functions evaluateImg() + and accumulate() are implemented in C++ to speedup evaluation + """ + + def evaluate(self): + """ + Run per image evaluation on given images and store results in self.evalImgs_cpp, a + datastructure that isn't readable from Python but is used by a c++ implementation of + accumulate(). Unlike the original COCO PythonAPI, we don't populate the datastructure + self.evalImgs because this datastructure is a computational bottleneck. + :return: None + """ + tic = time.time() + + p = self.params + # add backward compatibility if useSegm is specified in params + if p.useSegm is not None: + p.iouType = "segm" if p.useSegm == 1 else "bbox" + logger.info("Evaluate annotation type *{}*".format(p.iouType)) + p.imgIds = list(np.unique(p.imgIds)) + if p.useCats: + p.catIds = list(np.unique(p.catIds)) + p.maxDets = sorted(p.maxDets) + self.params = p + + self._prepare() # bottleneck + + # loop through images, area range, max detection number + catIds = p.catIds if p.useCats else [-1] + + if p.iouType == "segm" or p.iouType == "bbox": + computeIoU = self.computeIoU + elif p.iouType == "keypoints": + computeIoU = self.computeOks + self.ious = { + (imgId, catId): computeIoU(imgId, catId) for imgId in p.imgIds for catId in catIds + } # bottleneck + + maxDet = p.maxDets[-1] + + # <<<< Beginning of code differences with original COCO API + def convert_instances_to_cpp(instances, is_det=False): + # Convert annotations for a list of instances in an image to a format that's fast + # to access in C++ + instances_cpp = [] + for instance in instances: + instance_cpp = _C.InstanceAnnotation( + int(instance["id"]), + instance["score"] if is_det else instance.get("score", 0.0), + instance["area"], + bool(instance.get("iscrowd", 0)), + bool(instance.get("ignore", 0)), + ) + instances_cpp.append(instance_cpp) + return instances_cpp + + # Convert GT annotations, detections, and IOUs to a format that's fast to access in C++ + ground_truth_instances = [ + [convert_instances_to_cpp(self._gts[imgId, catId]) for catId in p.catIds] + for imgId in p.imgIds + ] + detected_instances = [ + [convert_instances_to_cpp(self._dts[imgId, catId], is_det=True) for catId in p.catIds] + for imgId in p.imgIds + ] + ious = [[self.ious[imgId, catId] for catId in catIds] for imgId in p.imgIds] + + if not p.useCats: + # For each image, flatten per-category lists into a single list + ground_truth_instances = [[[o for c in i for o in c]] for i in ground_truth_instances] + detected_instances = [[[o for c in i for o in c]] for i in detected_instances] + + # Call C++ implementation of self.evaluateImgs() + self._evalImgs_cpp = _C.COCOevalEvaluateImages( + p.areaRng, maxDet, p.iouThrs, ious, ground_truth_instances, detected_instances + ) + self._evalImgs = None + + self._paramsEval = copy.deepcopy(self.params) + toc = time.time() + logger.info("COCOeval_opt.evaluate() finished in {:0.2f} seconds.".format(toc - tic)) + # >>>> End of code differences with original COCO API + + def accumulate(self): + """ + Accumulate per image evaluation results and store the result in self.eval. Does not + support changing parameter settings from those used by self.evaluate() + """ + logger.info("Accumulating evaluation results...") + tic = time.time() + assert hasattr( + self, "_evalImgs_cpp" + ), "evaluate() must be called before accmulate() is called." + + self.eval = _C.COCOevalAccumulate(self._paramsEval, self._evalImgs_cpp) + + # recall is num_iou_thresholds X num_categories X num_area_ranges X num_max_detections + self.eval["recall"] = np.array(self.eval["recall"]).reshape( + self.eval["counts"][:1] + self.eval["counts"][2:] + ) + + # precision and scores are num_iou_thresholds X num_recall_thresholds X num_categories X + # num_area_ranges X num_max_detections + self.eval["precision"] = np.array(self.eval["precision"]).reshape(self.eval["counts"]) + self.eval["scores"] = np.array(self.eval["scores"]).reshape(self.eval["counts"]) + toc = time.time() + logger.info("COCOeval_opt.accumulate() finished in {:0.2f} seconds.".format(toc - tic)) diff --git a/detectron2/evaluation/flickr30k_evaluation.py b/detectron2/evaluation/flickr30k_evaluation.py new file mode 100644 index 0000000000000000000000000000000000000000..244f813773267c4a1a9257af86feab4212c0a85f --- /dev/null +++ b/detectron2/evaluation/flickr30k_evaluation.py @@ -0,0 +1,299 @@ +import logging +import numpy as np +import os +from collections import OrderedDict +from detectron2.config import global_cfg as cfg +import torch +from fvcore.common.file_io import PathManager +from detectron2.structures.boxes import pairwise_iou + +from detectron2.utils.comm import all_gather, is_main_process, synchronize +import pickle +from .evaluator import DatasetEvaluator +import json +from detectron2.structures import Boxes +import html +import ftfy +import regex as re + +PATTN = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE) + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + +def whitespace_clean(text): + text = re.sub(r'\s+', ' ', text) + text = text.strip() + return text + + +class FLICKR30KEvaluator(DatasetEvaluator): + + """ + Evaluate semantic segmentation + """ + + def __init__(self, dataset_name, distributed=True, output_dir=None): + """ + Args: + dataset_name (str): name of the dataset to be evaluated. + distributed (True): if True, will collect results from all ranks for evaluation. + Otherwise, will evaluate the results in the current process. + num_classes (int): number of classes + ignore_label (int): value in semantic segmentation ground truth. Predictions for the + corresponding pixels should be ignored. + output_dir (str): an output directory to dump results. + """ + self._dataset_name = dataset_name + self._distributed = distributed + self._output_dir = output_dir + + self._cpu_device = torch.device("cpu") + self._logger = logging.getLogger(__name__) + self.gt_boxes = json.load(open("/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/flickr30k_processed/bounding_boxes_test.json")) + self.gt_sents = json.load(open("/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/flickr30k_processed/sentences_test.json")) + + def reset(self): + self._predictions = {} + + def process(self, inputs, outputs): + """ + Args: + inputs: the inputs to a model. + It is a list of dicts. Each dict corresponds to an image and + contains keys like "height", "width", "file_name", "image_id". + outputs: the outputs of a model. It is either list of semantic segmentation predictions + (Tensor [H, W]) or list of dicts with key "sem_seg" that contains semantic + segmentation prediction in the same format. + """ + assert len(inputs) == 1 # batch = 1 during inference + dataset_name, img_id, (img_height, img_width), all_str2id_links = inputs[0][-1] + img_id = img_id.split('/')[-1] + match_scores, processed_results = outputs + match_scores = match_scores.to(self._cpu_device) + pred_boxes = processed_results[0]['instances'].proposal_boxes.to(self._cpu_device) + + self._predictions.update({img_id: [img_height, img_width, all_str2id_links, match_scores, pred_boxes]}) + + def merge_gt_boxes(self, box_anno): + gt_boxes = [] + phrase_ids = [] + scene_box_ids = box_anno['scene'] + for k, v in box_anno['boxes'].items(): + if k in scene_box_ids: # important: remove scene boxes, otherwise the number of each phrase type cannot match paper + continue + phrase_ids.append(k) + if len(v) == 1: + gt_boxes.append(v[0]) + else: + # when a phrase respond to multiple regions, we take the union of them as paper given + v = np.array(v) + box = [v[:, 0].min(), v[:, 1].min(), v[:, 2].max(), v[:, 3].max()] + gt_boxes.append(box) + gt_boxes = np.array(gt_boxes) + return phrase_ids, gt_boxes + + def find_ground_box(self, match_scores, all_str2id_links, sentences, gt_phrase_ids): + """ Given matching matrix between region feats and token feats, find the box that grounds a phrase + """ + num_box = match_scores.size(0) + num_cap = int(match_scores.size(1) / 77) + all_phrase_score = [] + all_phrase_ids = [] + for i in range(num_cap): # per sentence + this_score = match_scores[:, i*77:(i+1)*77] # [#boxes, 77] + input_ids = [iitem for item in all_str2id_links[i] for iitem in item[1]] + input_tokens = [item[0] for item in all_str2id_links[i]] + phrases = sentences[i]['phrases'] + for j, phrase in enumerate(phrases): # per phrase + if phrase['phrase_id'] not in gt_phrase_ids: # no gt box for this phrase, skip + continue + # locate the word + words = whitespace_clean(basic_clean(phrase['phrase'])).lower() # phrase['phrase'].lower().replace("-"," ").split() + words = re.findall(PATTN, words) + first_word_index = None # phrase['first_word_index'] + for idx in range(len(input_tokens) - len(words) + 1): # search start word of this phrase + if input_tokens[idx : idx + len(words)] == words: # NOTE: key step for alignment btw model prediction and annotation + first_word_index = idx + break + if first_word_index is None: + print("Fail to find phrase [{}] in input tokens [{}]".format(words, input_tokens)) + start_wd_ind = first_word_index + end_wd_ind = first_word_index + len(words) + if len(words) != len(phrase['phrase'].split()): + pass # print('tokens: {} <--> phrase: {}'.format(words, phrase['phrase'])) + # locate the token + start_tk_ind = 0 + for k_i, k in enumerate(range(0, start_wd_ind)): + start_tk_ind += len(all_str2id_links[i][k][1]) + token_cnt = 0 + for k_i, k in enumerate(range(start_wd_ind, end_wd_ind)): + if all_str2id_links[i][k][0] != words[k_i]: + print("Word not matched: {} in model output but {} in annotation".format(all_str2id_links[i][k][0], words[k_i])) + else: + token_cnt += len(all_str2id_links[i][k][1]) # ith sentence, kth word, and its tokens + end_tk_ind = start_tk_ind + token_cnt + # sanity check + phrase_ids1 = [iitem for item in all_str2id_links[i][start_wd_ind:end_wd_ind] for iitem in item[1]] # way 1: use word index to accumulate token ids in a phrase + phrase_ids2 = input_ids[start_tk_ind:end_tk_ind] # way 2: use token index to directly index token ids in a phrase + if phrase_ids1 != phrase_ids2: + print("Santity check: {} from word {} in token".format(phrase_ids1, phrase_ids2)) + # index similarity score + phrase_score = this_score[:, start_tk_ind:end_tk_ind] + phrase_score = phrase_score.mean(dim=1) # phrase_score.max(dim=1)[0] # + all_phrase_score.append(phrase_score) + all_phrase_ids.append(phrase['phrase_id']) + phrase_score_tensor = torch.cat(all_phrase_score) + phrase_score_tensor = phrase_score_tensor.view(len(all_phrase_ids), num_box) # NOTE: this should be [#phrases, #object proposals] + + return phrase_score_tensor, all_phrase_ids + + def evaluate(self): + """ + Evaluates Referring Segmentation IoU: + """ + + if self._distributed: + synchronize() + + self._predictions = all_gather(self._predictions) + + if not is_main_process(): + return + + all_prediction = {} + for p in self._predictions: + all_prediction.update(p) + else: + all_prediction = self._predictions + + if len(all_prediction) < 30: # resume inference results + save_path = "/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/flickr30k_processed/grounding_results/grounding_{}_imgs.npy".format(1000) + all_prediction = np.load(save_path, allow_pickle=True).tolist() + self._logger.info('Resume from {}'.format(save_path)) + else: # new run + save_path = "/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/flickr30k_processed/grounding_results/grounding_{}_imgs.npy".format(len(all_prediction)) + np.save(save_path, all_prediction) + self._logger.info('Save results to {}'.format(save_path)) + self._logger.info('Got {} images!'.format(len(all_prediction))) + + image_unique_ids = list(all_prediction.keys()) + image_evaled = [] + + total_num = 0 + recall_num = 0 + num_type = {} + recall_type = {} + acc_type = {} + recall_topk_num = {5:0, 10:0} + point_recall_num = 0 + EVAL_THRESH = 0.5 + type_cnts = {} + + for img_sent_id in image_unique_ids: + if img_sent_id not in self.gt_boxes: + continue + else: + image_evaled.append(img_sent_id) + # results from model + result = all_prediction[img_sent_id] + phrase_ids = None + phrase_types = [] # phrase type: each phrase belongs to a coarse object concept + pred_boxes = None # an object proposal selected by model for each phrase + img_height, img_width, all_str2id_links = result[0], result[1], result[2] # all_str2id_links: each word and its tokenized ids + match_scores = result[3] # matching score [#object proposals, #tokens] + precomp_boxes = result[4] # object proposals from offline module + # annotation from dataset + sentences = self.gt_sents[img_sent_id] + box_anno = self.gt_boxes[img_sent_id] + # sanity check and box merging + assert box_anno['height'] == img_height, box_anno['width'] == img_width + gt_phrase_ids, gt_boxes = self.merge_gt_boxes(box_anno) # merged if multiple boxes for the same phrase + if len(gt_phrase_ids) == 0: # no gt box for this image + continue + for sent_item in sentences: + for phrase_item in sent_item['phrases']: + if phrase_item['phrase_id'] in gt_phrase_ids: + phrase_types.append(phrase_item['phrase_type']) + + # merge similarity scores from token level to phrase level, and find the box that grounds the phrase + phrase_score_tensor, all_phrase_ids = self.find_ground_box(match_scores, all_str2id_links, sentences, gt_phrase_ids) + pred_boxes_ind = torch.argmax(phrase_score_tensor, dim=1) + pred_boxes = precomp_boxes[pred_boxes_ind] + pred_similarity = phrase_score_tensor # .t() # pred_similarity: matching score [#phrases, #object proposals] + + # get single target/gt box for each phrase + # 1. any gt box that can be matched as target + # refer to (https://github.com/BigRedT/info-ground/blob/22ae6d6ec8b38df473e73034fc895ebf97d39897/exp/ground/eval_flickr_phrase_loc.py#L90) + phrase_boxes = [box_anno['boxes'][p_id] for p_id in all_phrase_ids] + targets = [] + for pr_b, pd_b in zip(phrase_boxes, pred_boxes): + matched = False + for single_b in pr_b: + this_iou = pairwise_iou(Boxes(torch.from_numpy(np.array([single_b])).float()), Boxes(pd_b.view(1,-1))) + if (this_iou >= EVAL_THRESH).sum() > 0: + targets.append(single_b) + matched = True + break + if not matched: + targets.append(single_b) + targets = Boxes(torch.from_numpy(np.array(targets)).float()) + # 2. union box as target + # target_ind = np.array([gt_phrase_ids.index(p_id) for p_id in all_phrase_ids]) + # targets = gt_boxes[target_ind] # ground-truth boxes for each phrase in each sentence + # targets = Boxes(torch.from_numpy(targets).float()) + assert len(phrase_types) == len(targets) + + # single predicted box for each phrase + ious = pairwise_iou(targets, pred_boxes) # this function will change the target_boxes into cuda mode + iou = ious.numpy().diagonal() + total_num += iou.shape[0] + recall_num += int((iou >= EVAL_THRESH).sum()) # 0.5 + + # metric of point (can be ignored) + pred_boxes_tensor = pred_boxes.tensor + pred_center = (pred_boxes_tensor[:, :2] + pred_boxes_tensor[:, 2:]) / 2.0 + pred_center = pred_center.repeat(1, 2) ## x_c, y_c, x_c, y_c + targets_tensor = targets.tensor + fall_tensor = targets_tensor - pred_center + fall_tensor = (fall_tensor[:, :2] <= 0).float().sum(1) + (fall_tensor[:, 2:] >= 0).float().sum(1) + point_recall_num += (fall_tensor == 4).float().numpy().sum() + + # detailed accuracy across different phrase types + for pid, p_type in enumerate(phrase_types): + p_type = p_type[0] + num_type[p_type] = num_type.setdefault(p_type, 0) + 1 + recall_type[p_type] = recall_type.setdefault(p_type, 0) + (iou[pid] >= EVAL_THRESH) + + # metric of recall when multiple predicted boxes for each phrase + ious_top = pairwise_iou(targets, precomp_boxes).cpu() + for k in [5, 10]: + top_k = torch.topk(pred_similarity, k=k, dim=1)[0][:, [-1]] + pred_similarity_topk = (pred_similarity >= top_k).float() + ious_top_k = (ious_top * pred_similarity_topk).numpy() + recall_topk_num[k] += int(((ious_top_k >= EVAL_THRESH).sum(1) > 0).sum()) + + acc = recall_num / total_num + acc_top5 = recall_topk_num[5] / total_num + acc_top10 = recall_topk_num[10] / total_num + point_acc = point_recall_num / total_num + + # details about each coarse type of phrase + for type, type_num in num_type.items(): + acc_type[type] = recall_type[type] / type_num + + # if self._output_dir: + # PathManager.mkdirs(self._output_dir) + # file_path = os.path.join(self._output_dir, "prediction_{}.pkl".format(str(acc).replace('.', '_')[:6])) + # with PathManager.open(file_path, "wb") as f: + # pickle.dump(all_prediction, f) + + del all_prediction + self._logger.info('evaluation on {} expression instances, detailed_iou: {}'.format(len(image_evaled), acc_type)) + self._logger.info('Evaluate Pointing Accuracy: PointAcc:{}'.format(point_acc)) + results = OrderedDict({"acc": acc, "acc_top5": acc_top5, "acc_top10": acc_top10}) + self._logger.info(results) + self._logger.info(num_type) + return results \ No newline at end of file diff --git a/detectron2/evaluation/lvis_evaluation.py b/detectron2/evaluation/lvis_evaluation.py new file mode 100644 index 0000000000000000000000000000000000000000..81f08e40cf61f0c451e63565debac7f6877b99d9 --- /dev/null +++ b/detectron2/evaluation/lvis_evaluation.py @@ -0,0 +1,358 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import itertools +import json +import logging +import os +import pickle +from collections import OrderedDict +import torch + +import detectron2.utils.comm as comm +from detectron2.config import CfgNode +from detectron2.data import MetadataCatalog +from detectron2.structures import Boxes, BoxMode, pairwise_iou +from detectron2.utils.file_io import PathManager +from detectron2.utils.logger import create_small_table + +from .coco_evaluation import instances_to_coco_json +from .evaluator import DatasetEvaluator + + +class LVISEvaluator(DatasetEvaluator): + """ + Evaluate object proposal and instance detection/segmentation outputs using + LVIS's metrics and evaluation API. + """ + + def __init__(self, dataset_name, tasks=None, distributed=True, output_dir=None): + """ + Args: + dataset_name (str): name of the dataset to be evaluated. + It must have the following corresponding metadata: + "json_file": the path to the LVIS format annotation + tasks (tuple[str]): tasks that can be evaluated under the given + configuration. A task is one of "bbox", "segm". + By default, will infer this automatically from predictions. + distributed (True): if True, will collect results from all ranks for evaluation. + Otherwise, will evaluate the results in the current process. + output_dir (str): optional, an output directory to dump results. + """ + from lvis import LVIS + + self._logger = logging.getLogger(__name__) + + if tasks is not None and isinstance(tasks, CfgNode): + self._logger.warn( + "COCO Evaluator instantiated using config, this is deprecated behavior." + " Please pass in explicit arguments instead." + ) + self._tasks = None # Infering it from predictions should be better + else: + self._tasks = tasks + + self._distributed = distributed + self._output_dir = output_dir + + self._cpu_device = torch.device("cpu") + + self._metadata = MetadataCatalog.get(dataset_name) + json_file = PathManager.get_local_path(self._metadata.json_file) + self._lvis_api = LVIS(json_file) + # Test set json files do not contain annotations (evaluation must be + # performed using the LVIS evaluation server). + self._do_evaluation = len(self._lvis_api.get_ann_ids()) > 0 + + def reset(self): + self._predictions = [] + + def process(self, inputs, outputs): + """ + Args: + inputs: the inputs to a LVIS model (e.g., GeneralizedRCNN). + It is a list of dict. Each dict corresponds to an image and + contains keys like "height", "width", "file_name", "image_id". + outputs: the outputs of a LVIS model. It is a list of dicts with key + "instances" that contains :class:`Instances`. + """ + for input, output in zip(inputs, outputs): + prediction = {"image_id": input["image_id"]} + + if "instances" in output: + instances = output["instances"].to(self._cpu_device) + prediction["instances"] = instances_to_coco_json(instances, input["image_id"]) + if "proposals" in output: + prediction["proposals"] = output["proposals"].to(self._cpu_device) + self._predictions.append(prediction) + + def evaluate(self): + if self._distributed: + comm.synchronize() + predictions = comm.gather(self._predictions, dst=0) + predictions = list(itertools.chain(*predictions)) + + if not comm.is_main_process(): + return + else: + predictions = self._predictions + + if len(predictions) == 0: + self._logger.warning("[LVISEvaluator] Did not receive valid predictions.") + return {} + + if self._output_dir: + PathManager.mkdirs(self._output_dir) + file_path = os.path.join(self._output_dir, "instances_predictions.pth") + with PathManager.open(file_path, "wb") as f: + torch.save(predictions, f) + + self._results = OrderedDict() + if "proposals" in predictions[0]: + self._eval_box_proposals(predictions) + if "instances" in predictions[0]: + self._eval_predictions(predictions) + # Copy so the caller can do whatever with results + return copy.deepcopy(self._results) + + def _tasks_from_predictions(self, predictions): + for pred in predictions: + if "segmentation" in pred: + return ("bbox", "segm") + return ("bbox",) + + def _eval_predictions(self, predictions): + """ + Evaluate predictions. Fill self._results with the metrics of the tasks. + + Args: + predictions (list[dict]): list of outputs from the model + """ + self._logger.info("Preparing results in the LVIS format ...") + lvis_results = list(itertools.chain(*[x["instances"] for x in predictions])) + tasks = self._tasks or self._tasks_from_predictions(lvis_results) + + # LVIS evaluator can be used to evaluate results for COCO dataset categories. + # In this case `_metadata` variable will have a field with COCO-specific category mapping. + if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"): + reverse_id_mapping = { + v: k for k, v in self._metadata.thing_dataset_id_to_contiguous_id.items() + } + for result in lvis_results: + result["category_id"] = reverse_id_mapping[result["category_id"]] + else: + # unmap the category ids for LVIS (from 0-indexed to 1-indexed) + for result in lvis_results: + result["category_id"] += 1 + + if self._output_dir: + file_path = os.path.join(self._output_dir, "lvis_instances_results.json") + self._logger.info("Saving results to {}".format(file_path)) + with PathManager.open(file_path, "w") as f: + f.write(json.dumps(lvis_results)) + f.flush() + + if not self._do_evaluation: + self._logger.info("Annotations are not available for evaluation.") + return + + self._logger.info("Evaluating predictions ...") + for task in sorted(tasks): + res = _evaluate_predictions_on_lvis( + self._lvis_api, lvis_results, task, class_names=self._metadata.get("thing_classes") + ) + self._results[task] = res + + def _eval_box_proposals(self, predictions): + """ + Evaluate the box proposals in predictions. + Fill self._results with the metrics for "box_proposals" task. + """ + if self._output_dir: + # Saving generated box proposals to file. + # Predicted box_proposals are in XYXY_ABS mode. + bbox_mode = BoxMode.XYXY_ABS.value + ids, boxes, objectness_logits = [], [], [] + for prediction in predictions: + ids.append(prediction["image_id"]) + boxes.append(prediction["proposals"].proposal_boxes.tensor.numpy()) + objectness_logits.append(prediction["proposals"].objectness_logits.numpy()) + + proposal_data = { + "boxes": boxes, + "objectness_logits": objectness_logits, + "ids": ids, + "bbox_mode": bbox_mode, + } + with PathManager.open(os.path.join(self._output_dir, "box_proposals.pkl"), "wb") as f: + pickle.dump(proposal_data, f) + + if not self._do_evaluation: + self._logger.info("Annotations are not available for evaluation.") + return + + self._logger.info("Evaluating bbox proposals ...") + res = {} + areas = {"all": "", "small": "s", "medium": "m", "large": "l"} + for limit in [100, 1000]: + for area, suffix in areas.items(): + stats = _evaluate_box_proposals(predictions, self._lvis_api, area=area, limit=limit) + key = "AR{}@{:d}".format(suffix, limit) + res[key] = float(stats["ar"].item() * 100) + self._logger.info("Proposal metrics: \n" + create_small_table(res)) + self._results["box_proposals"] = res + + +# inspired from Detectron: +# https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L255 # noqa +def _evaluate_box_proposals(dataset_predictions, lvis_api, thresholds=None, area="all", limit=None): + """ + Evaluate detection proposal recall metrics. This function is a much + faster alternative to the official LVIS API recall evaluation code. However, + it produces slightly different results. + """ + # Record max overlap value for each gt box + # Return vector of overlap values + areas = { + "all": 0, + "small": 1, + "medium": 2, + "large": 3, + "96-128": 4, + "128-256": 5, + "256-512": 6, + "512-inf": 7, + } + area_ranges = [ + [0 ** 2, 1e5 ** 2], # all + [0 ** 2, 32 ** 2], # small + [32 ** 2, 96 ** 2], # medium + [96 ** 2, 1e5 ** 2], # large + [96 ** 2, 128 ** 2], # 96-128 + [128 ** 2, 256 ** 2], # 128-256 + [256 ** 2, 512 ** 2], # 256-512 + [512 ** 2, 1e5 ** 2], + ] # 512-inf + assert area in areas, "Unknown area range: {}".format(area) + area_range = area_ranges[areas[area]] + gt_overlaps = [] + num_pos = 0 + + for prediction_dict in dataset_predictions: + predictions = prediction_dict["proposals"] + + # sort predictions in descending order + # TODO maybe remove this and make it explicit in the documentation + inds = predictions.objectness_logits.sort(descending=True)[1] + predictions = predictions[inds] + + ann_ids = lvis_api.get_ann_ids(img_ids=[prediction_dict["image_id"]]) + anno = lvis_api.load_anns(ann_ids) + gt_boxes = [ + BoxMode.convert(obj["bbox"], BoxMode.XYWH_ABS, BoxMode.XYXY_ABS) for obj in anno + ] + gt_boxes = torch.as_tensor(gt_boxes).reshape(-1, 4) # guard against no boxes + gt_boxes = Boxes(gt_boxes) + gt_areas = torch.as_tensor([obj["area"] for obj in anno]) + + if len(gt_boxes) == 0 or len(predictions) == 0: + continue + + valid_gt_inds = (gt_areas >= area_range[0]) & (gt_areas <= area_range[1]) + gt_boxes = gt_boxes[valid_gt_inds] + + num_pos += len(gt_boxes) + + if len(gt_boxes) == 0: + continue + + if limit is not None and len(predictions) > limit: + predictions = predictions[:limit] + + overlaps = pairwise_iou(predictions.proposal_boxes, gt_boxes) + + _gt_overlaps = torch.zeros(len(gt_boxes)) + for j in range(min(len(predictions), len(gt_boxes))): + # find which proposal box maximally covers each gt box + # and get the iou amount of coverage for each gt box + max_overlaps, argmax_overlaps = overlaps.max(dim=0) + + # find which gt box is 'best' covered (i.e. 'best' = most iou) + gt_ovr, gt_ind = max_overlaps.max(dim=0) + assert gt_ovr >= 0 + # find the proposal box that covers the best covered gt box + box_ind = argmax_overlaps[gt_ind] + # record the iou coverage of this gt box + _gt_overlaps[j] = overlaps[box_ind, gt_ind] + assert _gt_overlaps[j] == gt_ovr + # mark the proposal box and the gt box as used + overlaps[box_ind, :] = -1 + overlaps[:, gt_ind] = -1 + + # append recorded iou coverage level + gt_overlaps.append(_gt_overlaps) + gt_overlaps = ( + torch.cat(gt_overlaps, dim=0) if len(gt_overlaps) else torch.zeros(0, dtype=torch.float32) + ) + gt_overlaps, _ = torch.sort(gt_overlaps) + + if thresholds is None: + step = 0.05 + thresholds = torch.arange(0.5, 0.95 + 1e-5, step, dtype=torch.float32) + recalls = torch.zeros_like(thresholds) + # compute recall for each iou threshold + for i, t in enumerate(thresholds): + recalls[i] = (gt_overlaps >= t).float().sum() / float(num_pos) + # ar = 2 * np.trapz(recalls, thresholds) + ar = recalls.mean() + return { + "ar": ar, + "recalls": recalls, + "thresholds": thresholds, + "gt_overlaps": gt_overlaps, + "num_pos": num_pos, + } + + +def _evaluate_predictions_on_lvis(lvis_gt, lvis_results, iou_type, class_names=None): + """ + Args: + iou_type (str): + kpt_oks_sigmas (list[float]): + class_names (None or list[str]): if provided, will use it to predict + per-category AP. + + Returns: + a dict of {metric name: score} + """ + metrics = { + "bbox": ["AP", "AP50", "AP75", "APs", "APm", "APl", "APr", "APc", "APf"], + "segm": ["AP", "AP50", "AP75", "APs", "APm", "APl", "APr", "APc", "APf"], + }[iou_type] + + logger = logging.getLogger(__name__) + + if len(lvis_results) == 0: # TODO: check if needed + logger.warn("No predictions from the model!") + return {metric: float("nan") for metric in metrics} + + if iou_type == "segm": + lvis_results = copy.deepcopy(lvis_results) + # When evaluating mask AP, if the results contain bbox, LVIS API will + # use the box area as the area of the instance, instead of the mask area. + # This leads to a different definition of small/medium/large. + # We remove the bbox field to let mask AP use mask area. + for c in lvis_results: + c.pop("bbox", None) + + from lvis import LVISEval, LVISResults + + lvis_results = LVISResults(lvis_gt, lvis_results) + lvis_eval = LVISEval(lvis_gt, lvis_results, iou_type) + lvis_eval.run() + lvis_eval.print_results() + + # Pull the standard metrics from the LVIS results + results = lvis_eval.get_results() + results = {metric: float(results[metric] * 100) for metric in metrics} + logger.info("Evaluation results for {}: \n".format(iou_type) + create_small_table(results)) + return results diff --git a/detectron2/evaluation/panoptic_evaluation.py b/detectron2/evaluation/panoptic_evaluation.py new file mode 100644 index 0000000000000000000000000000000000000000..9fb3462b7f9abf6feaa499976bfed526ebd17e31 --- /dev/null +++ b/detectron2/evaluation/panoptic_evaluation.py @@ -0,0 +1,199 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import contextlib +import io +import itertools +import json +import logging +import numpy as np +import os +import tempfile +from collections import OrderedDict +from typing import Optional +from PIL import Image +from tabulate import tabulate + +from detectron2.data import MetadataCatalog +from detectron2.utils import comm +from detectron2.utils.file_io import PathManager + +from .evaluator import DatasetEvaluator + +logger = logging.getLogger(__name__) + + +class COCOPanopticEvaluator(DatasetEvaluator): + """ + Evaluate Panoptic Quality metrics on COCO using PanopticAPI. + It saves panoptic segmentation prediction in `output_dir` + + It contains a synchronize call and has to be called from all workers. + """ + + def __init__(self, dataset_name: str, output_dir: Optional[str] = None): + """ + Args: + dataset_name: name of the dataset + output_dir: output directory to save results for evaluation. + """ + self._metadata = MetadataCatalog.get(dataset_name) + self._thing_contiguous_id_to_dataset_id = { + v: k for k, v in self._metadata.thing_dataset_id_to_contiguous_id.items() + } + self._stuff_contiguous_id_to_dataset_id = { + v: k for k, v in self._metadata.stuff_dataset_id_to_contiguous_id.items() + } + + self._output_dir = output_dir + if self._output_dir is not None: + PathManager.mkdirs(self._output_dir) + + def reset(self): + self._predictions = [] + + def _convert_category_id(self, segment_info): + isthing = segment_info.pop("isthing", None) + if isthing is None: + # the model produces panoptic category id directly. No more conversion needed + return segment_info + if isthing is True: + segment_info["category_id"] = self._thing_contiguous_id_to_dataset_id[ + segment_info["category_id"] + ] + else: + segment_info["category_id"] = self._stuff_contiguous_id_to_dataset_id[ + segment_info["category_id"] + ] + return segment_info + + def process(self, inputs, outputs): + from panopticapi.utils import id2rgb + + for input, output in zip(inputs, outputs): + panoptic_img, segments_info = output["panoptic_seg"] + panoptic_img = panoptic_img.cpu().numpy() + if segments_info is None: + # If "segments_info" is None, we assume "panoptic_img" is a + # H*W int32 image storing the panoptic_id in the format of + # category_id * label_divisor + instance_id. We reserve -1 for + # VOID label, and add 1 to panoptic_img since the official + # evaluation script uses 0 for VOID label. + label_divisor = self._metadata.label_divisor + segments_info = [] + for panoptic_label in np.unique(panoptic_img): + if panoptic_label == -1: + # VOID region. + continue + pred_class = panoptic_label // label_divisor + isthing = ( + pred_class in self._metadata.thing_dataset_id_to_contiguous_id.values() + ) + segments_info.append( + { + "id": int(panoptic_label) + 1, + "category_id": int(pred_class), + "isthing": bool(isthing), + } + ) + # Official evaluation script uses 0 for VOID label. + panoptic_img += 1 + + file_name = os.path.basename(input["file_name"]) + file_name_png = os.path.splitext(file_name)[0] + ".png" + with io.BytesIO() as out: + Image.fromarray(id2rgb(panoptic_img)).save(out, format="PNG") + segments_info = [self._convert_category_id(x) for x in segments_info] + self._predictions.append( + { + "image_id": input["image_id"], + "file_name": file_name_png, + "png_string": out.getvalue(), + "segments_info": segments_info, + } + ) + + def evaluate(self): + comm.synchronize() + + self._predictions = comm.gather(self._predictions) + self._predictions = list(itertools.chain(*self._predictions)) + if not comm.is_main_process(): + return + + # PanopticApi requires local files + gt_json = PathManager.get_local_path(self._metadata.panoptic_json) + gt_folder = PathManager.get_local_path(self._metadata.panoptic_root) + + with tempfile.TemporaryDirectory(prefix="panoptic_eval") as pred_dir: + logger.info("Writing all panoptic predictions to {} ...".format(pred_dir)) + for p in self._predictions: + with open(os.path.join(pred_dir, p["file_name"]), "wb") as f: + f.write(p.pop("png_string")) + + with open(gt_json, "r") as f: + json_data = json.load(f) + json_data["annotations"] = self._predictions + + output_dir = self._output_dir or pred_dir + predictions_json = os.path.join(output_dir, "predictions.json") + with PathManager.open(predictions_json, "w") as f: + f.write(json.dumps(json_data)) + + from panopticapi.evaluation import pq_compute + + with contextlib.redirect_stdout(io.StringIO()): + pq_res = pq_compute( + gt_json, + PathManager.get_local_path(predictions_json), + gt_folder=gt_folder, + pred_folder=pred_dir, + ) + + res = {} + res["PQ"] = 100 * pq_res["All"]["pq"] + res["SQ"] = 100 * pq_res["All"]["sq"] + res["RQ"] = 100 * pq_res["All"]["rq"] + res["PQ_th"] = 100 * pq_res["Things"]["pq"] + res["SQ_th"] = 100 * pq_res["Things"]["sq"] + res["RQ_th"] = 100 * pq_res["Things"]["rq"] + res["PQ_st"] = 100 * pq_res["Stuff"]["pq"] + res["SQ_st"] = 100 * pq_res["Stuff"]["sq"] + res["RQ_st"] = 100 * pq_res["Stuff"]["rq"] + + results = OrderedDict({"panoptic_seg": res}) + _print_panoptic_results(pq_res) + + return results + + +def _print_panoptic_results(pq_res): + headers = ["", "PQ", "SQ", "RQ", "#categories"] + data = [] + for name in ["All", "Things", "Stuff"]: + row = [name] + [pq_res[name][k] * 100 for k in ["pq", "sq", "rq"]] + [pq_res[name]["n"]] + data.append(row) + table = tabulate( + data, headers=headers, tablefmt="pipe", floatfmt=".3f", stralign="center", numalign="center" + ) + logger.info("Panoptic Evaluation Results:\n" + table) + + +if __name__ == "__main__": + from detectron2.utils.logger import setup_logger + + logger = setup_logger() + import argparse + + parser = argparse.ArgumentParser() + parser.add_argument("--gt-json") + parser.add_argument("--gt-dir") + parser.add_argument("--pred-json") + parser.add_argument("--pred-dir") + args = parser.parse_args() + + from panopticapi.evaluation import pq_compute + + with contextlib.redirect_stdout(io.StringIO()): + pq_res = pq_compute( + args.gt_json, args.pred_json, gt_folder=args.gt_dir, pred_folder=args.pred_dir + ) + _print_panoptic_results(pq_res) diff --git a/detectron2/evaluation/pascal_voc_evaluation.py b/detectron2/evaluation/pascal_voc_evaluation.py new file mode 100644 index 0000000000000000000000000000000000000000..1d1abcde2f87bb5f103e73cb364aaabbecb6e619 --- /dev/null +++ b/detectron2/evaluation/pascal_voc_evaluation.py @@ -0,0 +1,300 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import logging +import numpy as np +import os +import tempfile +import xml.etree.ElementTree as ET +from collections import OrderedDict, defaultdict +from functools import lru_cache +import torch + +from detectron2.data import MetadataCatalog +from detectron2.utils import comm +from detectron2.utils.file_io import PathManager + +from .evaluator import DatasetEvaluator + + +class PascalVOCDetectionEvaluator(DatasetEvaluator): + """ + Evaluate Pascal VOC style AP for Pascal VOC dataset. + It contains a synchronization, therefore has to be called from all ranks. + + Note that the concept of AP can be implemented in different ways and may not + produce identical results. This class mimics the implementation of the official + Pascal VOC Matlab API, and should produce similar but not identical results to the + official API. + """ + + def __init__(self, dataset_name): + """ + Args: + dataset_name (str): name of the dataset, e.g., "voc_2007_test" + """ + self._dataset_name = dataset_name + meta = MetadataCatalog.get(dataset_name) + + # Too many tiny files, download all to local for speed. + annotation_dir_local = PathManager.get_local_path( + os.path.join(meta.dirname, "Annotations/") + ) + self._anno_file_template = os.path.join(annotation_dir_local, "{}.xml") + self._image_set_path = os.path.join(meta.dirname, "ImageSets", "Main", meta.split + ".txt") + self._class_names = meta.thing_classes + assert meta.year in [2007, 2012], meta.year + self._is_2007 = meta.year == 2007 + self._cpu_device = torch.device("cpu") + self._logger = logging.getLogger(__name__) + + def reset(self): + self._predictions = defaultdict(list) # class name -> list of prediction strings + + def process(self, inputs, outputs): + for input, output in zip(inputs, outputs): + image_id = input["image_id"] + instances = output["instances"].to(self._cpu_device) + boxes = instances.pred_boxes.tensor.numpy() + scores = instances.scores.tolist() + classes = instances.pred_classes.tolist() + for box, score, cls in zip(boxes, scores, classes): + xmin, ymin, xmax, ymax = box + # The inverse of data loading logic in `datasets/pascal_voc.py` + xmin += 1 + ymin += 1 + self._predictions[cls].append( + f"{image_id} {score:.3f} {xmin:.1f} {ymin:.1f} {xmax:.1f} {ymax:.1f}" + ) + + def evaluate(self): + """ + Returns: + dict: has a key "segm", whose value is a dict of "AP", "AP50", and "AP75". + """ + all_predictions = comm.gather(self._predictions, dst=0) + if not comm.is_main_process(): + return + predictions = defaultdict(list) + for predictions_per_rank in all_predictions: + for clsid, lines in predictions_per_rank.items(): + predictions[clsid].extend(lines) + del all_predictions + + self._logger.info( + "Evaluating {} using {} metric. " + "Note that results do not use the official Matlab API.".format( + self._dataset_name, 2007 if self._is_2007 else 2012 + ) + ) + + with tempfile.TemporaryDirectory(prefix="pascal_voc_eval_") as dirname: + res_file_template = os.path.join(dirname, "{}.txt") + + aps = defaultdict(list) # iou -> ap per class + for cls_id, cls_name in enumerate(self._class_names): + lines = predictions.get(cls_id, [""]) + + with open(res_file_template.format(cls_name), "w") as f: + f.write("\n".join(lines)) + + for thresh in range(50, 100, 5): + rec, prec, ap = voc_eval( + res_file_template, + self._anno_file_template, + self._image_set_path, + cls_name, + ovthresh=thresh / 100.0, + use_07_metric=self._is_2007, + ) + aps[thresh].append(ap * 100) + + ret = OrderedDict() + mAP = {iou: np.mean(x) for iou, x in aps.items()} + ret["bbox"] = {"AP": np.mean(list(mAP.values())), "AP50": mAP[50], "AP75": mAP[75]} + return ret + + +############################################################################## +# +# Below code is modified from +# https://github.com/rbgirshick/py-faster-rcnn/blob/master/lib/datasets/voc_eval.py +# -------------------------------------------------------- +# Fast/er R-CNN +# Licensed under The MIT License [see LICENSE for details] +# Written by Bharath Hariharan +# -------------------------------------------------------- + +"""Python implementation of the PASCAL VOC devkit's AP evaluation code.""" + + +@lru_cache(maxsize=None) +def parse_rec(filename): + """Parse a PASCAL VOC xml file.""" + with PathManager.open(filename) as f: + tree = ET.parse(f) + objects = [] + for obj in tree.findall("object"): + obj_struct = {} + obj_struct["name"] = obj.find("name").text + obj_struct["pose"] = obj.find("pose").text + obj_struct["truncated"] = int(obj.find("truncated").text) + obj_struct["difficult"] = int(obj.find("difficult").text) + bbox = obj.find("bndbox") + obj_struct["bbox"] = [ + int(bbox.find("xmin").text), + int(bbox.find("ymin").text), + int(bbox.find("xmax").text), + int(bbox.find("ymax").text), + ] + objects.append(obj_struct) + + return objects + + +def voc_ap(rec, prec, use_07_metric=False): + """Compute VOC AP given precision and recall. If use_07_metric is true, uses + the VOC 07 11-point method (default:False). + """ + if use_07_metric: + # 11 point metric + ap = 0.0 + for t in np.arange(0.0, 1.1, 0.1): + if np.sum(rec >= t) == 0: + p = 0 + else: + p = np.max(prec[rec >= t]) + ap = ap + p / 11.0 + else: + # correct AP calculation + # first append sentinel values at the end + mrec = np.concatenate(([0.0], rec, [1.0])) + mpre = np.concatenate(([0.0], prec, [0.0])) + + # compute the precision envelope + for i in range(mpre.size - 1, 0, -1): + mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i]) + + # to calculate area under PR curve, look for points + # where X axis (recall) changes value + i = np.where(mrec[1:] != mrec[:-1])[0] + + # and sum (\Delta recall) * prec + ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) + return ap + + +def voc_eval(detpath, annopath, imagesetfile, classname, ovthresh=0.5, use_07_metric=False): + """rec, prec, ap = voc_eval(detpath, + annopath, + imagesetfile, + classname, + [ovthresh], + [use_07_metric]) + + Top level function that does the PASCAL VOC evaluation. + + detpath: Path to detections + detpath.format(classname) should produce the detection results file. + annopath: Path to annotations + annopath.format(imagename) should be the xml annotations file. + imagesetfile: Text file containing the list of images, one image per line. + classname: Category name (duh) + [ovthresh]: Overlap threshold (default = 0.5) + [use_07_metric]: Whether to use VOC07's 11 point AP computation + (default False) + """ + # assumes detections are in detpath.format(classname) + # assumes annotations are in annopath.format(imagename) + # assumes imagesetfile is a text file with each line an image name + + # first load gt + # read list of images + with PathManager.open(imagesetfile, "r") as f: + lines = f.readlines() + imagenames = [x.strip() for x in lines] + + # load annots + recs = {} + for imagename in imagenames: + recs[imagename] = parse_rec(annopath.format(imagename)) + + # extract gt objects for this class + class_recs = {} + npos = 0 + for imagename in imagenames: + R = [obj for obj in recs[imagename] if obj["name"] == classname] + bbox = np.array([x["bbox"] for x in R]) + difficult = np.array([x["difficult"] for x in R]).astype(np.bool) + # difficult = np.array([False for x in R]).astype(np.bool) # treat all "difficult" as GT + det = [False] * len(R) + npos = npos + sum(~difficult) + class_recs[imagename] = {"bbox": bbox, "difficult": difficult, "det": det} + + # read dets + detfile = detpath.format(classname) + with open(detfile, "r") as f: + lines = f.readlines() + + splitlines = [x.strip().split(" ") for x in lines] + image_ids = [x[0] for x in splitlines] + confidence = np.array([float(x[1]) for x in splitlines]) + BB = np.array([[float(z) for z in x[2:]] for x in splitlines]).reshape(-1, 4) + + # sort by confidence + sorted_ind = np.argsort(-confidence) + BB = BB[sorted_ind, :] + image_ids = [image_ids[x] for x in sorted_ind] + + # go down dets and mark TPs and FPs + nd = len(image_ids) + tp = np.zeros(nd) + fp = np.zeros(nd) + for d in range(nd): + R = class_recs[image_ids[d]] + bb = BB[d, :].astype(float) + ovmax = -np.inf + BBGT = R["bbox"].astype(float) + + if BBGT.size > 0: + # compute overlaps + # intersection + ixmin = np.maximum(BBGT[:, 0], bb[0]) + iymin = np.maximum(BBGT[:, 1], bb[1]) + ixmax = np.minimum(BBGT[:, 2], bb[2]) + iymax = np.minimum(BBGT[:, 3], bb[3]) + iw = np.maximum(ixmax - ixmin + 1.0, 0.0) + ih = np.maximum(iymax - iymin + 1.0, 0.0) + inters = iw * ih + + # union + uni = ( + (bb[2] - bb[0] + 1.0) * (bb[3] - bb[1] + 1.0) + + (BBGT[:, 2] - BBGT[:, 0] + 1.0) * (BBGT[:, 3] - BBGT[:, 1] + 1.0) + - inters + ) + + overlaps = inters / uni + ovmax = np.max(overlaps) + jmax = np.argmax(overlaps) + + if ovmax > ovthresh: + if not R["difficult"][jmax]: + if not R["det"][jmax]: + tp[d] = 1.0 + R["det"][jmax] = 1 + else: + fp[d] = 1.0 + else: + fp[d] = 1.0 + + # compute precision recall + fp = np.cumsum(fp) + tp = np.cumsum(tp) + rec = tp / float(npos) + # avoid divide by zero in case the first detection matches a difficult + # ground truth + prec = tp / np.maximum(tp + fp, np.finfo(np.float64).eps) + ap = voc_ap(rec, prec, use_07_metric) + + return rec, prec, ap diff --git a/detectron2/evaluation/rotated_coco_evaluation.py b/detectron2/evaluation/rotated_coco_evaluation.py new file mode 100644 index 0000000000000000000000000000000000000000..ea6d1b381dcf106339a03f08577df673ad439c46 --- /dev/null +++ b/detectron2/evaluation/rotated_coco_evaluation.py @@ -0,0 +1,207 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import itertools +import json +import numpy as np +import os +import torch +from pycocotools.cocoeval import COCOeval, maskUtils + +from detectron2.structures import BoxMode, RotatedBoxes, pairwise_iou_rotated +from detectron2.utils.file_io import PathManager + +from .coco_evaluation import COCOEvaluator + + +class RotatedCOCOeval(COCOeval): + @staticmethod + def is_rotated(box_list): + if type(box_list) == np.ndarray: + return box_list.shape[1] == 5 + elif type(box_list) == list: + if box_list == []: # cannot decide the box_dim + return False + return np.all( + np.array( + [ + (len(obj) == 5) and ((type(obj) == list) or (type(obj) == np.ndarray)) + for obj in box_list + ] + ) + ) + return False + + @staticmethod + def boxlist_to_tensor(boxlist, output_box_dim): + if type(boxlist) == np.ndarray: + box_tensor = torch.from_numpy(boxlist) + elif type(boxlist) == list: + if boxlist == []: + return torch.zeros((0, output_box_dim), dtype=torch.float32) + else: + box_tensor = torch.FloatTensor(boxlist) + else: + raise Exception("Unrecognized boxlist type") + + input_box_dim = box_tensor.shape[1] + if input_box_dim != output_box_dim: + if input_box_dim == 4 and output_box_dim == 5: + box_tensor = BoxMode.convert(box_tensor, BoxMode.XYWH_ABS, BoxMode.XYWHA_ABS) + else: + raise Exception( + "Unable to convert from {}-dim box to {}-dim box".format( + input_box_dim, output_box_dim + ) + ) + return box_tensor + + def compute_iou_dt_gt(self, dt, gt, is_crowd): + if self.is_rotated(dt) or self.is_rotated(gt): + # TODO: take is_crowd into consideration + assert all(c == 0 for c in is_crowd) + dt = RotatedBoxes(self.boxlist_to_tensor(dt, output_box_dim=5)) + gt = RotatedBoxes(self.boxlist_to_tensor(gt, output_box_dim=5)) + return pairwise_iou_rotated(dt, gt) + else: + # This is the same as the classical COCO evaluation + return maskUtils.iou(dt, gt, is_crowd) + + def computeIoU(self, imgId, catId): + p = self.params + if p.useCats: + gt = self._gts[imgId, catId] + dt = self._dts[imgId, catId] + else: + gt = [_ for cId in p.catIds for _ in self._gts[imgId, cId]] + dt = [_ for cId in p.catIds for _ in self._dts[imgId, cId]] + if len(gt) == 0 and len(dt) == 0: + return [] + inds = np.argsort([-d["score"] for d in dt], kind="mergesort") + dt = [dt[i] for i in inds] + if len(dt) > p.maxDets[-1]: + dt = dt[0 : p.maxDets[-1]] + + assert p.iouType == "bbox", "unsupported iouType for iou computation" + + g = [g["bbox"] for g in gt] + d = [d["bbox"] for d in dt] + + # compute iou between each dt and gt region + iscrowd = [int(o["iscrowd"]) for o in gt] + + # Note: this function is copied from cocoeval.py in cocoapi + # and the major difference is here. + ious = self.compute_iou_dt_gt(d, g, iscrowd) + return ious + + +class RotatedCOCOEvaluator(COCOEvaluator): + """ + Evaluate object proposal/instance detection outputs using COCO-like metrics and APIs, + with rotated boxes support. + Note: this uses IOU only and does not consider angle differences. + """ + + def process(self, inputs, outputs): + """ + Args: + inputs: the inputs to a COCO model (e.g., GeneralizedRCNN). + It is a list of dict. Each dict corresponds to an image and + contains keys like "height", "width", "file_name", "image_id". + outputs: the outputs of a COCO model. It is a list of dicts with key + "instances" that contains :class:`Instances`. + """ + for input, output in zip(inputs, outputs): + prediction = {"image_id": input["image_id"]} + + if "instances" in output: + instances = output["instances"].to(self._cpu_device) + + prediction["instances"] = self.instances_to_json(instances, input["image_id"]) + if "proposals" in output: + prediction["proposals"] = output["proposals"].to(self._cpu_device) + self._predictions.append(prediction) + + def instances_to_json(self, instances, img_id): + num_instance = len(instances) + if num_instance == 0: + return [] + + boxes = instances.pred_boxes.tensor.numpy() + if boxes.shape[1] == 4: + boxes = BoxMode.convert(boxes, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS) + boxes = boxes.tolist() + scores = instances.scores.tolist() + classes = instances.pred_classes.tolist() + + results = [] + for k in range(num_instance): + result = { + "image_id": img_id, + "category_id": classes[k], + "bbox": boxes[k], + "score": scores[k], + } + + results.append(result) + return results + + def _eval_predictions(self, predictions, img_ids=None): # img_ids: unused + """ + Evaluate predictions on the given tasks. + Fill self._results with the metrics of the tasks. + """ + self._logger.info("Preparing results for COCO format ...") + coco_results = list(itertools.chain(*[x["instances"] for x in predictions])) + + # unmap the category ids for COCO + if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"): + reverse_id_mapping = { + v: k for k, v in self._metadata.thing_dataset_id_to_contiguous_id.items() + } + for result in coco_results: + result["category_id"] = reverse_id_mapping[result["category_id"]] + + if self._output_dir: + file_path = os.path.join(self._output_dir, "coco_instances_results.json") + self._logger.info("Saving results to {}".format(file_path)) + with PathManager.open(file_path, "w") as f: + f.write(json.dumps(coco_results)) + f.flush() + + if not self._do_evaluation: + self._logger.info("Annotations are not available for evaluation.") + return + + self._logger.info("Evaluating predictions ...") + + assert self._tasks is None or set(self._tasks) == { + "bbox" + }, "[RotatedCOCOEvaluator] Only bbox evaluation is supported" + coco_eval = ( + self._evaluate_predictions_on_coco(self._coco_api, coco_results) + if len(coco_results) > 0 + else None # cocoapi does not handle empty results very well + ) + + task = "bbox" + res = self._derive_coco_results( + coco_eval, task, class_names=self._metadata.get("thing_classes") + ) + self._results[task] = res + + def _evaluate_predictions_on_coco(self, coco_gt, coco_results): + """ + Evaluate the coco results using COCOEval API. + """ + assert len(coco_results) > 0 + + coco_dt = coco_gt.loadRes(coco_results) + + # Only bbox is supported for now + coco_eval = RotatedCOCOeval(coco_gt, coco_dt, iouType="bbox") + + coco_eval.evaluate() + coco_eval.accumulate() + coco_eval.summarize() + + return coco_eval diff --git a/detectron2/evaluation/sem_seg_evaluation.py b/detectron2/evaluation/sem_seg_evaluation.py new file mode 100644 index 0000000000000000000000000000000000000000..7a19db71562ef47569dc7f77ec616af85447f0ec --- /dev/null +++ b/detectron2/evaluation/sem_seg_evaluation.py @@ -0,0 +1,184 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import itertools +import json +import logging +import numpy as np +import os +from collections import OrderedDict +import PIL.Image as Image +import pycocotools.mask as mask_util +import torch + +from detectron2.data import DatasetCatalog, MetadataCatalog +from detectron2.utils.comm import all_gather, is_main_process, synchronize +from detectron2.utils.file_io import PathManager + +from .evaluator import DatasetEvaluator + + +class SemSegEvaluator(DatasetEvaluator): + """ + Evaluate semantic segmentation metrics. + """ + + def __init__( + self, + dataset_name, + distributed=True, + output_dir=None, + *, + num_classes=None, + ignore_label=None, + ): + """ + Args: + dataset_name (str): name of the dataset to be evaluated. + distributed (bool): if True, will collect results from all ranks for evaluation. + Otherwise, will evaluate the results in the current process. + output_dir (str): an output directory to dump results. + num_classes, ignore_label: deprecated argument + """ + self._logger = logging.getLogger(__name__) + if num_classes is not None: + self._logger.warn( + "SemSegEvaluator(num_classes) is deprecated! It should be obtained from metadata." + ) + if ignore_label is not None: + self._logger.warn( + "SemSegEvaluator(ignore_label) is deprecated! It should be obtained from metadata." + ) + self._dataset_name = dataset_name + self._distributed = distributed + self._output_dir = output_dir + + self._cpu_device = torch.device("cpu") + + self.input_file_to_gt_file = { + dataset_record["file_name"]: dataset_record["sem_seg_file_name"] + for dataset_record in DatasetCatalog.get(dataset_name) + } + + meta = MetadataCatalog.get(dataset_name) + # Dict that maps contiguous training ids to COCO category ids + try: + c2d = meta.stuff_dataset_id_to_contiguous_id + self._contiguous_id_to_dataset_id = {v: k for k, v in c2d.items()} + except AttributeError: + self._contiguous_id_to_dataset_id = None + self._class_names = meta.stuff_classes + self._num_classes = len(meta.stuff_classes) + if num_classes is not None: + assert self._num_classes == num_classes, f"{self._num_classes} != {num_classes}" + self._ignore_label = ignore_label if ignore_label is not None else meta.ignore_label + + def reset(self): + self._conf_matrix = np.zeros((self._num_classes + 1, self._num_classes + 1), dtype=np.int64) + self._predictions = [] + + def process(self, inputs, outputs): + """ + Args: + inputs: the inputs to a model. + It is a list of dicts. Each dict corresponds to an image and + contains keys like "height", "width", "file_name". + outputs: the outputs of a model. It is either list of semantic segmentation predictions + (Tensor [H, W]) or list of dicts with key "sem_seg" that contains semantic + segmentation prediction in the same format. + """ + for input, output in zip(inputs, outputs): + output = output["sem_seg"].argmax(dim=0).to(self._cpu_device) + pred = np.array(output, dtype=np.int) + with PathManager.open(self.input_file_to_gt_file[input["file_name"]], "rb") as f: + gt = np.array(Image.open(f), dtype=np.int) + + gt[gt == self._ignore_label] = self._num_classes + + self._conf_matrix += np.bincount( + (self._num_classes + 1) * pred.reshape(-1) + gt.reshape(-1), + minlength=self._conf_matrix.size, + ).reshape(self._conf_matrix.shape) + + self._predictions.extend(self.encode_json_sem_seg(pred, input["file_name"])) + + def evaluate(self): + """ + Evaluates standard semantic segmentation metrics (http://cocodataset.org/#stuff-eval): + + * Mean intersection-over-union averaged across classes (mIoU) + * Frequency Weighted IoU (fwIoU) + * Mean pixel accuracy averaged across classes (mACC) + * Pixel Accuracy (pACC) + """ + if self._distributed: + synchronize() + conf_matrix_list = all_gather(self._conf_matrix) + self._predictions = all_gather(self._predictions) + self._predictions = list(itertools.chain(*self._predictions)) + if not is_main_process(): + return + + self._conf_matrix = np.zeros_like(self._conf_matrix) + for conf_matrix in conf_matrix_list: + self._conf_matrix += conf_matrix + + if self._output_dir: + PathManager.mkdirs(self._output_dir) + file_path = os.path.join(self._output_dir, "sem_seg_predictions.json") + with PathManager.open(file_path, "w") as f: + f.write(json.dumps(self._predictions)) + + acc = np.full(self._num_classes, np.nan, dtype=np.float) + iou = np.full(self._num_classes, np.nan, dtype=np.float) + tp = self._conf_matrix.diagonal()[:-1].astype(np.float) + pos_gt = np.sum(self._conf_matrix[:-1, :-1], axis=0).astype(np.float) + class_weights = pos_gt / np.sum(pos_gt) + pos_pred = np.sum(self._conf_matrix[:-1, :-1], axis=1).astype(np.float) + acc_valid = pos_gt > 0 + acc[acc_valid] = tp[acc_valid] / pos_gt[acc_valid] + iou_valid = (pos_gt + pos_pred) > 0 + union = pos_gt + pos_pred - tp + iou[acc_valid] = tp[acc_valid] / union[acc_valid] + macc = np.sum(acc[acc_valid]) / np.sum(acc_valid) + miou = np.sum(iou[acc_valid]) / np.sum(iou_valid) + fiou = np.sum(iou[acc_valid] * class_weights[acc_valid]) + pacc = np.sum(tp) / np.sum(pos_gt) + + res = {} + res["mIoU"] = 100 * miou + res["fwIoU"] = 100 * fiou + for i, name in enumerate(self._class_names): + res["IoU-{}".format(name)] = 100 * iou[i] + res["mACC"] = 100 * macc + res["pACC"] = 100 * pacc + for i, name in enumerate(self._class_names): + res["ACC-{}".format(name)] = 100 * acc[i] + + if self._output_dir: + file_path = os.path.join(self._output_dir, "sem_seg_evaluation.pth") + with PathManager.open(file_path, "wb") as f: + torch.save(res, f) + results = OrderedDict({"sem_seg": res}) + self._logger.info(results) + return results + + def encode_json_sem_seg(self, sem_seg, input_file_name): + """ + Convert semantic segmentation to COCO stuff format with segments encoded as RLEs. + See http://cocodataset.org/#format-results + """ + json_list = [] + for label in np.unique(sem_seg): + if self._contiguous_id_to_dataset_id is not None: + assert ( + label in self._contiguous_id_to_dataset_id + ), "Label {} is not in the metadata info for {}".format(label, self._dataset_name) + dataset_id = self._contiguous_id_to_dataset_id[label] + else: + dataset_id = int(label) + mask = (sem_seg == label).astype(np.uint8) + mask_rle = mask_util.encode(np.array(mask[:, :, None], order="F"))[0] + mask_rle["counts"] = mask_rle["counts"].decode("utf-8") + json_list.append( + {"file_name": input_file_name, "category_id": dataset_id, "segmentation": mask_rle} + ) + return json_list diff --git a/detectron2/evaluation/testing.py b/detectron2/evaluation/testing.py new file mode 100644 index 0000000000000000000000000000000000000000..9e5ae625bb0593fc20739dd3ea549157e4df4f3d --- /dev/null +++ b/detectron2/evaluation/testing.py @@ -0,0 +1,85 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import numpy as np +import pprint +import sys +from collections.abc import Mapping + + +def print_csv_format(results): + """ + Print main metrics in a format similar to Detectron, + so that they are easy to copypaste into a spreadsheet. + + Args: + results (OrderedDict[dict]): task_name -> {metric -> score} + unordered dict can also be printed, but in arbitrary order + """ + assert isinstance(results, Mapping) or not len(results), results + logger = logging.getLogger(__name__) + for task, res in results.items(): + if isinstance(res, Mapping): + # Don't print "AP-category" metrics since they are usually not tracked. + important_res = [(k, v) for k, v in res.items() if "-" not in k] + logger.info("copypaste: Task: {}".format(task)) + logger.info("copypaste: " + ",".join([k[0] for k in important_res])) + logger.info("copypaste: " + ",".join(["{0:.4f}".format(k[1]) for k in important_res])) + else: + logger.info(f"copypaste: {task}={res}") + + +def verify_results(cfg, results): + """ + Args: + results (OrderedDict[dict]): task_name -> {metric -> score} + + Returns: + bool: whether the verification succeeds or not + """ + expected_results = cfg.TEST.EXPECTED_RESULTS + if not len(expected_results): + return True + + ok = True + for task, metric, expected, tolerance in expected_results: + actual = results[task].get(metric, None) + if actual is None: + ok = False + continue + if not np.isfinite(actual): + ok = False + continue + diff = abs(actual - expected) + if diff > tolerance: + ok = False + + logger = logging.getLogger(__name__) + if not ok: + logger.error("Result verification failed!") + logger.error("Expected Results: " + str(expected_results)) + logger.error("Actual Results: " + pprint.pformat(results)) + + sys.exit(1) + else: + logger.info("Results verification passed.") + return ok + + +def flatten_results_dict(results): + """ + Expand a hierarchical dict of scalars into a flat dict of scalars. + If results[k1][k2][k3] = v, the returned dict will have the entry + {"k1/k2/k3": v}. + + Args: + results (dict): + """ + r = {} + for k, v in results.items(): + if isinstance(v, Mapping): + v = flatten_results_dict(v) + for kk, vv in v.items(): + r[k + "/" + kk] = vv + else: + r[k] = v + return r diff --git a/detectron2/export/README.md b/detectron2/export/README.md new file mode 100644 index 0000000000000000000000000000000000000000..9fcd33513fb81ef3aeb4d3c8d9732324dffa2646 --- /dev/null +++ b/detectron2/export/README.md @@ -0,0 +1,13 @@ + +This directory contains code to prepare a detectron2 model for deployment. +Currently it supports exporting a detectron2 model to Caffe2 format through ONNX. + +Please see [documentation](https://detectron2.readthedocs.io/tutorials/deployment.html) for its usage. + + +### Acknowledgements + +Thanks to Mobile Vision team at Facebook for developing the Caffe2 conversion tools. + +Thanks to Computing Platform Department - PAI team at Alibaba Group (@bddpqq, @chenbohua3) who +help export Detectron2 models to TorchScript. diff --git a/detectron2/export/__init__.py b/detectron2/export/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..78c27d64fa42760eeacd14d241cf28d58e3da490 --- /dev/null +++ b/detectron2/export/__init__.py @@ -0,0 +1,7 @@ +# -*- coding: utf-8 -*- + +from .api import * +from .flatten import TracingAdapter +from .torchscript import scripting_with_instances, dump_torchscript_IR + +__all__ = [k for k in globals().keys() if not k.startswith("_")] diff --git a/detectron2/export/api.py b/detectron2/export/api.py new file mode 100644 index 0000000000000000000000000000000000000000..e80989231ea5233e40f48a76e375a5a3c39208b1 --- /dev/null +++ b/detectron2/export/api.py @@ -0,0 +1,273 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import logging +import os +import torch +from caffe2.proto import caffe2_pb2 +from torch import nn + +from detectron2.config import CfgNode +from detectron2.utils.file_io import PathManager + +from .caffe2_inference import ProtobufDetectionModel +from .caffe2_modeling import META_ARCH_CAFFE2_EXPORT_TYPE_MAP, convert_batched_inputs_to_c2_format +from .shared import get_pb_arg_vali, get_pb_arg_vals, save_graph + +__all__ = [ + "add_export_config", + "export_caffe2_model", + "Caffe2Model", + "export_onnx_model", + "Caffe2Tracer", +] + + +def add_export_config(cfg): + """ + Add options needed by caffe2 export. + + Args: + cfg (CfgNode): a detectron2 config + + Returns: + CfgNode: + an updated config with new options that will be used by :class:`Caffe2Tracer`. + """ + is_frozen = cfg.is_frozen() + cfg.defrost() + cfg.EXPORT_CAFFE2 = CfgNode() + cfg.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT = False + if is_frozen: + cfg.freeze() + return cfg + + +class Caffe2Tracer: + """ + Make a detectron2 model traceable with Caffe2 operators. + This class creates a traceable version of a detectron2 model which: + + 1. Rewrite parts of the model using ops in Caffe2. Note that some ops do + not have GPU implementation in Caffe2. + 2. Remove post-processing and only produce raw layer outputs + + After making a traceable model, the class provide methods to export such a + model to different deployment formats. + Exported graph produced by this class take two input tensors: + + 1. (1, C, H, W) float "data" which is an image (usually in [0, 255]). + (H, W) often has to be padded to multiple of 32 (depend on the model + architecture). + 2. 1x3 float "im_info", each row of which is (height, width, 1.0). + Height and width are true image shapes before padding. + + The class currently only supports models using builtin meta architectures. + Batch inference is not supported, and contributions are welcome. + """ + + def __init__(self, cfg: CfgNode, model: nn.Module, inputs): + """ + Args: + cfg (CfgNode): a detectron2 config, with extra export-related options + added by :func:`add_export_config`. It's used to construct + caffe2-compatible model. + model (nn.Module): An original pytorch model. Must be among a few official models + in detectron2 that can be converted to become caffe2-compatible automatically. + Weights have to be already loaded to this model. + inputs: sample inputs that the given model takes for inference. + Will be used to trace the model. For most models, random inputs with + no detected objects will not work as they lead to wrong traces. + """ + assert isinstance(cfg, CfgNode), cfg + assert isinstance(model, torch.nn.Module), type(model) + + if "EXPORT_CAFFE2" not in cfg: + cfg = add_export_config(cfg) # will just the defaults + # TODO make it support custom models, by passing in c2 model directly + C2MetaArch = META_ARCH_CAFFE2_EXPORT_TYPE_MAP[cfg.MODEL.META_ARCHITECTURE] + self.traceable_model = C2MetaArch(cfg, copy.deepcopy(model)) + self.inputs = inputs + self.traceable_inputs = self.traceable_model.get_caffe2_inputs(inputs) + + def export_caffe2(self): + """ + Export the model to Caffe2's protobuf format. + The returned object can be saved with its :meth:`.save_protobuf()` method. + The result can be loaded and executed using Caffe2 runtime. + + Returns: + :class:`Caffe2Model` + """ + from .caffe2_export import export_caffe2_detection_model + + predict_net, init_net = export_caffe2_detection_model( + self.traceable_model, self.traceable_inputs + ) + return Caffe2Model(predict_net, init_net) + + def export_onnx(self): + """ + Export the model to ONNX format. + Note that the exported model contains custom ops only available in caffe2, therefore it + cannot be directly executed by other runtime (such as onnxruntime or TensorRT). + Post-processing or transformation passes may be applied on the model to accommodate + different runtimes, but we currently do not provide support for them. + + Returns: + onnx.ModelProto: an onnx model. + """ + from .caffe2_export import export_onnx_model as export_onnx_model_impl + + return export_onnx_model_impl(self.traceable_model, (self.traceable_inputs,)) + + def export_torchscript(self): + """ + Export the model to a ``torch.jit.TracedModule`` by tracing. + The returned object can be saved to a file by ``.save()``. + + Returns: + torch.jit.TracedModule: a torch TracedModule + """ + logger = logging.getLogger(__name__) + logger.info("Tracing the model with torch.jit.trace ...") + with torch.no_grad(): + return torch.jit.trace(self.traceable_model, (self.traceable_inputs,)) + + +class Caffe2Model(nn.Module): + """ + A wrapper around the traced model in Caffe2's protobuf format. + The exported graph has different inputs/outputs from the original Pytorch + model, as explained in :class:`Caffe2Tracer`. This class wraps around the + exported graph to simulate the same interface as the original Pytorch model. + It also provides functions to save/load models in Caffe2's format.' + + Examples: + :: + c2_model = Caffe2Tracer(cfg, torch_model, inputs).export_caffe2() + inputs = [{"image": img_tensor_CHW}] + outputs = c2_model(inputs) + orig_outputs = torch_model(inputs) + """ + + def __init__(self, predict_net, init_net): + super().__init__() + self.eval() # always in eval mode + self._predict_net = predict_net + self._init_net = init_net + self._predictor = None + + __init__.__HIDE_SPHINX_DOC__ = True + + @property + def predict_net(self): + """ + caffe2.core.Net: the underlying caffe2 predict net + """ + return self._predict_net + + @property + def init_net(self): + """ + caffe2.core.Net: the underlying caffe2 init net + """ + return self._init_net + + def save_protobuf(self, output_dir): + """ + Save the model as caffe2's protobuf format. + It saves the following files: + + * "model.pb": definition of the graph. Can be visualized with + tools like `netron `_. + * "model_init.pb": model parameters + * "model.pbtxt": human-readable definition of the graph. Not + needed for deployment. + + Args: + output_dir (str): the output directory to save protobuf files. + """ + logger = logging.getLogger(__name__) + logger.info("Saving model to {} ...".format(output_dir)) + if not PathManager.exists(output_dir): + PathManager.mkdirs(output_dir) + + with PathManager.open(os.path.join(output_dir, "model.pb"), "wb") as f: + f.write(self._predict_net.SerializeToString()) + with PathManager.open(os.path.join(output_dir, "model.pbtxt"), "w") as f: + f.write(str(self._predict_net)) + with PathManager.open(os.path.join(output_dir, "model_init.pb"), "wb") as f: + f.write(self._init_net.SerializeToString()) + + def save_graph(self, output_file, inputs=None): + """ + Save the graph as SVG format. + + Args: + output_file (str): a SVG file + inputs: optional inputs given to the model. + If given, the inputs will be used to run the graph to record + shape of every tensor. The shape information will be + saved together with the graph. + """ + from .caffe2_export import run_and_save_graph + + if inputs is None: + save_graph(self._predict_net, output_file, op_only=False) + else: + size_divisibility = get_pb_arg_vali(self._predict_net, "size_divisibility", 0) + device = get_pb_arg_vals(self._predict_net, "device", b"cpu").decode("ascii") + inputs = convert_batched_inputs_to_c2_format(inputs, size_divisibility, device) + inputs = [x.cpu().numpy() for x in inputs] + run_and_save_graph(self._predict_net, self._init_net, inputs, output_file) + + @staticmethod + def load_protobuf(dir): + """ + Args: + dir (str): a directory used to save Caffe2Model with + :meth:`save_protobuf`. + The files "model.pb" and "model_init.pb" are needed. + + Returns: + Caffe2Model: the caffe2 model loaded from this directory. + """ + predict_net = caffe2_pb2.NetDef() + with PathManager.open(os.path.join(dir, "model.pb"), "rb") as f: + predict_net.ParseFromString(f.read()) + + init_net = caffe2_pb2.NetDef() + with PathManager.open(os.path.join(dir, "model_init.pb"), "rb") as f: + init_net.ParseFromString(f.read()) + + return Caffe2Model(predict_net, init_net) + + def __call__(self, inputs): + """ + An interface that wraps around a Caffe2 model and mimics detectron2's models' + input/output format. See details about the format at :doc:`/tutorials/models`. + This is used to compare the outputs of caffe2 model with its original torch model. + + Due to the extra conversion between Pytorch/Caffe2, this method is not meant for + benchmark. Because of the conversion, this method also has dependency + on detectron2 in order to convert to detectron2's output format. + """ + if self._predictor is None: + self._predictor = ProtobufDetectionModel(self._predict_net, self._init_net) + return self._predictor(inputs) + + +def export_caffe2_model(cfg, model, inputs): + logger = logging.getLogger(__name__) + logger.warning( + "export_caffe2_model() is deprecated. Please use `Caffe2Tracer().export_caffe2() instead." + ) + return Caffe2Tracer(cfg, model, inputs).export_caffe2() + + +def export_onnx_model(cfg, model, inputs): + logger = logging.getLogger(__name__) + logger.warning( + "export_caffe2_model() is deprecated. Please use `Caffe2Tracer().export_onnx() instead." + ) + return Caffe2Tracer(cfg, model, inputs).export_onnx() diff --git a/detectron2/export/c10.py b/detectron2/export/c10.py new file mode 100644 index 0000000000000000000000000000000000000000..ffb47c6cf19ae07f334b751ccadd071ebbd25e2e --- /dev/null +++ b/detectron2/export/c10.py @@ -0,0 +1,527 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import math +import torch +import torch.nn.functional as F + +from detectron2.layers import cat +from detectron2.layers.roi_align_rotated import ROIAlignRotated +from detectron2.modeling import poolers +from detectron2.modeling.proposal_generator import rpn +from detectron2.modeling.roi_heads.mask_head import mask_rcnn_inference +from detectron2.structures import Boxes, ImageList, Instances, Keypoints + +from .shared import alias, to_device + + +""" +This file contains caffe2-compatible implementation of several detectron2 components. +""" + + +class Caffe2Boxes(Boxes): + """ + Representing a list of detectron2.structures.Boxes from minibatch, each box + is represented by a 5d vector (batch index + 4 coordinates), or a 6d vector + (batch index + 5 coordinates) for RotatedBoxes. + """ + + def __init__(self, tensor): + assert isinstance(tensor, torch.Tensor) + assert tensor.dim() == 2 and tensor.size(-1) in [4, 5, 6], tensor.size() + # TODO: make tensor immutable when dim is Nx5 for Boxes, + # and Nx6 for RotatedBoxes? + self.tensor = tensor + + +# TODO clean up this class, maybe just extend Instances +class InstancesList(object): + """ + Tensor representation of a list of Instances object for a batch of images. + + When dealing with a batch of images with Caffe2 ops, a list of bboxes + (instances) are usually represented by single Tensor with size + (sigma(Ni), 5) or (sigma(Ni), 4) plus a batch split Tensor. This class is + for providing common functions to convert between these two representations. + """ + + def __init__(self, im_info, indices, extra_fields=None): + # [N, 3] -> (H, W, Scale) + self.im_info = im_info + # [N,] -> indice of batch to which the instance belongs + self.indices = indices + # [N, ...] + self.batch_extra_fields = extra_fields or {} + + self.image_size = self.im_info + + def get_fields(self): + """like `get_fields` in the Instances object, + but return each field in tensor representations""" + ret = {} + for k, v in self.batch_extra_fields.items(): + # if isinstance(v, torch.Tensor): + # tensor_rep = v + # elif isinstance(v, (Boxes, Keypoints)): + # tensor_rep = v.tensor + # else: + # raise ValueError("Can't find tensor representation for: {}".format()) + ret[k] = v + return ret + + def has(self, name): + return name in self.batch_extra_fields + + def set(self, name, value): + data_len = len(value) + if len(self.batch_extra_fields): + assert ( + len(self) == data_len + ), "Adding a field of length {} to a Instances of length {}".format(data_len, len(self)) + self.batch_extra_fields[name] = value + + def __setattr__(self, name, val): + if name in ["im_info", "indices", "batch_extra_fields", "image_size"]: + super().__setattr__(name, val) + else: + self.set(name, val) + + def __getattr__(self, name): + if name not in self.batch_extra_fields: + raise AttributeError("Cannot find field '{}' in the given Instances!".format(name)) + return self.batch_extra_fields[name] + + def __len__(self): + return len(self.indices) + + def flatten(self): + ret = [] + for _, v in self.batch_extra_fields.items(): + if isinstance(v, (Boxes, Keypoints)): + ret.append(v.tensor) + else: + ret.append(v) + return ret + + @staticmethod + def to_d2_instances_list(instances_list): + """ + Convert InstancesList to List[Instances]. The input `instances_list` can + also be a List[Instances], in this case this method is a non-op. + """ + if not isinstance(instances_list, InstancesList): + assert all(isinstance(x, Instances) for x in instances_list) + return instances_list + + ret = [] + for i, info in enumerate(instances_list.im_info): + instances = Instances(torch.Size([int(info[0].item()), int(info[1].item())])) + + ids = instances_list.indices == i + for k, v in instances_list.batch_extra_fields.items(): + if isinstance(v, torch.Tensor): + instances.set(k, v[ids]) + continue + elif isinstance(v, Boxes): + instances.set(k, v[ids, -4:]) + continue + + target_type, tensor_source = v + assert isinstance(tensor_source, torch.Tensor) + assert tensor_source.shape[0] == instances_list.indices.shape[0] + tensor_source = tensor_source[ids] + + if issubclass(target_type, Boxes): + instances.set(k, Boxes(tensor_source[:, -4:])) + elif issubclass(target_type, Keypoints): + instances.set(k, Keypoints(tensor_source)) + elif issubclass(target_type, torch.Tensor): + instances.set(k, tensor_source) + else: + raise ValueError("Can't handle targe type: {}".format(target_type)) + + ret.append(instances) + return ret + + +class Caffe2Compatible(object): + """ + A model can inherit this class to indicate that it can be traced and deployed with caffe2. + """ + + def _get_tensor_mode(self): + return self._tensor_mode + + def _set_tensor_mode(self, v): + self._tensor_mode = v + + tensor_mode = property(_get_tensor_mode, _set_tensor_mode) + """ + If true, the model expects C2-style tensor only inputs/outputs format. + """ + + +class Caffe2RPN(Caffe2Compatible, rpn.RPN): + def _generate_proposals( + self, images, objectness_logits_pred, anchor_deltas_pred, gt_instances=None + ): + assert isinstance(images, ImageList) + if self.tensor_mode: + im_info = images.image_sizes + else: + im_info = torch.tensor([[im_sz[0], im_sz[1], 1.0] for im_sz in images.image_sizes]).to( + images.tensor.device + ) + assert isinstance(im_info, torch.Tensor) + + rpn_rois_list = [] + rpn_roi_probs_list = [] + for scores, bbox_deltas, cell_anchors_tensor, feat_stride in zip( + objectness_logits_pred, + anchor_deltas_pred, + iter(self.anchor_generator.cell_anchors), + self.anchor_generator.strides, + ): + scores = scores.detach() + bbox_deltas = bbox_deltas.detach() + + rpn_rois, rpn_roi_probs = torch.ops._caffe2.GenerateProposals( + scores, + bbox_deltas, + im_info, + cell_anchors_tensor, + spatial_scale=1.0 / feat_stride, + pre_nms_topN=self.pre_nms_topk[self.training], + post_nms_topN=self.post_nms_topk[self.training], + nms_thresh=self.nms_thresh, + min_size=self.min_box_size, + # correct_transform_coords=True, # deprecated argument + angle_bound_on=True, # Default + angle_bound_lo=-180, + angle_bound_hi=180, + clip_angle_thresh=1.0, # Default + legacy_plus_one=False, + ) + rpn_rois_list.append(rpn_rois) + rpn_roi_probs_list.append(rpn_roi_probs) + + # For FPN in D2, in RPN all proposals from different levels are concated + # together, ranked and picked by top post_nms_topk. Then in ROIPooler + # it calculates level_assignments and calls the RoIAlign from + # the corresponding level. + + if len(objectness_logits_pred) == 1: + rpn_rois = rpn_rois_list[0] + rpn_roi_probs = rpn_roi_probs_list[0] + else: + assert len(rpn_rois_list) == len(rpn_roi_probs_list) + rpn_post_nms_topN = self.post_nms_topk[self.training] + + device = rpn_rois_list[0].device + input_list = [to_device(x, "cpu") for x in (rpn_rois_list + rpn_roi_probs_list)] + + # TODO remove this after confirming rpn_max_level/rpn_min_level + # is not needed in CollectRpnProposals. + feature_strides = list(self.anchor_generator.strides) + rpn_min_level = int(math.log2(feature_strides[0])) + rpn_max_level = int(math.log2(feature_strides[-1])) + assert (rpn_max_level - rpn_min_level + 1) == len( + rpn_rois_list + ), "CollectRpnProposals requires continuous levels" + + rpn_rois = torch.ops._caffe2.CollectRpnProposals( + input_list, + # NOTE: in current implementation, rpn_max_level and rpn_min_level + # are not needed, only the subtraction of two matters and it + # can be infer from the number of inputs. Keep them now for + # consistency. + rpn_max_level=2 + len(rpn_rois_list) - 1, + rpn_min_level=2, + rpn_post_nms_topN=rpn_post_nms_topN, + ) + rpn_rois = to_device(rpn_rois, device) + rpn_roi_probs = [] + + proposals = self.c2_postprocess(im_info, rpn_rois, rpn_roi_probs, self.tensor_mode) + return proposals, {} + + def forward(self, images, features, gt_instances=None): + assert not self.training + features = [features[f] for f in self.in_features] + objectness_logits_pred, anchor_deltas_pred = self.rpn_head(features) + return self._generate_proposals( + images, + objectness_logits_pred, + anchor_deltas_pred, + gt_instances, + ) + + @staticmethod + def c2_postprocess(im_info, rpn_rois, rpn_roi_probs, tensor_mode): + proposals = InstancesList( + im_info=im_info, + indices=rpn_rois[:, 0], + extra_fields={ + "proposal_boxes": Caffe2Boxes(rpn_rois), + "objectness_logits": (torch.Tensor, rpn_roi_probs), + }, + ) + if not tensor_mode: + proposals = InstancesList.to_d2_instances_list(proposals) + else: + proposals = [proposals] + return proposals + + +class Caffe2ROIPooler(Caffe2Compatible, poolers.ROIPooler): + @staticmethod + def c2_preprocess(box_lists): + assert all(isinstance(x, Boxes) for x in box_lists) + if all(isinstance(x, Caffe2Boxes) for x in box_lists): + # input is pure-tensor based + assert len(box_lists) == 1 + pooler_fmt_boxes = box_lists[0].tensor + else: + pooler_fmt_boxes = poolers.convert_boxes_to_pooler_format(box_lists) + return pooler_fmt_boxes + + def forward(self, x, box_lists): + assert not self.training + + pooler_fmt_boxes = self.c2_preprocess(box_lists) + num_level_assignments = len(self.level_poolers) + + if num_level_assignments == 1: + if isinstance(self.level_poolers[0], ROIAlignRotated): + c2_roi_align = torch.ops._caffe2.RoIAlignRotated + aligned = True + else: + c2_roi_align = torch.ops._caffe2.RoIAlign + aligned = self.level_poolers[0].aligned + + out = c2_roi_align( + x[0], + pooler_fmt_boxes, + order="NCHW", + spatial_scale=float(self.level_poolers[0].spatial_scale), + pooled_h=int(self.output_size[0]), + pooled_w=int(self.output_size[1]), + sampling_ratio=int(self.level_poolers[0].sampling_ratio), + aligned=aligned, + ) + return out + + device = pooler_fmt_boxes.device + assert ( + self.max_level - self.min_level + 1 == 4 + ), "Currently DistributeFpnProposals only support 4 levels" + fpn_outputs = torch.ops._caffe2.DistributeFpnProposals( + to_device(pooler_fmt_boxes, "cpu"), + roi_canonical_scale=self.canonical_box_size, + roi_canonical_level=self.canonical_level, + roi_max_level=self.max_level, + roi_min_level=self.min_level, + legacy_plus_one=False, + ) + fpn_outputs = [to_device(x, device) for x in fpn_outputs] + + rois_fpn_list = fpn_outputs[:-1] + rois_idx_restore_int32 = fpn_outputs[-1] + + roi_feat_fpn_list = [] + for roi_fpn, x_level, pooler in zip(rois_fpn_list, x, self.level_poolers): + if isinstance(pooler, ROIAlignRotated): + c2_roi_align = torch.ops._caffe2.RoIAlignRotated + aligned = True + else: + c2_roi_align = torch.ops._caffe2.RoIAlign + aligned = bool(pooler.aligned) + + roi_feat_fpn = c2_roi_align( + x_level, + roi_fpn, + order="NCHW", + spatial_scale=float(pooler.spatial_scale), + pooled_h=int(self.output_size[0]), + pooled_w=int(self.output_size[1]), + sampling_ratio=int(pooler.sampling_ratio), + aligned=aligned, + ) + roi_feat_fpn_list.append(roi_feat_fpn) + + roi_feat_shuffled = cat(roi_feat_fpn_list, dim=0) + assert roi_feat_shuffled.numel() > 0 and rois_idx_restore_int32.numel() > 0, ( + "Caffe2 export requires tracing with a model checkpoint + input that can produce valid" + " detections. But no detections were obtained with the given checkpoint and input!" + ) + roi_feat = torch.ops._caffe2.BatchPermutation(roi_feat_shuffled, rois_idx_restore_int32) + return roi_feat + + +class Caffe2FastRCNNOutputsInference: + def __init__(self, tensor_mode): + self.tensor_mode = tensor_mode # whether the output is caffe2 tensor mode + + def __call__(self, box_predictor, predictions, proposals): + """equivalent to FastRCNNOutputLayers.inference""" + num_classes = box_predictor.num_classes + score_thresh = box_predictor.test_score_thresh + nms_thresh = box_predictor.test_nms_thresh + topk_per_image = box_predictor.test_topk_per_image + is_rotated = len(box_predictor.box2box_transform.weights) == 5 + + if is_rotated: + box_dim = 5 + assert box_predictor.box2box_transform.weights[4] == 1, ( + "The weights for Rotated BBoxTransform in C2 have only 4 dimensions," + + " thus enforcing the angle weight to be 1 for now" + ) + box2box_transform_weights = box_predictor.box2box_transform.weights[:4] + else: + box_dim = 4 + box2box_transform_weights = box_predictor.box2box_transform.weights + + class_logits, box_regression = predictions + if num_classes + 1 == class_logits.shape[1]: + class_prob = F.softmax(class_logits, -1) + else: + assert num_classes == class_logits.shape[1] + class_prob = F.sigmoid(class_logits) + # BoxWithNMSLimit will infer num_classes from the shape of the class_prob + # So append a zero column as placeholder for the background class + class_prob = torch.cat((class_prob, torch.zeros(class_prob.shape[0], 1)), dim=1) + + assert box_regression.shape[1] % box_dim == 0 + cls_agnostic_bbox_reg = box_regression.shape[1] // box_dim == 1 + + input_tensor_mode = proposals[0].proposal_boxes.tensor.shape[1] == box_dim + 1 + + rois = type(proposals[0].proposal_boxes).cat([p.proposal_boxes for p in proposals]) + device, dtype = rois.tensor.device, rois.tensor.dtype + if input_tensor_mode: + im_info = proposals[0].image_size + rois = rois.tensor + else: + im_info = torch.tensor( + [[sz[0], sz[1], 1.0] for sz in [x.image_size for x in proposals]] + ) + batch_ids = cat( + [ + torch.full((b, 1), i, dtype=dtype, device=device) + for i, b in enumerate(len(p) for p in proposals) + ], + dim=0, + ) + rois = torch.cat([batch_ids, rois.tensor], dim=1) + + roi_pred_bbox, roi_batch_splits = torch.ops._caffe2.BBoxTransform( + to_device(rois, "cpu"), + to_device(box_regression, "cpu"), + to_device(im_info, "cpu"), + weights=box2box_transform_weights, + apply_scale=True, + rotated=is_rotated, + angle_bound_on=True, + angle_bound_lo=-180, + angle_bound_hi=180, + clip_angle_thresh=1.0, + legacy_plus_one=False, + ) + roi_pred_bbox = to_device(roi_pred_bbox, device) + roi_batch_splits = to_device(roi_batch_splits, device) + + nms_outputs = torch.ops._caffe2.BoxWithNMSLimit( + to_device(class_prob, "cpu"), + to_device(roi_pred_bbox, "cpu"), + to_device(roi_batch_splits, "cpu"), + score_thresh=float(score_thresh), + nms=float(nms_thresh), + detections_per_im=int(topk_per_image), + soft_nms_enabled=False, + soft_nms_method="linear", + soft_nms_sigma=0.5, + soft_nms_min_score_thres=0.001, + rotated=is_rotated, + cls_agnostic_bbox_reg=cls_agnostic_bbox_reg, + input_boxes_include_bg_cls=False, + output_classes_include_bg_cls=False, + legacy_plus_one=False, + ) + roi_score_nms = to_device(nms_outputs[0], device) + roi_bbox_nms = to_device(nms_outputs[1], device) + roi_class_nms = to_device(nms_outputs[2], device) + roi_batch_splits_nms = to_device(nms_outputs[3], device) + roi_keeps_nms = to_device(nms_outputs[4], device) + roi_keeps_size_nms = to_device(nms_outputs[5], device) + if not self.tensor_mode: + roi_class_nms = roi_class_nms.to(torch.int64) + + roi_batch_ids = cat( + [ + torch.full((b, 1), i, dtype=dtype, device=device) + for i, b in enumerate(int(x.item()) for x in roi_batch_splits_nms) + ], + dim=0, + ) + + roi_class_nms = alias(roi_class_nms, "class_nms") + roi_score_nms = alias(roi_score_nms, "score_nms") + roi_bbox_nms = alias(roi_bbox_nms, "bbox_nms") + roi_batch_splits_nms = alias(roi_batch_splits_nms, "batch_splits_nms") + roi_keeps_nms = alias(roi_keeps_nms, "keeps_nms") + roi_keeps_size_nms = alias(roi_keeps_size_nms, "keeps_size_nms") + + results = InstancesList( + im_info=im_info, + indices=roi_batch_ids[:, 0], + extra_fields={ + "pred_boxes": Caffe2Boxes(roi_bbox_nms), + "scores": roi_score_nms, + "pred_classes": roi_class_nms, + }, + ) + + if not self.tensor_mode: + results = InstancesList.to_d2_instances_list(results) + batch_splits = roi_batch_splits_nms.int().tolist() + kept_indices = list(roi_keeps_nms.to(torch.int64).split(batch_splits)) + else: + results = [results] + kept_indices = [roi_keeps_nms] + + return results, kept_indices + + +class Caffe2MaskRCNNInference: + def __call__(self, pred_mask_logits, pred_instances): + """equivalent to mask_head.mask_rcnn_inference""" + if all(isinstance(x, InstancesList) for x in pred_instances): + assert len(pred_instances) == 1 + mask_probs_pred = pred_mask_logits.sigmoid() + mask_probs_pred = alias(mask_probs_pred, "mask_fcn_probs") + pred_instances[0].pred_masks = mask_probs_pred + else: + mask_rcnn_inference(pred_mask_logits, pred_instances) + + +class Caffe2KeypointRCNNInference: + def __init__(self, use_heatmap_max_keypoint): + self.use_heatmap_max_keypoint = use_heatmap_max_keypoint + + def __call__(self, pred_keypoint_logits, pred_instances): + # just return the keypoint heatmap for now, + # there will be option to call HeatmapMaxKeypointOp + output = alias(pred_keypoint_logits, "kps_score") + if all(isinstance(x, InstancesList) for x in pred_instances): + assert len(pred_instances) == 1 + if self.use_heatmap_max_keypoint: + device = output.device + output = torch.ops._caffe2.HeatmapMaxKeypoint( + to_device(output, "cpu"), + pred_instances[0].pred_boxes.tensor, + should_output_softmax=True, # worth make it configerable? + ) + output = to_device(output, device) + output = alias(output, "keypoints_out") + pred_instances[0].pred_keypoints = output + return pred_keypoint_logits diff --git a/detectron2/export/caffe2_export.py b/detectron2/export/caffe2_export.py new file mode 100644 index 0000000000000000000000000000000000000000..74ac123a7aed6cd77d6d833446a831d9048745b2 --- /dev/null +++ b/detectron2/export/caffe2_export.py @@ -0,0 +1,207 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import copy +import io +import logging +import numpy as np +from typing import List +import onnx +import torch +from caffe2.proto import caffe2_pb2 +from caffe2.python import core +from caffe2.python.onnx.backend import Caffe2Backend +from tabulate import tabulate +from termcolor import colored +from torch.onnx import OperatorExportTypes + +from .shared import ( + ScopedWS, + construct_init_net_from_params, + fuse_alias_placeholder, + fuse_copy_between_cpu_and_gpu, + get_params_from_init_net, + group_norm_replace_aten_with_caffe2, + infer_device_type, + remove_dead_end_ops, + remove_reshape_for_fc, + save_graph, +) + +logger = logging.getLogger(__name__) + + +def export_onnx_model(model, inputs): + """ + Trace and export a model to onnx format. + + Args: + model (nn.Module): + inputs (tuple[args]): the model will be called by `model(*inputs)` + + Returns: + an onnx model + """ + assert isinstance(model, torch.nn.Module) + + # make sure all modules are in eval mode, onnx may change the training state + # of the module if the states are not consistent + def _check_eval(module): + assert not module.training + + model.apply(_check_eval) + + # Export the model to ONNX + with torch.no_grad(): + with io.BytesIO() as f: + torch.onnx.export( + model, + inputs, + f, + operator_export_type=OperatorExportTypes.ONNX_ATEN_FALLBACK, + # verbose=True, # NOTE: uncomment this for debugging + # export_params=True, + ) + onnx_model = onnx.load_from_string(f.getvalue()) + + # Apply ONNX's Optimization + all_passes = onnx.optimizer.get_available_passes() + passes = ["fuse_bn_into_conv"] + assert all(p in all_passes for p in passes) + onnx_model = onnx.optimizer.optimize(onnx_model, passes) + return onnx_model + + +def _op_stats(net_def): + type_count = {} + for t in [op.type for op in net_def.op]: + type_count[t] = type_count.get(t, 0) + 1 + type_count_list = sorted(type_count.items(), key=lambda kv: kv[0]) # alphabet + type_count_list = sorted(type_count_list, key=lambda kv: -kv[1]) # count + return "\n".join("{:>4}x {}".format(count, name) for name, count in type_count_list) + + +def _assign_device_option( + predict_net: caffe2_pb2.NetDef, init_net: caffe2_pb2.NetDef, tensor_inputs: List[torch.Tensor] +): + """ + ONNX exported network doesn't have concept of device, assign necessary + device option for each op in order to make it runable on GPU runtime. + """ + + def _get_device_type(torch_tensor): + assert torch_tensor.device.type in ["cpu", "cuda"] + assert torch_tensor.device.index == 0 + return torch_tensor.device.type + + def _assign_op_device_option(net_proto, net_ssa, blob_device_types): + for op, ssa_i in zip(net_proto.op, net_ssa): + if op.type in ["CopyCPUToGPU", "CopyGPUToCPU"]: + op.device_option.CopyFrom(core.DeviceOption(caffe2_pb2.CUDA, 0)) + else: + devices = [blob_device_types[b] for b in ssa_i[0] + ssa_i[1]] + assert all(d == devices[0] for d in devices) + if devices[0] == "cuda": + op.device_option.CopyFrom(core.DeviceOption(caffe2_pb2.CUDA, 0)) + + # update ops in predict_net + predict_net_input_device_types = { + (name, 0): _get_device_type(tensor) + for name, tensor in zip(predict_net.external_input, tensor_inputs) + } + predict_net_device_types = infer_device_type( + predict_net, known_status=predict_net_input_device_types, device_name_style="pytorch" + ) + predict_net_ssa, _ = core.get_ssa(predict_net) + _assign_op_device_option(predict_net, predict_net_ssa, predict_net_device_types) + + # update ops in init_net + init_net_ssa, versions = core.get_ssa(init_net) + init_net_output_device_types = { + (name, versions[name]): predict_net_device_types[(name, 0)] + for name in init_net.external_output + } + init_net_device_types = infer_device_type( + init_net, known_status=init_net_output_device_types, device_name_style="pytorch" + ) + _assign_op_device_option(init_net, init_net_ssa, init_net_device_types) + + +def export_caffe2_detection_model(model: torch.nn.Module, tensor_inputs: List[torch.Tensor]): + """ + Export a caffe2-compatible Detectron2 model to caffe2 format via ONNX. + + Arg: + model: a caffe2-compatible version of detectron2 model, defined in caffe2_modeling.py + tensor_inputs: a list of tensors that caffe2 model takes as input. + """ + model = copy.deepcopy(model) + assert isinstance(model, torch.nn.Module) + assert hasattr(model, "encode_additional_info") + + # Export via ONNX + logger.info( + "Exporting a {} model via ONNX ...".format(type(model).__name__) + + " Some warnings from ONNX are expected and are usually not to worry about." + ) + onnx_model = export_onnx_model(model, (tensor_inputs,)) + # Convert ONNX model to Caffe2 protobuf + init_net, predict_net = Caffe2Backend.onnx_graph_to_caffe2_net(onnx_model) + ops_table = [[op.type, op.input, op.output] for op in predict_net.op] + table = tabulate(ops_table, headers=["type", "input", "output"], tablefmt="pipe") + logger.info( + "ONNX export Done. Exported predict_net (before optimizations):\n" + colored(table, "cyan") + ) + + # Apply protobuf optimization + fuse_alias_placeholder(predict_net, init_net) + if any(t.device.type != "cpu" for t in tensor_inputs): + fuse_copy_between_cpu_and_gpu(predict_net) + remove_dead_end_ops(init_net) + _assign_device_option(predict_net, init_net, tensor_inputs) + params, device_options = get_params_from_init_net(init_net) + predict_net, params = remove_reshape_for_fc(predict_net, params) + init_net = construct_init_net_from_params(params, device_options) + group_norm_replace_aten_with_caffe2(predict_net) + + # Record necessary information for running the pb model in Detectron2 system. + model.encode_additional_info(predict_net, init_net) + + logger.info("Operators used in predict_net: \n{}".format(_op_stats(predict_net))) + logger.info("Operators used in init_net: \n{}".format(_op_stats(init_net))) + + return predict_net, init_net + + +def run_and_save_graph(predict_net, init_net, tensor_inputs, graph_save_path): + """ + Run the caffe2 model on given inputs, recording the shape and draw the graph. + + predict_net/init_net: caffe2 model. + tensor_inputs: a list of tensors that caffe2 model takes as input. + graph_save_path: path for saving graph of exported model. + """ + + logger.info("Saving graph of ONNX exported model to {} ...".format(graph_save_path)) + save_graph(predict_net, graph_save_path, op_only=False) + + # Run the exported Caffe2 net + logger.info("Running ONNX exported model ...") + with ScopedWS("__ws_tmp__", True) as ws: + ws.RunNetOnce(init_net) + initialized_blobs = set(ws.Blobs()) + uninitialized = [inp for inp in predict_net.external_input if inp not in initialized_blobs] + for name, blob in zip(uninitialized, tensor_inputs): + ws.FeedBlob(name, blob) + + try: + ws.RunNetOnce(predict_net) + except RuntimeError as e: + logger.warning("Encountered RuntimeError: \n{}".format(str(e))) + + ws_blobs = {b: ws.FetchBlob(b) for b in ws.Blobs()} + blob_sizes = {b: ws_blobs[b].shape for b in ws_blobs if isinstance(ws_blobs[b], np.ndarray)} + + logger.info("Saving graph with blob shapes to {} ...".format(graph_save_path)) + save_graph(predict_net, graph_save_path, op_only=False, blob_sizes=blob_sizes) + + return ws_blobs diff --git a/detectron2/export/caffe2_inference.py b/detectron2/export/caffe2_inference.py new file mode 100644 index 0000000000000000000000000000000000000000..deb886c0417285ed1d5ad85eb941fa1ac757cdab --- /dev/null +++ b/detectron2/export/caffe2_inference.py @@ -0,0 +1,161 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import logging +import numpy as np +from itertools import count +import torch +from caffe2.proto import caffe2_pb2 +from caffe2.python import core + +from .caffe2_modeling import META_ARCH_CAFFE2_EXPORT_TYPE_MAP, convert_batched_inputs_to_c2_format +from .shared import ScopedWS, get_pb_arg_vali, get_pb_arg_vals, infer_device_type + +logger = logging.getLogger(__name__) + + +# ===== ref: mobile-vision predictor's 'Caffe2Wrapper' class ====== +class ProtobufModel(torch.nn.Module): + """ + Wrapper of a caffe2's protobuf model. + It works just like nn.Module, but running caffe2 under the hood. + Input/Output are tuple[tensor] that match the caffe2 net's external_input/output. + """ + + _ids = count(0) + + def __init__(self, predict_net, init_net): + logger.info(f"Initializing ProtobufModel for: {predict_net.name} ...") + super().__init__() + assert isinstance(predict_net, caffe2_pb2.NetDef) + assert isinstance(init_net, caffe2_pb2.NetDef) + # create unique temporary workspace for each instance + self.ws_name = "__tmp_ProtobufModel_{}__".format(next(self._ids)) + self.net = core.Net(predict_net) + + logger.info("Running init_net once to fill the parameters ...") + with ScopedWS(self.ws_name, is_reset=True, is_cleanup=False) as ws: + ws.RunNetOnce(init_net) + uninitialized_external_input = [] + for blob in self.net.Proto().external_input: + if blob not in ws.Blobs(): + uninitialized_external_input.append(blob) + ws.CreateBlob(blob) + ws.CreateNet(self.net) + + self._error_msgs = set() + self._input_blobs = uninitialized_external_input + + def _infer_output_devices(self, inputs): + """ + Returns: + list[str]: list of device for each external output + """ + + def _get_device_type(torch_tensor): + assert torch_tensor.device.type in ["cpu", "cuda"] + assert torch_tensor.device.index == 0 + return torch_tensor.device.type + + predict_net = self.net.Proto() + input_device_types = { + (name, 0): _get_device_type(tensor) for name, tensor in zip(self._input_blobs, inputs) + } + device_type_map = infer_device_type( + predict_net, known_status=input_device_types, device_name_style="pytorch" + ) + ssa, versions = core.get_ssa(predict_net) + versioned_outputs = [(name, versions[name]) for name in predict_net.external_output] + output_devices = [device_type_map[outp] for outp in versioned_outputs] + return output_devices + + def forward(self, inputs): + """ + Args: + inputs (tuple[torch.Tensor]) + + Returns: + tuple[torch.Tensor] + """ + assert len(inputs) == len(self._input_blobs), ( + f"Length of inputs ({len(inputs)}) " + f"doesn't match the required input blobs: {self._input_blobs}" + ) + + with ScopedWS(self.ws_name, is_reset=False, is_cleanup=False) as ws: + for b, tensor in zip(self._input_blobs, inputs): + ws.FeedBlob(b, tensor) + + try: + ws.RunNet(self.net.Proto().name) + except RuntimeError as e: + if not str(e) in self._error_msgs: + self._error_msgs.add(str(e)) + logger.warning("Encountered new RuntimeError: \n{}".format(str(e))) + logger.warning("Catch the error and use partial results.") + + c2_outputs = [ws.FetchBlob(b) for b in self.net.Proto().external_output] + # Remove outputs of current run, this is necessary in order to + # prevent fetching the result from previous run if the model fails + # in the middle. + for b in self.net.Proto().external_output: + # Needs to create uninitialized blob to make the net runable. + # This is "equivalent" to: ws.RemoveBlob(b) then ws.CreateBlob(b), + # but there'no such API. + ws.FeedBlob(b, f"{b}, a C++ native class of type nullptr (uninitialized).") + + # Cast output to torch.Tensor on the desired device + output_devices = ( + self._infer_output_devices(inputs) + if any(t.device.type != "cpu" for t in inputs) + else ["cpu" for _ in self.net.Proto().external_output] + ) + + outputs = [] + for name, c2_output, device in zip( + self.net.Proto().external_output, c2_outputs, output_devices + ): + if not isinstance(c2_output, np.ndarray): + raise RuntimeError( + "Invalid output for blob {}, received: {}".format(name, c2_output) + ) + outputs.append(torch.tensor(c2_output).to(device=device)) + return tuple(outputs) + + +class ProtobufDetectionModel(torch.nn.Module): + """ + A class works just like a pytorch meta arch in terms of inference, but running + caffe2 model under the hood. + """ + + def __init__(self, predict_net, init_net, *, convert_outputs=None): + """ + Args: + predict_net, init_net (core.Net): caffe2 nets + convert_outptus (callable): a function that converts caffe2 + outputs to the same format of the original pytorch model. + By default, use the one defined in the caffe2 meta_arch. + """ + super().__init__() + self.protobuf_model = ProtobufModel(predict_net, init_net) + self.size_divisibility = get_pb_arg_vali(predict_net, "size_divisibility", 0) + self.device = get_pb_arg_vals(predict_net, "device", b"cpu").decode("ascii") + + if convert_outputs is None: + meta_arch = get_pb_arg_vals(predict_net, "meta_architecture", b"GeneralizedRCNN") + meta_arch = META_ARCH_CAFFE2_EXPORT_TYPE_MAP[meta_arch.decode("ascii")] + self._convert_outputs = meta_arch.get_outputs_converter(predict_net, init_net) + else: + self._convert_outputs = convert_outputs + + def _convert_inputs(self, batched_inputs): + # currently all models convert inputs in the same way + return convert_batched_inputs_to_c2_format( + batched_inputs, self.size_divisibility, self.device + ) + + def forward(self, batched_inputs): + c2_inputs = self._convert_inputs(batched_inputs) + c2_results = self.protobuf_model(c2_inputs) + c2_results = dict(zip(self.protobuf_model.net.Proto().external_output, c2_results)) + return self._convert_outputs(batched_inputs, c2_inputs, c2_results) diff --git a/detectron2/export/caffe2_modeling.py b/detectron2/export/caffe2_modeling.py new file mode 100644 index 0000000000000000000000000000000000000000..7a9fc78164c32f6709245d3a456af19ffde7c497 --- /dev/null +++ b/detectron2/export/caffe2_modeling.py @@ -0,0 +1,503 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import functools +import io +import struct +import types +import torch + +from detectron2.modeling import meta_arch +from detectron2.modeling.box_regression import Box2BoxTransform +from detectron2.modeling.meta_arch.panoptic_fpn import combine_semantic_and_instance_outputs +from detectron2.modeling.meta_arch.retinanet import permute_to_N_HWA_K +from detectron2.modeling.postprocessing import detector_postprocess, sem_seg_postprocess +from detectron2.modeling.roi_heads import keypoint_head +from detectron2.structures import Boxes, ImageList, Instances, RotatedBoxes + +from .c10 import Caffe2Compatible +from .caffe2_patch import ROIHeadsPatcher, patch_generalized_rcnn +from .shared import ( + alias, + check_set_pb_arg, + get_pb_arg_floats, + get_pb_arg_valf, + get_pb_arg_vali, + get_pb_arg_vals, + mock_torch_nn_functional_interpolate, +) + + +def assemble_rcnn_outputs_by_name(image_sizes, tensor_outputs, force_mask_on=False): + """ + A function to assemble caffe2 model's outputs (i.e. Dict[str, Tensor]) + to detectron2's format (i.e. list of Instances instance). + This only works when the model follows the Caffe2 detectron's naming convention. + + Args: + image_sizes (List[List[int, int]]): [H, W] of every image. + tensor_outputs (Dict[str, Tensor]): external_output to its tensor. + + force_mask_on (Bool): if true, the it make sure there'll be pred_masks even + if the mask is not found from tensor_outputs (usually due to model crash) + """ + + results = [Instances(image_size) for image_size in image_sizes] + + batch_splits = tensor_outputs.get("batch_splits", None) + if batch_splits: + raise NotImplementedError() + assert len(image_sizes) == 1 + result = results[0] + + bbox_nms = tensor_outputs["bbox_nms"] + score_nms = tensor_outputs["score_nms"] + class_nms = tensor_outputs["class_nms"] + # Detection will always success because Conv support 0-batch + assert bbox_nms is not None + assert score_nms is not None + assert class_nms is not None + if bbox_nms.shape[1] == 5: + result.pred_boxes = RotatedBoxes(bbox_nms) + else: + result.pred_boxes = Boxes(bbox_nms) + result.scores = score_nms + result.pred_classes = class_nms.to(torch.int64) + + mask_fcn_probs = tensor_outputs.get("mask_fcn_probs", None) + if mask_fcn_probs is not None: + # finish the mask pred + mask_probs_pred = mask_fcn_probs + num_masks = mask_probs_pred.shape[0] + class_pred = result.pred_classes + indices = torch.arange(num_masks, device=class_pred.device) + mask_probs_pred = mask_probs_pred[indices, class_pred][:, None] + result.pred_masks = mask_probs_pred + elif force_mask_on: + # NOTE: there's no way to know the height/width of mask here, it won't be + # used anyway when batch size is 0, so just set them to 0. + result.pred_masks = torch.zeros([0, 1, 0, 0], dtype=torch.uint8) + + keypoints_out = tensor_outputs.get("keypoints_out", None) + kps_score = tensor_outputs.get("kps_score", None) + if keypoints_out is not None: + # keypoints_out: [N, 4, #kypoints], where 4 is in order of (x, y, score, prob) + keypoints_tensor = keypoints_out + # NOTE: it's possible that prob is not calculated if "should_output_softmax" + # is set to False in HeatmapMaxKeypoint, so just using raw score, seems + # it doesn't affect mAP. TODO: check more carefully. + keypoint_xyp = keypoints_tensor.transpose(1, 2)[:, :, [0, 1, 2]] + result.pred_keypoints = keypoint_xyp + elif kps_score is not None: + # keypoint heatmap to sparse data structure + pred_keypoint_logits = kps_score + keypoint_head.keypoint_rcnn_inference(pred_keypoint_logits, [result]) + + return results + + +def _cast_to_f32(f64): + return struct.unpack("f", struct.pack("f", f64))[0] + + +def set_caffe2_compatible_tensor_mode(model, enable=True): + def _fn(m): + if isinstance(m, Caffe2Compatible): + m.tensor_mode = enable + + model.apply(_fn) + + +def convert_batched_inputs_to_c2_format(batched_inputs, size_divisibility, device): + """ + See get_caffe2_inputs() below. + """ + assert all(isinstance(x, dict) for x in batched_inputs) + assert all(x["image"].dim() == 3 for x in batched_inputs) + + images = [x["image"] for x in batched_inputs] + images = ImageList.from_tensors(images, size_divisibility) + + im_info = [] + for input_per_image, image_size in zip(batched_inputs, images.image_sizes): + target_height = input_per_image.get("height", image_size[0]) + target_width = input_per_image.get("width", image_size[1]) # noqa + # NOTE: The scale inside im_info is kept as convention and for providing + # post-processing information if further processing is needed. For + # current Caffe2 model definitions that don't include post-processing inside + # the model, this number is not used. + # NOTE: There can be a slight difference between width and height + # scales, using a single number can results in numerical difference + # compared with D2's post-processing. + scale = target_height / image_size[0] + im_info.append([image_size[0], image_size[1], scale]) + im_info = torch.Tensor(im_info) + + return images.tensor.to(device), im_info.to(device) + + +class Caffe2MetaArch(Caffe2Compatible, torch.nn.Module): + """ + Base class for caffe2-compatible implementation of a meta architecture. + The forward is traceable and its traced graph can be converted to caffe2 + graph through ONNX. + """ + + def __init__(self, cfg, torch_model): + """ + Args: + cfg (CfgNode): + torch_model (nn.Module): the detectron2 model (meta_arch) to be + converted. + """ + super().__init__() + self._wrapped_model = torch_model + self.eval() + set_caffe2_compatible_tensor_mode(self, True) + + def get_caffe2_inputs(self, batched_inputs): + """ + Convert pytorch-style structured inputs to caffe2-style inputs that + are tuples of tensors. + + Args: + batched_inputs (list[dict]): inputs to a detectron2 model + in its standard format. Each dict has "image" (CHW tensor), and optionally + "height" and "width". + + Returns: + tuple[Tensor]: + tuple of tensors that will be the inputs to the + :meth:`forward` method. For existing models, the first + is an NCHW tensor (padded and batched); the second is + a im_info Nx3 tensor, where the rows are + (height, width, unused legacy parameter) + """ + return convert_batched_inputs_to_c2_format( + batched_inputs, + self._wrapped_model.backbone.size_divisibility, + self._wrapped_model.device, + ) + + def encode_additional_info(self, predict_net, init_net): + """ + Save extra metadata that will be used by inference in the output protobuf. + """ + pass + + def forward(self, inputs): + """ + Run the forward in caffe2-style. It has to use caffe2-compatible ops + and the method will be used for tracing. + + Args: + inputs (tuple[Tensor]): inputs defined by :meth:`get_caffe2_input`. + They will be the inputs of the converted caffe2 graph. + + Returns: + tuple[Tensor]: output tensors. They will be the outputs of the + converted caffe2 graph. + """ + raise NotImplementedError + + def _caffe2_preprocess_image(self, inputs): + """ + Caffe2 implementation of preprocess_image, which is called inside each MetaArch's forward. + It normalizes the input images, and the final caffe2 graph assumes the + inputs have been batched already. + """ + data, im_info = inputs + data = alias(data, "data") + im_info = alias(im_info, "im_info") + mean, std = self._wrapped_model.pixel_mean, self._wrapped_model.pixel_std + normalized_data = (data - mean) / std + normalized_data = alias(normalized_data, "normalized_data") + + # Pack (data, im_info) into ImageList which is recognized by self.inference. + images = ImageList(tensor=normalized_data, image_sizes=im_info) + return images + + @staticmethod + def get_outputs_converter(predict_net, init_net): + """ + Creates a function that converts outputs of the caffe2 model to + detectron2's standard format. + The function uses information in `predict_net` and `init_net` that are + available at inferene time. Therefore the function logic can be used in inference. + + The returned function has the following signature: + + def convert(batched_inputs, c2_inputs, c2_results) -> detectron2_outputs + + Where + + * batched_inputs (list[dict]): the original input format of the meta arch + * c2_inputs (tuple[Tensor]): the caffe2 inputs. + * c2_results (dict[str, Tensor]): the caffe2 output format, + corresponding to the outputs of the :meth:`forward` function. + * detectron2_outputs: the original output format of the meta arch. + + This function can be used to compare the outputs of the original meta arch and + the converted caffe2 graph. + + Returns: + callable: a callable of the above signature. + """ + raise NotImplementedError + + +class Caffe2GeneralizedRCNN(Caffe2MetaArch): + def __init__(self, cfg, torch_model): + assert isinstance(torch_model, meta_arch.GeneralizedRCNN) + torch_model = patch_generalized_rcnn(torch_model) + super().__init__(cfg, torch_model) + + self.roi_heads_patcher = ROIHeadsPatcher( + self._wrapped_model.roi_heads, cfg.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT + ) + + def encode_additional_info(self, predict_net, init_net): + size_divisibility = self._wrapped_model.backbone.size_divisibility + check_set_pb_arg(predict_net, "size_divisibility", "i", size_divisibility) + check_set_pb_arg( + predict_net, "device", "s", str.encode(str(self._wrapped_model.device), "ascii") + ) + check_set_pb_arg(predict_net, "meta_architecture", "s", b"GeneralizedRCNN") + + @mock_torch_nn_functional_interpolate() + def forward(self, inputs): + if not self.tensor_mode: + return self._wrapped_model.inference(inputs) + images = self._caffe2_preprocess_image(inputs) + features = self._wrapped_model.backbone(images.tensor) + proposals, _ = self._wrapped_model.proposal_generator(images, features) + with self.roi_heads_patcher.mock_roi_heads(): + detector_results, _ = self._wrapped_model.roi_heads(images, features, proposals) + return tuple(detector_results[0].flatten()) + + @staticmethod + def get_outputs_converter(predict_net, init_net): + def f(batched_inputs, c2_inputs, c2_results): + _, im_info = c2_inputs + image_sizes = [[int(im[0]), int(im[1])] for im in im_info] + results = assemble_rcnn_outputs_by_name(image_sizes, c2_results) + return meta_arch.GeneralizedRCNN._postprocess(results, batched_inputs, image_sizes) + + return f + + +class Caffe2PanopticFPN(Caffe2MetaArch): + def __init__(self, cfg, torch_model): + assert isinstance(torch_model, meta_arch.PanopticFPN) + torch_model = patch_generalized_rcnn(torch_model) + super().__init__(cfg, torch_model) + + self.roi_heads_patcher = ROIHeadsPatcher( + self._wrapped_model.roi_heads, cfg.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT + ) + + @mock_torch_nn_functional_interpolate() + def forward(self, inputs): + assert self.tensor_mode + images = self._caffe2_preprocess_image(inputs) + features = self._wrapped_model.backbone(images.tensor) + + sem_seg_results, _ = self._wrapped_model.sem_seg_head(features) + sem_seg_results = alias(sem_seg_results, "sem_seg") + + proposals, _ = self._wrapped_model.proposal_generator(images, features) + + with self.roi_heads_patcher.mock_roi_heads(self.tensor_mode): + detector_results, _ = self._wrapped_model.roi_heads(images, features, proposals) + + return tuple(detector_results[0].flatten()) + (sem_seg_results,) + + def encode_additional_info(self, predict_net, init_net): + size_divisibility = self._wrapped_model.backbone.size_divisibility + check_set_pb_arg(predict_net, "size_divisibility", "i", size_divisibility) + check_set_pb_arg( + predict_net, "device", "s", str.encode(str(self._wrapped_model.device), "ascii") + ) + check_set_pb_arg(predict_net, "meta_architecture", "s", b"PanopticFPN") + + # Inference parameters: + check_set_pb_arg( + predict_net, + "combine_overlap_threshold", + "f", + _cast_to_f32(self._wrapped_model.combine_overlap_thresh), + ) + check_set_pb_arg( + predict_net, + "combine_stuff_area_limit", + "i", + self._wrapped_model.combine_stuff_area_thresh, + ) + check_set_pb_arg( + predict_net, + "combine_instances_confidence_threshold", + "f", + _cast_to_f32(self._wrapped_model.combine_instances_score_thresh), + ) + + @staticmethod + def get_outputs_converter(predict_net, init_net): + combine_overlap_threshold = get_pb_arg_valf(predict_net, "combine_overlap_threshold", None) + combine_stuff_area_limit = get_pb_arg_vali(predict_net, "combine_stuff_area_limit", None) + combine_instances_confidence_threshold = get_pb_arg_valf( + predict_net, "combine_instances_confidence_threshold", None + ) + + def f(batched_inputs, c2_inputs, c2_results): + _, im_info = c2_inputs + image_sizes = [[int(im[0]), int(im[1])] for im in im_info] + detector_results = assemble_rcnn_outputs_by_name( + image_sizes, c2_results, force_mask_on=True + ) + sem_seg_results = c2_results["sem_seg"] + + # copied from meta_arch/panoptic_fpn.py ... + processed_results = [] + for sem_seg_result, detector_result, input_per_image, image_size in zip( + sem_seg_results, detector_results, batched_inputs, image_sizes + ): + height = input_per_image.get("height", image_size[0]) + width = input_per_image.get("width", image_size[1]) + sem_seg_r = sem_seg_postprocess(sem_seg_result, image_size, height, width) + detector_r = detector_postprocess(detector_result, height, width) + + processed_results.append({"sem_seg": sem_seg_r, "instances": detector_r}) + + panoptic_r = combine_semantic_and_instance_outputs( + detector_r, + sem_seg_r.argmax(dim=0), + combine_overlap_threshold, + combine_stuff_area_limit, + combine_instances_confidence_threshold, + ) + processed_results[-1]["panoptic_seg"] = panoptic_r + return processed_results + + return f + + +class Caffe2RetinaNet(Caffe2MetaArch): + def __init__(self, cfg, torch_model): + assert isinstance(torch_model, meta_arch.RetinaNet) + super().__init__(cfg, torch_model) + + @mock_torch_nn_functional_interpolate() + def forward(self, inputs): + assert self.tensor_mode + images = self._caffe2_preprocess_image(inputs) + + # explicitly return the images sizes to avoid removing "im_info" by ONNX + # since it's not used in the forward path + return_tensors = [images.image_sizes] + + features = self._wrapped_model.backbone(images.tensor) + features = [features[f] for f in self._wrapped_model.head_in_features] + for i, feature_i in enumerate(features): + features[i] = alias(feature_i, "feature_{}".format(i), is_backward=True) + return_tensors.append(features[i]) + + pred_logits, pred_anchor_deltas = self._wrapped_model.head(features) + for i, (box_cls_i, box_delta_i) in enumerate(zip(pred_logits, pred_anchor_deltas)): + return_tensors.append(alias(box_cls_i, "box_cls_{}".format(i))) + return_tensors.append(alias(box_delta_i, "box_delta_{}".format(i))) + + return tuple(return_tensors) + + def encode_additional_info(self, predict_net, init_net): + size_divisibility = self._wrapped_model.backbone.size_divisibility + check_set_pb_arg(predict_net, "size_divisibility", "i", size_divisibility) + check_set_pb_arg( + predict_net, "device", "s", str.encode(str(self._wrapped_model.device), "ascii") + ) + check_set_pb_arg(predict_net, "meta_architecture", "s", b"RetinaNet") + + # Inference parameters: + check_set_pb_arg( + predict_net, "score_threshold", "f", _cast_to_f32(self._wrapped_model.test_score_thresh) + ) + check_set_pb_arg( + predict_net, "topk_candidates", "i", self._wrapped_model.test_topk_candidates + ) + check_set_pb_arg( + predict_net, "nms_threshold", "f", _cast_to_f32(self._wrapped_model.test_nms_thresh) + ) + check_set_pb_arg( + predict_net, + "max_detections_per_image", + "i", + self._wrapped_model.max_detections_per_image, + ) + + check_set_pb_arg( + predict_net, + "bbox_reg_weights", + "floats", + [_cast_to_f32(w) for w in self._wrapped_model.box2box_transform.weights], + ) + self._encode_anchor_generator_cfg(predict_net) + + def _encode_anchor_generator_cfg(self, predict_net): + # serialize anchor_generator for future use + serialized_anchor_generator = io.BytesIO() + torch.save(self._wrapped_model.anchor_generator, serialized_anchor_generator) + # Ideally we can put anchor generating inside the model, then we don't + # need to store this information. + bytes = serialized_anchor_generator.getvalue() + check_set_pb_arg(predict_net, "serialized_anchor_generator", "s", bytes) + + @staticmethod + def get_outputs_converter(predict_net, init_net): + self = types.SimpleNamespace() + serialized_anchor_generator = io.BytesIO( + get_pb_arg_vals(predict_net, "serialized_anchor_generator", None) + ) + self.anchor_generator = torch.load(serialized_anchor_generator) + bbox_reg_weights = get_pb_arg_floats(predict_net, "bbox_reg_weights", None) + self.box2box_transform = Box2BoxTransform(weights=tuple(bbox_reg_weights)) + self.test_score_thresh = get_pb_arg_valf(predict_net, "score_threshold", None) + self.test_topk_candidates = get_pb_arg_vali(predict_net, "topk_candidates", None) + self.test_nms_thresh = get_pb_arg_valf(predict_net, "nms_threshold", None) + self.max_detections_per_image = get_pb_arg_vali( + predict_net, "max_detections_per_image", None + ) + + # hack to reuse inference code from RetinaNet + self.inference = functools.partial(meta_arch.RetinaNet.inference, self) + self.inference_single_image = functools.partial( + meta_arch.RetinaNet.inference_single_image, self + ) + + def f(batched_inputs, c2_inputs, c2_results): + _, im_info = c2_inputs + image_sizes = [[int(im[0]), int(im[1])] for im in im_info] + + num_features = len([x for x in c2_results.keys() if x.startswith("box_cls_")]) + pred_logits = [c2_results["box_cls_{}".format(i)] for i in range(num_features)] + pred_anchor_deltas = [c2_results["box_delta_{}".format(i)] for i in range(num_features)] + + # For each feature level, feature should have the same batch size and + # spatial dimension as the box_cls and box_delta. + dummy_features = [x.clone()[:, 0:0, :, :] for x in pred_logits] + anchors = self.anchor_generator(dummy_features) + + # self.num_classess can be inferred + self.num_classes = pred_logits[0].shape[1] // (pred_anchor_deltas[0].shape[1] // 4) + + pred_logits = [permute_to_N_HWA_K(x, self.num_classes) for x in pred_logits] + pred_anchor_deltas = [permute_to_N_HWA_K(x, 4) for x in pred_anchor_deltas] + + results = self.inference(anchors, pred_logits, pred_anchor_deltas, image_sizes) + return meta_arch.GeneralizedRCNN._postprocess(results, batched_inputs, image_sizes) + + return f + + +META_ARCH_CAFFE2_EXPORT_TYPE_MAP = { + "GeneralizedRCNN": Caffe2GeneralizedRCNN, + "PanopticFPN": Caffe2PanopticFPN, + "RetinaNet": Caffe2RetinaNet, +} diff --git a/detectron2/export/caffe2_patch.py b/detectron2/export/caffe2_patch.py new file mode 100644 index 0000000000000000000000000000000000000000..c9eee594a27cdec29ce5f2b6f7730171eda3805e --- /dev/null +++ b/detectron2/export/caffe2_patch.py @@ -0,0 +1,152 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import contextlib +from unittest import mock +import torch + +from detectron2.modeling import poolers +from detectron2.modeling.proposal_generator import rpn +from detectron2.modeling.roi_heads import keypoint_head, mask_head +from detectron2.modeling.roi_heads.fast_rcnn import FastRCNNOutputLayers + +from .c10 import ( + Caffe2Compatible, + Caffe2FastRCNNOutputsInference, + Caffe2KeypointRCNNInference, + Caffe2MaskRCNNInference, + Caffe2ROIPooler, + Caffe2RPN, +) + + +class GenericMixin(object): + pass + + +class Caffe2CompatibleConverter(object): + """ + A GenericUpdater which implements the `create_from` interface, by modifying + module object and assign it with another class replaceCls. + """ + + def __init__(self, replaceCls): + self.replaceCls = replaceCls + + def create_from(self, module): + # update module's class to the new class + assert isinstance(module, torch.nn.Module) + if issubclass(self.replaceCls, GenericMixin): + # replaceCls should act as mixin, create a new class on-the-fly + new_class = type( + "{}MixedWith{}".format(self.replaceCls.__name__, module.__class__.__name__), + (self.replaceCls, module.__class__), + {}, # {"new_method": lambda self: ...}, + ) + module.__class__ = new_class + else: + # replaceCls is complete class, this allow arbitrary class swap + module.__class__ = self.replaceCls + + # initialize Caffe2Compatible + if isinstance(module, Caffe2Compatible): + module.tensor_mode = False + + return module + + +def patch(model, target, updater, *args, **kwargs): + """ + recursively (post-order) update all modules with the target type and its + subclasses, make a initialization/composition/inheritance/... via the + updater.create_from. + """ + for name, module in model.named_children(): + model._modules[name] = patch(module, target, updater, *args, **kwargs) + if isinstance(model, target): + return updater.create_from(model, *args, **kwargs) + return model + + +def patch_generalized_rcnn(model): + ccc = Caffe2CompatibleConverter + model = patch(model, rpn.RPN, ccc(Caffe2RPN)) + model = patch(model, poolers.ROIPooler, ccc(Caffe2ROIPooler)) + + return model + + +@contextlib.contextmanager +def mock_fastrcnn_outputs_inference( + tensor_mode, check=True, box_predictor_type=FastRCNNOutputLayers +): + with mock.patch.object( + box_predictor_type, + "inference", + autospec=True, + side_effect=Caffe2FastRCNNOutputsInference(tensor_mode), + ) as mocked_func: + yield + if check: + assert mocked_func.call_count > 0 + + +@contextlib.contextmanager +def mock_mask_rcnn_inference(tensor_mode, patched_module, check=True): + with mock.patch( + "{}.mask_rcnn_inference".format(patched_module), side_effect=Caffe2MaskRCNNInference() + ) as mocked_func: + yield + if check: + assert mocked_func.call_count > 0 + + +@contextlib.contextmanager +def mock_keypoint_rcnn_inference(tensor_mode, patched_module, use_heatmap_max_keypoint, check=True): + with mock.patch( + "{}.keypoint_rcnn_inference".format(patched_module), + side_effect=Caffe2KeypointRCNNInference(use_heatmap_max_keypoint), + ) as mocked_func: + yield + if check: + assert mocked_func.call_count > 0 + + +class ROIHeadsPatcher: + def __init__(self, heads, use_heatmap_max_keypoint): + self.heads = heads + self.use_heatmap_max_keypoint = use_heatmap_max_keypoint + + @contextlib.contextmanager + def mock_roi_heads(self, tensor_mode=True): + """ + Patching several inference functions inside ROIHeads and its subclasses + + Args: + tensor_mode (bool): whether the inputs/outputs are caffe2's tensor + format or not. Default to True. + """ + # NOTE: this requries the `keypoint_rcnn_inference` and `mask_rcnn_inference` + # are called inside the same file as BaseXxxHead due to using mock.patch. + kpt_heads_mod = keypoint_head.BaseKeypointRCNNHead.__module__ + mask_head_mod = mask_head.BaseMaskRCNNHead.__module__ + + mock_ctx_managers = [ + mock_fastrcnn_outputs_inference( + tensor_mode=tensor_mode, + check=True, + box_predictor_type=type(self.heads.box_predictor), + ) + ] + if getattr(self.heads, "keypoint_on", False): + mock_ctx_managers += [ + mock_keypoint_rcnn_inference( + tensor_mode, kpt_heads_mod, self.use_heatmap_max_keypoint + ) + ] + if getattr(self.heads, "mask_on", False): + mock_ctx_managers += [mock_mask_rcnn_inference(tensor_mode, mask_head_mod)] + + with contextlib.ExitStack() as stack: # python 3.3+ + for mgr in mock_ctx_managers: + stack.enter_context(mgr) + yield diff --git a/detectron2/export/flatten.py b/detectron2/export/flatten.py new file mode 100644 index 0000000000000000000000000000000000000000..5d229719b56bf3b57727f3751dbb9af1b6d173f1 --- /dev/null +++ b/detectron2/export/flatten.py @@ -0,0 +1,327 @@ +import collections +from dataclasses import dataclass +from typing import Callable, List, Optional, Tuple +import torch +from torch import nn + +from detectron2.structures import Boxes, Instances, ROIMasks +from detectron2.utils.registry import _convert_target_to_string, locate + +from .torchscript_patch import patch_builtin_len + + +@dataclass +class Schema: + """ + A Schema defines how to flatten a possibly hierarchical object into tuple of + primitive objects, so it can be used as inputs/outputs of PyTorch's tracing. + + PyTorch does not support tracing a function that produces rich output + structures (e.g. dict, Instances, Boxes). To trace such a function, we + flatten the rich object into tuple of tensors, and return this tuple of tensors + instead. Meanwhile, we also need to know how to "rebuild" the original object + from the flattened results, so we can evaluate the flattened results. + A Schema defines how to flatten an object, and while flattening it, it records + necessary schemas so that the object can be rebuilt using the flattened outputs. + + The flattened object and the schema object is returned by ``.flatten`` classmethod. + Then the original object can be rebuilt with the ``__call__`` method of schema. + + A Schema is a dataclass that can be serialized easily. + """ + + # inspired by FetchMapper in tensorflow/python/client/session.py + + @classmethod + def flatten(cls, obj): + raise NotImplementedError + + def __call__(self, values): + raise NotImplementedError + + @staticmethod + def _concat(values): + ret = () + sizes = [] + for v in values: + assert isinstance(v, tuple), "Flattened results must be a tuple" + ret = ret + v + sizes.append(len(v)) + return ret, sizes + + @staticmethod + def _split(values, sizes): + if len(sizes): + expected_len = sum(sizes) + assert ( + len(values) == expected_len + ), f"Values has length {len(values)} but expect length {expected_len}." + ret = [] + for k in range(len(sizes)): + begin, end = sum(sizes[:k]), sum(sizes[: k + 1]) + ret.append(values[begin:end]) + return ret + + +@dataclass +class ListSchema(Schema): + schemas: List[Schema] # the schemas that define how to flatten each element in the list + sizes: List[int] # the flattened length of each element + + def __call__(self, values): + values = self._split(values, self.sizes) + if len(values) != len(self.schemas): + raise ValueError( + f"Values has length {len(values)} but schemas " f"has length {len(self.schemas)}!" + ) + values = [m(v) for m, v in zip(self.schemas, values)] + return list(values) + + @classmethod + def flatten(cls, obj): + res = [flatten_to_tuple(k) for k in obj] + values, sizes = cls._concat([k[0] for k in res]) + return values, cls([k[1] for k in res], sizes) + + +@dataclass +class TupleSchema(ListSchema): + def __call__(self, values): + return tuple(super().__call__(values)) + + +@dataclass +class IdentitySchema(Schema): + def __call__(self, values): + return values[0] + + @classmethod + def flatten(cls, obj): + return (obj,), cls() + + +@dataclass +class DictSchema(ListSchema): + keys: List[str] + + def __call__(self, values): + values = super().__call__(values) + return dict(zip(self.keys, values)) + + @classmethod + def flatten(cls, obj): + for k in obj.keys(): + if not isinstance(k, str): + raise KeyError("Only support flattening dictionaries if keys are str.") + keys = sorted(obj.keys()) + values = [obj[k] for k in keys] + ret, schema = ListSchema.flatten(values) + return ret, cls(schema.schemas, schema.sizes, keys) + + +@dataclass +class InstancesSchema(DictSchema): + def __call__(self, values): + image_size, fields = values[-1], values[:-1] + fields = super().__call__(fields) + return Instances(image_size, **fields) + + @classmethod + def flatten(cls, obj): + ret, schema = super().flatten(obj.get_fields()) + size = obj.image_size + if not isinstance(size, torch.Tensor): + size = torch.tensor(size) + return ret + (size,), schema + + +@dataclass +class TensorWrapSchema(Schema): + """ + For classes that are simple wrapper of tensors, e.g. + Boxes, RotatedBoxes, BitMasks + """ + + class_name: str + + def __call__(self, values): + return locate(self.class_name)(values[0]) + + @classmethod + def flatten(cls, obj): + return (obj.tensor,), cls(_convert_target_to_string(type(obj))) + + +# if more custom structures needed in the future, can allow +# passing in extra schemas for custom types +def flatten_to_tuple(obj): + """ + Flatten an object so it can be used for PyTorch tracing. + Also returns how to rebuild the original object from the flattened outputs. + + Returns: + res (tuple): the flattened results that can be used as tracing outputs + schema: an object with a ``__call__`` method such that ``schema(res) == obj``. + It is a pure dataclass that can be serialized. + """ + schemas = [ + ((str, bytes), IdentitySchema), + (list, ListSchema), + (tuple, TupleSchema), + (collections.abc.Mapping, DictSchema), + (Instances, InstancesSchema), + ((Boxes, ROIMasks), TensorWrapSchema), + ] + for klass, schema in schemas: + if isinstance(obj, klass): + F = schema + break + else: + F = IdentitySchema + + return F.flatten(obj) + + +class TracingAdapter(nn.Module): + """ + A model may take rich input/output format (e.g. dict or custom classes), + but `torch.jit.trace` requires tuple of tensors as input/output. + This adapter flattens input/output format of a model so it becomes traceable. + + It also records the necessary schema to rebuild model's inputs/outputs from flattened + inputs/outputs. + + Example: + :: + outputs = model(inputs) # inputs/outputs may be rich structure + adapter = TracingAdapter(model, inputs) + + # can now trace the model, with adapter.flattened_inputs, or another + # tuple of tensors with the same length and meaning + traced = torch.jit.trace(adapter, adapter.flattened_inputs) + + # traced model can only produce flattened outputs (tuple of tensors) + flattened_outputs = traced(*adapter.flattened_inputs) + # adapter knows the schema to convert it back (new_outputs == outputs) + new_outputs = adapter.outputs_schema(flattened_outputs) + """ + + flattened_inputs: Tuple[torch.Tensor] = None + """ + Flattened version of inputs given to this class's constructor. + """ + + inputs_schema: Schema = None + """ + Schema of the inputs given to this class's constructor. + """ + + outputs_schema: Schema = None + """ + Schema of the output produced by calling the given model with inputs. + """ + + def __init__( + self, + model: nn.Module, + inputs, + inference_func: Optional[Callable] = None, + allow_non_tensor: bool = False, + ): + """ + Args: + model: an nn.Module + inputs: An input argument or a tuple of input arguments used to call model. + After flattening, it has to only consist of tensors. + inference_func: a callable that takes (model, *inputs), calls the + model with inputs, and return outputs. By default it + is ``lambda model, *inputs: model(*inputs)``. Can be override + if you need to call the model differently. + allow_non_tensor: allow inputs/outputs to contain non-tensor objects. + This option will filter out non-tensor objects to make the + model traceable, but ``inputs_schema``/``outputs_schema`` cannot be + used anymore because inputs/outputs cannot be rebuilt from pure tensors. + This is useful when you're only interested in the single trace of + execution (e.g. for flop count), but not interested in + generalizing the traced graph to new inputs. + """ + super().__init__() + if isinstance(model, (nn.parallel.distributed.DistributedDataParallel, nn.DataParallel)): + model = model.module + self.model = model + if not isinstance(inputs, tuple): + inputs = (inputs,) + self.inputs = inputs + self.allow_non_tensor = allow_non_tensor + + if inference_func is None: + inference_func = lambda model, *inputs: model(*inputs) # noqa + self.inference_func = inference_func + + self.flattened_inputs, self.inputs_schema = flatten_to_tuple(inputs) + + if all(isinstance(x, torch.Tensor) for x in self.flattened_inputs): + return + if self.allow_non_tensor: + self.flattened_inputs = tuple( + [x for x in self.flattened_inputs if isinstance(x, torch.Tensor)] + ) + self.inputs_schema = None + else: + for input in self.flattened_inputs: + if not isinstance(input, torch.Tensor): + raise ValueError( + "Inputs for tracing must only contain tensors. " + f"Got a {type(input)} instead." + ) + + def forward(self, *args: torch.Tensor): + with torch.no_grad(), patch_builtin_len(): + if self.inputs_schema is not None: + inputs_orig_format = self.inputs_schema(args) + else: + if args != self.flattened_inputs: + raise ValueError( + "TracingAdapter does not contain valid inputs_schema." + " So it cannot generalize to other inputs and must be" + " traced with `.flattened_inputs`." + ) + inputs_orig_format = self.inputs + + outputs = self.inference_func(self.model, *inputs_orig_format) + flattened_outputs, schema = flatten_to_tuple(outputs) + + flattened_output_tensors = tuple( + [x for x in flattened_outputs if isinstance(x, torch.Tensor)] + ) + if len(flattened_output_tensors) < len(flattened_outputs): + if self.allow_non_tensor: + flattened_outputs = flattened_output_tensors + self.outputs_schema = None + else: + raise ValueError( + "Model cannot be traced because some model outputs " + "cannot flatten to tensors." + ) + else: # schema is valid + if self.outputs_schema is None: + self.outputs_schema = schema + else: + assert self.outputs_schema == schema, ( + "Model should always return outputs with the same " + "structure so it can be traced!" + ) + return flattened_outputs + + def _create_wrapper(self, traced_model): + """ + Return a function that has an input/output interface the same as the + original model, but it calls the given traced model under the hood. + """ + + def forward(*args): + flattened_inputs, _ = flatten_to_tuple(args) + flattened_outputs = traced_model(*flattened_inputs) + return self.outputs_schema(flattened_outputs) + + return forward diff --git a/detectron2/export/shared.py b/detectron2/export/shared.py new file mode 100644 index 0000000000000000000000000000000000000000..2d0f7bf3999064a68f28a1207d65a2de7ae98c0a --- /dev/null +++ b/detectron2/export/shared.py @@ -0,0 +1,1034 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import collections +import contextlib +import copy +import functools +import logging +import numpy as np +import os +from typing import Any, Callable, Dict, List, Optional, Tuple, Union +from unittest import mock +import caffe2.python.utils as putils +import torch +import torch.nn.functional as F +from caffe2.proto import caffe2_pb2 +from caffe2.python import core, net_drawer, workspace +from torch.nn.functional import interpolate as interp + +logger = logging.getLogger(__name__) + + +# ==== torch/utils_toffee/cast.py ======================================= + + +def to_device(t, device_str): + """ + This function is a replacement of .to(another_device) such that it allows the + casting to be traced properly by explicitly calling the underlying copy ops. + It also avoids introducing unncessary op when casting to the same device. + """ + src = t.device + dst = torch.device(device_str) + + if src == dst: + return t + elif src.type == "cuda" and dst.type == "cpu": + return torch.ops._caffe2.CopyGPUToCPU(t) + elif src.type == "cpu" and dst.type == "cuda": + return torch.ops._caffe2.CopyCPUToGPU(t) + else: + raise RuntimeError("Can't cast tensor from device {} to device {}".format(src, dst)) + + +# ==== torch/utils_toffee/interpolate.py ======================================= + + +# Note: borrowed from vision/detection/fair/detectron/detectron/modeling/detector.py +def BilinearInterpolation(tensor_in, up_scale): + assert up_scale % 2 == 0, "Scale should be even" + + def upsample_filt(size): + factor = (size + 1) // 2 + if size % 2 == 1: + center = factor - 1 + else: + center = factor - 0.5 + + og = np.ogrid[:size, :size] + return (1 - abs(og[0] - center) / factor) * (1 - abs(og[1] - center) / factor) + + kernel_size = int(up_scale) * 2 + bil_filt = upsample_filt(kernel_size) + + dim = int(tensor_in.shape[1]) + kernel = np.zeros((dim, dim, kernel_size, kernel_size), dtype=np.float32) + kernel[range(dim), range(dim), :, :] = bil_filt + + tensor_out = F.conv_transpose2d( + tensor_in, + weight=to_device(torch.Tensor(kernel), tensor_in.device), + bias=None, + stride=int(up_scale), + padding=int(up_scale / 2), + ) + + return tensor_out + + +# NOTE: ONNX is incompatible with traced torch.nn.functional.interpolate if +# using dynamic `scale_factor` rather than static `size`. (T43166860) +# NOTE: Caffe2 Int8 conversion might not be able to quantize `size` properly. +def onnx_compatibale_interpolate( + input, size=None, scale_factor=None, mode="nearest", align_corners=None +): + # NOTE: The input dimensions are interpreted in the form: + # `mini-batch x channels x [optional depth] x [optional height] x width`. + if size is None and scale_factor is not None: + if input.dim() == 4: + if isinstance(scale_factor, (int, float)): + height_scale, width_scale = (scale_factor, scale_factor) + else: + assert isinstance(scale_factor, (tuple, list)) + assert len(scale_factor) == 2 + height_scale, width_scale = scale_factor + + assert not align_corners, "No matching C2 op for align_corners == True" + if mode == "nearest": + return torch.ops._caffe2.ResizeNearest( + input, order="NCHW", width_scale=width_scale, height_scale=height_scale + ) + elif mode == "bilinear": + logger.warning( + "Use F.conv_transpose2d for bilinear interpolate" + " because there's no such C2 op, this may cause significant" + " slowdown and the boundary pixels won't be as same as" + " using F.interpolate due to padding." + ) + assert height_scale == width_scale + return BilinearInterpolation(input, up_scale=height_scale) + logger.warning("Output size is not static, it might cause ONNX conversion issue") + + return interp(input, size, scale_factor, mode, align_corners) + + +@contextlib.contextmanager +def mock_torch_nn_functional_interpolate(): + if torch.onnx.is_in_onnx_export(): + with mock.patch( + "torch.nn.functional.interpolate", side_effect=onnx_compatibale_interpolate + ): + yield + else: + yield + + +# ==== torch/utils_caffe2/ws_utils.py ========================================== + + +class ScopedWS(object): + def __init__(self, ws_name, is_reset, is_cleanup=False): + self.ws_name = ws_name + self.is_reset = is_reset + self.is_cleanup = is_cleanup + self.org_ws = "" + + def __enter__(self): + self.org_ws = workspace.CurrentWorkspace() + if self.ws_name is not None: + workspace.SwitchWorkspace(self.ws_name, True) + if self.is_reset: + workspace.ResetWorkspace() + + return workspace + + def __exit__(self, *args): + if self.is_cleanup: + workspace.ResetWorkspace() + if self.ws_name is not None: + workspace.SwitchWorkspace(self.org_ws) + + +def fetch_any_blob(name): + bb = None + try: + bb = workspace.FetchBlob(name) + except TypeError: + bb = workspace.FetchInt8Blob(name) + except Exception as e: + logger.error("Get blob {} error: {}".format(name, e)) + + return bb + + +# ==== torch/utils_caffe2/protobuf.py ========================================== + + +def get_pb_arg(pb, arg_name): + for x in pb.arg: + if x.name == arg_name: + return x + return None + + +def get_pb_arg_valf(pb, arg_name, default_val): + arg = get_pb_arg(pb, arg_name) + return arg.f if arg is not None else default_val + + +def get_pb_arg_floats(pb, arg_name, default_val): + arg = get_pb_arg(pb, arg_name) + return list(map(float, arg.floats)) if arg is not None else default_val + + +def get_pb_arg_ints(pb, arg_name, default_val): + arg = get_pb_arg(pb, arg_name) + return list(map(int, arg.ints)) if arg is not None else default_val + + +def get_pb_arg_vali(pb, arg_name, default_val): + arg = get_pb_arg(pb, arg_name) + return arg.i if arg is not None else default_val + + +def get_pb_arg_vals(pb, arg_name, default_val): + arg = get_pb_arg(pb, arg_name) + return arg.s if arg is not None else default_val + + +def get_pb_arg_valstrings(pb, arg_name, default_val): + arg = get_pb_arg(pb, arg_name) + return list(arg.strings) if arg is not None else default_val + + +def check_set_pb_arg(pb, arg_name, arg_attr, arg_value, allow_override=False): + arg = get_pb_arg(pb, arg_name) + if arg is None: + arg = putils.MakeArgument(arg_name, arg_value) + assert hasattr(arg, arg_attr) + pb.arg.extend([arg]) + if allow_override and getattr(arg, arg_attr) != arg_value: + logger.warning( + "Override argument {}: {} -> {}".format(arg_name, getattr(arg, arg_attr), arg_value) + ) + setattr(arg, arg_attr, arg_value) + else: + assert arg is not None + assert getattr(arg, arg_attr) == arg_value, "Existing value {}, new value {}".format( + getattr(arg, arg_attr), arg_value + ) + + +def _create_const_fill_op_from_numpy(name, tensor, device_option=None): + assert type(tensor) == np.ndarray + kTypeNameMapper = { + np.dtype("float32"): "GivenTensorFill", + np.dtype("int32"): "GivenTensorIntFill", + np.dtype("int64"): "GivenTensorInt64Fill", + np.dtype("uint8"): "GivenTensorStringFill", + } + + args_dict = {} + if tensor.dtype == np.dtype("uint8"): + args_dict.update({"values": [str(tensor.data)], "shape": [1]}) + else: + args_dict.update({"values": tensor, "shape": tensor.shape}) + + if device_option is not None: + args_dict["device_option"] = device_option + + return core.CreateOperator(kTypeNameMapper[tensor.dtype], [], [name], **args_dict) + + +def _create_const_fill_op_from_c2_int8_tensor(name, int8_tensor): + assert type(int8_tensor) == workspace.Int8Tensor + kTypeNameMapper = { + np.dtype("int32"): "Int8GivenIntTensorFill", + np.dtype("uint8"): "Int8GivenTensorFill", + } + + tensor = int8_tensor.data + assert tensor.dtype in [np.dtype("uint8"), np.dtype("int32")] + values = tensor.tobytes() if tensor.dtype == np.dtype("uint8") else tensor + + return core.CreateOperator( + kTypeNameMapper[tensor.dtype], + [], + [name], + values=values, + shape=tensor.shape, + Y_scale=int8_tensor.scale, + Y_zero_point=int8_tensor.zero_point, + ) + + +def create_const_fill_op( + name: str, + blob: Union[np.ndarray, workspace.Int8Tensor], + device_option: Optional[caffe2_pb2.DeviceOption] = None, +) -> caffe2_pb2.OperatorDef: + """ + Given a blob object, return the Caffe2 operator that creates this blob + as constant. Currently support NumPy tensor and Caffe2 Int8Tensor. + """ + + tensor_type = type(blob) + assert tensor_type in [ + np.ndarray, + workspace.Int8Tensor, + ], 'Error when creating const fill op for "{}", unsupported blob type: {}'.format( + name, type(blob) + ) + + if tensor_type == np.ndarray: + return _create_const_fill_op_from_numpy(name, blob, device_option) + elif tensor_type == workspace.Int8Tensor: + assert device_option is None + return _create_const_fill_op_from_c2_int8_tensor(name, blob) + + +def construct_init_net_from_params( + params: Dict[str, Any], device_options: Optional[Dict[str, caffe2_pb2.DeviceOption]] = None +) -> caffe2_pb2.NetDef: + """ + Construct the init_net from params dictionary + """ + init_net = caffe2_pb2.NetDef() + device_options = device_options or {} + for name, blob in params.items(): + if isinstance(blob, str): + logger.warning( + ( + "Blob {} with type {} is not supported in generating init net," + " skipped.".format(name, type(blob)) + ) + ) + continue + init_net.op.extend( + [create_const_fill_op(name, blob, device_option=device_options.get(name, None))] + ) + init_net.external_output.append(name) + return init_net + + +def get_producer_map(ssa): + """ + Return dict from versioned blob to (i, j), + where i is index of producer op, j is the index of output of that op. + """ + producer_map = {} + for i in range(len(ssa)): + outputs = ssa[i][1] + for j, outp in enumerate(outputs): + producer_map[outp] = (i, j) + return producer_map + + +def get_consumer_map(ssa): + """ + Return dict from versioned blob to list of (i, j), + where i is index of consumer op, j is the index of input of that op. + """ + consumer_map = collections.defaultdict(list) + for i in range(len(ssa)): + inputs = ssa[i][0] + for j, inp in enumerate(inputs): + consumer_map[inp].append((i, j)) + return consumer_map + + +def get_params_from_init_net( + init_net: caffe2_pb2.NetDef, +) -> [Dict[str, Any], Dict[str, caffe2_pb2.DeviceOption]]: + """ + Take the output blobs from init_net by running it. + Outputs: + params: dict from blob name to numpy array + device_options: dict from blob name to the device option of its creating op + """ + # NOTE: this assumes that the params is determined by producer op with the + # only exception be CopyGPUToCPU which is CUDA op but returns CPU tensor. + def _get_device_option(producer_op): + if producer_op.type == "CopyGPUToCPU": + return caffe2_pb2.DeviceOption() + else: + return producer_op.device_option + + with ScopedWS("__get_params_from_init_net__", is_reset=True, is_cleanup=True) as ws: + ws.RunNetOnce(init_net) + params = {b: fetch_any_blob(b) for b in init_net.external_output} + ssa, versions = core.get_ssa(init_net) + producer_map = get_producer_map(ssa) + device_options = { + b: _get_device_option(init_net.op[producer_map[(b, versions[b])][0]]) + for b in init_net.external_output + } + return params, device_options + + +def _updater_raise(op, input_types, output_types): + raise RuntimeError( + "Failed to apply updater for op {} given input_types {} and" + " output_types {}".format(op, input_types, output_types) + ) + + +def _generic_status_identifier( + predict_net: caffe2_pb2.NetDef, + status_updater: Callable, + known_status: Dict[Tuple[str, int], Any], +) -> Dict[Tuple[str, int], Any]: + """ + Statically infer the status of each blob, the status can be such as device type + (CPU/GPU), layout (NCHW/NHWC), data type (float32/int8), etc. "Blob" here + is versioned blob (Tuple[str, int]) in the format compatible with ssa. + Inputs: + predict_net: the caffe2 network + status_updater: a callable, given an op and the status of its input/output, + it returns the updated status of input/output. `None` is used for + representing unknown status. + known_status: a dict containing known status, used as initialization. + Outputs: + A dict mapping from versioned blob to its status + """ + ssa, versions = core.get_ssa(predict_net) + versioned_ext_input = [(b, 0) for b in predict_net.external_input] + versioned_ext_output = [(b, versions[b]) for b in predict_net.external_output] + all_versioned_blobs = set().union(*[set(x[0] + x[1]) for x in ssa]) + + allowed_vbs = all_versioned_blobs.union(versioned_ext_input).union(versioned_ext_output) + assert all(k in allowed_vbs for k in known_status) + assert all(v is not None for v in known_status.values()) + _known_status = copy.deepcopy(known_status) + + def _check_and_update(key, value): + assert value is not None + if key in _known_status: + if not _known_status[key] == value: + raise RuntimeError( + "Confilict status for {}, existing status {}, new status {}".format( + key, _known_status[key], value + ) + ) + _known_status[key] = value + + def _update_i(op, ssa_i): + versioned_inputs = ssa_i[0] + versioned_outputs = ssa_i[1] + + inputs_status = [_known_status.get(b, None) for b in versioned_inputs] + outputs_status = [_known_status.get(b, None) for b in versioned_outputs] + + new_inputs_status, new_outputs_status = status_updater(op, inputs_status, outputs_status) + + for versioned_blob, status in zip( + versioned_inputs + versioned_outputs, new_inputs_status + new_outputs_status + ): + if status is not None: + _check_and_update(versioned_blob, status) + + for op, ssa_i in zip(predict_net.op, ssa): + _update_i(op, ssa_i) + for op, ssa_i in zip(reversed(predict_net.op), reversed(ssa)): + _update_i(op, ssa_i) + + # NOTE: This strictly checks all the blob from predict_net must be assgined + # a known status. However sometimes it's impossible (eg. having deadend op), + # we may relax this constraint if + for k in all_versioned_blobs: + if k not in _known_status: + raise NotImplementedError( + "Can not infer the status for {}. Currently only support the case where" + " a single forward and backward pass can identify status for all blobs.".format(k) + ) + + return _known_status + + +def infer_device_type( + predict_net: caffe2_pb2.NetDef, + known_status: Dict[Tuple[str, int], Any], + device_name_style: str = "caffe2", +) -> Dict[Tuple[str, int], str]: + """Return the device type ("cpu" or "gpu"/"cuda") of each (versioned) blob""" + + assert device_name_style in ["caffe2", "pytorch"] + _CPU_STR = "cpu" + _GPU_STR = "gpu" if device_name_style == "caffe2" else "cuda" + + def _copy_cpu_to_gpu_updater(op, input_types, output_types): + if input_types[0] == _GPU_STR or output_types[0] == _CPU_STR: + _updater_raise(op, input_types, output_types) + return ([_CPU_STR], [_GPU_STR]) + + def _copy_gpu_to_cpu_updater(op, input_types, output_types): + if input_types[0] == _CPU_STR or output_types[0] == _GPU_STR: + _updater_raise(op, input_types, output_types) + return ([_GPU_STR], [_CPU_STR]) + + def _other_ops_updater(op, input_types, output_types): + non_none_types = [x for x in input_types + output_types if x is not None] + if len(non_none_types) > 0: + the_type = non_none_types[0] + if not all(x == the_type for x in non_none_types): + _updater_raise(op, input_types, output_types) + else: + the_type = None + return ([the_type for _ in op.input], [the_type for _ in op.output]) + + def _device_updater(op, *args, **kwargs): + return { + "CopyCPUToGPU": _copy_cpu_to_gpu_updater, + "CopyGPUToCPU": _copy_gpu_to_cpu_updater, + }.get(op.type, _other_ops_updater)(op, *args, **kwargs) + + return _generic_status_identifier(predict_net, _device_updater, known_status) + + +# ==== torch/utils_caffe2/vis.py =============================================== + + +def _modify_blob_names(ops, blob_rename_f): + ret = [] + + def _replace_list(blob_list, replaced_list): + del blob_list[:] + blob_list.extend(replaced_list) + + for x in ops: + cur = copy.deepcopy(x) + _replace_list(cur.input, list(map(blob_rename_f, cur.input))) + _replace_list(cur.output, list(map(blob_rename_f, cur.output))) + ret.append(cur) + + return ret + + +def _rename_blob(name, blob_sizes, blob_ranges): + def _list_to_str(bsize): + ret = ", ".join([str(x) for x in bsize]) + ret = "[" + ret + "]" + return ret + + ret = name + if blob_sizes is not None and name in blob_sizes: + ret += "\n" + _list_to_str(blob_sizes[name]) + if blob_ranges is not None and name in blob_ranges: + ret += "\n" + _list_to_str(blob_ranges[name]) + + return ret + + +# graph_name could not contain word 'graph' +def save_graph(net, file_name, graph_name="net", op_only=True, blob_sizes=None, blob_ranges=None): + blob_rename_f = functools.partial(_rename_blob, blob_sizes=blob_sizes, blob_ranges=blob_ranges) + return save_graph_base(net, file_name, graph_name, op_only, blob_rename_f) + + +def save_graph_base(net, file_name, graph_name="net", op_only=True, blob_rename_func=None): + graph = None + ops = net.op + if blob_rename_func is not None: + ops = _modify_blob_names(ops, blob_rename_func) + if not op_only: + graph = net_drawer.GetPydotGraph(ops, graph_name, rankdir="TB") + else: + graph = net_drawer.GetPydotGraphMinimal( + ops, graph_name, rankdir="TB", minimal_dependency=True + ) + + try: + par_dir = os.path.dirname(file_name) + if not os.path.exists(par_dir): + os.makedirs(par_dir) + + format = os.path.splitext(os.path.basename(file_name))[-1] + if format == ".png": + graph.write_png(file_name) + elif format == ".pdf": + graph.write_pdf(file_name) + elif format == ".svg": + graph.write_svg(file_name) + else: + print("Incorrect format {}".format(format)) + except Exception as e: + print("Error when writing graph to image {}".format(e)) + + return graph + + +# ==== torch/utils_toffee/aten_to_caffe2.py ==================================== + + +def group_norm_replace_aten_with_caffe2(predict_net: caffe2_pb2.NetDef): + """ + For ONNX exported model, GroupNorm will be represented as ATen op, + this can be a drop in replacement from ATen to GroupNorm + """ + count = 0 + for op in predict_net.op: + if op.type == "ATen": + op_name = get_pb_arg_vals(op, "operator", None) # return byte in py3 + if op_name and op_name.decode() == "group_norm": + op.arg.remove(get_pb_arg(op, "operator")) + + if get_pb_arg_vali(op, "cudnn_enabled", None): + op.arg.remove(get_pb_arg(op, "cudnn_enabled")) + + num_groups = get_pb_arg_vali(op, "num_groups", None) + if num_groups is not None: + op.arg.remove(get_pb_arg(op, "num_groups")) + check_set_pb_arg(op, "group", "i", num_groups) + + op.type = "GroupNorm" + count += 1 + if count > 1: + logger.info("Replaced {} ATen operator to GroupNormOp".format(count)) + + +# ==== torch/utils_toffee/alias.py ============================================= + + +def alias(x, name, is_backward=False): + if not torch.onnx.is_in_onnx_export(): + return x + assert isinstance(x, torch.Tensor) + return torch.ops._caffe2.AliasWithName(x, name, is_backward=is_backward) + + +def fuse_alias_placeholder(predict_net, init_net): + """Remove AliasWithName placeholder and rename the input/output of it""" + # First we finish all the re-naming + for i, op in enumerate(predict_net.op): + if op.type == "AliasWithName": + assert len(op.input) == 1 + assert len(op.output) == 1 + name = get_pb_arg_vals(op, "name", None).decode() + is_backward = bool(get_pb_arg_vali(op, "is_backward", 0)) + rename_op_input(predict_net, init_net, i, 0, name, from_producer=is_backward) + rename_op_output(predict_net, i, 0, name) + + # Remove AliasWithName, should be very safe since it's a non-op + new_ops = [] + for op in predict_net.op: + if op.type != "AliasWithName": + new_ops.append(op) + else: + # safety check + assert op.input == op.output + assert op.input[0] == op.arg[0].s.decode() + del predict_net.op[:] + predict_net.op.extend(new_ops) + + +# ==== torch/utils_caffe2/graph_transform.py =================================== + + +class IllegalGraphTransformError(ValueError): + """When a graph transform function call can't be executed.""" + + +def _rename_versioned_blob_in_proto( + proto: caffe2_pb2.NetDef, + old_name: str, + new_name: str, + version: int, + ssa: List[Tuple[List[Tuple[str, int]], List[Tuple[str, int]]]], + start_versions: Dict[str, int], + end_versions: Dict[str, int], +): + """In given proto, rename all blobs with matched version""" + # Operater list + for op, i_th_ssa in zip(proto.op, ssa): + versioned_inputs, versioned_outputs = i_th_ssa + for i in range(len(op.input)): + if versioned_inputs[i] == (old_name, version): + op.input[i] = new_name + for i in range(len(op.output)): + if versioned_outputs[i] == (old_name, version): + op.output[i] = new_name + # external_input + if start_versions.get(old_name, 0) == version: + for i in range(len(proto.external_input)): + if proto.external_input[i] == old_name: + proto.external_input[i] = new_name + # external_output + if end_versions.get(old_name, 0) == version: + for i in range(len(proto.external_output)): + if proto.external_output[i] == old_name: + proto.external_output[i] = new_name + + +def rename_op_input( + predict_net: caffe2_pb2.NetDef, + init_net: caffe2_pb2.NetDef, + op_id: int, + input_id: int, + new_name: str, + from_producer: bool = False, +): + """ + Rename the op_id-th operator in predict_net, change it's input_id-th input's + name to the new_name. It also does automatic re-route and change + external_input and init_net if necessary. + - It requires the input is only consumed by this op. + - This function modifies predict_net and init_net in-place. + - When from_producer is enable, this also updates other operators that consumes + the same input. Be cautious because may trigger unintended behavior. + """ + assert isinstance(predict_net, caffe2_pb2.NetDef) + assert isinstance(init_net, caffe2_pb2.NetDef) + + init_net_ssa, init_net_versions = core.get_ssa(init_net) + predict_net_ssa, predict_net_versions = core.get_ssa( + predict_net, copy.deepcopy(init_net_versions) + ) + + versioned_inputs, versioned_outputs = predict_net_ssa[op_id] + old_name, version = versioned_inputs[input_id] + + if from_producer: + producer_map = get_producer_map(predict_net_ssa) + if not (old_name, version) in producer_map: + raise NotImplementedError( + "Can't find producer, the input {} is probably from" + " init_net, this is not supported yet.".format(old_name) + ) + producer = producer_map[(old_name, version)] + rename_op_output(predict_net, producer[0], producer[1], new_name) + return + + def contain_targets(op_ssa): + return (old_name, version) in op_ssa[0] + + is_consumer = [contain_targets(op_ssa) for op_ssa in predict_net_ssa] + if sum(is_consumer) > 1: + raise IllegalGraphTransformError( + ( + "Input '{}' of operator(#{}) are consumed by other ops, please use" + + " rename_op_output on the producer instead. Offending op: \n{}" + ).format(old_name, op_id, predict_net.op[op_id]) + ) + + # update init_net + _rename_versioned_blob_in_proto( + init_net, old_name, new_name, version, init_net_ssa, {}, init_net_versions + ) + # update predict_net + _rename_versioned_blob_in_proto( + predict_net, + old_name, + new_name, + version, + predict_net_ssa, + init_net_versions, + predict_net_versions, + ) + + +def rename_op_output(predict_net: caffe2_pb2.NetDef, op_id: int, output_id: int, new_name: str): + """ + Rename the op_id-th operator in predict_net, change it's output_id-th input's + name to the new_name. It also does automatic re-route and change + external_output and if necessary. + - It allows multiple consumers of its output. + - This function modifies predict_net in-place, doesn't need init_net. + """ + assert isinstance(predict_net, caffe2_pb2.NetDef) + + ssa, blob_versions = core.get_ssa(predict_net) + + versioned_inputs, versioned_outputs = ssa[op_id] + old_name, version = versioned_outputs[output_id] + + # update predict_net + _rename_versioned_blob_in_proto( + predict_net, old_name, new_name, version, ssa, {}, blob_versions + ) + + +def get_sub_graph_external_input_output( + predict_net: caffe2_pb2.NetDef, sub_graph_op_indices: List[int] +) -> Tuple[List[Tuple[str, int]], List[Tuple[str, int]]]: + """ + Return the list of external input/output of sub-graph, + each element is tuple of the name and corresponding version in predict_net. + + external input/output is defined the same way as caffe2 NetDef. + """ + ssa, versions = core.get_ssa(predict_net) + + all_inputs = [] + all_outputs = [] + for op_id in sub_graph_op_indices: + all_inputs += [inp for inp in ssa[op_id][0] if inp not in all_inputs] + all_outputs += list(ssa[op_id][1]) # ssa output won't repeat + + # for versioned blobs, external inputs are just those blob in all_inputs + # but not in all_outputs + ext_inputs = [inp for inp in all_inputs if inp not in all_outputs] + + # external outputs are essentially outputs of this subgraph that are used + # outside of this sub-graph (including predict_net.external_output) + all_other_inputs = sum( + (ssa[i][0] for i in range(len(ssa)) if i not in sub_graph_op_indices), + [(outp, versions[outp]) for outp in predict_net.external_output], + ) + ext_outputs = [outp for outp in all_outputs if outp in set(all_other_inputs)] + + return ext_inputs, ext_outputs + + +class DiGraph: + """A DAG representation of caffe2 graph, each vertice is a versioned blob.""" + + def __init__(self): + self.vertices = set() + self.graph = collections.defaultdict(list) + + def add_edge(self, u, v): + self.graph[u].append(v) + self.vertices.add(u) + self.vertices.add(v) + + # grab from https://www.geeksforgeeks.org/find-paths-given-source-destination/ + def get_all_paths(self, s, d): + visited = {k: False for k in self.vertices} + path = [] + all_paths = [] + + def _get_all_paths_util(graph, u, d, visited, path): + visited[u] = True + path.append(u) + if u == d: + all_paths.append(copy.deepcopy(path)) + else: + for i in graph[u]: + if not visited[i]: + _get_all_paths_util(graph, i, d, visited, path) + path.pop() + visited[u] = False + + _get_all_paths_util(self.graph, s, d, visited, path) + return all_paths + + @staticmethod + def from_ssa(ssa): + graph = DiGraph() + for op_id in range(len(ssa)): + for inp in ssa[op_id][0]: + for outp in ssa[op_id][1]: + graph.add_edge(inp, outp) + return graph + + +def _get_dependency_chain(ssa, versioned_target, versioned_source): + """ + Return the index list of relevant operator to produce target blob from source blob, + if there's no dependency, return empty list. + """ + + # finding all paths between nodes can be O(N!), thus we can only search + # in the subgraph using the op starting from the first consumer of source blob + # to the producer of the target blob. + consumer_map = get_consumer_map(ssa) + producer_map = get_producer_map(ssa) + start_op = min(x[0] for x in consumer_map[versioned_source]) - 15 + end_op = ( + producer_map[versioned_target][0] + 15 if versioned_target in producer_map else start_op + ) + sub_graph_ssa = ssa[start_op : end_op + 1] + if len(sub_graph_ssa) > 30: + logger.warning( + "Subgraph bebetween {} and {} is large (from op#{} to op#{}), it" + " might take non-trival time to find all paths between them.".format( + versioned_source, versioned_target, start_op, end_op + ) + ) + + dag = DiGraph.from_ssa(sub_graph_ssa) + paths = dag.get_all_paths(versioned_source, versioned_target) # include two ends + ops_in_paths = [[producer_map[blob][0] for blob in path[1:]] for path in paths] + return sorted(set().union(*[set(ops) for ops in ops_in_paths])) + + +def identify_reshape_sub_graph(predict_net: caffe2_pb2.NetDef) -> List[List[int]]: + """ + Idenfity the reshape sub-graph in a protobuf. + The reshape sub-graph is defined as matching the following pattern: + + (input_blob) -> Op_1 -> ... -> Op_N -> (new_shape) -─┐ + └-------------------------------------------> Reshape -> (output_blob) + + Return: + List of sub-graphs, each sub-graph is represented as a list of indices + of the relavent ops, [Op_1, Op_2, ..., Op_N, Reshape] + """ + + ssa, _ = core.get_ssa(predict_net) + + ret = [] + for i, op in enumerate(predict_net.op): + if op.type == "Reshape": + assert len(op.input) == 2 + input_ssa = ssa[i][0] + data_source = input_ssa[0] + shape_source = input_ssa[1] + op_indices = _get_dependency_chain(ssa, shape_source, data_source) + ret.append(op_indices + [i]) + return ret + + +def remove_reshape_for_fc(predict_net, params): + """ + In PyTorch nn.Linear has to take 2D tensor, this often leads to reshape + a 4D tensor to 2D by calling .view(). However this (dynamic) reshaping + doesn't work well with ONNX and Int8 tools, and cause using extra + ops (eg. ExpandDims) that might not be available on mobile. + Luckily Caffe2 supports 4D tensor for FC, so we can remove those reshape + after exporting ONNX model. + """ + from caffe2.python import core + + # find all reshape sub-graph that can be removed, which is now all Reshape + # sub-graph whose output is only consumed by FC. + # TODO: to make it safer, we may need the actually value to better determine + # if a Reshape before FC is removable. + reshape_sub_graphs = identify_reshape_sub_graph(predict_net) + sub_graphs_to_remove = [] + for reshape_sub_graph in reshape_sub_graphs: + reshape_op_id = reshape_sub_graph[-1] + assert predict_net.op[reshape_op_id].type == "Reshape" + ssa, _ = core.get_ssa(predict_net) + reshape_output = ssa[reshape_op_id][1][0] + consumers = [i for i in range(len(ssa)) if reshape_output in ssa[i][0]] + if all(predict_net.op[consumer].type == "FC" for consumer in consumers): + # safety check if the sub-graph is isolated, for this reshape sub-graph, + # it means it has one non-param external input and one external output. + ext_inputs, ext_outputs = get_sub_graph_external_input_output( + predict_net, reshape_sub_graph + ) + non_params_ext_inputs = [inp for inp in ext_inputs if inp[1] != 0] + if len(non_params_ext_inputs) == 1 and len(ext_outputs) == 1: + sub_graphs_to_remove.append(reshape_sub_graph) + + # perform removing subgraph by: + # 1: rename the Reshape's output to its input, then the graph can be + # seen as in-place itentify, meaning whose external input/output are the same. + # 2: simply remove those ops. + remove_op_ids = [] + params_to_remove = [] + for sub_graph in sub_graphs_to_remove: + logger.info( + "Remove Reshape sub-graph:\n{}".format( + "".join(["(#{:>4})\n{}".format(i, predict_net.op[i]) for i in sub_graph]) + ) + ) + reshape_op_id = sub_graph[-1] + new_reshap_output = predict_net.op[reshape_op_id].input[0] + rename_op_output(predict_net, reshape_op_id, 0, new_reshap_output) + ext_inputs, ext_outputs = get_sub_graph_external_input_output(predict_net, sub_graph) + non_params_ext_inputs = [inp for inp in ext_inputs if inp[1] != 0] + params_ext_inputs = [inp for inp in ext_inputs if inp[1] == 0] + assert len(non_params_ext_inputs) == 1 and len(ext_outputs) == 1 + assert ext_outputs[0][0] == non_params_ext_inputs[0][0] + assert ext_outputs[0][1] == non_params_ext_inputs[0][1] + 1 + remove_op_ids.extend(sub_graph) + params_to_remove.extend(params_ext_inputs) + + predict_net = copy.deepcopy(predict_net) + new_ops = [op for i, op in enumerate(predict_net.op) if i not in remove_op_ids] + del predict_net.op[:] + predict_net.op.extend(new_ops) + for versioned_params in params_to_remove: + name = versioned_params[0] + logger.info("Remove params: {} from init_net and predict_net.external_input".format(name)) + del params[name] + predict_net.external_input.remove(name) + + return predict_net, params + + +def fuse_copy_between_cpu_and_gpu(predict_net: caffe2_pb2.NetDef): + """ + In-place fuse extra copy ops between cpu/gpu for the following case: + a -CopyAToB-> b -CopyBToA> c1 -NextOp1-> d1 + -CopyBToA> c2 -NextOp2-> d2 + The fused network will look like: + a -NextOp1-> d1 + -NextOp2-> d2 + """ + + _COPY_OPS = ["CopyCPUToGPU", "CopyGPUToCPU"] + + def _fuse_once(predict_net): + ssa, blob_versions = core.get_ssa(predict_net) + consumer_map = get_consumer_map(ssa) + versioned_external_output = [ + (name, blob_versions[name]) for name in predict_net.external_output + ] + + for op_id, op in enumerate(predict_net.op): + if op.type in _COPY_OPS: + fw_copy_versioned_output = ssa[op_id][1][0] + consumer_ids = [x[0] for x in consumer_map[fw_copy_versioned_output]] + reverse_op_type = _COPY_OPS[1 - _COPY_OPS.index(op.type)] + + is_fusable = ( + len(consumer_ids) > 0 + and fw_copy_versioned_output not in versioned_external_output + and all( + predict_net.op[_op_id].type == reverse_op_type + and ssa[_op_id][1][0] not in versioned_external_output + for _op_id in consumer_ids + ) + ) + + if is_fusable: + for rv_copy_op_id in consumer_ids: + # making each NextOp uses "a" directly and removing Copy ops + rs_copy_versioned_output = ssa[rv_copy_op_id][1][0] + next_op_id, inp_id = consumer_map[rs_copy_versioned_output][0] + predict_net.op[next_op_id].input[inp_id] = op.input[0] + # remove CopyOps + new_ops = [ + op + for i, op in enumerate(predict_net.op) + if i != op_id and i not in consumer_ids + ] + del predict_net.op[:] + predict_net.op.extend(new_ops) + return True + + return False + + # _fuse_once returns False is nothing can be fused + while _fuse_once(predict_net): + pass + + +def remove_dead_end_ops(net_def: caffe2_pb2.NetDef): + """remove ops if its output is not used or not in external_output""" + ssa, versions = core.get_ssa(net_def) + versioned_external_output = [(name, versions[name]) for name in net_def.external_output] + consumer_map = get_consumer_map(ssa) + removed_op_ids = set() + + def _is_dead_end(versioned_blob): + return not ( + versioned_blob in versioned_external_output + or ( + len(consumer_map[versioned_blob]) > 0 + and all(x[0] not in removed_op_ids for x in consumer_map[versioned_blob]) + ) + ) + + for i, ssa_i in reversed(list(enumerate(ssa))): + versioned_outputs = ssa_i[1] + if all(_is_dead_end(outp) for outp in versioned_outputs): + removed_op_ids.add(i) + + # simply removing those deadend ops should have no effect to external_output + new_ops = [op for i, op in enumerate(net_def.op) if i not in removed_op_ids] + del net_def.op[:] + net_def.op.extend(new_ops) diff --git a/detectron2/export/torchscript.py b/detectron2/export/torchscript.py new file mode 100644 index 0000000000000000000000000000000000000000..480451a9df7ae42958b93150fd2156ee4029f2a7 --- /dev/null +++ b/detectron2/export/torchscript.py @@ -0,0 +1,133 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import os +import torch + +from detectron2.utils.env import TORCH_VERSION +from detectron2.utils.file_io import PathManager + +from .torchscript_patch import freeze_training_mode, patch_instances + +__all__ = ["scripting_with_instances", "dump_torchscript_IR"] + + +def scripting_with_instances(model, fields): + """ + Run :func:`torch.jit.script` on a model that uses the :class:`Instances` class. Since + attributes of :class:`Instances` are "dynamically" added in eager mode，it is difficult + for scripting to support it out of the box. This function is made to support scripting + a model that uses :class:`Instances`. It does the following: + + 1. Create a scriptable ``new_Instances`` class which behaves similarly to ``Instances``, + but with all attributes been "static". + The attributes need to be statically declared in the ``fields`` argument. + 2. Register ``new_Instances``, and force scripting compiler to + use it when trying to compile ``Instances``. + + After this function, the process will be reverted. User should be able to script another model + using different fields. + + Example: + Assume that ``Instances`` in the model consist of two attributes named + ``proposal_boxes`` and ``objectness_logits`` with type :class:`Boxes` and + :class:`Tensor` respectively during inference. You can call this function like: + :: + fields = {"proposal_boxes": Boxes, "objectness_logits": torch.Tensor} + torchscipt_model = scripting_with_instances(model, fields) + + Note: + It only support models in evaluation mode. + + Args: + model (nn.Module): The input model to be exported by scripting. + fields (Dict[str, type]): Attribute names and corresponding type that + ``Instances`` will use in the model. Note that all attributes used in ``Instances`` + need to be added, regardless of whether they are inputs/outputs of the model. + Data type not defined in detectron2 is not supported for now. + + Returns: + torch.jit.ScriptModule: the model in torchscript format + """ + assert TORCH_VERSION >= (1, 8), "This feature is not available in PyTorch < 1.8" + assert ( + not model.training + ), "Currently we only support exporting models in evaluation mode to torchscript" + + with freeze_training_mode(model), patch_instances(fields): + scripted_model = torch.jit.script(model) + return scripted_model + + +# alias for old name +export_torchscript_with_instances = scripting_with_instances + + +def dump_torchscript_IR(model, dir): + """ + Dump IR of a TracedModule/ScriptModule/Function in various format (code, graph, + inlined graph). Useful for debugging. + + Args: + model (TracedModule/ScriptModule/ScriptFUnction): traced or scripted module + dir (str): output directory to dump files. + """ + PathManager.mkdirs(dir) + + def _get_script_mod(mod): + if isinstance(mod, torch.jit.TracedModule): + return mod._actual_script_module + return mod + + # Dump pretty-printed code: https://pytorch.org/docs/stable/jit.html#inspecting-code + with PathManager.open(os.path.join(dir, "model_ts_code.txt"), "w") as f: + + def get_code(mod): + # Try a few ways to get code using private attributes. + try: + # This contains more information than just `mod.code` + return _get_script_mod(mod)._c.code + except AttributeError: + pass + try: + return mod.code + except AttributeError: + return None + + def dump_code(prefix, mod): + code = get_code(mod) + name = prefix or "root model" + if code is None: + f.write(f"Could not found code for {name} (type={mod.original_name})\n") + f.write("\n") + else: + f.write(f"\nCode for {name}, type={mod.original_name}:\n") + f.write(code) + f.write("\n") + f.write("-" * 80) + + for name, m in mod.named_children(): + dump_code(prefix + "." + name, m) + + if isinstance(model, torch.jit.ScriptFunction): + f.write(get_code(model)) + else: + dump_code("", model) + + def _get_graph(model): + try: + # Recursively dump IR of all modules + return _get_script_mod(model)._c.dump_to_str(True, False, False) + except AttributeError: + return model.graph.str() + + with PathManager.open(os.path.join(dir, "model_ts_IR.txt"), "w") as f: + f.write(_get_graph(model)) + + # Dump IR of the entire graph (all submodules inlined) + with PathManager.open(os.path.join(dir, "model_ts_IR_inlined.txt"), "w") as f: + f.write(str(model.inlined_graph)) + + if not isinstance(model, torch.jit.ScriptFunction): + # Dump the model structure in pytorch style + with PathManager.open(os.path.join(dir, "model.txt"), "w") as f: + f.write(str(model)) diff --git a/detectron2/export/torchscript_patch.py b/detectron2/export/torchscript_patch.py new file mode 100644 index 0000000000000000000000000000000000000000..618e7e0c4bd58e4fc1dc7c3d0e69c1b2ae73089e --- /dev/null +++ b/detectron2/export/torchscript_patch.py @@ -0,0 +1,377 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import os +import sys +import tempfile +from contextlib import ExitStack, contextmanager +from copy import deepcopy +from unittest import mock +import torch +from torch import nn + +# need some explicit imports due to https://github.com/pytorch/pytorch/issues/38964 +import detectron2 # noqa F401 +from detectron2.structures import Boxes, Instances +from detectron2.utils.env import _import_file + +_counter = 0 + + +def _clear_jit_cache(): + from torch.jit._recursive import concrete_type_store + from torch.jit._state import _jit_caching_layer + + concrete_type_store.type_store.clear() # for modules + _jit_caching_layer.clear() # for free functions + + +def _add_instances_conversion_methods(newInstances): + """ + Add from_instances methods to the scripted Instances class. + """ + cls_name = newInstances.__name__ + + @torch.jit.unused + def from_instances(instances: Instances): + """ + Create scripted Instances from original Instances + """ + fields = instances.get_fields() + image_size = instances.image_size + ret = newInstances(image_size) + for name, val in fields.items(): + assert hasattr(ret, f"_{name}"), f"No attribute named {name} in {cls_name}" + setattr(ret, name, deepcopy(val)) + return ret + + newInstances.from_instances = from_instances + + +@contextmanager +def patch_instances(fields): + """ + A contextmanager, under which the Instances class in detectron2 is replaced + by a statically-typed scriptable class, defined by `fields`. + See more in `scripting_with_instances`. + """ + + with tempfile.TemporaryDirectory(prefix="detectron2") as dir, tempfile.NamedTemporaryFile( + mode="w", encoding="utf-8", suffix=".py", dir=dir, delete=False + ) as f: + try: + # Objects that use Instances should not reuse previously-compiled + # results in cache, because `Instances` could be a new class each time. + _clear_jit_cache() + + cls_name, s = _gen_instance_module(fields) + f.write(s) + f.flush() + f.close() + + module = _import(f.name) + new_instances = getattr(module, cls_name) + _ = torch.jit.script(new_instances) + # let torchscript think Instances was scripted already + Instances.__torch_script_class__ = True + # let torchscript find new_instances when looking for the jit type of Instances + Instances._jit_override_qualname = torch._jit_internal._qualified_name(new_instances) + + _add_instances_conversion_methods(new_instances) + yield new_instances + finally: + try: + del Instances.__torch_script_class__ + del Instances._jit_override_qualname + except AttributeError: + pass + sys.modules.pop(module.__name__) + + +def _gen_instance_class(fields): + """ + Args: + fields (dict[name: type]) + """ + + class _FieldType: + def __init__(self, name, type_): + assert isinstance(name, str), f"Field name must be str, got {name}" + self.name = name + self.type_ = type_ + self.annotation = f"{type_.__module__}.{type_.__name__}" + + fields = [_FieldType(k, v) for k, v in fields.items()] + + def indent(level, s): + return " " * 4 * level + s + + lines = [] + + global _counter + _counter += 1 + + cls_name = "ScriptedInstances{}".format(_counter) + + field_names = tuple(x.name for x in fields) + lines.append( + f""" +class {cls_name}: + def __init__(self, image_size: Tuple[int, int]): + self.image_size = image_size + self._field_names = {field_names} +""" + ) + + for f in fields: + lines.append( + indent(2, f"self._{f.name} = torch.jit.annotate(Optional[{f.annotation}], None)") + ) + + for f in fields: + lines.append( + f""" + @property + def {f.name}(self) -> {f.annotation}: + # has to use a local for type refinement + # https://pytorch.org/docs/stable/jit_language_reference.html#optional-type-refinement + t = self._{f.name} + assert t is not None + return t + + @{f.name}.setter + def {f.name}(self, value: {f.annotation}) -> None: + self._{f.name} = value +""" + ) + + # support method `__len__` + lines.append( + """ + def __len__(self) -> int: +""" + ) + for f in fields: + lines.append( + f""" + t = self._{f.name} + if t is not None: + return len(t) +""" + ) + lines.append( + """ + raise NotImplementedError("Empty Instances does not support __len__!") +""" + ) + + # support method `has` + lines.append( + """ + def has(self, name: str) -> bool: +""" + ) + for f in fields: + lines.append( + f""" + if name == "{f.name}": + return self._{f.name} is not None +""" + ) + lines.append( + """ + return False +""" + ) + + # support method `to` + lines.append( + f""" + def to(self, device: torch.device) -> "{cls_name}": + ret = {cls_name}(self.image_size) +""" + ) + for f in fields: + if hasattr(f.type_, "to"): + lines.append( + f""" + t = self._{f.name} + if t is not None: + ret._{f.name} = t.to(device) +""" + ) + else: + # For now, ignore fields that cannot be moved to devices. + # Maybe can support other tensor-like classes (e.g. __torch_function__) + pass + lines.append( + """ + return ret +""" + ) + + # support method `getitem` + lines.append( + f""" + def __getitem__(self, item) -> "{cls_name}": + ret = {cls_name}(self.image_size) +""" + ) + for f in fields: + lines.append( + f""" + t = self._{f.name} + if t is not None: + ret._{f.name} = t[item] +""" + ) + lines.append( + """ + return ret +""" + ) + + # support method `get_fields()` + lines.append( + """ + def get_fields(self) -> Dict[str, Tensor]: + ret = {} + """ + ) + for f in fields: + if f.type_ == Boxes: + stmt = "t.tensor" + elif f.type_ == torch.Tensor: + stmt = "t" + else: + stmt = f'assert False, "unsupported type {str(f.type_)}"' + lines.append( + f""" + t = self._{f.name} + if t is not None: + ret["{f.name}"] = {stmt} + """ + ) + lines.append( + """ + return ret""" + ) + return cls_name, os.linesep.join(lines) + + +def _gen_instance_module(fields): + # TODO: find a more automatic way to enable import of other classes + s = """ +from copy import deepcopy +import torch +from torch import Tensor +import typing +from typing import * + +import detectron2 +from detectron2.structures import Boxes, Instances + +""" + + cls_name, cls_def = _gen_instance_class(fields) + s += cls_def + return cls_name, s + + +def _import(path): + return _import_file( + "{}{}".format(sys.modules[__name__].__name__, _counter), path, make_importable=True + ) + + +@contextmanager +def patch_builtin_len(modules=()): + """ + Patch the builtin len() function of a few detectron2 modules + to use __len__ instead, because __len__ does not convert values to + integers and therefore is friendly to tracing. + + Args: + modules (list[stsr]): names of extra modules to patch len(), in + addition to those in detectron2. + """ + + def _new_len(obj): + return obj.__len__() + + with ExitStack() as stack: + MODULES = [ + "detectron2.modeling.roi_heads.fast_rcnn", + "detectron2.modeling.roi_heads.mask_head", + "detectron2.modeling.roi_heads.keypoint_head", + ] + list(modules) + ctxs = [stack.enter_context(mock.patch(mod + ".len")) for mod in MODULES] + for m in ctxs: + m.side_effect = _new_len + yield + + +def patch_nonscriptable_classes(): + """ + Apply patches on a few nonscriptable detectron2 classes. + Should not have side-effects on eager usage. + """ + # __prepare_scriptable__ can also be added to models for easier maintenance. + # But it complicates the clean model code. + + from detectron2.modeling.backbone import ResNet, FPN + + # Due to https://github.com/pytorch/pytorch/issues/36061, + # we change backbone to use ModuleList for scripting. + # (note: this changes param names in state_dict) + + def prepare_resnet(self): + ret = deepcopy(self) + ret.stages = nn.ModuleList(ret.stages) + for k in self.stage_names: + delattr(ret, k) + return ret + + ResNet.__prepare_scriptable__ = prepare_resnet + + def prepare_fpn(self): + ret = deepcopy(self) + ret.lateral_convs = nn.ModuleList(ret.lateral_convs) + ret.output_convs = nn.ModuleList(ret.output_convs) + for name, _ in self.named_children(): + if name.startswith("fpn_"): + delattr(ret, name) + return ret + + FPN.__prepare_scriptable__ = prepare_fpn + + # Annotate some attributes to be constants for the purpose of scripting, + # even though they are not constants in eager mode. + from detectron2.modeling.roi_heads import StandardROIHeads + + if hasattr(StandardROIHeads, "__annotations__"): + # copy first to avoid editing annotations of base class + StandardROIHeads.__annotations__ = deepcopy(StandardROIHeads.__annotations__) + StandardROIHeads.__annotations__["mask_on"] = torch.jit.Final[bool] + StandardROIHeads.__annotations__["keypoint_on"] = torch.jit.Final[bool] + + +# These patches are not supposed to have side-effects. +patch_nonscriptable_classes() + + +@contextmanager +def freeze_training_mode(model): + """ + A context manager that annotates the "training" attribute of every submodule + to constant, so that the training codepath in these modules can be + meta-compiled away. Upon exiting, the annotations are reverted. + """ + classes = {type(x) for x in model.modules()} + # __constants__ is the old way to annotate constants and not compatible + # with __annotations__ . + classes = {x for x in classes if not hasattr(x, "__constants__")} + for cls in classes: + cls.__annotations__["training"] = torch.jit.Final[bool] + yield + for cls in classes: + cls.__annotations__["training"] = bool diff --git a/detectron2/layers/__init__.py b/detectron2/layers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c8bd1fb024d1cb911dda3f8a77f7ec3ad2e63798 --- /dev/null +++ b/detectron2/layers/__init__.py @@ -0,0 +1,22 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .batch_norm import FrozenBatchNorm2d, get_norm, NaiveSyncBatchNorm +from .deform_conv import DeformConv, ModulatedDeformConv +from .mask_ops import paste_masks_in_image +from .nms import batched_nms, batched_nms_rotated, nms, nms_rotated +from .roi_align import ROIAlign, roi_align +from .roi_align_rotated import ROIAlignRotated, roi_align_rotated +from .shape_spec import ShapeSpec +from .wrappers import ( + BatchNorm2d, + Conv2d, + ConvTranspose2d, + cat, + interpolate, + Linear, + nonzero_tuple, + cross_entropy, +) +from .blocks import CNNBlockBase, DepthwiseSeparableConv2d +from .aspp import ASPP + +__all__ = [k for k in globals().keys() if not k.startswith("_")] diff --git a/detectron2/layers/__pycache__/__init__.cpython-39.pyc b/detectron2/layers/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3c563bb15954199c697f55209fead1001b816a34 Binary files /dev/null and b/detectron2/layers/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/aspp.cpython-39.pyc b/detectron2/layers/__pycache__/aspp.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f566c7174b47c47424a8e42d94742034a5e09f14 Binary files /dev/null and b/detectron2/layers/__pycache__/aspp.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/batch_norm.cpython-39.pyc b/detectron2/layers/__pycache__/batch_norm.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6b8418956d20304d60ddec51fab9aabf0232a022 Binary files /dev/null and b/detectron2/layers/__pycache__/batch_norm.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/blocks.cpython-39.pyc b/detectron2/layers/__pycache__/blocks.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..51f5325e5a5f5c419273217abd190b28d2d38747 Binary files /dev/null and b/detectron2/layers/__pycache__/blocks.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/deform_conv.cpython-39.pyc b/detectron2/layers/__pycache__/deform_conv.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..87d8f9b622d854d9c6f4cae8e8eafa8cbf4639c8 Binary files /dev/null and b/detectron2/layers/__pycache__/deform_conv.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/mask_ops.cpython-39.pyc b/detectron2/layers/__pycache__/mask_ops.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c45abcd5ded745ec3682fe5bc9900b3badc67586 Binary files /dev/null and b/detectron2/layers/__pycache__/mask_ops.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/nms.cpython-39.pyc b/detectron2/layers/__pycache__/nms.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8d05be9d5901cb900db8904af6e129b977d2c9ab Binary files /dev/null and b/detectron2/layers/__pycache__/nms.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/roi_align.cpython-39.pyc b/detectron2/layers/__pycache__/roi_align.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..712996736ee3b81966e91e92188074d91dd2b9e0 Binary files /dev/null and b/detectron2/layers/__pycache__/roi_align.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/roi_align_rotated.cpython-39.pyc b/detectron2/layers/__pycache__/roi_align_rotated.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ad55e32d500c9eb5b12706a516c5f4d09420b835 Binary files /dev/null and b/detectron2/layers/__pycache__/roi_align_rotated.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/rotated_boxes.cpython-39.pyc b/detectron2/layers/__pycache__/rotated_boxes.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f892317c0b67fe5a3c548f600e1205cb72ebc358 Binary files /dev/null and b/detectron2/layers/__pycache__/rotated_boxes.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/shape_spec.cpython-39.pyc b/detectron2/layers/__pycache__/shape_spec.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8d36f1394607dd352736cda493f0776a9f02a62a Binary files /dev/null and b/detectron2/layers/__pycache__/shape_spec.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/soft_nms.cpython-39.pyc b/detectron2/layers/__pycache__/soft_nms.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..f642d9dc3b2efe6c2a0dce21f8398e0b03af6e49 Binary files /dev/null and b/detectron2/layers/__pycache__/soft_nms.cpython-39.pyc differ diff --git a/detectron2/layers/__pycache__/wrappers.cpython-39.pyc b/detectron2/layers/__pycache__/wrappers.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7ca63688645e2cc2a2201ec1b5a79288afc9cc36 Binary files /dev/null and b/detectron2/layers/__pycache__/wrappers.cpython-39.pyc differ diff --git a/detectron2/layers/aspp.py b/detectron2/layers/aspp.py new file mode 100644 index 0000000000000000000000000000000000000000..14861aa9ede4fea6a69a49f189bcab997b558148 --- /dev/null +++ b/detectron2/layers/aspp.py @@ -0,0 +1,144 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +from copy import deepcopy +import fvcore.nn.weight_init as weight_init +import torch +from torch import nn +from torch.nn import functional as F + +from .batch_norm import get_norm +from .blocks import DepthwiseSeparableConv2d +from .wrappers import Conv2d + + +class ASPP(nn.Module): + """ + Atrous Spatial Pyramid Pooling (ASPP). + """ + + def __init__( + self, + in_channels, + out_channels, + dilations, + *, + norm, + activation, + pool_kernel_size=None, + dropout: float = 0.0, + use_depthwise_separable_conv=False, + ): + """ + Args: + in_channels (int): number of input channels for ASPP. + out_channels (int): number of output channels. + dilations (list): a list of 3 dilations in ASPP. + norm (str or callable): normalization for all conv layers. + See :func:`layers.get_norm` for supported format. norm is + applied to all conv layers except the conv following + global average pooling. + activation (callable): activation function. + pool_kernel_size (tuple, list): the average pooling size (kh, kw) + for image pooling layer in ASPP. If set to None, it always + performs global average pooling. If not None, it must be + divisible by the shape of inputs in forward(). It is recommended + to use a fixed input feature size in training, and set this + option to match this size, so that it performs global average + pooling in training, and the size of the pooling window stays + consistent in inference. + dropout (float): apply dropout on the output of ASPP. It is used in + the official DeepLab implementation with a rate of 0.1: + https://github.com/tensorflow/models/blob/21b73d22f3ed05b650e85ac50849408dd36de32e/research/deeplab/model.py#L532 # noqa + use_depthwise_separable_conv (bool): use DepthwiseSeparableConv2d + for 3x3 convs in ASPP, proposed in :paper:`DeepLabV3+`. + """ + super(ASPP, self).__init__() + assert len(dilations) == 3, "ASPP expects 3 dilations, got {}".format(len(dilations)) + self.pool_kernel_size = pool_kernel_size + self.dropout = dropout + use_bias = norm == "" + self.convs = nn.ModuleList() + # conv 1x1 + self.convs.append( + Conv2d( + in_channels, + out_channels, + kernel_size=1, + bias=use_bias, + norm=get_norm(norm, out_channels), + activation=deepcopy(activation), + ) + ) + weight_init.c2_xavier_fill(self.convs[-1]) + # atrous convs + for dilation in dilations: + if use_depthwise_separable_conv: + self.convs.append( + DepthwiseSeparableConv2d( + in_channels, + out_channels, + kernel_size=3, + padding=dilation, + dilation=dilation, + norm1=norm, + activation1=deepcopy(activation), + norm2=norm, + activation2=deepcopy(activation), + ) + ) + else: + self.convs.append( + Conv2d( + in_channels, + out_channels, + kernel_size=3, + padding=dilation, + dilation=dilation, + bias=use_bias, + norm=get_norm(norm, out_channels), + activation=deepcopy(activation), + ) + ) + weight_init.c2_xavier_fill(self.convs[-1]) + # image pooling + # We do not add BatchNorm because the spatial resolution is 1x1, + # the original TF implementation has BatchNorm. + if pool_kernel_size is None: + image_pooling = nn.Sequential( + nn.AdaptiveAvgPool2d(1), + Conv2d(in_channels, out_channels, 1, bias=True, activation=deepcopy(activation)), + ) + else: + image_pooling = nn.Sequential( + nn.AvgPool2d(kernel_size=pool_kernel_size, stride=1), + Conv2d(in_channels, out_channels, 1, bias=True, activation=deepcopy(activation)), + ) + weight_init.c2_xavier_fill(image_pooling[1]) + self.convs.append(image_pooling) + + self.project = Conv2d( + 5 * out_channels, + out_channels, + kernel_size=1, + bias=use_bias, + norm=get_norm(norm, out_channels), + activation=deepcopy(activation), + ) + weight_init.c2_xavier_fill(self.project) + + def forward(self, x): + size = x.shape[-2:] + if self.pool_kernel_size is not None: + if size[0] % self.pool_kernel_size[0] or size[1] % self.pool_kernel_size[1]: + raise ValueError( + "`pool_kernel_size` must be divisible by the shape of inputs. " + "Input size: {} `pool_kernel_size`: {}".format(size, self.pool_kernel_size) + ) + res = [] + for conv in self.convs: + res.append(conv(x)) + res[-1] = F.interpolate(res[-1], size=size, mode="bilinear", align_corners=False) + res = torch.cat(res, dim=1) + res = self.project(res) + res = F.dropout(res, self.dropout, training=self.training) if self.dropout > 0 else res + return res diff --git a/detectron2/layers/batch_norm.py b/detectron2/layers/batch_norm.py new file mode 100644 index 0000000000000000000000000000000000000000..975346249b8e9be4b710308c5c11846c2afc6319 --- /dev/null +++ b/detectron2/layers/batch_norm.py @@ -0,0 +1,243 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import torch +import torch.distributed as dist +from fvcore.nn.distributed import differentiable_all_reduce +from torch import nn +from torch.nn import functional as F + +from detectron2.utils import comm, env + +from .wrappers import BatchNorm2d + + +class FrozenBatchNorm2d(nn.Module): + """ + BatchNorm2d where the batch statistics and the affine parameters are fixed. + + It contains non-trainable buffers called + "weight" and "bias", "running_mean", "running_var", + initialized to perform identity transformation. + + The pre-trained backbone models from Caffe2 only contain "weight" and "bias", + which are computed from the original four parameters of BN. + The affine transform `x * weight + bias` will perform the equivalent + computation of `(x - running_mean) / sqrt(running_var) * weight + bias`. + When loading a backbone model from Caffe2, "running_mean" and "running_var" + will be left unchanged as identity transformation. + + Other pre-trained backbone models may contain all 4 parameters. + + The forward is implemented by `F.batch_norm(..., training=False)`. + """ + + _version = 3 + + def __init__(self, num_features, eps=1e-5): + super().__init__() + self.num_features = num_features + self.eps = eps + self.register_buffer("weight", torch.ones(num_features)) + self.register_buffer("bias", torch.zeros(num_features)) + self.register_buffer("running_mean", torch.zeros(num_features)) + self.register_buffer("running_var", torch.ones(num_features) - eps) + + def forward(self, x): + if x.requires_grad: + # When gradients are needed, F.batch_norm will use extra memory + # because its backward op computes gradients for weight/bias as well. + scale = self.weight * (self.running_var + self.eps).rsqrt() + bias = self.bias - self.running_mean * scale + scale = scale.reshape(1, -1, 1, 1) + bias = bias.reshape(1, -1, 1, 1) + out_dtype = x.dtype # may be half + return x * scale.to(out_dtype) + bias.to(out_dtype) + else: + # When gradients are not needed, F.batch_norm is a single fused op + # and provide more optimization opportunities. + return F.batch_norm( + x, + self.running_mean, + self.running_var, + self.weight, + self.bias, + training=False, + eps=self.eps, + ) + + def _load_from_state_dict( + self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs + ): + version = local_metadata.get("version", None) + + if version is None or version < 2: + # when use offline modules, avoid overwriting running mean and var for loaded weights + skip_reset = False + for k_n in state_dict: # checkpoint weights + if 'ignore_others' in k_n: #if 'offline' in k_n: + skip_reset = True + if not skip_reset: + # No running_mean/var in early versions + # This will silent the warnings + if prefix + "running_mean" not in state_dict: + state_dict[prefix + "running_mean"] = torch.zeros_like(self.running_mean) + if prefix + "running_var" not in state_dict: + state_dict[prefix + "running_var"] = torch.ones_like(self.running_var) + + # NOTE: if a checkpoint is trained with BatchNorm and loaded (together with + # version number) to FrozenBatchNorm, running_var will be wrong. One solution + # is to remove the version number from the checkpoint. + if version is not None and version < 3: + logger = logging.getLogger(__name__) + logger.info("FrozenBatchNorm {} is upgraded to version 3.".format(prefix.rstrip("."))) + # In version < 3, running_var are used without +eps. + state_dict[prefix + "running_var"] -= self.eps + + super()._load_from_state_dict( + state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs + ) + + def __repr__(self): + return "FrozenBatchNorm2d(num_features={}, eps={})".format(self.num_features, self.eps) + + @classmethod + def convert_frozen_batchnorm(cls, module): + """ + Convert all BatchNorm/SyncBatchNorm in module into FrozenBatchNorm. + + Args: + module (torch.nn.Module): + + Returns: + If module is BatchNorm/SyncBatchNorm, returns a new module. + Otherwise, in-place convert module and return it. + + Similar to convert_sync_batchnorm in + https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/batchnorm.py + """ + bn_module = nn.modules.batchnorm + bn_module = (bn_module.BatchNorm2d, bn_module.SyncBatchNorm) + res = module + if isinstance(module, bn_module): + res = cls(module.num_features) + if module.affine: + res.weight.data = module.weight.data.clone().detach() + res.bias.data = module.bias.data.clone().detach() + res.running_mean.data = module.running_mean.data + res.running_var.data = module.running_var.data + res.eps = module.eps + else: + for name, child in module.named_children(): + new_child = cls.convert_frozen_batchnorm(child) + if new_child is not child: + res.add_module(name, new_child) + return res + + +def get_norm(norm, out_channels): + """ + Args: + norm (str or callable): either one of BN, SyncBN, FrozenBN, GN; + or a callable that takes a channel number and returns + the normalization layer as a nn.Module. + + Returns: + nn.Module or None: the normalization layer + """ + if norm is None: + return None + if isinstance(norm, str): + if len(norm) == 0: + return None + norm = { + "BN": BatchNorm2d, + # Fixed in https://github.com/pytorch/pytorch/pull/36382 + "SyncBN": NaiveSyncBatchNorm if env.TORCH_VERSION <= (1, 5) else nn.SyncBatchNorm, + "FrozenBN": FrozenBatchNorm2d, + "GN": lambda channels: nn.GroupNorm(32, channels), + # for debugging: + "nnSyncBN": nn.SyncBatchNorm, + "naiveSyncBN": NaiveSyncBatchNorm, + }[norm] + return norm(out_channels) + + +class NaiveSyncBatchNorm(BatchNorm2d): + """ + In PyTorch<=1.5, ``nn.SyncBatchNorm`` has incorrect gradient + when the batch size on each worker is different. + (e.g., when scale augmentation is used, or when it is applied to mask head). + + This is a slower but correct alternative to `nn.SyncBatchNorm`. + + Note: + There isn't a single definition of Sync BatchNorm. + + When ``stats_mode==""``, this module computes overall statistics by using + statistics of each worker with equal weight. The result is true statistics + of all samples (as if they are all on one worker) only when all workers + have the same (N, H, W). This mode does not support inputs with zero batch size. + + When ``stats_mode=="N"``, this module computes overall statistics by weighting + the statistics of each worker by their ``N``. The result is true statistics + of all samples (as if they are all on one worker) only when all workers + have the same (H, W). It is slower than ``stats_mode==""``. + + Even though the result of this module may not be the true statistics of all samples, + it may still be reasonable because it might be preferrable to assign equal weights + to all workers, regardless of their (H, W) dimension, instead of putting larger weight + on larger images. From preliminary experiments, little difference is found between such + a simplified implementation and an accurate computation of overall mean & variance. + """ + + def __init__(self, *args, stats_mode="", **kwargs): + super().__init__(*args, **kwargs) + assert stats_mode in ["", "N"] + self._stats_mode = stats_mode + + def forward(self, input): + if comm.get_world_size() == 1 or not self.training: + return super().forward(input) + + B, C = input.shape[0], input.shape[1] + + half_input = input.dtype == torch.float16 + if half_input: + # fp16 does not have good enough numerics for the reduction here + input = input.float() + mean = torch.mean(input, dim=[0, 2, 3]) + meansqr = torch.mean(input * input, dim=[0, 2, 3]) + + if self._stats_mode == "": + assert B > 0, 'SyncBatchNorm(stats_mode="") does not support zero batch size.' + vec = torch.cat([mean, meansqr], dim=0) + vec = differentiable_all_reduce(vec) * (1.0 / dist.get_world_size()) + mean, meansqr = torch.split(vec, C) + momentum = self.momentum + else: + if B == 0: + vec = torch.zeros([2 * C + 1], device=mean.device, dtype=mean.dtype) + vec = vec + input.sum() # make sure there is gradient w.r.t input + else: + vec = torch.cat( + [mean, meansqr, torch.ones([1], device=mean.device, dtype=mean.dtype)], dim=0 + ) + vec = differentiable_all_reduce(vec * B) + + total_batch = vec[-1].detach() + momentum = total_batch.clamp(max=1) * self.momentum # no update if total_batch is 0 + mean, meansqr, _ = torch.split(vec / total_batch.clamp(min=1), C) # avoid div-by-zero + + var = meansqr - mean * mean + invstd = torch.rsqrt(var + self.eps) + scale = self.weight * invstd + bias = self.bias - mean * scale + scale = scale.reshape(1, -1, 1, 1) + bias = bias.reshape(1, -1, 1, 1) + + self.running_mean += momentum * (mean.detach() - self.running_mean) + self.running_var += momentum * (var.detach() - self.running_var) + ret = input * scale + bias + if half_input: + ret = ret.half() + return ret diff --git a/detectron2/layers/blocks.py b/detectron2/layers/blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..1995a4bf7339e8deb7eaaffda4f819dda55e7ac7 --- /dev/null +++ b/detectron2/layers/blocks.py @@ -0,0 +1,111 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import fvcore.nn.weight_init as weight_init +from torch import nn + +from .batch_norm import FrozenBatchNorm2d, get_norm +from .wrappers import Conv2d + + +""" +CNN building blocks. +""" + + +class CNNBlockBase(nn.Module): + """ + A CNN block is assumed to have input channels, output channels and a stride. + The input and output of `forward()` method must be NCHW tensors. + The method can perform arbitrary computation but must match the given + channels and stride specification. + + Attribute: + in_channels (int): + out_channels (int): + stride (int): + """ + + def __init__(self, in_channels, out_channels, stride): + """ + The `__init__` method of any subclass should also contain these arguments. + + Args: + in_channels (int): + out_channels (int): + stride (int): + """ + super().__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.stride = stride + + def freeze(self): + """ + Make this block not trainable. + This method sets all parameters to `requires_grad=False`, + and convert all BatchNorm layers to FrozenBatchNorm + + Returns: + the block itself + """ + for p in self.parameters(): + p.requires_grad = False + FrozenBatchNorm2d.convert_frozen_batchnorm(self) + return self + + +class DepthwiseSeparableConv2d(nn.Module): + """ + A kxk depthwise convolution + a 1x1 convolution. + + In :paper:`xception`, norm & activation are applied on the second conv. + :paper:`mobilenet` uses norm & activation on both convs. + """ + + def __init__( + self, + in_channels, + out_channels, + kernel_size=3, + padding=1, + dilation=1, + *, + norm1=None, + activation1=None, + norm2=None, + activation2=None, + ): + """ + Args: + norm1, norm2 (str or callable): normalization for the two conv layers. + activation1, activation2 (callable(Tensor) -> Tensor): activation + function for the two conv layers. + """ + super().__init__() + self.depthwise = Conv2d( + in_channels, + in_channels, + kernel_size=kernel_size, + padding=padding, + dilation=dilation, + groups=in_channels, + bias=not norm1, + norm=get_norm(norm1, in_channels), + activation=activation1, + ) + self.pointwise = Conv2d( + in_channels, + out_channels, + kernel_size=1, + bias=not norm2, + norm=get_norm(norm2, out_channels), + activation=activation2, + ) + + # default initialization + weight_init.c2_msra_fill(self.depthwise) + weight_init.c2_msra_fill(self.pointwise) + + def forward(self, x): + return self.pointwise(self.depthwise(x)) diff --git a/detectron2/layers/csrc/README.md b/detectron2/layers/csrc/README.md new file mode 100644 index 0000000000000000000000000000000000000000..778ed3da0bae89820831bcd8a72ff7b9cad8d4dd --- /dev/null +++ b/detectron2/layers/csrc/README.md @@ -0,0 +1,7 @@ + + +To add a new Op: + +1. Create a new directory +2. Implement new ops there +3. Delcare its Python interface in `vision.cpp`. diff --git a/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated.h b/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated.h new file mode 100644 index 0000000000000000000000000000000000000000..7ceb5185e0305de554644dc082866e41416520b8 --- /dev/null +++ b/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated.h @@ -0,0 +1,115 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#pragma once +#include + +namespace detectron2 { + +at::Tensor ROIAlignRotated_forward_cpu( + const at::Tensor& input, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int sampling_ratio); + +at::Tensor ROIAlignRotated_backward_cpu( + const at::Tensor& grad, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int batch_size, + const int channels, + const int height, + const int width, + const int sampling_ratio); + +#if defined(WITH_CUDA) || defined(WITH_HIP) +at::Tensor ROIAlignRotated_forward_cuda( + const at::Tensor& input, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int sampling_ratio); + +at::Tensor ROIAlignRotated_backward_cuda( + const at::Tensor& grad, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int batch_size, + const int channels, + const int height, + const int width, + const int sampling_ratio); +#endif + +// Interface for Python +inline at::Tensor ROIAlignRotated_forward( + const at::Tensor& input, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int sampling_ratio) { + if (input.is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + return ROIAlignRotated_forward_cuda( + input, + rois, + spatial_scale, + pooled_height, + pooled_width, + sampling_ratio); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + return ROIAlignRotated_forward_cpu( + input, rois, spatial_scale, pooled_height, pooled_width, sampling_ratio); +} + +inline at::Tensor ROIAlignRotated_backward( + const at::Tensor& grad, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int batch_size, + const int channels, + const int height, + const int width, + const int sampling_ratio) { + if (grad.is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + return ROIAlignRotated_backward_cuda( + grad, + rois, + spatial_scale, + pooled_height, + pooled_width, + batch_size, + channels, + height, + width, + sampling_ratio); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + return ROIAlignRotated_backward_cpu( + grad, + rois, + spatial_scale, + pooled_height, + pooled_width, + batch_size, + channels, + height, + width, + sampling_ratio); +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated_cpu.cpp b/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated_cpu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..2a3d3056cc71a4acaafb570739a9dd247a7eb1ed --- /dev/null +++ b/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated_cpu.cpp @@ -0,0 +1,522 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#include +#include "ROIAlignRotated.h" + +// Note: this implementation originates from the Caffe2 ROIAlignRotated Op +// and PyTorch ROIAlign (non-rotated) Op implementations. +// The key difference between this implementation and those ones is +// we don't do "legacy offset" in this version, as there aren't many previous +// works, if any, using the "legacy" ROIAlignRotated Op. +// This would make the interface a bit cleaner. + +namespace detectron2 { + +namespace { +template +struct PreCalc { + int pos1; + int pos2; + int pos3; + int pos4; + T w1; + T w2; + T w3; + T w4; +}; + +template +void pre_calc_for_bilinear_interpolate( + const int height, + const int width, + const int pooled_height, + const int pooled_width, + const int iy_upper, + const int ix_upper, + T roi_start_h, + T roi_start_w, + T bin_size_h, + T bin_size_w, + int roi_bin_grid_h, + int roi_bin_grid_w, + T roi_center_h, + T roi_center_w, + T cos_theta, + T sin_theta, + std::vector>& pre_calc) { + int pre_calc_index = 0; + for (int ph = 0; ph < pooled_height; ph++) { + for (int pw = 0; pw < pooled_width; pw++) { + for (int iy = 0; iy < iy_upper; iy++) { + const T yy = roi_start_h + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); // e.g., 0.5, 1.5 + for (int ix = 0; ix < ix_upper; ix++) { + const T xx = roi_start_w + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + + // Rotate by theta around the center and translate + // In image space, (y, x) is the order for Right Handed System, + // and this is essentially multiplying the point by a rotation matrix + // to rotate it counterclockwise through angle theta. + T y = yy * cos_theta - xx * sin_theta + roi_center_h; + T x = yy * sin_theta + xx * cos_theta + roi_center_w; + // deal with: inverse elements are out of feature map boundary + if (y < -1.0 || y > height || x < -1.0 || x > width) { + // empty + PreCalc pc; + pc.pos1 = 0; + pc.pos2 = 0; + pc.pos3 = 0; + pc.pos4 = 0; + pc.w1 = 0; + pc.w2 = 0; + pc.w3 = 0; + pc.w4 = 0; + pre_calc[pre_calc_index] = pc; + pre_calc_index += 1; + continue; + } + + if (y < 0) { + y = 0; + } + if (x < 0) { + x = 0; + } + + int y_low = (int)y; + int x_low = (int)x; + int y_high; + int x_high; + + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = (T)y_low; + } else { + y_high = y_low + 1; + } + + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = (T)x_low; + } else { + x_high = x_low + 1; + } + + T ly = y - y_low; + T lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + T w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + + // save weights and indices + PreCalc pc; + pc.pos1 = y_low * width + x_low; + pc.pos2 = y_low * width + x_high; + pc.pos3 = y_high * width + x_low; + pc.pos4 = y_high * width + x_high; + pc.w1 = w1; + pc.w2 = w2; + pc.w3 = w3; + pc.w4 = w4; + pre_calc[pre_calc_index] = pc; + + pre_calc_index += 1; + } + } + } + } +} + +template +void bilinear_interpolate_gradient( + const int height, + const int width, + T y, + T x, + T& w1, + T& w2, + T& w3, + T& w4, + int& x_low, + int& x_high, + int& y_low, + int& y_high) { + // deal with cases that inverse elements are out of feature map boundary + if (y < -1.0 || y > height || x < -1.0 || x > width) { + // empty + w1 = w2 = w3 = w4 = 0.; + x_low = x_high = y_low = y_high = -1; + return; + } + + if (y < 0) { + y = 0; + } + + if (x < 0) { + x = 0; + } + + y_low = (int)y; + x_low = (int)x; + + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = (T)y_low; + } else { + y_high = y_low + 1; + } + + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = (T)x_low; + } else { + x_high = x_low + 1; + } + + T ly = y - y_low; + T lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + + // reference in forward + // T v1 = input[y_low * width + x_low]; + // T v2 = input[y_low * width + x_high]; + // T v3 = input[y_high * width + x_low]; + // T v4 = input[y_high * width + x_high]; + // T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + + w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + + return; +} + +template +inline void add(T* address, const T& val) { + *address += val; +} + +} // namespace + +template +void ROIAlignRotatedForward( + const int nthreads, + const T* input, + const T& spatial_scale, + const int channels, + const int height, + const int width, + const int pooled_height, + const int pooled_width, + const int sampling_ratio, + const T* rois, + T* output) { + int n_rois = nthreads / channels / pooled_width / pooled_height; + // (n, c, ph, pw) is an element in the pooled output + // can be parallelized using omp + // #pragma omp parallel for num_threads(32) + for (int n = 0; n < n_rois; n++) { + int index_n = n * channels * pooled_width * pooled_height; + + const T* current_roi = rois + n * 6; + int roi_batch_ind = current_roi[0]; + + // Do not use rounding; this implementation detail is critical + // ROIAlignRotated supports align == true, i.e., continuous coordinate + // by default, thus the 0.5 offset + T offset = (T)0.5; + T roi_center_w = current_roi[1] * spatial_scale - offset; + T roi_center_h = current_roi[2] * spatial_scale - offset; + T roi_width = current_roi[3] * spatial_scale; + T roi_height = current_roi[4] * spatial_scale; + T theta = current_roi[5] * M_PI / 180.0; + T cos_theta = cos(theta); + T sin_theta = sin(theta); + + AT_ASSERTM( + roi_width >= 0 && roi_height >= 0, + "ROIs in ROIAlignRotated do not have non-negative size!"); + + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + + // We use roi_bin_grid to sample the grid and mimic integral + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); // e.g., = 2 + int roi_bin_grid_w = + (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); + + // We do average (integral) pooling inside a bin + const T count = std::max(roi_bin_grid_h * roi_bin_grid_w, 1); // e.g. = 4 + + // we want to precalculate indices and weights shared by all channels, + // this is the key point of optimization + std::vector> pre_calc( + roi_bin_grid_h * roi_bin_grid_w * pooled_width * pooled_height); + + // roi_start_h and roi_start_w are computed wrt the center of RoI (x, y). + // Appropriate translation needs to be applied after. + T roi_start_h = -roi_height / 2.0; + T roi_start_w = -roi_width / 2.0; + + pre_calc_for_bilinear_interpolate( + height, + width, + pooled_height, + pooled_width, + roi_bin_grid_h, + roi_bin_grid_w, + roi_start_h, + roi_start_w, + bin_size_h, + bin_size_w, + roi_bin_grid_h, + roi_bin_grid_w, + roi_center_h, + roi_center_w, + cos_theta, + sin_theta, + pre_calc); + + for (int c = 0; c < channels; c++) { + int index_n_c = index_n + c * pooled_width * pooled_height; + const T* offset_input = + input + (roi_batch_ind * channels + c) * height * width; + int pre_calc_index = 0; + + for (int ph = 0; ph < pooled_height; ph++) { + for (int pw = 0; pw < pooled_width; pw++) { + int index = index_n_c + ph * pooled_width + pw; + + T output_val = 0.; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + PreCalc pc = pre_calc[pre_calc_index]; + output_val += pc.w1 * offset_input[pc.pos1] + + pc.w2 * offset_input[pc.pos2] + + pc.w3 * offset_input[pc.pos3] + pc.w4 * offset_input[pc.pos4]; + + pre_calc_index += 1; + } + } + output_val /= count; + + output[index] = output_val; + } // for pw + } // for ph + } // for c + } // for n +} + +template +void ROIAlignRotatedBackward( + const int nthreads, + // may not be contiguous. should index using n_stride, etc + const T* grad_output, + const T& spatial_scale, + const int channels, + const int height, + const int width, + const int pooled_height, + const int pooled_width, + const int sampling_ratio, + T* grad_input, + const T* rois, + const int n_stride, + const int c_stride, + const int h_stride, + const int w_stride) { + for (int index = 0; index < nthreads; index++) { + // (n, c, ph, pw) is an element in the pooled output + int pw = index % pooled_width; + int ph = (index / pooled_width) % pooled_height; + int c = (index / pooled_width / pooled_height) % channels; + int n = index / pooled_width / pooled_height / channels; + + const T* current_roi = rois + n * 6; + int roi_batch_ind = current_roi[0]; + + // Do not use rounding; this implementation detail is critical + // ROIAlignRotated supports align == true, i.e., continuous coordinate + // by default, thus the 0.5 offset + T offset = (T)0.5; + T roi_center_w = current_roi[1] * spatial_scale - offset; + T roi_center_h = current_roi[2] * spatial_scale - offset; + T roi_width = current_roi[3] * spatial_scale; + T roi_height = current_roi[4] * spatial_scale; + T theta = current_roi[5] * M_PI / 180.0; + T cos_theta = cos(theta); + T sin_theta = sin(theta); + + AT_ASSERTM( + roi_width >= 0 && roi_height >= 0, + "ROIs in ROIAlignRotated do not have non-negative size!"); + + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + + T* offset_grad_input = + grad_input + ((roi_batch_ind * channels + c) * height * width); + + int output_offset = n * n_stride + c * c_stride; + const T* offset_grad_output = grad_output + output_offset; + const T grad_output_this_bin = + offset_grad_output[ph * h_stride + pw * w_stride]; + + // We use roi_bin_grid to sample the grid and mimic integral + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); // e.g., = 2 + int roi_bin_grid_w = + (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); + + // roi_start_h and roi_start_w are computed wrt the center of RoI (x, y). + // Appropriate translation needs to be applied after. + T roi_start_h = -roi_height / 2.0; + T roi_start_w = -roi_width / 2.0; + + // We do average (integral) pooling inside a bin + const T count = roi_bin_grid_h * roi_bin_grid_w; // e.g. = 4 + + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + const T yy = roi_start_h + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); // e.g., 0.5, 1.5 + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T xx = roi_start_w + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + + // Rotate by theta around the center and translate + T y = yy * cos_theta - xx * sin_theta + roi_center_h; + T x = yy * sin_theta + xx * cos_theta + roi_center_w; + + T w1, w2, w3, w4; + int x_low, x_high, y_low, y_high; + + bilinear_interpolate_gradient( + height, width, y, x, w1, w2, w3, w4, x_low, x_high, y_low, y_high); + + T g1 = grad_output_this_bin * w1 / count; + T g2 = grad_output_this_bin * w2 / count; + T g3 = grad_output_this_bin * w3 / count; + T g4 = grad_output_this_bin * w4 / count; + + if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) { + // atomic add is not needed for now since it is single threaded + add(offset_grad_input + y_low * width + x_low, static_cast(g1)); + add(offset_grad_input + y_low * width + x_high, static_cast(g2)); + add(offset_grad_input + y_high * width + x_low, static_cast(g3)); + add(offset_grad_input + y_high * width + x_high, static_cast(g4)); + } // if + } // ix + } // iy + } // for +} // ROIAlignRotatedBackward + +at::Tensor ROIAlignRotated_forward_cpu( + const at::Tensor& input, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int sampling_ratio) { + AT_ASSERTM(input.device().is_cpu(), "input must be a CPU tensor"); + AT_ASSERTM(rois.device().is_cpu(), "rois must be a CPU tensor"); + + at::TensorArg input_t{input, "input", 1}, rois_t{rois, "rois", 2}; + + at::CheckedFrom c = "ROIAlign_forward_cpu"; + at::checkAllSameType(c, {input_t, rois_t}); + + auto num_rois = rois.size(0); + auto channels = input.size(1); + auto height = input.size(2); + auto width = input.size(3); + + at::Tensor output = at::zeros( + {num_rois, channels, pooled_height, pooled_width}, input.options()); + + auto output_size = num_rois * pooled_height * pooled_width * channels; + + if (output.numel() == 0) { + return output; + } + + auto input_ = input.contiguous(), rois_ = rois.contiguous(); + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + input.scalar_type(), "ROIAlignRotated_forward", [&] { + ROIAlignRotatedForward( + output_size, + input_.data_ptr(), + spatial_scale, + channels, + height, + width, + pooled_height, + pooled_width, + sampling_ratio, + rois_.data_ptr(), + output.data_ptr()); + }); + return output; +} + +at::Tensor ROIAlignRotated_backward_cpu( + const at::Tensor& grad, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int batch_size, + const int channels, + const int height, + const int width, + const int sampling_ratio) { + AT_ASSERTM(grad.device().is_cpu(), "grad must be a CPU tensor"); + AT_ASSERTM(rois.device().is_cpu(), "rois must be a CPU tensor"); + + at::TensorArg grad_t{grad, "grad", 1}, rois_t{rois, "rois", 2}; + + at::CheckedFrom c = "ROIAlignRotated_backward_cpu"; + at::checkAllSameType(c, {grad_t, rois_t}); + + at::Tensor grad_input = + at::zeros({batch_size, channels, height, width}, grad.options()); + + // handle possibly empty gradients + if (grad.numel() == 0) { + return grad_input; + } + + // get stride values to ensure indexing into gradients is correct. + int n_stride = grad.stride(0); + int c_stride = grad.stride(1); + int h_stride = grad.stride(2); + int w_stride = grad.stride(3); + + auto rois_ = rois.contiguous(); + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + grad.scalar_type(), "ROIAlignRotated_forward", [&] { + ROIAlignRotatedBackward( + grad.numel(), + grad.data_ptr(), + spatial_scale, + channels, + height, + width, + pooled_height, + pooled_width, + sampling_ratio, + grad_input.data_ptr(), + rois_.data_ptr(), + n_stride, + c_stride, + h_stride, + w_stride); + }); + return grad_input; +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated_cuda.cu b/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated_cuda.cu new file mode 100644 index 0000000000000000000000000000000000000000..fca186519143b168a912c880a4cf495a0a5a9322 --- /dev/null +++ b/detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated_cuda.cu @@ -0,0 +1,443 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#include +#include +#include +#include + +// TODO make it in a common file +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \ + i += blockDim.x * gridDim.x) + +// Note: this implementation originates from the Caffe2 ROIAlignRotated Op +// and PyTorch ROIAlign (non-rotated) Op implementations. +// The key difference between this implementation and those ones is +// we don't do "legacy offset" in this version, as there aren't many previous +// works, if any, using the "legacy" ROIAlignRotated Op. +// This would make the interface a bit cleaner. + +namespace detectron2 { + +namespace { + +template +__device__ T bilinear_interpolate( + const T* input, + const int height, + const int width, + T y, + T x) { + // deal with cases that inverse elements are out of feature map boundary + if (y < -1.0 || y > height || x < -1.0 || x > width) { + // empty + return 0; + } + + if (y < 0) { + y = 0; + } + + if (x < 0) { + x = 0; + } + + int y_low = (int)y; + int x_low = (int)x; + int y_high; + int x_high; + + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = (T)y_low; + } else { + y_high = y_low + 1; + } + + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = (T)x_low; + } else { + x_high = x_low + 1; + } + + T ly = y - y_low; + T lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + // do bilinear interpolation + T v1 = input[y_low * width + x_low]; + T v2 = input[y_low * width + x_high]; + T v3 = input[y_high * width + x_low]; + T v4 = input[y_high * width + x_high]; + T w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + + T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + + return val; +} + +template +__device__ void bilinear_interpolate_gradient( + const int height, + const int width, + T y, + T x, + T& w1, + T& w2, + T& w3, + T& w4, + int& x_low, + int& x_high, + int& y_low, + int& y_high) { + // deal with cases that inverse elements are out of feature map boundary + if (y < -1.0 || y > height || x < -1.0 || x > width) { + // empty + w1 = w2 = w3 = w4 = 0.; + x_low = x_high = y_low = y_high = -1; + return; + } + + if (y < 0) { + y = 0; + } + + if (x < 0) { + x = 0; + } + + y_low = (int)y; + x_low = (int)x; + + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = (T)y_low; + } else { + y_high = y_low + 1; + } + + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = (T)x_low; + } else { + x_high = x_low + 1; + } + + T ly = y - y_low; + T lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + + // reference in forward + // T v1 = input[y_low * width + x_low]; + // T v2 = input[y_low * width + x_high]; + // T v3 = input[y_high * width + x_low]; + // T v4 = input[y_high * width + x_high]; + // T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + + w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + + return; +} + +} // namespace + +template +__global__ void RoIAlignRotatedForward( + const int nthreads, + const T* input, + const T spatial_scale, + const int channels, + const int height, + const int width, + const int pooled_height, + const int pooled_width, + const int sampling_ratio, + const T* rois, + T* top_data) { + CUDA_1D_KERNEL_LOOP(index, nthreads) { + // (n, c, ph, pw) is an element in the pooled output + int pw = index % pooled_width; + int ph = (index / pooled_width) % pooled_height; + int c = (index / pooled_width / pooled_height) % channels; + int n = index / pooled_width / pooled_height / channels; + + const T* current_roi = rois + n * 6; + int roi_batch_ind = current_roi[0]; + + // Do not use rounding; this implementation detail is critical + // ROIAlignRotated supports align == true, i.e., continuous coordinate + // by default, thus the 0.5 offset + T offset = (T)0.5; + T roi_center_w = current_roi[1] * spatial_scale - offset; + T roi_center_h = current_roi[2] * spatial_scale - offset; + T roi_width = current_roi[3] * spatial_scale; + T roi_height = current_roi[4] * spatial_scale; + T theta = current_roi[5] * M_PI / 180.0; + T cos_theta = cos(theta); + T sin_theta = sin(theta); + + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + + const T* offset_input = + input + (roi_batch_ind * channels + c) * height * width; + + // We use roi_bin_grid to sample the grid and mimic integral + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); // e.g., = 2 + int roi_bin_grid_w = + (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); + + // roi_start_h and roi_start_w are computed wrt the center of RoI (x, y). + // Appropriate translation needs to be applied after. + T roi_start_h = -roi_height / 2.0; + T roi_start_w = -roi_width / 2.0; + + // We do average (inte gral) pooling inside a bin + const T count = max(roi_bin_grid_h * roi_bin_grid_w, 1); // e.g. = 4 + + T output_val = 0.; + for (int iy = 0; iy < roi_bin_grid_h; iy++) // e.g., iy = 0, 1 + { + const T yy = roi_start_h + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); // e.g., 0.5, 1.5 + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T xx = roi_start_w + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + + // Rotate by theta around the center and translate + T y = yy * cos_theta - xx * sin_theta + roi_center_h; + T x = yy * sin_theta + xx * cos_theta + roi_center_w; + + T val = bilinear_interpolate(offset_input, height, width, y, x); + output_val += val; + } + } + output_val /= count; + + top_data[index] = output_val; + } +} + +template +__global__ void RoIAlignRotatedBackwardFeature( + const int nthreads, + const T* top_diff, + const int num_rois, + const T spatial_scale, + const int channels, + const int height, + const int width, + const int pooled_height, + const int pooled_width, + const int sampling_ratio, + T* bottom_diff, + const T* rois) { + CUDA_1D_KERNEL_LOOP(index, nthreads) { + // (n, c, ph, pw) is an element in the pooled output + int pw = index % pooled_width; + int ph = (index / pooled_width) % pooled_height; + int c = (index / pooled_width / pooled_height) % channels; + int n = index / pooled_width / pooled_height / channels; + + const T* current_roi = rois + n * 6; + int roi_batch_ind = current_roi[0]; + + // Do not use rounding; this implementation detail is critical + // ROIAlignRotated supports align == true, i.e., continuous coordinate + // by default, thus the 0.5 offset + T offset = (T)0.5; + T roi_center_w = current_roi[1] * spatial_scale - offset; + T roi_center_h = current_roi[2] * spatial_scale - offset; + T roi_width = current_roi[3] * spatial_scale; + T roi_height = current_roi[4] * spatial_scale; + T theta = current_roi[5] * M_PI / 180.0; + T cos_theta = cos(theta); + T sin_theta = sin(theta); + + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + + T* offset_bottom_diff = + bottom_diff + (roi_batch_ind * channels + c) * height * width; + + int top_offset = (n * channels + c) * pooled_height * pooled_width; + const T* offset_top_diff = top_diff + top_offset; + const T top_diff_this_bin = offset_top_diff[ph * pooled_width + pw]; + + // We use roi_bin_grid to sample the grid and mimic integral + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); // e.g., = 2 + int roi_bin_grid_w = + (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); + + // roi_start_h and roi_start_w are computed wrt the center of RoI (x, y). + // Appropriate translation needs to be applied after. + T roi_start_h = -roi_height / 2.0; + T roi_start_w = -roi_width / 2.0; + + // We do average (integral) pooling inside a bin + const T count = roi_bin_grid_h * roi_bin_grid_w; // e.g. = 4 + + for (int iy = 0; iy < roi_bin_grid_h; iy++) // e.g., iy = 0, 1 + { + const T yy = roi_start_h + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); // e.g., 0.5, 1.5 + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T xx = roi_start_w + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + + // Rotate by theta around the center and translate + T y = yy * cos_theta - xx * sin_theta + roi_center_h; + T x = yy * sin_theta + xx * cos_theta + roi_center_w; + + T w1, w2, w3, w4; + int x_low, x_high, y_low, y_high; + + bilinear_interpolate_gradient( + height, width, y, x, w1, w2, w3, w4, x_low, x_high, y_low, y_high); + + T g1 = top_diff_this_bin * w1 / count; + T g2 = top_diff_this_bin * w2 / count; + T g3 = top_diff_this_bin * w3 / count; + T g4 = top_diff_this_bin * w4 / count; + + if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) { + atomicAdd( + offset_bottom_diff + y_low * width + x_low, static_cast(g1)); + atomicAdd( + offset_bottom_diff + y_low * width + x_high, static_cast(g2)); + atomicAdd( + offset_bottom_diff + y_high * width + x_low, static_cast(g3)); + atomicAdd( + offset_bottom_diff + y_high * width + x_high, static_cast(g4)); + } // if + } // ix + } // iy + } // CUDA_1D_KERNEL_LOOP +} // RoIAlignRotatedBackward + +at::Tensor ROIAlignRotated_forward_cuda( + const at::Tensor& input, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int sampling_ratio) { + AT_ASSERTM(input.device().is_cuda(), "input must be a CUDA tensor"); + AT_ASSERTM(rois.device().is_cuda(), "rois must be a CUDA tensor"); + at::TensorArg input_t{input, "input", 1}, rois_t{rois, "rois", 2}; + + at::CheckedFrom c = "ROIAlignRotated_forward_cuda"; + at::checkAllSameGPU(c, {input_t, rois_t}); + at::checkAllSameType(c, {input_t, rois_t}); + at::cuda::CUDAGuard device_guard(input.device()); + + auto num_rois = rois.size(0); + auto channels = input.size(1); + auto height = input.size(2); + auto width = input.size(3); + + auto output = at::empty( + {num_rois, channels, pooled_height, pooled_width}, input.options()); + auto output_size = num_rois * pooled_height * pooled_width * channels; + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + dim3 grid(std::min( + at::cuda::ATenCeilDiv( + static_cast(output_size), static_cast(512)), + static_cast(4096))); + dim3 block(512); + + if (output.numel() == 0) { + AT_CUDA_CHECK(cudaGetLastError()); + return output; + } + + auto input_ = input.contiguous(), rois_ = rois.contiguous(); + AT_DISPATCH_FLOATING_TYPES( + input.scalar_type(), "ROIAlignRotated_forward", [&] { + RoIAlignRotatedForward<<>>( + output_size, + input_.data_ptr(), + spatial_scale, + channels, + height, + width, + pooled_height, + pooled_width, + sampling_ratio, + rois_.data_ptr(), + output.data_ptr()); + }); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + return output; +} + +// TODO remove the dependency on input and use instead its sizes -> save memory +at::Tensor ROIAlignRotated_backward_cuda( + const at::Tensor& grad, + const at::Tensor& rois, + const float spatial_scale, + const int pooled_height, + const int pooled_width, + const int batch_size, + const int channels, + const int height, + const int width, + const int sampling_ratio) { + AT_ASSERTM(grad.device().is_cuda(), "grad must be a CUDA tensor"); + AT_ASSERTM(rois.device().is_cuda(), "rois must be a CUDA tensor"); + + at::TensorArg grad_t{grad, "grad", 1}, rois_t{rois, "rois", 2}; + at::CheckedFrom c = "ROIAlign_backward_cuda"; + at::checkAllSameGPU(c, {grad_t, rois_t}); + at::checkAllSameType(c, {grad_t, rois_t}); + at::cuda::CUDAGuard device_guard(grad.device()); + + auto num_rois = rois.size(0); + auto grad_input = + at::zeros({batch_size, channels, height, width}, grad.options()); + + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + dim3 grid(std::min( + at::cuda::ATenCeilDiv( + static_cast(grad.numel()), static_cast(512)), + static_cast(4096))); + dim3 block(512); + + // handle possibly empty gradients + if (grad.numel() == 0) { + AT_CUDA_CHECK(cudaGetLastError()); + return grad_input; + } + + auto grad_ = grad.contiguous(), rois_ = rois.contiguous(); + AT_DISPATCH_FLOATING_TYPES( + grad.scalar_type(), "ROIAlignRotated_backward", [&] { + RoIAlignRotatedBackwardFeature<<>>( + grad.numel(), + grad_.data_ptr(), + num_rois, + spatial_scale, + channels, + height, + width, + pooled_height, + pooled_width, + sampling_ratio, + grad_input.data_ptr(), + rois_.data_ptr()); + }); + AT_CUDA_CHECK(cudaGetLastError()); + return grad_input; +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated.h b/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated.h new file mode 100644 index 0000000000000000000000000000000000000000..b65888b1be11881a776827b5212f08b8f63138f9 --- /dev/null +++ b/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated.h @@ -0,0 +1,35 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#pragma once +#include + +namespace detectron2 { + +at::Tensor box_iou_rotated_cpu( + const at::Tensor& boxes1, + const at::Tensor& boxes2); + +#if defined(WITH_CUDA) || defined(WITH_HIP) +at::Tensor box_iou_rotated_cuda( + const at::Tensor& boxes1, + const at::Tensor& boxes2); +#endif + +// Interface for Python +// inline is needed to prevent multiple function definitions when this header is +// included by different cpps +inline at::Tensor box_iou_rotated( + const at::Tensor& boxes1, + const at::Tensor& boxes2) { + assert(boxes1.device().is_cuda() == boxes2.device().is_cuda()); + if (boxes1.device().is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + return box_iou_rotated_cuda(boxes1.contiguous(), boxes2.contiguous()); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + + return box_iou_rotated_cpu(boxes1.contiguous(), boxes2.contiguous()); +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cpu.cpp b/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cpu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c843487b5fa4e8077dd27402ec99009266ddda8d --- /dev/null +++ b/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cpu.cpp @@ -0,0 +1,39 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#include "box_iou_rotated.h" +#include "box_iou_rotated_utils.h" + +namespace detectron2 { + +template +void box_iou_rotated_cpu_kernel( + const at::Tensor& boxes1, + const at::Tensor& boxes2, + at::Tensor& ious) { + auto num_boxes1 = boxes1.size(0); + auto num_boxes2 = boxes2.size(0); + + for (int i = 0; i < num_boxes1; i++) { + for (int j = 0; j < num_boxes2; j++) { + ious[i * num_boxes2 + j] = single_box_iou_rotated( + boxes1[i].data_ptr(), boxes2[j].data_ptr()); + } + } +} + +at::Tensor box_iou_rotated_cpu( + // input must be contiguous: + const at::Tensor& boxes1, + const at::Tensor& boxes2) { + auto num_boxes1 = boxes1.size(0); + auto num_boxes2 = boxes2.size(0); + at::Tensor ious = + at::empty({num_boxes1 * num_boxes2}, boxes1.options().dtype(at::kFloat)); + + box_iou_rotated_cpu_kernel(boxes1, boxes2, ious); + + // reshape from 1d array to 2d array + auto shape = std::vector{num_boxes1, num_boxes2}; + return ious.reshape(shape); +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cuda.cu b/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cuda.cu new file mode 100644 index 0000000000000000000000000000000000000000..952710e53041187907fbd113f8d0d0fa24134a86 --- /dev/null +++ b/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cuda.cu @@ -0,0 +1,130 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#include +#include +#include +#include +#include "box_iou_rotated_utils.h" + +namespace detectron2 { + +// 2D block with 32 * 16 = 512 threads per block +const int BLOCK_DIM_X = 32; +const int BLOCK_DIM_Y = 16; + +template +__global__ void box_iou_rotated_cuda_kernel( + const int n_boxes1, + const int n_boxes2, + const T* dev_boxes1, + const T* dev_boxes2, + T* dev_ious) { + const int row_start = blockIdx.x * blockDim.x; + const int col_start = blockIdx.y * blockDim.y; + + const int row_size = min(n_boxes1 - row_start, blockDim.x); + const int col_size = min(n_boxes2 - col_start, blockDim.y); + + __shared__ float block_boxes1[BLOCK_DIM_X * 5]; + __shared__ float block_boxes2[BLOCK_DIM_Y * 5]; + + // It's safe to copy using threadIdx.x since BLOCK_DIM_X >= BLOCK_DIM_Y + if (threadIdx.x < row_size && threadIdx.y == 0) { + block_boxes1[threadIdx.x * 5 + 0] = + dev_boxes1[(row_start + threadIdx.x) * 5 + 0]; + block_boxes1[threadIdx.x * 5 + 1] = + dev_boxes1[(row_start + threadIdx.x) * 5 + 1]; + block_boxes1[threadIdx.x * 5 + 2] = + dev_boxes1[(row_start + threadIdx.x) * 5 + 2]; + block_boxes1[threadIdx.x * 5 + 3] = + dev_boxes1[(row_start + threadIdx.x) * 5 + 3]; + block_boxes1[threadIdx.x * 5 + 4] = + dev_boxes1[(row_start + threadIdx.x) * 5 + 4]; + } + + if (threadIdx.x < col_size && threadIdx.y == 0) { + block_boxes2[threadIdx.x * 5 + 0] = + dev_boxes2[(col_start + threadIdx.x) * 5 + 0]; + block_boxes2[threadIdx.x * 5 + 1] = + dev_boxes2[(col_start + threadIdx.x) * 5 + 1]; + block_boxes2[threadIdx.x * 5 + 2] = + dev_boxes2[(col_start + threadIdx.x) * 5 + 2]; + block_boxes2[threadIdx.x * 5 + 3] = + dev_boxes2[(col_start + threadIdx.x) * 5 + 3]; + block_boxes2[threadIdx.x * 5 + 4] = + dev_boxes2[(col_start + threadIdx.x) * 5 + 4]; + } + __syncthreads(); + + if (threadIdx.x < row_size && threadIdx.y < col_size) { + int offset = (row_start + threadIdx.x) * n_boxes2 + col_start + threadIdx.y; + dev_ious[offset] = single_box_iou_rotated( + block_boxes1 + threadIdx.x * 5, block_boxes2 + threadIdx.y * 5); + } +} + +at::Tensor box_iou_rotated_cuda( + // input must be contiguous + const at::Tensor& boxes1, + const at::Tensor& boxes2) { + using scalar_t = float; + AT_ASSERTM( + boxes1.scalar_type() == at::kFloat, "boxes1 must be a float tensor"); + AT_ASSERTM( + boxes2.scalar_type() == at::kFloat, "boxes2 must be a float tensor"); + AT_ASSERTM(boxes1.is_cuda(), "boxes1 must be a CUDA tensor"); + AT_ASSERTM(boxes2.is_cuda(), "boxes2 must be a CUDA tensor"); + at::cuda::CUDAGuard device_guard(boxes1.device()); + + auto num_boxes1 = boxes1.size(0); + auto num_boxes2 = boxes2.size(0); + + at::Tensor ious = + at::empty({num_boxes1 * num_boxes2}, boxes1.options().dtype(at::kFloat)); + + bool transpose = false; + if (num_boxes1 > 0 && num_boxes2 > 0) { + scalar_t *data1 = boxes1.data_ptr(), + *data2 = boxes2.data_ptr(); + + if (num_boxes2 > 65535 * BLOCK_DIM_Y) { + AT_ASSERTM( + num_boxes1 <= 65535 * BLOCK_DIM_Y, + "Too many boxes for box_iou_rotated_cuda!"); + // x dim is allowed to be large, but y dim cannot, + // so we transpose the two to avoid "invalid configuration argument" + // error. We assume one of them is small. Otherwise the result is hard to + // fit in memory anyway. + std::swap(num_boxes1, num_boxes2); + std::swap(data1, data2); + transpose = true; + } + + const int blocks_x = + at::cuda::ATenCeilDiv(static_cast(num_boxes1), BLOCK_DIM_X); + const int blocks_y = + at::cuda::ATenCeilDiv(static_cast(num_boxes2), BLOCK_DIM_Y); + + dim3 blocks(blocks_x, blocks_y); + dim3 threads(BLOCK_DIM_X, BLOCK_DIM_Y); + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + box_iou_rotated_cuda_kernel<<>>( + num_boxes1, + num_boxes2, + data1, + data2, + (scalar_t*)ious.data_ptr()); + + AT_CUDA_CHECK(cudaGetLastError()); + } + + // reshape from 1d array to 2d array + auto shape = std::vector{num_boxes1, num_boxes2}; + if (transpose) { + return ious.view(shape).t(); + } else { + return ious.view(shape); + } +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_utils.h b/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_utils.h new file mode 100644 index 0000000000000000000000000000000000000000..b54a5dde2ca11a74d29c4d8adb7fe1634f5baf9c --- /dev/null +++ b/detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_utils.h @@ -0,0 +1,370 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#pragma once + +#include +#include + +#if defined(__CUDACC__) || __HCC__ == 1 || __HIP__ == 1 +// Designates functions callable from the host (CPU) and the device (GPU) +#define HOST_DEVICE __host__ __device__ +#define HOST_DEVICE_INLINE HOST_DEVICE __forceinline__ +#else +#include +#define HOST_DEVICE +#define HOST_DEVICE_INLINE HOST_DEVICE inline +#endif + +namespace detectron2 { + +namespace { + +template +struct RotatedBox { + T x_ctr, y_ctr, w, h, a; +}; + +template +struct Point { + T x, y; + HOST_DEVICE_INLINE Point(const T& px = 0, const T& py = 0) : x(px), y(py) {} + HOST_DEVICE_INLINE Point operator+(const Point& p) const { + return Point(x + p.x, y + p.y); + } + HOST_DEVICE_INLINE Point& operator+=(const Point& p) { + x += p.x; + y += p.y; + return *this; + } + HOST_DEVICE_INLINE Point operator-(const Point& p) const { + return Point(x - p.x, y - p.y); + } + HOST_DEVICE_INLINE Point operator*(const T coeff) const { + return Point(x * coeff, y * coeff); + } +}; + +template +HOST_DEVICE_INLINE T dot_2d(const Point& A, const Point& B) { + return A.x * B.x + A.y * B.y; +} + +// R: result type. can be different from input type +template +HOST_DEVICE_INLINE R cross_2d(const Point& A, const Point& B) { + return static_cast(A.x) * static_cast(B.y) - + static_cast(B.x) * static_cast(A.y); +} + +template +HOST_DEVICE_INLINE void get_rotated_vertices( + const RotatedBox& box, + Point (&pts)[4]) { + // M_PI / 180. == 0.01745329251 + double theta = box.a * 0.01745329251; + T cosTheta2 = (T)cos(theta) * 0.5f; + T sinTheta2 = (T)sin(theta) * 0.5f; + + // y: top --> down; x: left --> right + pts[0].x = box.x_ctr + sinTheta2 * box.h + cosTheta2 * box.w; + pts[0].y = box.y_ctr + cosTheta2 * box.h - sinTheta2 * box.w; + pts[1].x = box.x_ctr - sinTheta2 * box.h + cosTheta2 * box.w; + pts[1].y = box.y_ctr - cosTheta2 * box.h - sinTheta2 * box.w; + pts[2].x = 2 * box.x_ctr - pts[0].x; + pts[2].y = 2 * box.y_ctr - pts[0].y; + pts[3].x = 2 * box.x_ctr - pts[1].x; + pts[3].y = 2 * box.y_ctr - pts[1].y; +} + +template +HOST_DEVICE_INLINE int get_intersection_points( + const Point (&pts1)[4], + const Point (&pts2)[4], + Point (&intersections)[24]) { + // Line vector + // A line from p1 to p2 is: p1 + (p2-p1)*t, t=[0,1] + Point vec1[4], vec2[4]; + for (int i = 0; i < 4; i++) { + vec1[i] = pts1[(i + 1) % 4] - pts1[i]; + vec2[i] = pts2[(i + 1) % 4] - pts2[i]; + } + + // When computing the intersection area, it doesn't hurt if we have + // more (duplicated/approximate) intersections/vertices than needed, + // while it can cause drastic difference if we miss an intersection/vertex. + // Therefore, we add an epsilon to relax the comparisons between + // the float point numbers that decide the intersection points. + double EPS = 1e-5; + + // Line test - test all line combos for intersection + int num = 0; // number of intersections + for (int i = 0; i < 4; i++) { + for (int j = 0; j < 4; j++) { + // Solve for 2x2 Ax=b + T det = cross_2d(vec2[j], vec1[i]); + + // This takes care of parallel lines + if (fabs(det) <= 1e-14) { + continue; + } + + auto vec12 = pts2[j] - pts1[i]; + + T t1 = cross_2d(vec2[j], vec12) / det; + T t2 = cross_2d(vec1[i], vec12) / det; + + if (t1 > -EPS && t1 < 1.0f + EPS && t2 > -EPS && t2 < 1.0f + EPS) { + intersections[num++] = pts1[i] + vec1[i] * t1; + } + } + } + + // Check for vertices of rect1 inside rect2 + { + const auto& AB = vec2[0]; + const auto& DA = vec2[3]; + auto ABdotAB = dot_2d(AB, AB); + auto ADdotAD = dot_2d(DA, DA); + for (int i = 0; i < 4; i++) { + // assume ABCD is the rectangle, and P is the point to be judged + // P is inside ABCD iff. P's projection on AB lies within AB + // and P's projection on AD lies within AD + + auto AP = pts1[i] - pts2[0]; + + auto APdotAB = dot_2d(AP, AB); + auto APdotAD = -dot_2d(AP, DA); + + if ((APdotAB > -EPS) && (APdotAD > -EPS) && (APdotAB < ABdotAB + EPS) && + (APdotAD < ADdotAD + EPS)) { + intersections[num++] = pts1[i]; + } + } + } + + // Reverse the check - check for vertices of rect2 inside rect1 + { + const auto& AB = vec1[0]; + const auto& DA = vec1[3]; + auto ABdotAB = dot_2d(AB, AB); + auto ADdotAD = dot_2d(DA, DA); + for (int i = 0; i < 4; i++) { + auto AP = pts2[i] - pts1[0]; + + auto APdotAB = dot_2d(AP, AB); + auto APdotAD = -dot_2d(AP, DA); + + if ((APdotAB > -EPS) && (APdotAD > -EPS) && (APdotAB < ABdotAB + EPS) && + (APdotAD < ADdotAD + EPS)) { + intersections[num++] = pts2[i]; + } + } + } + + return num; +} + +template +HOST_DEVICE_INLINE int convex_hull_graham( + const Point (&p)[24], + const int& num_in, + Point (&q)[24], + bool shift_to_zero = false) { + assert(num_in >= 2); + + // Step 1: + // Find point with minimum y + // if more than 1 points have the same minimum y, + // pick the one with the minimum x. + int t = 0; + for (int i = 1; i < num_in; i++) { + if (p[i].y < p[t].y || (p[i].y == p[t].y && p[i].x < p[t].x)) { + t = i; + } + } + auto& start = p[t]; // starting point + + // Step 2: + // Subtract starting point from every points (for sorting in the next step) + for (int i = 0; i < num_in; i++) { + q[i] = p[i] - start; + } + + // Swap the starting point to position 0 + auto tmp = q[0]; + q[0] = q[t]; + q[t] = tmp; + + // Step 3: + // Sort point 1 ~ num_in according to their relative cross-product values + // (essentially sorting according to angles) + // If the angles are the same, sort according to their distance to origin + T dist[24]; +#if defined(__CUDACC__) || __HCC__ == 1 || __HIP__ == 1 + // compute distance to origin before sort, and sort them together with the + // points + for (int i = 0; i < num_in; i++) { + dist[i] = dot_2d(q[i], q[i]); + } + + // CUDA version + // In the future, we can potentially use thrust + // for sorting here to improve speed (though not guaranteed) + for (int i = 1; i < num_in - 1; i++) { + for (int j = i + 1; j < num_in; j++) { + T crossProduct = cross_2d(q[i], q[j]); + if ((crossProduct < -1e-6) || + (fabs(crossProduct) < 1e-6 && dist[i] > dist[j])) { + auto q_tmp = q[i]; + q[i] = q[j]; + q[j] = q_tmp; + auto dist_tmp = dist[i]; + dist[i] = dist[j]; + dist[j] = dist_tmp; + } + } + } +#else + // CPU version + std::sort( + q + 1, q + num_in, [](const Point& A, const Point& B) -> bool { + T temp = cross_2d(A, B); + if (fabs(temp) < 1e-6) { + return dot_2d(A, A) < dot_2d(B, B); + } else { + return temp > 0; + } + }); + // compute distance to origin after sort, since the points are now different. + for (int i = 0; i < num_in; i++) { + dist[i] = dot_2d(q[i], q[i]); + } +#endif + + // Step 4: + // Make sure there are at least 2 points (that don't overlap with each other) + // in the stack + int k; // index of the non-overlapped second point + for (k = 1; k < num_in; k++) { + if (dist[k] > 1e-8) { + break; + } + } + if (k == num_in) { + // We reach the end, which means the convex hull is just one point + q[0] = p[t]; + return 1; + } + q[1] = q[k]; + int m = 2; // 2 points in the stack + // Step 5: + // Finally we can start the scanning process. + // When a non-convex relationship between the 3 points is found + // (either concave shape or duplicated points), + // we pop the previous point from the stack + // until the 3-point relationship is convex again, or + // until the stack only contains two points + for (int i = k + 1; i < num_in; i++) { + while (m > 1) { + auto q1 = q[i] - q[m - 2], q2 = q[m - 1] - q[m - 2]; + // cross_2d() uses FMA and therefore computes round(round(q1.x*q2.y) - + // q2.x*q1.y) So it may not return 0 even when q1==q2. Therefore we + // compare round(q1.x*q2.y) and round(q2.x*q1.y) directly. (round means + // round to nearest floating point). + if (q1.x * q2.y >= q2.x * q1.y) + m--; + else + break; + } + // Using double also helps, but float can solve the issue for now. + // while (m > 1 && cross_2d(q[i] - q[m - 2], q[m - 1] - q[m - 2]) + // >= 0) { + // m--; + // } + q[m++] = q[i]; + } + + // Step 6 (Optional): + // In general sense we need the original coordinates, so we + // need to shift the points back (reverting Step 2) + // But if we're only interested in getting the area/perimeter of the shape + // We can simply return. + if (!shift_to_zero) { + for (int i = 0; i < m; i++) { + q[i] += start; + } + } + + return m; +} + +template +HOST_DEVICE_INLINE T polygon_area(const Point (&q)[24], const int& m) { + if (m <= 2) { + return 0; + } + + T area = 0; + for (int i = 1; i < m - 1; i++) { + area += fabs(cross_2d(q[i] - q[0], q[i + 1] - q[0])); + } + + return area / 2.0; +} + +template +HOST_DEVICE_INLINE T rotated_boxes_intersection( + const RotatedBox& box1, + const RotatedBox& box2) { + // There are up to 4 x 4 + 4 + 4 = 24 intersections (including dups) returned + // from rotated_rect_intersection_pts + Point intersectPts[24], orderedPts[24]; + + Point pts1[4]; + Point pts2[4]; + get_rotated_vertices(box1, pts1); + get_rotated_vertices(box2, pts2); + + int num = get_intersection_points(pts1, pts2, intersectPts); + + if (num <= 2) { + return 0.0; + } + + // Convex Hull to order the intersection points in clockwise order and find + // the contour area. + int num_convex = convex_hull_graham(intersectPts, num, orderedPts, true); + return polygon_area(orderedPts, num_convex); +} + +} // namespace + +template +HOST_DEVICE_INLINE T +single_box_iou_rotated(T const* const box1_raw, T const* const box2_raw) { + // shift center to the middle point to achieve higher precision in result + RotatedBox box1, box2; + auto center_shift_x = (box1_raw[0] + box2_raw[0]) / 2.0; + auto center_shift_y = (box1_raw[1] + box2_raw[1]) / 2.0; + box1.x_ctr = box1_raw[0] - center_shift_x; + box1.y_ctr = box1_raw[1] - center_shift_y; + box1.w = box1_raw[2]; + box1.h = box1_raw[3]; + box1.a = box1_raw[4]; + box2.x_ctr = box2_raw[0] - center_shift_x; + box2.y_ctr = box2_raw[1] - center_shift_y; + box2.w = box2_raw[2]; + box2.h = box2_raw[3]; + box2.a = box2_raw[4]; + + T area1 = box1.w * box1.h; + T area2 = box2.w * box2.h; + if (area1 < 1e-14 || area2 < 1e-14) { + return 0.f; + } + + T intersection = rotated_boxes_intersection(box1, box2); + T iou = intersection / (area1 + area2 - intersection); + return iou; +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/cocoeval/cocoeval.cpp b/detectron2/layers/csrc/cocoeval/cocoeval.cpp new file mode 100644 index 0000000000000000000000000000000000000000..0a5b7b907c06720fefc77b0dfd921b8ec3ecf2be --- /dev/null +++ b/detectron2/layers/csrc/cocoeval/cocoeval.cpp @@ -0,0 +1,507 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#include "cocoeval.h" +#include +#include +#include +#include + +using namespace pybind11::literals; + +namespace detectron2 { + +namespace COCOeval { + +// Sort detections from highest score to lowest, such that +// detection_instances[detection_sorted_indices[t]] >= +// detection_instances[detection_sorted_indices[t+1]]. Use stable_sort to match +// original COCO API +void SortInstancesByDetectionScore( + const std::vector& detection_instances, + std::vector* detection_sorted_indices) { + detection_sorted_indices->resize(detection_instances.size()); + std::iota( + detection_sorted_indices->begin(), detection_sorted_indices->end(), 0); + std::stable_sort( + detection_sorted_indices->begin(), + detection_sorted_indices->end(), + [&detection_instances](size_t j1, size_t j2) { + return detection_instances[j1].score > detection_instances[j2].score; + }); +} + +// Partition the ground truth objects based on whether or not to ignore them +// based on area +void SortInstancesByIgnore( + const std::array& area_range, + const std::vector& ground_truth_instances, + std::vector* ground_truth_sorted_indices, + std::vector* ignores) { + ignores->clear(); + ignores->reserve(ground_truth_instances.size()); + for (auto o : ground_truth_instances) { + ignores->push_back( + o.ignore || o.area < area_range[0] || o.area > area_range[1]); + } + + ground_truth_sorted_indices->resize(ground_truth_instances.size()); + std::iota( + ground_truth_sorted_indices->begin(), + ground_truth_sorted_indices->end(), + 0); + std::stable_sort( + ground_truth_sorted_indices->begin(), + ground_truth_sorted_indices->end(), + [&ignores](size_t j1, size_t j2) { + return (int)(*ignores)[j1] < (int)(*ignores)[j2]; + }); +} + +// For each IOU threshold, greedily match each detected instance to a ground +// truth instance (if possible) and store the results +void MatchDetectionsToGroundTruth( + const std::vector& detection_instances, + const std::vector& detection_sorted_indices, + const std::vector& ground_truth_instances, + const std::vector& ground_truth_sorted_indices, + const std::vector& ignores, + const std::vector>& ious, + const std::vector& iou_thresholds, + const std::array& area_range, + ImageEvaluation* results) { + // Initialize memory to store return data matches and ignore + const int num_iou_thresholds = iou_thresholds.size(); + const int num_ground_truth = ground_truth_sorted_indices.size(); + const int num_detections = detection_sorted_indices.size(); + std::vector ground_truth_matches( + num_iou_thresholds * num_ground_truth, 0); + std::vector& detection_matches = results->detection_matches; + std::vector& detection_ignores = results->detection_ignores; + std::vector& ground_truth_ignores = results->ground_truth_ignores; + detection_matches.resize(num_iou_thresholds * num_detections, 0); + detection_ignores.resize(num_iou_thresholds * num_detections, false); + ground_truth_ignores.resize(num_ground_truth); + for (auto g = 0; g < num_ground_truth; ++g) { + ground_truth_ignores[g] = ignores[ground_truth_sorted_indices[g]]; + } + + for (auto t = 0; t < num_iou_thresholds; ++t) { + for (auto d = 0; d < num_detections; ++d) { + // information about best match so far (match=-1 -> unmatched) + double best_iou = std::min(iou_thresholds[t], 1 - 1e-10); + int match = -1; + for (auto g = 0; g < num_ground_truth; ++g) { + // if this ground truth instance is already matched and not a + // crowd, it cannot be matched to another detection + if (ground_truth_matches[t * num_ground_truth + g] > 0 && + !ground_truth_instances[ground_truth_sorted_indices[g]].is_crowd) { + continue; + } + + // if detected instance matched to a regular ground truth + // instance, we can break on the first ground truth instance + // tagged as ignore (because they are sorted by the ignore tag) + if (match >= 0 && !ground_truth_ignores[match] && + ground_truth_ignores[g]) { + break; + } + + // if IOU overlap is the best so far, store the match appropriately + if (ious[d][ground_truth_sorted_indices[g]] >= best_iou) { + best_iou = ious[d][ground_truth_sorted_indices[g]]; + match = g; + } + } + // if match was made, store id of match for both detection and + // ground truth + if (match >= 0) { + detection_ignores[t * num_detections + d] = ground_truth_ignores[match]; + detection_matches[t * num_detections + d] = + ground_truth_instances[ground_truth_sorted_indices[match]].id; + ground_truth_matches[t * num_ground_truth + match] = + detection_instances[detection_sorted_indices[d]].id; + } + + // set unmatched detections outside of area range to ignore + const InstanceAnnotation& detection = + detection_instances[detection_sorted_indices[d]]; + detection_ignores[t * num_detections + d] = + detection_ignores[t * num_detections + d] || + (detection_matches[t * num_detections + d] == 0 && + (detection.area < area_range[0] || detection.area > area_range[1])); + } + } + + // store detection score results + results->detection_scores.resize(detection_sorted_indices.size()); + for (size_t d = 0; d < detection_sorted_indices.size(); ++d) { + results->detection_scores[d] = + detection_instances[detection_sorted_indices[d]].score; + } +} + +std::vector EvaluateImages( + const std::vector>& area_ranges, + int max_detections, + const std::vector& iou_thresholds, + const ImageCategoryInstances>& image_category_ious, + const ImageCategoryInstances& + image_category_ground_truth_instances, + const ImageCategoryInstances& + image_category_detection_instances) { + const int num_area_ranges = area_ranges.size(); + const int num_images = image_category_ground_truth_instances.size(); + const int num_categories = + image_category_ious.size() > 0 ? image_category_ious[0].size() : 0; + std::vector detection_sorted_indices; + std::vector ground_truth_sorted_indices; + std::vector ignores; + std::vector results_all( + num_images * num_area_ranges * num_categories); + + // Store results for each image, category, and area range combination. Results + // for each IOU threshold are packed into the same ImageEvaluation object + for (auto i = 0; i < num_images; ++i) { + for (auto c = 0; c < num_categories; ++c) { + const std::vector& ground_truth_instances = + image_category_ground_truth_instances[i][c]; + const std::vector& detection_instances = + image_category_detection_instances[i][c]; + + SortInstancesByDetectionScore( + detection_instances, &detection_sorted_indices); + if ((int)detection_sorted_indices.size() > max_detections) { + detection_sorted_indices.resize(max_detections); + } + + for (size_t a = 0; a < area_ranges.size(); ++a) { + SortInstancesByIgnore( + area_ranges[a], + ground_truth_instances, + &ground_truth_sorted_indices, + &ignores); + + MatchDetectionsToGroundTruth( + detection_instances, + detection_sorted_indices, + ground_truth_instances, + ground_truth_sorted_indices, + ignores, + image_category_ious[i][c], + iou_thresholds, + area_ranges[a], + &results_all + [c * num_area_ranges * num_images + a * num_images + i]); + } + } + } + + return results_all; +} + +// Convert a python list to a vector +template +std::vector list_to_vec(const py::list& l) { + std::vector v(py::len(l)); + for (int i = 0; i < (int)py::len(l); ++i) { + v[i] = l[i].cast(); + } + return v; +} + +// Helper function to Accumulate() +// Considers the evaluation results applicable to a particular category, area +// range, and max_detections parameter setting, which begin at +// evaluations[evaluation_index]. Extracts a sorted list of length n of all +// applicable detection instances concatenated across all images in the dataset, +// which are represented by the outputs evaluation_indices, detection_scores, +// image_detection_indices, and detection_sorted_indices--all of which are +// length n. evaluation_indices[i] stores the applicable index into +// evaluations[] for instance i, which has detection score detection_score[i], +// and is the image_detection_indices[i]'th of the list of detections +// for the image containing i. detection_sorted_indices[] defines a sorted +// permutation of the 3 other outputs +int BuildSortedDetectionList( + const std::vector& evaluations, + const int64_t evaluation_index, + const int64_t num_images, + const int max_detections, + std::vector* evaluation_indices, + std::vector* detection_scores, + std::vector* detection_sorted_indices, + std::vector* image_detection_indices) { + assert(evaluations.size() >= evaluation_index + num_images); + + // Extract a list of object instances of the applicable category, area + // range, and max detections requirements such that they can be sorted + image_detection_indices->clear(); + evaluation_indices->clear(); + detection_scores->clear(); + image_detection_indices->reserve(num_images * max_detections); + evaluation_indices->reserve(num_images * max_detections); + detection_scores->reserve(num_images * max_detections); + int num_valid_ground_truth = 0; + for (auto i = 0; i < num_images; ++i) { + const ImageEvaluation& evaluation = evaluations[evaluation_index + i]; + + for (int d = 0; + d < (int)evaluation.detection_scores.size() && d < max_detections; + ++d) { // detected instances + evaluation_indices->push_back(evaluation_index + i); + image_detection_indices->push_back(d); + detection_scores->push_back(evaluation.detection_scores[d]); + } + for (auto ground_truth_ignore : evaluation.ground_truth_ignores) { + if (!ground_truth_ignore) { + ++num_valid_ground_truth; + } + } + } + + // Sort detections by decreasing score, using stable sort to match + // python implementation + detection_sorted_indices->resize(detection_scores->size()); + std::iota( + detection_sorted_indices->begin(), detection_sorted_indices->end(), 0); + std::stable_sort( + detection_sorted_indices->begin(), + detection_sorted_indices->end(), + [&detection_scores](size_t j1, size_t j2) { + return (*detection_scores)[j1] > (*detection_scores)[j2]; + }); + + return num_valid_ground_truth; +} + +// Helper function to Accumulate() +// Compute a precision recall curve given a sorted list of detected instances +// encoded in evaluations, evaluation_indices, detection_scores, +// detection_sorted_indices, image_detection_indices (see +// BuildSortedDetectionList()). Using vectors precisions and recalls +// and temporary storage, output the results into precisions_out, recalls_out, +// and scores_out, which are large buffers containing many precion/recall curves +// for all possible parameter settings, with precisions_out_index and +// recalls_out_index defining the applicable indices to store results. +void ComputePrecisionRecallCurve( + const int64_t precisions_out_index, + const int64_t precisions_out_stride, + const int64_t recalls_out_index, + const std::vector& recall_thresholds, + const int iou_threshold_index, + const int num_iou_thresholds, + const int num_valid_ground_truth, + const std::vector& evaluations, + const std::vector& evaluation_indices, + const std::vector& detection_scores, + const std::vector& detection_sorted_indices, + const std::vector& image_detection_indices, + std::vector* precisions, + std::vector* recalls, + std::vector* precisions_out, + std::vector* scores_out, + std::vector* recalls_out) { + assert(recalls_out->size() > recalls_out_index); + + // Compute precision/recall for each instance in the sorted list of detections + int64_t true_positives_sum = 0, false_positives_sum = 0; + precisions->clear(); + recalls->clear(); + precisions->reserve(detection_sorted_indices.size()); + recalls->reserve(detection_sorted_indices.size()); + assert(!evaluations.empty() || detection_sorted_indices.empty()); + for (auto detection_sorted_index : detection_sorted_indices) { + const ImageEvaluation& evaluation = + evaluations[evaluation_indices[detection_sorted_index]]; + const auto num_detections = + evaluation.detection_matches.size() / num_iou_thresholds; + const auto detection_index = iou_threshold_index * num_detections + + image_detection_indices[detection_sorted_index]; + assert(evaluation.detection_matches.size() > detection_index); + assert(evaluation.detection_ignores.size() > detection_index); + const int64_t detection_match = + evaluation.detection_matches[detection_index]; + const bool detection_ignores = + evaluation.detection_ignores[detection_index]; + const auto true_positive = detection_match > 0 && !detection_ignores; + const auto false_positive = detection_match == 0 && !detection_ignores; + if (true_positive) { + ++true_positives_sum; + } + if (false_positive) { + ++false_positives_sum; + } + + const double recall = + static_cast(true_positives_sum) / num_valid_ground_truth; + recalls->push_back(recall); + const int64_t num_valid_detections = + true_positives_sum + false_positives_sum; + const double precision = num_valid_detections > 0 + ? static_cast(true_positives_sum) / num_valid_detections + : 0.0; + precisions->push_back(precision); + } + + (*recalls_out)[recalls_out_index] = !recalls->empty() ? recalls->back() : 0; + + for (int64_t i = static_cast(precisions->size()) - 1; i > 0; --i) { + if ((*precisions)[i] > (*precisions)[i - 1]) { + (*precisions)[i - 1] = (*precisions)[i]; + } + } + + // Sample the per instance precision/recall list at each recall threshold + for (size_t r = 0; r < recall_thresholds.size(); ++r) { + // first index in recalls >= recall_thresholds[r] + std::vector::iterator low = std::lower_bound( + recalls->begin(), recalls->end(), recall_thresholds[r]); + size_t precisions_index = low - recalls->begin(); + + const auto results_ind = precisions_out_index + r * precisions_out_stride; + assert(results_ind < precisions_out->size()); + assert(results_ind < scores_out->size()); + if (precisions_index < precisions->size()) { + (*precisions_out)[results_ind] = (*precisions)[precisions_index]; + (*scores_out)[results_ind] = + detection_scores[detection_sorted_indices[precisions_index]]; + } else { + (*precisions_out)[results_ind] = 0; + (*scores_out)[results_ind] = 0; + } + } +} +py::dict Accumulate( + const py::object& params, + const std::vector& evaluations) { + const std::vector recall_thresholds = + list_to_vec(params.attr("recThrs")); + const std::vector max_detections = + list_to_vec(params.attr("maxDets")); + const int num_iou_thresholds = py::len(params.attr("iouThrs")); + const int num_recall_thresholds = py::len(params.attr("recThrs")); + const int num_categories = params.attr("useCats").cast() == 1 + ? py::len(params.attr("catIds")) + : 1; + const int num_area_ranges = py::len(params.attr("areaRng")); + const int num_max_detections = py::len(params.attr("maxDets")); + const int num_images = py::len(params.attr("imgIds")); + + std::vector precisions_out( + num_iou_thresholds * num_recall_thresholds * num_categories * + num_area_ranges * num_max_detections, + -1); + std::vector recalls_out( + num_iou_thresholds * num_categories * num_area_ranges * + num_max_detections, + -1); + std::vector scores_out( + num_iou_thresholds * num_recall_thresholds * num_categories * + num_area_ranges * num_max_detections, + -1); + + // Consider the list of all detected instances in the entire dataset in one + // large list. evaluation_indices, detection_scores, + // image_detection_indices, and detection_sorted_indices all have the same + // length as this list, such that each entry corresponds to one detected + // instance + std::vector evaluation_indices; // indices into evaluations[] + std::vector detection_scores; // detection scores of each instance + std::vector detection_sorted_indices; // sorted indices of all + // instances in the dataset + std::vector + image_detection_indices; // indices into the list of detected instances in + // the same image as each instance + std::vector precisions, recalls; + + for (auto c = 0; c < num_categories; ++c) { + for (auto a = 0; a < num_area_ranges; ++a) { + for (auto m = 0; m < num_max_detections; ++m) { + // The COCO PythonAPI assumes evaluations[] (the return value of + // COCOeval::EvaluateImages() is one long list storing results for each + // combination of category, area range, and image id, with categories in + // the outermost loop and images in the innermost loop. + const int64_t evaluations_index = + c * num_area_ranges * num_images + a * num_images; + int num_valid_ground_truth = BuildSortedDetectionList( + evaluations, + evaluations_index, + num_images, + max_detections[m], + &evaluation_indices, + &detection_scores, + &detection_sorted_indices, + &image_detection_indices); + + if (num_valid_ground_truth == 0) { + continue; + } + + for (auto t = 0; t < num_iou_thresholds; ++t) { + // recalls_out is a flattened vectors representing a + // num_iou_thresholds X num_categories X num_area_ranges X + // num_max_detections matrix + const int64_t recalls_out_index = + t * num_categories * num_area_ranges * num_max_detections + + c * num_area_ranges * num_max_detections + + a * num_max_detections + m; + + // precisions_out and scores_out are flattened vectors + // representing a num_iou_thresholds X num_recall_thresholds X + // num_categories X num_area_ranges X num_max_detections matrix + const int64_t precisions_out_stride = + num_categories * num_area_ranges * num_max_detections; + const int64_t precisions_out_index = t * num_recall_thresholds * + num_categories * num_area_ranges * num_max_detections + + c * num_area_ranges * num_max_detections + + a * num_max_detections + m; + + ComputePrecisionRecallCurve( + precisions_out_index, + precisions_out_stride, + recalls_out_index, + recall_thresholds, + t, + num_iou_thresholds, + num_valid_ground_truth, + evaluations, + evaluation_indices, + detection_scores, + detection_sorted_indices, + image_detection_indices, + &precisions, + &recalls, + &precisions_out, + &scores_out, + &recalls_out); + } + } + } + } + + time_t rawtime; + struct tm local_time; + std::array buffer; + time(&rawtime); +#ifdef _WIN32 + localtime_s(&local_time, &rawtime); +#else + localtime_r(&rawtime, &local_time); +#endif + strftime( + buffer.data(), 200, "%Y-%m-%d %H:%num_max_detections:%S", &local_time); + return py::dict( + "params"_a = params, + "counts"_a = std::vector( + {num_iou_thresholds, + num_recall_thresholds, + num_categories, + num_area_ranges, + num_max_detections}), + "date"_a = buffer, + "precision"_a = precisions_out, + "recall"_a = recalls_out, + "scores"_a = scores_out); +} + +} // namespace COCOeval + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/cocoeval/cocoeval.h b/detectron2/layers/csrc/cocoeval/cocoeval.h new file mode 100644 index 0000000000000000000000000000000000000000..db246e49a026b7cd989b305f4d3d98100be3c912 --- /dev/null +++ b/detectron2/layers/csrc/cocoeval/cocoeval.h @@ -0,0 +1,88 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#pragma once + +#include +#include +#include +#include +#include + +namespace py = pybind11; + +namespace detectron2 { + +namespace COCOeval { + +// Annotation data for a single object instance in an image +struct InstanceAnnotation { + InstanceAnnotation( + uint64_t id, + double score, + double area, + bool is_crowd, + bool ignore) + : id{id}, score{score}, area{area}, is_crowd{is_crowd}, ignore{ignore} {} + uint64_t id; + double score = 0.; + double area = 0.; + bool is_crowd = false; + bool ignore = false; +}; + +// Stores intermediate results for evaluating detection results for a single +// image that has D detected instances and G ground truth instances. This stores +// matches between detected and ground truth instances +struct ImageEvaluation { + // For each of the D detected instances, the id of the matched ground truth + // instance, or 0 if unmatched + std::vector detection_matches; + + // The detection score of each of the D detected instances + std::vector detection_scores; + + // Marks whether or not each of G instances was ignored from evaluation (e.g., + // because it's outside area_range) + std::vector ground_truth_ignores; + + // Marks whether or not each of D instances was ignored from evaluation (e.g., + // because it's outside aRng) + std::vector detection_ignores; +}; + +template +using ImageCategoryInstances = std::vector>>; + +// C++ implementation of COCO API cocoeval.py::COCOeval.evaluateImg(). For each +// combination of image, category, area range settings, and IOU thresholds to +// evaluate, it matches detected instances to ground truth instances and stores +// the results into a vector of ImageEvaluation results, which will be +// interpreted by the COCOeval::Accumulate() function to produce precion-recall +// curves. The parameters of nested vectors have the following semantics: +// image_category_ious[i][c][d][g] is the intersection over union of the d'th +// detected instance and g'th ground truth instance of +// category category_ids[c] in image image_ids[i] +// image_category_ground_truth_instances[i][c] is a vector of ground truth +// instances in image image_ids[i] of category category_ids[c] +// image_category_detection_instances[i][c] is a vector of detected +// instances in image image_ids[i] of category category_ids[c] +std::vector EvaluateImages( + const std::vector>& area_ranges, // vector of 2-tuples + int max_detections, + const std::vector& iou_thresholds, + const ImageCategoryInstances>& image_category_ious, + const ImageCategoryInstances& + image_category_ground_truth_instances, + const ImageCategoryInstances& + image_category_detection_instances); + +// C++ implementation of COCOeval.accumulate(), which generates precision +// recall curves for each set of category, IOU threshold, detection area range, +// and max number of detections parameters. It is assumed that the parameter +// evaluations is the return value of the functon COCOeval::EvaluateImages(), +// which was called with the same parameter settings params +py::dict Accumulate( + const py::object& params, + const std::vector& evalutations); + +} // namespace COCOeval +} // namespace detectron2 diff --git a/detectron2/layers/csrc/cuda_version.cu b/detectron2/layers/csrc/cuda_version.cu new file mode 100644 index 0000000000000000000000000000000000000000..6dfe1b90c1f65c443681813fd3e3386c9faa3360 --- /dev/null +++ b/detectron2/layers/csrc/cuda_version.cu @@ -0,0 +1,26 @@ +// Copyright (c) Facebook, Inc. and its affiliates. + +#include + +namespace detectron2 { +int get_cudart_version() { +// Not a ROCM platform: Either HIP is not used, or +// it is used, but platform is not ROCM (i.e. it is CUDA) +#if !defined(__HIP_PLATFORM_HCC__) + return CUDART_VERSION; +#else + int version = 0; + +#if HIP_VERSION_MAJOR != 0 + // Create a convention similar to that of CUDA, as assumed by other + // parts of the code. + + version = HIP_VERSION_MINOR; + version += (HIP_VERSION_MAJOR * 100); +#else + hipRuntimeGetVersion(&version); +#endif + return version; +#endif +} +} // namespace detectron2 diff --git a/detectron2/layers/csrc/deformable/deform_conv.h b/detectron2/layers/csrc/deformable/deform_conv.h new file mode 100644 index 0000000000000000000000000000000000000000..ec8c6c2fdb0274aefb86523894174f9ca58bbb43 --- /dev/null +++ b/detectron2/layers/csrc/deformable/deform_conv.h @@ -0,0 +1,377 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#pragma once +#include + +namespace detectron2 { + +#if defined(WITH_CUDA) || defined(WITH_HIP) +int deform_conv_forward_cuda( + at::Tensor input, + at::Tensor weight, + at::Tensor offset, + at::Tensor output, + at::Tensor columns, + at::Tensor ones, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + int im2col_step); + +int deform_conv_backward_input_cuda( + at::Tensor input, + at::Tensor offset, + at::Tensor gradOutput, + at::Tensor gradInput, + at::Tensor gradOffset, + at::Tensor weight, + at::Tensor columns, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + int im2col_step); + +int deform_conv_backward_parameters_cuda( + at::Tensor input, + at::Tensor offset, + at::Tensor gradOutput, + at::Tensor gradWeight, // at::Tensor gradBias, + at::Tensor columns, + at::Tensor ones, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + float scale, + int im2col_step); + +void modulated_deform_conv_cuda_forward( + at::Tensor input, + at::Tensor weight, + at::Tensor bias, + at::Tensor ones, + at::Tensor offset, + at::Tensor mask, + at::Tensor output, + at::Tensor columns, + int kernel_h, + int kernel_w, + const int stride_h, + const int stride_w, + const int pad_h, + const int pad_w, + const int dilation_h, + const int dilation_w, + const int group, + const int deformable_group, + const bool with_bias); + +void modulated_deform_conv_cuda_backward( + at::Tensor input, + at::Tensor weight, + at::Tensor bias, + at::Tensor ones, + at::Tensor offset, + at::Tensor mask, + at::Tensor columns, + at::Tensor grad_input, + at::Tensor grad_weight, + at::Tensor grad_bias, + at::Tensor grad_offset, + at::Tensor grad_mask, + at::Tensor grad_output, + int kernel_h, + int kernel_w, + int stride_h, + int stride_w, + int pad_h, + int pad_w, + int dilation_h, + int dilation_w, + int group, + int deformable_group, + const bool with_bias); + +#endif + +inline int deform_conv_forward( + at::Tensor input, + at::Tensor weight, + at::Tensor offset, + at::Tensor output, + at::Tensor columns, + at::Tensor ones, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + int im2col_step) { + if (input.is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + TORCH_CHECK(weight.is_cuda(), "weight tensor is not on GPU!"); + TORCH_CHECK(offset.is_cuda(), "offset tensor is not on GPU!"); + return deform_conv_forward_cuda( + input, + weight, + offset, + output, + columns, + ones, + kW, + kH, + dW, + dH, + padW, + padH, + dilationW, + dilationH, + group, + deformable_group, + im2col_step); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + AT_ERROR("Not implemented on the CPU"); +} + +inline int deform_conv_backward_input( + at::Tensor input, + at::Tensor offset, + at::Tensor gradOutput, + at::Tensor gradInput, + at::Tensor gradOffset, + at::Tensor weight, + at::Tensor columns, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + int im2col_step) { + if (gradOutput.is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + TORCH_CHECK(input.is_cuda(), "input tensor is not on GPU!"); + TORCH_CHECK(weight.is_cuda(), "weight tensor is not on GPU!"); + TORCH_CHECK(offset.is_cuda(), "offset tensor is not on GPU!"); + return deform_conv_backward_input_cuda( + input, + offset, + gradOutput, + gradInput, + gradOffset, + weight, + columns, + kW, + kH, + dW, + dH, + padW, + padH, + dilationW, + dilationH, + group, + deformable_group, + im2col_step); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + AT_ERROR("Not implemented on the CPU"); +} + +inline int deform_conv_backward_filter( + at::Tensor input, + at::Tensor offset, + at::Tensor gradOutput, + at::Tensor gradWeight, // at::Tensor gradBias, + at::Tensor columns, + at::Tensor ones, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + float scale, + int im2col_step) { + if (gradOutput.is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + TORCH_CHECK(input.is_cuda(), "input tensor is not on GPU!"); + TORCH_CHECK(offset.is_cuda(), "offset tensor is not on GPU!"); + return deform_conv_backward_parameters_cuda( + input, + offset, + gradOutput, + gradWeight, + columns, + ones, + kW, + kH, + dW, + dH, + padW, + padH, + dilationW, + dilationH, + group, + deformable_group, + scale, + im2col_step); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + AT_ERROR("Not implemented on the CPU"); +} + +inline void modulated_deform_conv_forward( + at::Tensor input, + at::Tensor weight, + at::Tensor bias, + at::Tensor ones, + at::Tensor offset, + at::Tensor mask, + at::Tensor output, + at::Tensor columns, + int kernel_h, + int kernel_w, + const int stride_h, + const int stride_w, + const int pad_h, + const int pad_w, + const int dilation_h, + const int dilation_w, + const int group, + const int deformable_group, + const bool with_bias) { + if (input.is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + TORCH_CHECK(weight.is_cuda(), "weight tensor is not on GPU!"); + TORCH_CHECK(bias.is_cuda(), "bias tensor is not on GPU!"); + TORCH_CHECK(offset.is_cuda(), "offset tensor is not on GPU!"); + return modulated_deform_conv_cuda_forward( + input, + weight, + bias, + ones, + offset, + mask, + output, + columns, + kernel_h, + kernel_w, + stride_h, + stride_w, + pad_h, + pad_w, + dilation_h, + dilation_w, + group, + deformable_group, + with_bias); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + AT_ERROR("Not implemented on the CPU"); +} + +inline void modulated_deform_conv_backward( + at::Tensor input, + at::Tensor weight, + at::Tensor bias, + at::Tensor ones, + at::Tensor offset, + at::Tensor mask, + at::Tensor columns, + at::Tensor grad_input, + at::Tensor grad_weight, + at::Tensor grad_bias, + at::Tensor grad_offset, + at::Tensor grad_mask, + at::Tensor grad_output, + int kernel_h, + int kernel_w, + int stride_h, + int stride_w, + int pad_h, + int pad_w, + int dilation_h, + int dilation_w, + int group, + int deformable_group, + const bool with_bias) { + if (grad_output.is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + TORCH_CHECK(input.is_cuda(), "input tensor is not on GPU!"); + TORCH_CHECK(weight.is_cuda(), "weight tensor is not on GPU!"); + TORCH_CHECK(bias.is_cuda(), "bias tensor is not on GPU!"); + TORCH_CHECK(offset.is_cuda(), "offset tensor is not on GPU!"); + return modulated_deform_conv_cuda_backward( + input, + weight, + bias, + ones, + offset, + mask, + columns, + grad_input, + grad_weight, + grad_bias, + grad_offset, + grad_mask, + grad_output, + kernel_h, + kernel_w, + stride_h, + stride_w, + pad_h, + pad_w, + dilation_h, + dilation_w, + group, + deformable_group, + with_bias); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + AT_ERROR("Not implemented on the CPU"); +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/deformable/deform_conv_cuda.cu b/detectron2/layers/csrc/deformable/deform_conv_cuda.cu new file mode 100644 index 0000000000000000000000000000000000000000..2072bb856ec40b61c3826cead2fb7bb7c971a089 --- /dev/null +++ b/detectron2/layers/csrc/deformable/deform_conv_cuda.cu @@ -0,0 +1,1223 @@ +// Copyright (c) Facebook, Inc. and its affiliates. + +// modified from +// https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda.cpp +// Original license: Apache 2.0 + +// modify from +// https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda.c +// Original license: Apache 2.0 + +#include + +#include "deform_conv.h" + +#include +#include + +namespace detectron2 { + +void deformable_im2col( + const at::Tensor data_im, + const at::Tensor data_offset, + const int channels, + const int height, + const int width, + const int ksize_h, + const int ksize_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int parallel_imgs, + const int deformable_group, + at::Tensor data_col); + +void deformable_col2im( + const at::Tensor data_col, + const at::Tensor data_offset, + const int channels, + const int height, + const int width, + const int ksize_h, + const int ksize_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int parallel_imgs, + const int deformable_group, + at::Tensor grad_im); + +void deformable_col2im_coord( + const at::Tensor data_col, + const at::Tensor data_im, + const at::Tensor data_offset, + const int channels, + const int height, + const int width, + const int ksize_h, + const int ksize_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int parallel_imgs, + const int deformable_group, + at::Tensor grad_offset); + +void modulated_deformable_im2col_cuda( + const at::Tensor data_im, + const at::Tensor data_offset, + const at::Tensor data_mask, + const int batch_size, + const int channels, + const int height_im, + const int width_im, + const int height_col, + const int width_col, + const int kernel_h, + const int kenerl_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int deformable_group, + at::Tensor data_col); + +void modulated_deformable_col2im_cuda( + const at::Tensor data_col, + const at::Tensor data_offset, + const at::Tensor data_mask, + const int batch_size, + const int channels, + const int height_im, + const int width_im, + const int height_col, + const int width_col, + const int kernel_h, + const int kenerl_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int deformable_group, + at::Tensor grad_im); + +void modulated_deformable_col2im_coord_cuda( + const at::Tensor data_col, + const at::Tensor data_im, + const at::Tensor data_offset, + const at::Tensor data_mask, + const int batch_size, + const int channels, + const int height_im, + const int width_im, + const int height_col, + const int width_col, + const int kernel_h, + const int kenerl_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int deformable_group, + at::Tensor grad_offset, + at::Tensor grad_mask); + +void shape_check( + at::Tensor input, + at::Tensor offset, + at::Tensor* gradOutput, + at::Tensor weight, + int kH, + int kW, + int dH, + int dW, + int padH, + int padW, + int dilationH, + int dilationW, + int group, + int deformable_group) { + TORCH_CHECK( + weight.ndimension() == 4, + "4D weight tensor (nOutputPlane,nInputPlane,kH,kW) expected, " + "but got: %s", + weight.ndimension()); + + TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous"); + + TORCH_CHECK( + kW > 0 && kH > 0, + "kernel size should be greater than zero, but got kH: %d kW: %d", + kH, + kW); + + TORCH_CHECK( + (weight.size(2) == kH && weight.size(3) == kW), + "kernel size should be consistent with weight, ", + "but got kH: %d kW: %d weight.size(2): %d, weight.size(3): %d", + kH, + kW, + weight.size(2), + weight.size(3)); + + TORCH_CHECK( + dW > 0 && dH > 0, + "stride should be greater than zero, but got dH: %d dW: %d", + dH, + dW); + + TORCH_CHECK( + dilationW > 0 && dilationH > 0, + "dilation should be greater than 0, but got dilationH: %d dilationW: %d", + dilationH, + dilationW); + + int ndim = input.ndimension(); + int dimf = 0; + int dimh = 1; + int dimw = 2; + + if (ndim == 4) { + dimf++; + dimh++; + dimw++; + } + + TORCH_CHECK( + ndim == 3 || ndim == 4, + "3D or 4D input tensor expected but got: %s", + ndim); + + long nInputPlane = weight.size(1) * group; + long inputHeight = input.size(dimh); + long inputWidth = input.size(dimw); + long nOutputPlane = weight.size(0); + long outputHeight = + (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1; + long outputWidth = + (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1; + + TORCH_CHECK( + nInputPlane % deformable_group == 0, + "input channels must divide deformable group size"); + + if (outputWidth < 1 || outputHeight < 1) + AT_ERROR( + "Given input size: (%ld x %ld x %ld). " + "Calculated output size: (%ld x %ld x %ld). Output size is too small", + nInputPlane, + inputHeight, + inputWidth, + nOutputPlane, + outputHeight, + outputWidth); + + TORCH_CHECK( + input.size(1) == nInputPlane, + "invalid number of input planes, expected: %d, but got: %d", + nInputPlane, + input.size(1)); + + TORCH_CHECK( + (inputHeight + 2 * padH >= kH && inputWidth + 2 * padW >= kW), + "input image is smaller than kernel"); + + TORCH_CHECK( + (offset.size(2) == outputHeight && offset.size(3) == outputWidth), + "invalid spatial size of offset, expected height: %d width: %d, but " + "got height: %d width: %d", + outputHeight, + outputWidth, + offset.size(2), + offset.size(3)); + + TORCH_CHECK( + (offset.size(1) == deformable_group * 2 * kH * kW), + "invalid number of channels of offset"); + + if (gradOutput != NULL) { + TORCH_CHECK( + gradOutput->size(dimf) == nOutputPlane, + "invalid number of gradOutput planes, expected: %d, but got: %d", + nOutputPlane, + gradOutput->size(dimf)); + + TORCH_CHECK( + (gradOutput->size(dimh) == outputHeight && + gradOutput->size(dimw) == outputWidth), + "invalid size of gradOutput, expected height: %d width: %d , but " + "got height: %d width: %d", + outputHeight, + outputWidth, + gradOutput->size(dimh), + gradOutput->size(dimw)); + } +} + +int deform_conv_forward_cuda( + at::Tensor input, + at::Tensor weight, + at::Tensor offset, + at::Tensor output, + at::Tensor columns, + at::Tensor ones, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + int im2col_step) { + // todo: resize columns to include im2col: done + // todo: add im2col_step as input + // todo: add new output buffer and transpose it to output (or directly + // transpose output) todo: possibly change data indexing because of + // parallel_imgs + + shape_check( + input, + offset, + NULL, + weight, + kH, + kW, + dH, + dW, + padH, + padW, + dilationH, + dilationW, + group, + deformable_group); + + input = input.contiguous(); + offset = offset.contiguous(); + weight = weight.contiguous(); + + int batch = 1; + if (input.ndimension() == 3) { + // Force batch + batch = 0; + input.unsqueeze_(0); + offset.unsqueeze_(0); + } + + // todo: assert batchsize dividable by im2col_step + + long batchSize = input.size(0); + long nInputPlane = input.size(1); + long inputHeight = input.size(2); + long inputWidth = input.size(3); + + long nOutputPlane = weight.size(0); + + long outputWidth = + (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1; + long outputHeight = + (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1; + + TORCH_CHECK((offset.size(0) == batchSize), "invalid batch size of offset"); + + output = output.view( + {batchSize / im2col_step, + im2col_step, + nOutputPlane, + outputHeight, + outputWidth}); + columns = at::zeros( + {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth}, + input.options()); + + if (ones.ndimension() != 2 || + ones.size(0) * ones.size(1) < outputHeight * outputWidth) { + ones = at::ones({outputHeight, outputWidth}, input.options()); + } + + input = input.view( + {batchSize / im2col_step, + im2col_step, + nInputPlane, + inputHeight, + inputWidth}); + offset = offset.view( + {batchSize / im2col_step, + im2col_step, + deformable_group * 2 * kH * kW, + outputHeight, + outputWidth}); + + at::Tensor output_buffer = at::zeros( + {batchSize / im2col_step, + nOutputPlane, + im2col_step * outputHeight, + outputWidth}, + output.options()); + + output_buffer = output_buffer.view( + {output_buffer.size(0), + group, + output_buffer.size(1) / group, + output_buffer.size(2), + output_buffer.size(3)}); + + for (int elt = 0; elt < batchSize / im2col_step; elt++) { + deformable_im2col( + input[elt], + offset[elt], + nInputPlane, + inputHeight, + inputWidth, + kH, + kW, + padH, + padW, + dH, + dW, + dilationH, + dilationW, + im2col_step, + deformable_group, + columns); + + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + weight = weight.view( + {group, + weight.size(0) / group, + weight.size(1), + weight.size(2), + weight.size(3)}); + + for (int g = 0; g < group; g++) { + output_buffer[elt][g] = output_buffer[elt][g] + .flatten(1) + .addmm_(weight[g].flatten(1), columns[g]) + .view_as(output_buffer[elt][g]); + } + } + + output_buffer = output_buffer.view( + {output_buffer.size(0), + output_buffer.size(1) * output_buffer.size(2), + output_buffer.size(3), + output_buffer.size(4)}); + + output_buffer = output_buffer.view( + {batchSize / im2col_step, + nOutputPlane, + im2col_step, + outputHeight, + outputWidth}); + output_buffer.transpose_(1, 2); + output.copy_(output_buffer); + output = output.view({batchSize, nOutputPlane, outputHeight, outputWidth}); + + input = input.view({batchSize, nInputPlane, inputHeight, inputWidth}); + offset = offset.view( + {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + if (batch == 0) { + output = output.view({nOutputPlane, outputHeight, outputWidth}); + input = input.view({nInputPlane, inputHeight, inputWidth}); + offset = offset.view({offset.size(1), offset.size(2), offset.size(3)}); + } + + return 1; +} + +int deform_conv_backward_input_cuda( + at::Tensor input, + at::Tensor offset, + at::Tensor gradOutput, + at::Tensor gradInput, + at::Tensor gradOffset, + at::Tensor weight, + at::Tensor columns, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + int im2col_step) { + shape_check( + input, + offset, + &gradOutput, + weight, + kH, + kW, + dH, + dW, + padH, + padW, + dilationH, + dilationW, + group, + deformable_group); + + input = input.contiguous(); + offset = offset.contiguous(); + gradOutput = gradOutput.contiguous(); + weight = weight.contiguous(); + + int batch = 1; + + if (input.ndimension() == 3) { + // Force batch + batch = 0; + input = input.view({1, input.size(0), input.size(1), input.size(2)}); + offset = offset.view({1, offset.size(0), offset.size(1), offset.size(2)}); + gradOutput = gradOutput.view( + {1, gradOutput.size(0), gradOutput.size(1), gradOutput.size(2)}); + } + + long batchSize = input.size(0); + long nInputPlane = input.size(1); + long inputHeight = input.size(2); + long inputWidth = input.size(3); + + long nOutputPlane = weight.size(0); + + long outputWidth = + (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1; + long outputHeight = + (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1; + + TORCH_CHECK((offset.size(0) == batchSize), 3, "invalid batch size of offset"); + gradInput = gradInput.view({batchSize, nInputPlane, inputHeight, inputWidth}); + columns = at::zeros( + {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth}, + input.options()); + + // change order of grad output + gradOutput = gradOutput.view( + {batchSize / im2col_step, + im2col_step, + nOutputPlane, + outputHeight, + outputWidth}); + gradOutput.transpose_(1, 2); + + gradInput = gradInput.view( + {batchSize / im2col_step, + im2col_step, + nInputPlane, + inputHeight, + inputWidth}); + input = input.view( + {batchSize / im2col_step, + im2col_step, + nInputPlane, + inputHeight, + inputWidth}); + gradOffset = gradOffset.view( + {batchSize / im2col_step, + im2col_step, + deformable_group * 2 * kH * kW, + outputHeight, + outputWidth}); + offset = offset.view( + {batchSize / im2col_step, + im2col_step, + deformable_group * 2 * kH * kW, + outputHeight, + outputWidth}); + + for (int elt = 0; elt < batchSize / im2col_step; elt++) { + // divide into groups + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + weight = weight.view( + {group, + weight.size(0) / group, + weight.size(1), + weight.size(2), + weight.size(3)}); + gradOutput = gradOutput.view( + {gradOutput.size(0), + group, + gradOutput.size(1) / group, + gradOutput.size(2), + gradOutput.size(3), + gradOutput.size(4)}); + + for (int g = 0; g < group; g++) { + columns[g] = columns[g].addmm_( + weight[g].flatten(1).transpose(0, 1), + gradOutput[elt][g].flatten(1), + 0.0f, + 1.0f); + } + + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + gradOutput = gradOutput.view( + {gradOutput.size(0), + gradOutput.size(1) * gradOutput.size(2), + gradOutput.size(3), + gradOutput.size(4), + gradOutput.size(5)}); + + deformable_col2im_coord( + columns, + input[elt], + offset[elt], + nInputPlane, + inputHeight, + inputWidth, + kH, + kW, + padH, + padW, + dH, + dW, + dilationH, + dilationW, + im2col_step, + deformable_group, + gradOffset[elt]); + + deformable_col2im( + columns, + offset[elt], + nInputPlane, + inputHeight, + inputWidth, + kH, + kW, + padH, + padW, + dH, + dW, + dilationH, + dilationW, + im2col_step, + deformable_group, + gradInput[elt]); + } + + gradOutput.transpose_(1, 2); + gradOutput = + gradOutput.view({batchSize, nOutputPlane, outputHeight, outputWidth}); + + gradInput = gradInput.view({batchSize, nInputPlane, inputHeight, inputWidth}); + input = input.view({batchSize, nInputPlane, inputHeight, inputWidth}); + gradOffset = gradOffset.view( + {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + offset = offset.view( + {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + if (batch == 0) { + gradOutput = gradOutput.view({nOutputPlane, outputHeight, outputWidth}); + input = input.view({nInputPlane, inputHeight, inputWidth}); + gradInput = gradInput.view({nInputPlane, inputHeight, inputWidth}); + offset = offset.view({offset.size(1), offset.size(2), offset.size(3)}); + gradOffset = + gradOffset.view({offset.size(1), offset.size(2), offset.size(3)}); + } + + return 1; +} + +int deform_conv_backward_parameters_cuda( + at::Tensor input, + at::Tensor offset, + at::Tensor gradOutput, + at::Tensor gradWeight, // at::Tensor gradBias, + at::Tensor columns, + at::Tensor ones, + int kW, + int kH, + int dW, + int dH, + int padW, + int padH, + int dilationW, + int dilationH, + int group, + int deformable_group, + float scale, + int im2col_step) { + // todo: transpose and reshape outGrad + // todo: reshape columns + // todo: add im2col_step as input + + shape_check( + input, + offset, + &gradOutput, + gradWeight, + kH, + kW, + dH, + dW, + padH, + padW, + dilationH, + dilationW, + group, + deformable_group); + + input = input.contiguous(); + offset = offset.contiguous(); + gradOutput = gradOutput.contiguous(); + + int batch = 1; + + if (input.ndimension() == 3) { + // Force batch + batch = 0; + input = input.view( + at::IntList({1, input.size(0), input.size(1), input.size(2)})); + gradOutput = gradOutput.view( + {1, gradOutput.size(0), gradOutput.size(1), gradOutput.size(2)}); + } + + long batchSize = input.size(0); + long nInputPlane = input.size(1); + long inputHeight = input.size(2); + long inputWidth = input.size(3); + + long nOutputPlane = gradWeight.size(0); + + long outputWidth = + (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1; + long outputHeight = + (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1; + + TORCH_CHECK((offset.size(0) == batchSize), "invalid batch size of offset"); + + columns = at::zeros( + {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth}, + input.options()); + + gradOutput = gradOutput.view( + {batchSize / im2col_step, + im2col_step, + nOutputPlane, + outputHeight, + outputWidth}); + gradOutput.transpose_(1, 2); + + at::Tensor gradOutputBuffer = at::zeros_like(gradOutput); + gradOutputBuffer = gradOutputBuffer.view( + {batchSize / im2col_step, + nOutputPlane, + im2col_step, + outputHeight, + outputWidth}); + gradOutputBuffer.copy_(gradOutput); + // gradOutput is not contiguous, so we do reshape (instead of view) next + gradOutputBuffer = gradOutputBuffer.reshape( + {batchSize / im2col_step, + nOutputPlane, + im2col_step * outputHeight, + outputWidth}); + + gradOutput.transpose_(1, 2); + gradOutput = + gradOutput.view({batchSize, nOutputPlane, outputHeight, outputWidth}); + + input = input.view( + {batchSize / im2col_step, + im2col_step, + nInputPlane, + inputHeight, + inputWidth}); + offset = offset.view( + {batchSize / im2col_step, + im2col_step, + deformable_group * 2 * kH * kW, + outputHeight, + outputWidth}); + + for (int elt = 0; elt < batchSize / im2col_step; elt++) { + deformable_im2col( + input[elt], + offset[elt], + nInputPlane, + inputHeight, + inputWidth, + kH, + kW, + padH, + padW, + dH, + dW, + dilationH, + dilationW, + im2col_step, + deformable_group, + columns); + + // divide into group + gradOutputBuffer = gradOutputBuffer.view( + {gradOutputBuffer.size(0), + group, + gradOutputBuffer.size(1) / group, + gradOutputBuffer.size(2), + gradOutputBuffer.size(3)}); + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + gradWeight = gradWeight.view( + {group, + gradWeight.size(0) / group, + gradWeight.size(1), + gradWeight.size(2), + gradWeight.size(3)}); + + for (int g = 0; g < group; g++) { + gradWeight[g] = gradWeight[g] + .flatten(1) + .addmm_( + gradOutputBuffer[elt][g].flatten(1), + columns[g].transpose(1, 0), + 1.0, + scale) + .view_as(gradWeight[g]); + } + gradOutputBuffer = gradOutputBuffer.view( + {gradOutputBuffer.size(0), + gradOutputBuffer.size(1) * gradOutputBuffer.size(2), + gradOutputBuffer.size(3), + gradOutputBuffer.size(4)}); + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + gradWeight = gradWeight.view( + {gradWeight.size(0) * gradWeight.size(1), + gradWeight.size(2), + gradWeight.size(3), + gradWeight.size(4)}); + } + + input = input.view({batchSize, nInputPlane, inputHeight, inputWidth}); + offset = offset.view( + {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth}); + + if (batch == 0) { + gradOutput = gradOutput.view({nOutputPlane, outputHeight, outputWidth}); + input = input.view({nInputPlane, inputHeight, inputWidth}); + } + + return 1; +} + +void modulated_deform_conv_cuda_forward( + at::Tensor input, + at::Tensor weight, + at::Tensor bias, + at::Tensor ones, + at::Tensor offset, + at::Tensor mask, + at::Tensor output, + at::Tensor columns, + int kernel_h, + int kernel_w, + const int stride_h, + const int stride_w, + const int pad_h, + const int pad_w, + const int dilation_h, + const int dilation_w, + const int group, + const int deformable_group, + const bool with_bias) { + shape_check( + input, + offset, + NULL, + weight, + kernel_h, + kernel_w, + stride_h, + stride_w, + pad_h, + pad_w, + dilation_h, + dilation_w, + group, + deformable_group); + + TORCH_CHECK(input.is_contiguous(), "input tensor has to be contiguous"); + TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous"); + + const int batch = input.size(0); + const int channels = input.size(1); + const int height = input.size(2); + const int width = input.size(3); + + const int channels_out = weight.size(0); + const int channels_kernel = weight.size(1); + const int kernel_h_ = weight.size(2); + const int kernel_w_ = weight.size(3); + + if (kernel_h_ != kernel_h || kernel_w_ != kernel_w) + AT_ERROR( + "Input shape and kernel shape wont match: (%d x %d vs %d x %d).", + kernel_h_, + kernel_w, + kernel_h_, + kernel_w_); + if (channels != channels_kernel * group) + AT_ERROR( + "Input shape and kernel channels wont match: (%d vs %d).", + channels, + channels_kernel * group); + + const int height_out = + (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1; + const int width_out = + (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1; + + // mask shape check + TORCH_CHECK( + (mask.size(2) == height_out && mask.size(3) == width_out), + "invalid spatial size of mask, expected height: %d width: %d, but " + "got height: %d width: %d", + height_out, + width_out, + mask.size(2), + mask.size(3)); + + TORCH_CHECK( + (mask.size(1) == deformable_group * kernel_h * kernel_w), + "invalid number of channels of mask"); + + if (ones.ndimension() != 2 || + ones.size(0) * ones.size(1) < height_out * width_out) { + // Resize plane and fill with ones... + ones = at::ones({height_out, width_out}, input.options()); + } + + // resize output + output = output.view({batch, channels_out, height_out, width_out}).zero_(); + // resize temporary columns + columns = at::zeros( + {channels * kernel_h * kernel_w, 1 * height_out * width_out}, + input.options()); + + output = output.view( + {output.size(0), + group, + output.size(1) / group, + output.size(2), + output.size(3)}); + + for (int b = 0; b < batch; b++) { + modulated_deformable_im2col_cuda( + input[b], + offset[b], + mask[b], + 1, + channels, + height, + width, + height_out, + width_out, + kernel_h, + kernel_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + deformable_group, + columns); + + // divide into group + weight = weight.view( + {group, + weight.size(0) / group, + weight.size(1), + weight.size(2), + weight.size(3)}); + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + + for (int g = 0; g < group; g++) { + output[b][g] = output[b][g] + .flatten(1) + .addmm_(weight[g].flatten(1), columns[g]) + .view_as(output[b][g]); + } + + weight = weight.view( + {weight.size(0) * weight.size(1), + weight.size(2), + weight.size(3), + weight.size(4)}); + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + } + + output = output.view( + {output.size(0), + output.size(1) * output.size(2), + output.size(3), + output.size(4)}); + + if (with_bias) { + output += bias.view({1, bias.size(0), 1, 1}); + } +} + +void modulated_deform_conv_cuda_backward( + at::Tensor input, + at::Tensor weight, + at::Tensor bias, + at::Tensor ones, + at::Tensor offset, + at::Tensor mask, + at::Tensor columns, + at::Tensor grad_input, + at::Tensor grad_weight, + at::Tensor grad_bias, + at::Tensor grad_offset, + at::Tensor grad_mask, + at::Tensor grad_output, + int kernel_h, + int kernel_w, + int stride_h, + int stride_w, + int pad_h, + int pad_w, + int dilation_h, + int dilation_w, + int group, + int deformable_group, + const bool with_bias) { + shape_check( + input, + offset, + &grad_output, + weight, + kernel_h, + kernel_w, + stride_h, + stride_w, + pad_h, + pad_w, + dilation_h, + dilation_w, + group, + deformable_group); + + TORCH_CHECK(input.is_contiguous(), "input tensor has to be contiguous"); + TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous"); + + const int batch = input.size(0); + const int channels = input.size(1); + const int height = input.size(2); + const int width = input.size(3); + + const int channels_kernel = weight.size(1); + const int kernel_h_ = weight.size(2); + const int kernel_w_ = weight.size(3); + if (kernel_h_ != kernel_h || kernel_w_ != kernel_w) + AT_ERROR( + "Input shape and kernel shape wont match: (%d x %d vs %d x %d).", + kernel_h_, + kernel_w, + kernel_h_, + kernel_w_); + if (channels != channels_kernel * group) + AT_ERROR( + "Input shape and kernel channels wont match: (%d vs %d).", + channels, + channels_kernel * group); + + const int height_out = + (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1; + const int width_out = + (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1; + + // mask shape check + TORCH_CHECK( + (mask.size(2) == height_out && mask.size(3) == width_out), + "invalid spatial size of mask, expected height: %d width: %d, but " + "got height: %d width: %d", + height_out, + width_out, + mask.size(2), + mask.size(3)); + + TORCH_CHECK( + (mask.size(1) == deformable_group * kernel_h * kernel_w), + "invalid number of channels of mask"); + + if (ones.ndimension() != 2 || + ones.size(0) * ones.size(1) < height_out * width_out) { + // Resize plane and fill with ones... + ones = at::ones({height_out, width_out}, input.options()); + } + + grad_input = grad_input.view({batch, channels, height, width}); + columns = at::zeros( + {channels * kernel_h * kernel_w, height_out * width_out}, + input.options()); + + grad_output = grad_output.view( + {grad_output.size(0), + group, + grad_output.size(1) / group, + grad_output.size(2), + grad_output.size(3)}); + + for (int b = 0; b < batch; b++) { + // divide int group + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + weight = weight.view( + {group, + weight.size(0) / group, + weight.size(1), + weight.size(2), + weight.size(3)}); + + for (int g = 0; g < group; g++) { + columns[g].addmm_( + weight[g].flatten(1).transpose(0, 1), + grad_output[b][g].flatten(1), + 0.0f, + 1.0f); + } + + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + weight = weight.view( + {weight.size(0) * weight.size(1), + weight.size(2), + weight.size(3), + weight.size(4)}); + + // gradient w.r.t. input coordinate data + modulated_deformable_col2im_coord_cuda( + columns, + input[b], + offset[b], + mask[b], + 1, + channels, + height, + width, + height_out, + width_out, + kernel_h, + kernel_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + deformable_group, + grad_offset[b], + grad_mask[b]); + // gradient w.r.t. input data + modulated_deformable_col2im_cuda( + columns, + offset[b], + mask[b], + 1, + channels, + height, + width, + height_out, + width_out, + kernel_h, + kernel_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + deformable_group, + grad_input[b]); + + // gradient w.r.t. weight, dWeight should accumulate across the batch and + // group + modulated_deformable_im2col_cuda( + input[b], + offset[b], + mask[b], + 1, + channels, + height, + width, + height_out, + width_out, + kernel_h, + kernel_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + deformable_group, + columns); + + columns = columns.view({group, columns.size(0) / group, columns.size(1)}); + grad_weight = grad_weight.view( + {group, + grad_weight.size(0) / group, + grad_weight.size(1), + grad_weight.size(2), + grad_weight.size(3)}); + if (with_bias) + grad_bias = grad_bias.view({group, grad_bias.size(0) / group}); + + for (int g = 0; g < group; g++) { + grad_weight[g] = + grad_weight[g] + .flatten(1) + .addmm_(grad_output[b][g].flatten(1), columns[g].transpose(0, 1)) + .view_as(grad_weight[g]); + if (with_bias) { + grad_bias[g] = + grad_bias[g] + .view({-1, 1}) + .addmm_(grad_output[b][g].flatten(1), ones.view({-1, 1})) + .view(-1); + } + } + + columns = + columns.view({columns.size(0) * columns.size(1), columns.size(2)}); + grad_weight = grad_weight.view( + {grad_weight.size(0) * grad_weight.size(1), + grad_weight.size(2), + grad_weight.size(3), + grad_weight.size(4)}); + if (with_bias) + grad_bias = grad_bias.view({grad_bias.size(0) * grad_bias.size(1)}); + } + grad_output = grad_output.view( + {grad_output.size(0) * grad_output.size(1), + grad_output.size(2), + grad_output.size(3), + grad_output.size(4)}); +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/deformable/deform_conv_cuda_kernel.cu b/detectron2/layers/csrc/deformable/deform_conv_cuda_kernel.cu new file mode 100644 index 0000000000000000000000000000000000000000..f299c7add116685e9c87a187a85ea63f9f808867 --- /dev/null +++ b/detectron2/layers/csrc/deformable/deform_conv_cuda_kernel.cu @@ -0,0 +1,1288 @@ +// Copyright (c) Facebook, Inc. and its affiliates. + +// modified from +// https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/dcn/src/deform_conv_cuda_kernel.cu +// Original license: Apache 2.0 +// clang-format off + +// modify from +// https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda_kernel.cu + +/*! + ******************* BEGIN Caffe Copyright Notice and Disclaimer ***************** + * + * COPYRIGHT + * + * All contributions by the University of California: + * Copyright (c) 2014-2017 The Regents of the University of California (Regents) + * All rights reserved. + * + * All other contributions: + * Copyright (c) 2014-2017, the respective contributors + * All rights reserved. + * + * Caffe uses a shared copyright model: each contributor holds copyright over + * their contributions to Caffe. The project versioning records all such + * contribution and copyright details. If a contributor wants to further mark + * their specific copyright on a particular contribution, they should indicate + * their copyright solely in the commit message of the change when it is + * committed. + * + * LICENSE + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are met: + * + * 1. Redistributions of source code must retain the above copyright notice, this + * list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright notice, + * this list of conditions and the following disclaimer in the documentation + * and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" + *AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + *IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE + *FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + *DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR + *SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER + *CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, + *OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + *OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * CONTRIBUTION AGREEMENT + * + * By contributing to the BVLC/caffe repository through pull-request, comment, + * or otherwise, the contributor releases their content to the + * license and copyright terms herein. + * + ***************** END Caffe Copyright Notice and Disclaimer ********************* + * + * Copyright (c) 2018 Microsoft + * Licensed under The MIT License [see LICENSE for details] + * \file modulated_deformable_im2col.cuh + * \brief Function definitions of converting an image to + * column matrix based on kernel, padding, dilation, and offset. + * These functions are mainly used in deformable convolution operators. + * \ref: https://arxiv.org/abs/1703.06211 + * \author Yuwen Xiong, Haozhi Qi, Jifeng Dai, Xizhou Zhu, Han Hu, Dazhi Cheng + */ + +#include +#include +#include +#include +#include +#include + +using namespace at; + +#define CUDA_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + + +namespace { + +const int CUDA_NUM_THREADS = 1024; +const int kMaxGridNum = 65535; + +inline int GET_BLOCKS(const int N) { + return std::min(kMaxGridNum, (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS); +} + +} + +template +__device__ scalar_t deformable_im2col_bilinear( + const scalar_t* bottom_data, + const int data_width, + const int height, + const int width, + scalar_t h, + scalar_t w) { + int h_low = floor(h); + int w_low = floor(w); + int h_high = h_low + 1; + int w_high = w_low + 1; + + scalar_t lh = h - h_low; + scalar_t lw = w - w_low; + scalar_t hh = 1 - lh, hw = 1 - lw; + + scalar_t v1 = 0; + if (h_low >= 0 && w_low >= 0) + v1 = bottom_data[h_low * data_width + w_low]; + scalar_t v2 = 0; + if (h_low >= 0 && w_high <= width - 1) + v2 = bottom_data[h_low * data_width + w_high]; + scalar_t v3 = 0; + if (h_high <= height - 1 && w_low >= 0) + v3 = bottom_data[h_high * data_width + w_low]; + scalar_t v4 = 0; + if (h_high <= height - 1 && w_high <= width - 1) + v4 = bottom_data[h_high * data_width + w_high]; + + scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw; + + scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + return val; +} + +template +__device__ scalar_t get_gradient_weight( + scalar_t argmax_h, + scalar_t argmax_w, + const int h, + const int w, + const int height, + const int width) { + if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || + argmax_w >= width) { + // empty + return 0; + } + + int argmax_h_low = floor(argmax_h); + int argmax_w_low = floor(argmax_w); + int argmax_h_high = argmax_h_low + 1; + int argmax_w_high = argmax_w_low + 1; + + scalar_t weight = 0; + if (h == argmax_h_low && w == argmax_w_low) + weight = (h + 1 - argmax_h) * (w + 1 - argmax_w); + if (h == argmax_h_low && w == argmax_w_high) + weight = (h + 1 - argmax_h) * (argmax_w + 1 - w); + if (h == argmax_h_high && w == argmax_w_low) + weight = (argmax_h + 1 - h) * (w + 1 - argmax_w); + if (h == argmax_h_high && w == argmax_w_high) + weight = (argmax_h + 1 - h) * (argmax_w + 1 - w); + return weight; +} + +template +__device__ scalar_t get_coordinate_weight( + scalar_t argmax_h, + scalar_t argmax_w, + const int height, + const int width, + const scalar_t* im_data, + const int data_width, + const int bp_dir) { + if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || + argmax_w >= width) { + // empty + return 0; + } + + int argmax_h_low = floor(argmax_h); + int argmax_w_low = floor(argmax_w); + int argmax_h_high = argmax_h_low + 1; + int argmax_w_high = argmax_w_low + 1; + + scalar_t weight = 0; + + if (bp_dir == 0) { + if (argmax_h_low >= 0 && argmax_w_low >= 0) + weight += -1 * (argmax_w_low + 1 - argmax_w) * + im_data[argmax_h_low * data_width + argmax_w_low]; + if (argmax_h_low >= 0 && argmax_w_high <= width - 1) + weight += -1 * (argmax_w - argmax_w_low) * + im_data[argmax_h_low * data_width + argmax_w_high]; + if (argmax_h_high <= height - 1 && argmax_w_low >= 0) + weight += (argmax_w_low + 1 - argmax_w) * + im_data[argmax_h_high * data_width + argmax_w_low]; + if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1) + weight += (argmax_w - argmax_w_low) * + im_data[argmax_h_high * data_width + argmax_w_high]; + } else if (bp_dir == 1) { + if (argmax_h_low >= 0 && argmax_w_low >= 0) + weight += -1 * (argmax_h_low + 1 - argmax_h) * + im_data[argmax_h_low * data_width + argmax_w_low]; + if (argmax_h_low >= 0 && argmax_w_high <= width - 1) + weight += (argmax_h_low + 1 - argmax_h) * + im_data[argmax_h_low * data_width + argmax_w_high]; + if (argmax_h_high <= height - 1 && argmax_w_low >= 0) + weight += -1 * (argmax_h - argmax_h_low) * + im_data[argmax_h_high * data_width + argmax_w_low]; + if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1) + weight += (argmax_h - argmax_h_low) * + im_data[argmax_h_high * data_width + argmax_w_high]; + } + + return weight; +} + +template +__global__ void deformable_im2col_gpu_kernel( + const int n, + const scalar_t* data_im, + const scalar_t* data_offset, + const int height, + const int width, + const int kernel_h, + const int kernel_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, + const int num_channels, + const int deformable_group, + const int height_col, + const int width_col, + scalar_t* data_col) { + CUDA_KERNEL_LOOP(index, n) { + // index index of output matrix + const int w_col = index % width_col; + const int h_col = (index / width_col) % height_col; + const int b_col = (index / width_col / height_col) % batch_size; + const int c_im = (index / width_col / height_col) / batch_size; + const int c_col = c_im * kernel_h * kernel_w; + + // compute deformable group index + const int deformable_group_index = c_im / channel_per_deformable_group; + + const int h_in = h_col * stride_h - pad_h; + const int w_in = w_col * stride_w - pad_w; + scalar_t* data_col_ptr = data_col + + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col; + // const scalar_t* data_im_ptr = data_im + ((b_col * num_channels + c_im) * + // height + h_in) * width + w_in; + const scalar_t* data_im_ptr = + data_im + (b_col * num_channels + c_im) * height * width; + const scalar_t* data_offset_ptr = data_offset + + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * + kernel_w * height_col * width_col; + + for (int i = 0; i < kernel_h; ++i) { + for (int j = 0; j < kernel_w; ++j) { + const int data_offset_h_ptr = + ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col; + const int data_offset_w_ptr = + ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + + w_col; + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + scalar_t val = static_cast(0); + const scalar_t h_im = h_in + i * dilation_h + offset_h; + const scalar_t w_im = w_in + j * dilation_w + offset_w; + if (h_im > -1 && w_im > -1 && h_im < height && w_im < width) { + // const scalar_t map_h = i * dilation_h + offset_h; + // const scalar_t map_w = j * dilation_w + offset_w; + // const int cur_height = height - h_in; + // const int cur_width = width - w_in; + // val = deformable_im2col_bilinear(data_im_ptr, width, cur_height, + // cur_width, map_h, map_w); + val = deformable_im2col_bilinear( + data_im_ptr, width, height, width, h_im, w_im); + } + *data_col_ptr = val; + data_col_ptr += batch_size * height_col * width_col; + } + } + } +} + + +template +__global__ void deformable_col2im_gpu_kernel( + const int n, + const scalar_t* data_col, + const scalar_t* data_offset, + const int channels, + const int height, + const int width, + const int kernel_h, + const int kernel_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, + const int deformable_group, + const int height_col, + const int width_col, + scalar_t* grad_im) { + CUDA_KERNEL_LOOP(index, n) { + const int j = (index / width_col / height_col / batch_size) % kernel_w; + const int i = + (index / width_col / height_col / batch_size / kernel_w) % kernel_h; + const int c = + index / width_col / height_col / batch_size / kernel_w / kernel_h; + // compute the start and end of the output + + const int deformable_group_index = c / channel_per_deformable_group; + + int w_out = index % width_col; + int h_out = (index / width_col) % height_col; + int b = (index / width_col / height_col) % batch_size; + int w_in = w_out * stride_w - pad_w; + int h_in = h_out * stride_h - pad_h; + + const scalar_t* data_offset_ptr = data_offset + + (b * deformable_group + deformable_group_index) * 2 * kernel_h * + kernel_w * height_col * width_col; + const int data_offset_h_ptr = + ((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out; + const int data_offset_w_ptr = + ((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out; + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + const scalar_t cur_inv_h_data = h_in + i * dilation_h + offset_h; + const scalar_t cur_inv_w_data = w_in + j * dilation_w + offset_w; + + const scalar_t cur_top_grad = data_col[index]; + const int cur_h = (int)cur_inv_h_data; + const int cur_w = (int)cur_inv_w_data; + for (int dy = -2; dy <= 2; dy++) { + for (int dx = -2; dx <= 2; dx++) { + if (cur_h + dy >= 0 && cur_h + dy < height && cur_w + dx >= 0 && + cur_w + dx < width && abs(cur_inv_h_data - (cur_h + dy)) < 1 && + abs(cur_inv_w_data - (cur_w + dx)) < 1) { + int cur_bottom_grad_pos = + ((b * channels + c) * height + cur_h + dy) * width + cur_w + dx; + scalar_t weight = get_gradient_weight( + cur_inv_h_data, + cur_inv_w_data, + cur_h + dy, + cur_w + dx, + height, + width); + atomicAdd(grad_im + cur_bottom_grad_pos, weight * cur_top_grad); + } + } + } + } +} + + +template +__global__ void deformable_col2im_coord_gpu_kernel( + const int n, + const scalar_t* data_col, + const scalar_t* data_im, + const scalar_t* data_offset, + const int channels, + const int height, + const int width, + const int kernel_h, + const int kernel_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, + const int offset_channels, + const int deformable_group, + const int height_col, + const int width_col, + scalar_t* grad_offset) { + CUDA_KERNEL_LOOP(index, n) { + scalar_t val = 0; + int w = index % width_col; + int h = (index / width_col) % height_col; + int c = (index / width_col / height_col) % offset_channels; + int b = (index / width_col / height_col) / offset_channels; + // compute the start and end of the output + + const int deformable_group_index = c / (2 * kernel_h * kernel_w); + const int col_step = kernel_h * kernel_w; + int cnt = 0; + const scalar_t* data_col_ptr = data_col + + deformable_group_index * channel_per_deformable_group * batch_size * + width_col * height_col; + const scalar_t* data_im_ptr = data_im + + (b * deformable_group + deformable_group_index) * + channel_per_deformable_group / kernel_h / kernel_w * height * width; + const scalar_t* data_offset_ptr = data_offset + + (b * deformable_group + deformable_group_index) * 2 * kernel_h * + kernel_w * height_col * width_col; + + const int offset_c = c - deformable_group_index * 2 * kernel_h * kernel_w; + + for (int col_c = (offset_c / 2); col_c < channel_per_deformable_group; + col_c += col_step) { + const int col_pos = + (((col_c * batch_size + b) * height_col) + h) * width_col + w; + const int bp_dir = offset_c % 2; + + int j = (col_pos / width_col / height_col / batch_size) % kernel_w; + int i = + (col_pos / width_col / height_col / batch_size / kernel_w) % kernel_h; + int w_out = col_pos % width_col; + int h_out = (col_pos / width_col) % height_col; + int w_in = w_out * stride_w - pad_w; + int h_in = h_out * stride_h - pad_h; + const int data_offset_h_ptr = + (((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out); + const int data_offset_w_ptr = + (((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + + w_out); + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + scalar_t inv_h = h_in + i * dilation_h + offset_h; + scalar_t inv_w = w_in + j * dilation_w + offset_w; + if (inv_h <= -1 || inv_w <= -1 || inv_h >= height || inv_w >= width) { + inv_h = inv_w = -2; + } + const scalar_t weight = get_coordinate_weight( + inv_h, + inv_w, + height, + width, + data_im_ptr + cnt * height * width, + width, + bp_dir); + val += weight * data_col_ptr[col_pos]; + cnt += 1; + } + + grad_offset[index] = val; + } +} + + +namespace detectron2 { + +void deformable_im2col( + const at::Tensor data_im, + const at::Tensor data_offset, + const int channels, + const int height, + const int width, + const int ksize_h, + const int ksize_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int parallel_imgs, + const int deformable_group, + at::Tensor data_col) { + // num_axes should be smaller than block size + // todo: check parallel_imgs is correctly passed in + int height_col = + (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1; + int width_col = + (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1; + int num_kernels = channels * height_col * width_col * parallel_imgs; + int channel_per_deformable_group = channels / deformable_group; + + at::cuda::CUDAGuard device_guard(data_im.device()); + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_im.scalar_type(), "deformable_im2col_gpu", ([&] { + const scalar_t* data_im_ = data_im.data_ptr(); + const scalar_t* data_offset_ = data_offset.data_ptr(); + scalar_t* data_col_ = data_col.data_ptr(); + + deformable_im2col_gpu_kernel<<< + GET_BLOCKS(num_kernels), + CUDA_NUM_THREADS, + 0, + stream>>>( + num_kernels, + data_im_, + data_offset_, + height, + width, + ksize_h, + ksize_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + channel_per_deformable_group, + parallel_imgs, + channels, + deformable_group, + height_col, + width_col, + data_col_); + })); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) { + printf("error in deformable_im2col: %s\n", cudaGetErrorString(err)); + } +} + + +void deformable_col2im( + const at::Tensor data_col, + const at::Tensor data_offset, + const int channels, + const int height, + const int width, + const int ksize_h, + const int ksize_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int parallel_imgs, + const int deformable_group, + at::Tensor grad_im) { + // todo: make sure parallel_imgs is passed in correctly + int height_col = + (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1; + int width_col = + (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1; + int num_kernels = + channels * ksize_h * ksize_w * height_col * width_col * parallel_imgs; + int channel_per_deformable_group = channels / deformable_group; + + at::cuda::CUDAGuard device_guard(data_col.device()); + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_col.scalar_type(), "deformable_col2im_gpu", ([&] { + const scalar_t* data_col_ = data_col.data_ptr(); + const scalar_t* data_offset_ = data_offset.data_ptr(); + scalar_t* grad_im_ = grad_im.data_ptr(); + + deformable_col2im_gpu_kernel<<< + GET_BLOCKS(num_kernels), + CUDA_NUM_THREADS, + 0, + stream>>>( + num_kernels, + data_col_, + data_offset_, + channels, + height, + width, + ksize_h, + ksize_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + channel_per_deformable_group, + parallel_imgs, + deformable_group, + height_col, + width_col, + grad_im_); + })); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) { + printf("error in deformable_col2im: %s\n", cudaGetErrorString(err)); + } +} + + +void deformable_col2im_coord( + const at::Tensor data_col, + const at::Tensor data_im, + const at::Tensor data_offset, + const int channels, + const int height, + const int width, + const int ksize_h, + const int ksize_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int parallel_imgs, + const int deformable_group, + at::Tensor grad_offset) { + int height_col = + (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1; + int width_col = + (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1; + int num_kernels = height_col * width_col * 2 * ksize_h * ksize_w * + deformable_group * parallel_imgs; + int channel_per_deformable_group = + channels * ksize_h * ksize_w / deformable_group; + + at::cuda::CUDAGuard device_guard(data_col.device()); + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_col.scalar_type(), "deformable_col2im_coord_gpu", ([&] { + const scalar_t* data_col_ = data_col.data_ptr(); + const scalar_t* data_im_ = data_im.data_ptr(); + const scalar_t* data_offset_ = data_offset.data_ptr(); + scalar_t* grad_offset_ = grad_offset.data_ptr(); + + deformable_col2im_coord_gpu_kernel<<< + GET_BLOCKS(num_kernels), + CUDA_NUM_THREADS, + 0, + stream>>>( + num_kernels, + data_col_, + data_im_, + data_offset_, + channels, + height, + width, + ksize_h, + ksize_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + channel_per_deformable_group, + parallel_imgs, + 2 * ksize_h * ksize_w * deformable_group, + deformable_group, + height_col, + width_col, + grad_offset_); + })); +} + +} // namespace detectron2 + + +template +__device__ scalar_t dmcn_im2col_bilinear( + const scalar_t* bottom_data, + const int data_width, + const int height, + const int width, + scalar_t h, + scalar_t w) { + int h_low = floor(h); + int w_low = floor(w); + int h_high = h_low + 1; + int w_high = w_low + 1; + + scalar_t lh = h - h_low; + scalar_t lw = w - w_low; + scalar_t hh = 1 - lh, hw = 1 - lw; + + scalar_t v1 = 0; + if (h_low >= 0 && w_low >= 0) + v1 = bottom_data[h_low * data_width + w_low]; + scalar_t v2 = 0; + if (h_low >= 0 && w_high <= width - 1) + v2 = bottom_data[h_low * data_width + w_high]; + scalar_t v3 = 0; + if (h_high <= height - 1 && w_low >= 0) + v3 = bottom_data[h_high * data_width + w_low]; + scalar_t v4 = 0; + if (h_high <= height - 1 && w_high <= width - 1) + v4 = bottom_data[h_high * data_width + w_high]; + + scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw; + + scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + return val; +} + +template +__device__ scalar_t dmcn_get_gradient_weight( + scalar_t argmax_h, + scalar_t argmax_w, + const int h, + const int w, + const int height, + const int width) { + if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || + argmax_w >= width) { + // empty + return 0; + } + + int argmax_h_low = floor(argmax_h); + int argmax_w_low = floor(argmax_w); + int argmax_h_high = argmax_h_low + 1; + int argmax_w_high = argmax_w_low + 1; + + scalar_t weight = 0; + if (h == argmax_h_low && w == argmax_w_low) + weight = (h + 1 - argmax_h) * (w + 1 - argmax_w); + if (h == argmax_h_low && w == argmax_w_high) + weight = (h + 1 - argmax_h) * (argmax_w + 1 - w); + if (h == argmax_h_high && w == argmax_w_low) + weight = (argmax_h + 1 - h) * (w + 1 - argmax_w); + if (h == argmax_h_high && w == argmax_w_high) + weight = (argmax_h + 1 - h) * (argmax_w + 1 - w); + return weight; +} + +template +__device__ scalar_t dmcn_get_coordinate_weight( + scalar_t argmax_h, + scalar_t argmax_w, + const int height, + const int width, + const scalar_t* im_data, + const int data_width, + const int bp_dir) { + if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || + argmax_w >= width) { + // empty + return 0; + } + + int argmax_h_low = floor(argmax_h); + int argmax_w_low = floor(argmax_w); + int argmax_h_high = argmax_h_low + 1; + int argmax_w_high = argmax_w_low + 1; + + scalar_t weight = 0; + + if (bp_dir == 0) { + if (argmax_h_low >= 0 && argmax_w_low >= 0) + weight += -1 * (argmax_w_low + 1 - argmax_w) * + im_data[argmax_h_low * data_width + argmax_w_low]; + if (argmax_h_low >= 0 && argmax_w_high <= width - 1) + weight += -1 * (argmax_w - argmax_w_low) * + im_data[argmax_h_low * data_width + argmax_w_high]; + if (argmax_h_high <= height - 1 && argmax_w_low >= 0) + weight += (argmax_w_low + 1 - argmax_w) * + im_data[argmax_h_high * data_width + argmax_w_low]; + if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1) + weight += (argmax_w - argmax_w_low) * + im_data[argmax_h_high * data_width + argmax_w_high]; + } else if (bp_dir == 1) { + if (argmax_h_low >= 0 && argmax_w_low >= 0) + weight += -1 * (argmax_h_low + 1 - argmax_h) * + im_data[argmax_h_low * data_width + argmax_w_low]; + if (argmax_h_low >= 0 && argmax_w_high <= width - 1) + weight += (argmax_h_low + 1 - argmax_h) * + im_data[argmax_h_low * data_width + argmax_w_high]; + if (argmax_h_high <= height - 1 && argmax_w_low >= 0) + weight += -1 * (argmax_h - argmax_h_low) * + im_data[argmax_h_high * data_width + argmax_w_low]; + if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1) + weight += (argmax_h - argmax_h_low) * + im_data[argmax_h_high * data_width + argmax_w_high]; + } + + return weight; +} + +template +__global__ void modulated_deformable_im2col_gpu_kernel( + const int n, + const scalar_t* data_im, + const scalar_t* data_offset, + const scalar_t* data_mask, + const int height, + const int width, + const int kernel_h, + const int kernel_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, + const int num_channels, + const int deformable_group, + const int height_col, + const int width_col, + scalar_t* data_col) { + CUDA_KERNEL_LOOP(index, n) { + // index index of output matrix + const int w_col = index % width_col; + const int h_col = (index / width_col) % height_col; + const int b_col = (index / width_col / height_col) % batch_size; + const int c_im = (index / width_col / height_col) / batch_size; + const int c_col = c_im * kernel_h * kernel_w; + + // compute deformable group index + const int deformable_group_index = c_im / channel_per_deformable_group; + + const int h_in = h_col * stride_h - pad_h; + const int w_in = w_col * stride_w - pad_w; + + scalar_t* data_col_ptr = data_col + + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col; + // const float* data_im_ptr = data_im + ((b_col * num_channels + c_im) * + // height + h_in) * width + w_in; + const scalar_t* data_im_ptr = + data_im + (b_col * num_channels + c_im) * height * width; + const scalar_t* data_offset_ptr = data_offset + + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * + kernel_w * height_col * width_col; + + const scalar_t* data_mask_ptr = data_mask + + (b_col * deformable_group + deformable_group_index) * kernel_h * + kernel_w * height_col * width_col; + + for (int i = 0; i < kernel_h; ++i) { + for (int j = 0; j < kernel_w; ++j) { + const int data_offset_h_ptr = + ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col; + const int data_offset_w_ptr = + ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + + w_col; + const int data_mask_hw_ptr = + ((i * kernel_w + j) * height_col + h_col) * width_col + w_col; + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + const scalar_t mask = data_mask_ptr[data_mask_hw_ptr]; + scalar_t val = static_cast(0); + const scalar_t h_im = h_in + i * dilation_h + offset_h; + const scalar_t w_im = w_in + j * dilation_w + offset_w; + // if (h_im >= 0 && w_im >= 0 && h_im < height && w_im < width) { + if (h_im > -1 && w_im > -1 && h_im < height && w_im < width) { + // const float map_h = i * dilation_h + offset_h; + // const float map_w = j * dilation_w + offset_w; + // const int cur_height = height - h_in; + // const int cur_width = width - w_in; + // val = dmcn_im2col_bilinear(data_im_ptr, width, cur_height, + // cur_width, map_h, map_w); + val = dmcn_im2col_bilinear( + data_im_ptr, width, height, width, h_im, w_im); + } + *data_col_ptr = val * mask; + data_col_ptr += batch_size * height_col * width_col; + // data_col_ptr += height_col * width_col; + } + } + } +} + +template +__global__ void modulated_deformable_col2im_gpu_kernel( + const int n, + const scalar_t* data_col, + const scalar_t* data_offset, + const scalar_t* data_mask, + const int channels, + const int height, + const int width, + const int kernel_h, + const int kernel_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, + const int deformable_group, + const int height_col, + const int width_col, + scalar_t* grad_im) { + CUDA_KERNEL_LOOP(index, n) { + const int j = (index / width_col / height_col / batch_size) % kernel_w; + const int i = + (index / width_col / height_col / batch_size / kernel_w) % kernel_h; + const int c = + index / width_col / height_col / batch_size / kernel_w / kernel_h; + // compute the start and end of the output + + const int deformable_group_index = c / channel_per_deformable_group; + + int w_out = index % width_col; + int h_out = (index / width_col) % height_col; + int b = (index / width_col / height_col) % batch_size; + int w_in = w_out * stride_w - pad_w; + int h_in = h_out * stride_h - pad_h; + + const scalar_t* data_offset_ptr = data_offset + + (b * deformable_group + deformable_group_index) * 2 * kernel_h * + kernel_w * height_col * width_col; + const scalar_t* data_mask_ptr = data_mask + + (b * deformable_group + deformable_group_index) * kernel_h * kernel_w * + height_col * width_col; + const int data_offset_h_ptr = + ((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out; + const int data_offset_w_ptr = + ((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out; + const int data_mask_hw_ptr = + ((i * kernel_w + j) * height_col + h_out) * width_col + w_out; + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + const scalar_t mask = data_mask_ptr[data_mask_hw_ptr]; + const scalar_t cur_inv_h_data = h_in + i * dilation_h + offset_h; + const scalar_t cur_inv_w_data = w_in + j * dilation_w + offset_w; + + const scalar_t cur_top_grad = data_col[index] * mask; + const int cur_h = (int)cur_inv_h_data; + const int cur_w = (int)cur_inv_w_data; + for (int dy = -2; dy <= 2; dy++) { + for (int dx = -2; dx <= 2; dx++) { + if (cur_h + dy >= 0 && cur_h + dy < height && cur_w + dx >= 0 && + cur_w + dx < width && abs(cur_inv_h_data - (cur_h + dy)) < 1 && + abs(cur_inv_w_data - (cur_w + dx)) < 1) { + int cur_bottom_grad_pos = + ((b * channels + c) * height + cur_h + dy) * width + cur_w + dx; + scalar_t weight = dmcn_get_gradient_weight( + cur_inv_h_data, + cur_inv_w_data, + cur_h + dy, + cur_w + dx, + height, + width); + atomicAdd(grad_im + cur_bottom_grad_pos, weight * cur_top_grad); + } + } + } + } +} + +template +__global__ void modulated_deformable_col2im_coord_gpu_kernel( + const int n, + const scalar_t* data_col, + const scalar_t* data_im, + const scalar_t* data_offset, + const scalar_t* data_mask, + const int channels, + const int height, + const int width, + const int kernel_h, + const int kernel_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int channel_per_deformable_group, + const int batch_size, + const int offset_channels, + const int deformable_group, + const int height_col, + const int width_col, + scalar_t* grad_offset, + scalar_t* grad_mask) { + CUDA_KERNEL_LOOP(index, n) { + scalar_t val = 0, mval = 0; + int w = index % width_col; + int h = (index / width_col) % height_col; + int c = (index / width_col / height_col) % offset_channels; + int b = (index / width_col / height_col) / offset_channels; + // compute the start and end of the output + + const int deformable_group_index = c / (2 * kernel_h * kernel_w); + const int col_step = kernel_h * kernel_w; + int cnt = 0; + const scalar_t* data_col_ptr = data_col + + deformable_group_index * channel_per_deformable_group * batch_size * + width_col * height_col; + const scalar_t* data_im_ptr = data_im + + (b * deformable_group + deformable_group_index) * + channel_per_deformable_group / kernel_h / kernel_w * height * width; + const scalar_t* data_offset_ptr = data_offset + + (b * deformable_group + deformable_group_index) * 2 * kernel_h * + kernel_w * height_col * width_col; + const scalar_t* data_mask_ptr = data_mask + + (b * deformable_group + deformable_group_index) * kernel_h * kernel_w * + height_col * width_col; + + const int offset_c = c - deformable_group_index * 2 * kernel_h * kernel_w; + + for (int col_c = (offset_c / 2); col_c < channel_per_deformable_group; + col_c += col_step) { + const int col_pos = + (((col_c * batch_size + b) * height_col) + h) * width_col + w; + const int bp_dir = offset_c % 2; + + int j = (col_pos / width_col / height_col / batch_size) % kernel_w; + int i = + (col_pos / width_col / height_col / batch_size / kernel_w) % kernel_h; + int w_out = col_pos % width_col; + int h_out = (col_pos / width_col) % height_col; + int w_in = w_out * stride_w - pad_w; + int h_in = h_out * stride_h - pad_h; + const int data_offset_h_ptr = + (((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out); + const int data_offset_w_ptr = + (((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + + w_out); + const int data_mask_hw_ptr = + (((i * kernel_w + j) * height_col + h_out) * width_col + w_out); + const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr]; + const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr]; + const scalar_t mask = data_mask_ptr[data_mask_hw_ptr]; + scalar_t inv_h = h_in + i * dilation_h + offset_h; + scalar_t inv_w = w_in + j * dilation_w + offset_w; + if (inv_h <= -1 || inv_w <= -1 || inv_h >= height || inv_w >= width) { + inv_h = inv_w = -2; + } else { + mval += data_col_ptr[col_pos] * + dmcn_im2col_bilinear( + data_im_ptr + cnt * height * width, + width, + height, + width, + inv_h, + inv_w); + } + const scalar_t weight = dmcn_get_coordinate_weight( + inv_h, + inv_w, + height, + width, + data_im_ptr + cnt * height * width, + width, + bp_dir); + val += weight * data_col_ptr[col_pos] * mask; + cnt += 1; + } + // KERNEL_ASSIGN(grad_offset[index], offset_req, val); + grad_offset[index] = val; + if (offset_c % 2 == 0) + // KERNEL_ASSIGN(grad_mask[(((b * deformable_group + + // deformable_group_index) * kernel_h * kernel_w + offset_c / 2) * + // height_col + h) * width_col + w], mask_req, mval); + grad_mask + [(((b * deformable_group + deformable_group_index) * kernel_h * + kernel_w + + offset_c / 2) * + height_col + + h) * + width_col + + w] = mval; + } +} + + +namespace detectron2 { + +void modulated_deformable_im2col_cuda( + const at::Tensor data_im, + const at::Tensor data_offset, + const at::Tensor data_mask, + const int batch_size, + const int channels, + const int height_im, + const int width_im, + const int height_col, + const int width_col, + const int kernel_h, + const int kenerl_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int deformable_group, + at::Tensor data_col) { + // num_axes should be smaller than block size + const int channel_per_deformable_group = channels / deformable_group; + const int num_kernels = channels * batch_size * height_col * width_col; + + at::cuda::CUDAGuard device_guard(data_im.device()); + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_im.scalar_type(), "modulated_deformable_im2col_gpu", ([&] { + const scalar_t* data_im_ = data_im.data_ptr(); + const scalar_t* data_offset_ = data_offset.data_ptr(); + const scalar_t* data_mask_ = data_mask.data_ptr(); + scalar_t* data_col_ = data_col.data_ptr(); + + modulated_deformable_im2col_gpu_kernel<<< + GET_BLOCKS(num_kernels), + CUDA_NUM_THREADS, + 0, + stream>>>( + num_kernels, + data_im_, + data_offset_, + data_mask_, + height_im, + width_im, + kernel_h, + kenerl_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + channel_per_deformable_group, + batch_size, + channels, + deformable_group, + height_col, + width_col, + data_col_); + })); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) { + printf( + "error in modulated_deformable_im2col_cuda: %s\n", + cudaGetErrorString(err)); + } +} + +void modulated_deformable_col2im_cuda( + const at::Tensor data_col, + const at::Tensor data_offset, + const at::Tensor data_mask, + const int batch_size, + const int channels, + const int height_im, + const int width_im, + const int height_col, + const int width_col, + const int kernel_h, + const int kernel_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int deformable_group, + at::Tensor grad_im) { + const int channel_per_deformable_group = channels / deformable_group; + const int num_kernels = + channels * kernel_h * kernel_w * batch_size * height_col * width_col; + + at::cuda::CUDAGuard device_guard(data_col.device()); + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_col.scalar_type(), "modulated_deformable_col2im_gpu", ([&] { + const scalar_t* data_col_ = data_col.data_ptr(); + const scalar_t* data_offset_ = data_offset.data_ptr(); + const scalar_t* data_mask_ = data_mask.data_ptr(); + scalar_t* grad_im_ = grad_im.data_ptr(); + + modulated_deformable_col2im_gpu_kernel<<< + GET_BLOCKS(num_kernels), + CUDA_NUM_THREADS, + 0, + stream>>>( + num_kernels, + data_col_, + data_offset_, + data_mask_, + channels, + height_im, + width_im, + kernel_h, + kernel_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + channel_per_deformable_group, + batch_size, + deformable_group, + height_col, + width_col, + grad_im_); + })); + + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) { + printf( + "error in modulated_deformable_col2im_cuda: %s\n", + cudaGetErrorString(err)); + } +} + +void modulated_deformable_col2im_coord_cuda( + const at::Tensor data_col, + const at::Tensor data_im, + const at::Tensor data_offset, + const at::Tensor data_mask, + const int batch_size, + const int channels, + const int height_im, + const int width_im, + const int height_col, + const int width_col, + const int kernel_h, + const int kernel_w, + const int pad_h, + const int pad_w, + const int stride_h, + const int stride_w, + const int dilation_h, + const int dilation_w, + const int deformable_group, + at::Tensor grad_offset, + at::Tensor grad_mask) { + const int num_kernels = batch_size * height_col * width_col * 2 * kernel_h * + kernel_w * deformable_group; + const int channel_per_deformable_group = + channels * kernel_h * kernel_w / deformable_group; + + at::cuda::CUDAGuard device_guard(data_col.device()); + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + AT_DISPATCH_FLOATING_TYPES_AND_HALF( + data_col.scalar_type(), "modulated_deformable_col2im_coord_gpu", ([&] { + const scalar_t* data_col_ = data_col.data_ptr(); + const scalar_t* data_im_ = data_im.data_ptr(); + const scalar_t* data_offset_ = data_offset.data_ptr(); + const scalar_t* data_mask_ = data_mask.data_ptr(); + scalar_t* grad_offset_ = grad_offset.data_ptr(); + scalar_t* grad_mask_ = grad_mask.data_ptr(); + + modulated_deformable_col2im_coord_gpu_kernel<<< + GET_BLOCKS(num_kernels), + CUDA_NUM_THREADS, + 0, + stream>>>( + num_kernels, + data_col_, + data_im_, + data_offset_, + data_mask_, + channels, + height_im, + width_im, + kernel_h, + kernel_w, + pad_h, + pad_w, + stride_h, + stride_w, + dilation_h, + dilation_w, + channel_per_deformable_group, + batch_size, + 2 * kernel_h * kernel_w * deformable_group, + deformable_group, + height_col, + width_col, + grad_offset_, + grad_mask_); + })); + cudaError_t err = cudaGetLastError(); + if (err != cudaSuccess) { + printf( + "error in modulated_deformable_col2im_coord_cuda: %s\n", + cudaGetErrorString(err)); + } +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/nms_rotated/nms_rotated.h b/detectron2/layers/csrc/nms_rotated/nms_rotated.h new file mode 100644 index 0000000000000000000000000000000000000000..bd855e832afea4354885f5d8bfe94e204f51827e --- /dev/null +++ b/detectron2/layers/csrc/nms_rotated/nms_rotated.h @@ -0,0 +1,39 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#pragma once +#include + +namespace detectron2 { + +at::Tensor nms_rotated_cpu( + const at::Tensor& dets, + const at::Tensor& scores, + const double iou_threshold); + +#if defined(WITH_CUDA) || defined(WITH_HIP) +at::Tensor nms_rotated_cuda( + const at::Tensor& dets, + const at::Tensor& scores, + const double iou_threshold); +#endif + +// Interface for Python +// inline is needed to prevent multiple function definitions when this header is +// included by different cpps +inline at::Tensor nms_rotated( + const at::Tensor& dets, + const at::Tensor& scores, + const double iou_threshold) { + assert(dets.device().is_cuda() == scores.device().is_cuda()); + if (dets.device().is_cuda()) { +#if defined(WITH_CUDA) || defined(WITH_HIP) + return nms_rotated_cuda( + dets.contiguous(), scores.contiguous(), iou_threshold); +#else + AT_ERROR("Not compiled with GPU support"); +#endif + } + + return nms_rotated_cpu(dets.contiguous(), scores.contiguous(), iou_threshold); +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/nms_rotated/nms_rotated_cpu.cpp b/detectron2/layers/csrc/nms_rotated/nms_rotated_cpu.cpp new file mode 100644 index 0000000000000000000000000000000000000000..d7556e645b604aa83d86cc702b783fd8ecedffcc --- /dev/null +++ b/detectron2/layers/csrc/nms_rotated/nms_rotated_cpu.cpp @@ -0,0 +1,75 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#include "../box_iou_rotated/box_iou_rotated_utils.h" +#include "nms_rotated.h" + +namespace detectron2 { + +template +at::Tensor nms_rotated_cpu_kernel( + const at::Tensor& dets, + const at::Tensor& scores, + const double iou_threshold) { + // nms_rotated_cpu_kernel is modified from torchvision's nms_cpu_kernel, + // however, the code in this function is much shorter because + // we delegate the IoU computation for rotated boxes to + // the single_box_iou_rotated function in box_iou_rotated_utils.h + AT_ASSERTM(dets.device().is_cpu(), "dets must be a CPU tensor"); + AT_ASSERTM(scores.device().is_cpu(), "scores must be a CPU tensor"); + AT_ASSERTM( + dets.scalar_type() == scores.scalar_type(), + "dets should have the same type as scores"); + + if (dets.numel() == 0) { + return at::empty({0}, dets.options().dtype(at::kLong)); + } + + auto order_t = std::get<1>(scores.sort(0, /* descending=*/true)); + + auto ndets = dets.size(0); + at::Tensor suppressed_t = at::zeros({ndets}, dets.options().dtype(at::kByte)); + at::Tensor keep_t = at::zeros({ndets}, dets.options().dtype(at::kLong)); + + auto suppressed = suppressed_t.data_ptr(); + auto keep = keep_t.data_ptr(); + auto order = order_t.data_ptr(); + + int64_t num_to_keep = 0; + + for (int64_t _i = 0; _i < ndets; _i++) { + auto i = order[_i]; + if (suppressed[i] == 1) { + continue; + } + + keep[num_to_keep++] = i; + + for (int64_t _j = _i + 1; _j < ndets; _j++) { + auto j = order[_j]; + if (suppressed[j] == 1) { + continue; + } + + auto ovr = single_box_iou_rotated( + dets[i].data_ptr(), dets[j].data_ptr()); + if (ovr >= iou_threshold) { + suppressed[j] = 1; + } + } + } + return keep_t.narrow(/*dim=*/0, /*start=*/0, /*length=*/num_to_keep); +} + +at::Tensor nms_rotated_cpu( + // input must be contiguous + const at::Tensor& dets, + const at::Tensor& scores, + const double iou_threshold) { + auto result = at::empty({0}, dets.options()); + + AT_DISPATCH_FLOATING_TYPES(dets.scalar_type(), "nms_rotated", [&] { + result = nms_rotated_cpu_kernel(dets, scores, iou_threshold); + }); + return result; +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/nms_rotated/nms_rotated_cuda.cu b/detectron2/layers/csrc/nms_rotated/nms_rotated_cuda.cu new file mode 100644 index 0000000000000000000000000000000000000000..2a3db5c62e7a2da52ccf5bac980653c943d630fd --- /dev/null +++ b/detectron2/layers/csrc/nms_rotated/nms_rotated_cuda.cu @@ -0,0 +1,145 @@ +// Copyright (c) Facebook, Inc. and its affiliates. +#include +#include +#include +#include +#ifdef WITH_CUDA +#include "../box_iou_rotated/box_iou_rotated_utils.h" +#endif +// TODO avoid this when pytorch supports "same directory" hipification +#ifdef WITH_HIP +#include "box_iou_rotated/box_iou_rotated_utils.h" +#endif + +using namespace detectron2; + +namespace { +int const threadsPerBlock = sizeof(unsigned long long) * 8; +} + +template +__global__ void nms_rotated_cuda_kernel( + const int n_boxes, + const double iou_threshold, + const T* dev_boxes, + unsigned long long* dev_mask) { + // nms_rotated_cuda_kernel is modified from torchvision's nms_cuda_kernel + + const int row_start = blockIdx.y; + const int col_start = blockIdx.x; + + // if (row_start > col_start) return; + + const int row_size = + min(n_boxes - row_start * threadsPerBlock, threadsPerBlock); + const int col_size = + min(n_boxes - col_start * threadsPerBlock, threadsPerBlock); + + // Compared to nms_cuda_kernel, where each box is represented with 4 values + // (x1, y1, x2, y2), each rotated box is represented with 5 values + // (x_center, y_center, width, height, angle_degrees) here. + __shared__ T block_boxes[threadsPerBlock * 5]; + if (threadIdx.x < col_size) { + block_boxes[threadIdx.x * 5 + 0] = + dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 0]; + block_boxes[threadIdx.x * 5 + 1] = + dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 1]; + block_boxes[threadIdx.x * 5 + 2] = + dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 2]; + block_boxes[threadIdx.x * 5 + 3] = + dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 3]; + block_boxes[threadIdx.x * 5 + 4] = + dev_boxes[(threadsPerBlock * col_start + threadIdx.x) * 5 + 4]; + } + __syncthreads(); + + if (threadIdx.x < row_size) { + const int cur_box_idx = threadsPerBlock * row_start + threadIdx.x; + const T* cur_box = dev_boxes + cur_box_idx * 5; + int i = 0; + unsigned long long t = 0; + int start = 0; + if (row_start == col_start) { + start = threadIdx.x + 1; + } + for (i = start; i < col_size; i++) { + // Instead of devIoU used by original horizontal nms, here + // we use the single_box_iou_rotated function from box_iou_rotated_utils.h + if (single_box_iou_rotated(cur_box, block_boxes + i * 5) > + iou_threshold) { + t |= 1ULL << i; + } + } + const int col_blocks = at::cuda::ATenCeilDiv(n_boxes, threadsPerBlock); + dev_mask[cur_box_idx * col_blocks + col_start] = t; + } +} + +namespace detectron2 { + +at::Tensor nms_rotated_cuda( + // input must be contiguous + const at::Tensor& dets, + const at::Tensor& scores, + double iou_threshold) { + // using scalar_t = float; + AT_ASSERTM(dets.is_cuda(), "dets must be a CUDA tensor"); + AT_ASSERTM(scores.is_cuda(), "scores must be a CUDA tensor"); + at::cuda::CUDAGuard device_guard(dets.device()); + + auto order_t = std::get<1>(scores.sort(0, /* descending=*/true)); + auto dets_sorted = dets.index_select(0, order_t); + + auto dets_num = dets.size(0); + + const int col_blocks = + at::cuda::ATenCeilDiv(static_cast(dets_num), threadsPerBlock); + + at::Tensor mask = + at::empty({dets_num * col_blocks}, dets.options().dtype(at::kLong)); + + dim3 blocks(col_blocks, col_blocks); + dim3 threads(threadsPerBlock); + cudaStream_t stream = at::cuda::getCurrentCUDAStream(); + + AT_DISPATCH_FLOATING_TYPES( + dets_sorted.scalar_type(), "nms_rotated_kernel_cuda", [&] { + nms_rotated_cuda_kernel<<>>( + dets_num, + iou_threshold, + dets_sorted.data_ptr(), + (unsigned long long*)mask.data_ptr()); + }); + + at::Tensor mask_cpu = mask.to(at::kCPU); + unsigned long long* mask_host = + (unsigned long long*)mask_cpu.data_ptr(); + + std::vector remv(col_blocks); + memset(&remv[0], 0, sizeof(unsigned long long) * col_blocks); + + at::Tensor keep = + at::empty({dets_num}, dets.options().dtype(at::kLong).device(at::kCPU)); + int64_t* keep_out = keep.data_ptr(); + + int num_to_keep = 0; + for (int i = 0; i < dets_num; i++) { + int nblock = i / threadsPerBlock; + int inblock = i % threadsPerBlock; + + if (!(remv[nblock] & (1ULL << inblock))) { + keep_out[num_to_keep++] = i; + unsigned long long* p = mask_host + i * col_blocks; + for (int j = nblock; j < col_blocks; j++) { + remv[j] |= p[j]; + } + } + } + + AT_CUDA_CHECK(cudaGetLastError()); + return order_t.index( + {keep.narrow(/*dim=*/0, /*start=*/0, /*length=*/num_to_keep) + .to(order_t.device(), keep.scalar_type())}); +} + +} // namespace detectron2 diff --git a/detectron2/layers/csrc/vision.cpp b/detectron2/layers/csrc/vision.cpp new file mode 100644 index 0000000000000000000000000000000000000000..f6c049f7b4970b5ab88bf4bea5c5cf95897da0f7 --- /dev/null +++ b/detectron2/layers/csrc/vision.cpp @@ -0,0 +1,129 @@ +// Copyright (c) Facebook, Inc. and its affiliates. + +#include +#include "ROIAlignRotated/ROIAlignRotated.h" +#include "box_iou_rotated/box_iou_rotated.h" +#include "cocoeval/cocoeval.h" +#include "deformable/deform_conv.h" +#include "nms_rotated/nms_rotated.h" + +namespace detectron2 { + +#if defined(WITH_CUDA) || defined(WITH_HIP) +extern int get_cudart_version(); +#endif + +std::string get_cuda_version() { +#if defined(WITH_CUDA) || defined(WITH_HIP) + std::ostringstream oss; + +#if defined(WITH_CUDA) + oss << "CUDA "; +#else + oss << "HIP "; +#endif + + // copied from + // https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/cuda/detail/CUDAHooks.cpp#L231 + auto printCudaStyleVersion = [&](int v) { + oss << (v / 1000) << "." << (v / 10 % 100); + if (v % 10 != 0) { + oss << "." << (v % 10); + } + }; + printCudaStyleVersion(get_cudart_version()); + return oss.str(); +#else // neither CUDA nor HIP + return std::string("not available"); +#endif +} + +bool has_cuda() { +#if defined(WITH_CUDA) + return true; +#else + return false; +#endif +} + +// similar to +// https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/Version.cpp +std::string get_compiler_version() { + std::ostringstream ss; +#if defined(__GNUC__) +#ifndef __clang__ + +#if ((__GNUC__ <= 4) && (__GNUC_MINOR__ <= 8)) +#error "GCC >= 4.9 is required!" +#endif + + { ss << "GCC " << __GNUC__ << "." << __GNUC_MINOR__; } +#endif +#endif + +#if defined(__clang_major__) + { + ss << "clang " << __clang_major__ << "." << __clang_minor__ << "." + << __clang_patchlevel__; + } +#endif + +#if defined(_MSC_VER) + { ss << "MSVC " << _MSC_FULL_VER; } +#endif + return ss.str(); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("get_compiler_version", &get_compiler_version, "get_compiler_version"); + m.def("get_cuda_version", &get_cuda_version, "get_cuda_version"); + m.def("has_cuda", &has_cuda, "has_cuda"); + + m.def("box_iou_rotated", &box_iou_rotated, "IoU for rotated boxes"); + + m.def("deform_conv_forward", &deform_conv_forward, "deform_conv_forward"); + m.def( + "deform_conv_backward_input", + &deform_conv_backward_input, + "deform_conv_backward_input"); + m.def( + "deform_conv_backward_filter", + &deform_conv_backward_filter, + "deform_conv_backward_filter"); + m.def( + "modulated_deform_conv_forward", + &modulated_deform_conv_forward, + "modulated_deform_conv_forward"); + m.def( + "modulated_deform_conv_backward", + &modulated_deform_conv_backward, + "modulated_deform_conv_backward"); + + m.def("nms_rotated", &nms_rotated, "NMS for rotated boxes"); + + m.def( + "roi_align_rotated_forward", + &ROIAlignRotated_forward, + "Forward pass for Rotated ROI-Align Operator"); + m.def( + "roi_align_rotated_backward", + &ROIAlignRotated_backward, + "Backward pass for Rotated ROI-Align Operator"); + + m.def("COCOevalAccumulate", &COCOeval::Accumulate, "COCOeval::Accumulate"); + m.def( + "COCOevalEvaluateImages", + &COCOeval::EvaluateImages, + "COCOeval::EvaluateImages"); + pybind11::class_(m, "InstanceAnnotation") + .def(pybind11::init()); + pybind11::class_(m, "ImageEvaluation") + .def(pybind11::init<>()); +} + +#ifdef TORCH_LIBRARY +TORCH_LIBRARY(detectron2, m) { + m.def("nms_rotated", &nms_rotated); +} +#endif +} // namespace detectron2 diff --git a/detectron2/layers/deform_conv.py b/detectron2/layers/deform_conv.py new file mode 100644 index 0000000000000000000000000000000000000000..eca070f59645af4c9ccd003d99678f19538f355d --- /dev/null +++ b/detectron2/layers/deform_conv.py @@ -0,0 +1,501 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import math +from functools import lru_cache +import torch +from torch import nn +from torch.autograd import Function +from torch.autograd.function import once_differentiable +from torch.nn.modules.utils import _pair +from torchvision.ops import deform_conv2d + +from detectron2 import _C + +from .wrappers import _NewEmptyTensorOp + + +class _DeformConv(Function): + @staticmethod + def forward( + ctx, + input, + offset, + weight, + stride=1, + padding=0, + dilation=1, + groups=1, + deformable_groups=1, + im2col_step=64, + ): + if input is not None and input.dim() != 4: + raise ValueError( + "Expected 4D tensor as input, got {}D tensor instead.".format(input.dim()) + ) + ctx.stride = _pair(stride) + ctx.padding = _pair(padding) + ctx.dilation = _pair(dilation) + ctx.groups = groups + ctx.deformable_groups = deformable_groups + ctx.im2col_step = im2col_step + + ctx.save_for_backward(input, offset, weight) + + output = input.new_empty( + _DeformConv._output_size(input, weight, ctx.padding, ctx.dilation, ctx.stride) + ) + + ctx.bufs_ = [input.new_empty(0), input.new_empty(0)] # columns, ones + + if not input.is_cuda: + if deformable_groups != 1: + raise NotImplementedError( + "Deformable Conv with deformable_groups != 1 is not supported on CPUs!" + ) + return deform_conv2d( + input, offset, weight, stride=stride, padding=padding, dilation=dilation + ) + else: + cur_im2col_step = _DeformConv._cal_im2col_step(input.shape[0], ctx.im2col_step) + assert (input.shape[0] % cur_im2col_step) == 0, "im2col step must divide batchsize" + + _C.deform_conv_forward( + input, + weight, + offset, + output, + ctx.bufs_[0], + ctx.bufs_[1], + weight.size(3), + weight.size(2), + ctx.stride[1], + ctx.stride[0], + ctx.padding[1], + ctx.padding[0], + ctx.dilation[1], + ctx.dilation[0], + ctx.groups, + ctx.deformable_groups, + cur_im2col_step, + ) + return output + + @staticmethod + @once_differentiable + def backward(ctx, grad_output): + input, offset, weight = ctx.saved_tensors + + grad_input = grad_offset = grad_weight = None + + if not grad_output.is_cuda: + raise NotImplementedError("Deformable Conv is not supported on CPUs!") + else: + cur_im2col_step = _DeformConv._cal_im2col_step(input.shape[0], ctx.im2col_step) + assert (input.shape[0] % cur_im2col_step) == 0, "im2col step must divide batchsize" + + if ctx.needs_input_grad[0] or ctx.needs_input_grad[1]: + grad_input = torch.zeros_like(input) + grad_offset = torch.zeros_like(offset) + _C.deform_conv_backward_input( + input, + offset, + grad_output, + grad_input, + grad_offset, + weight, + ctx.bufs_[0], + weight.size(3), + weight.size(2), + ctx.stride[1], + ctx.stride[0], + ctx.padding[1], + ctx.padding[0], + ctx.dilation[1], + ctx.dilation[0], + ctx.groups, + ctx.deformable_groups, + cur_im2col_step, + ) + + if ctx.needs_input_grad[2]: + grad_weight = torch.zeros_like(weight) + _C.deform_conv_backward_filter( + input, + offset, + grad_output, + grad_weight, + ctx.bufs_[0], + ctx.bufs_[1], + weight.size(3), + weight.size(2), + ctx.stride[1], + ctx.stride[0], + ctx.padding[1], + ctx.padding[0], + ctx.dilation[1], + ctx.dilation[0], + ctx.groups, + ctx.deformable_groups, + 1, + cur_im2col_step, + ) + + return grad_input, grad_offset, grad_weight, None, None, None, None, None, None + + @staticmethod + def _output_size(input, weight, padding, dilation, stride): + channels = weight.size(0) + output_size = (input.size(0), channels) + for d in range(input.dim() - 2): + in_size = input.size(d + 2) + pad = padding[d] + kernel = dilation[d] * (weight.size(d + 2) - 1) + 1 + stride_ = stride[d] + output_size += ((in_size + (2 * pad) - kernel) // stride_ + 1,) + if not all(map(lambda s: s > 0, output_size)): + raise ValueError( + "convolution input is too small (output would be {})".format( + "x".join(map(str, output_size)) + ) + ) + return output_size + + @staticmethod + @lru_cache(maxsize=128) + def _cal_im2col_step(input_size, default_size): + """ + Calculate proper im2col step size, which should be divisible by input_size and not larger + than prefer_size. Meanwhile the step size should be as large as possible to be more + efficient. So we choose the largest one among all divisors of input_size which are smaller + than prefer_size. + :param input_size: input batch size . + :param default_size: default preferred im2col step size. + :return: the largest proper step size. + """ + if input_size <= default_size: + return input_size + best_step = 1 + for step in range(2, min(int(math.sqrt(input_size)) + 1, default_size)): + if input_size % step == 0: + if input_size // step <= default_size: + return input_size // step + best_step = step + + return best_step + + +class _ModulatedDeformConv(Function): + @staticmethod + def forward( + ctx, + input, + offset, + mask, + weight, + bias=None, + stride=1, + padding=0, + dilation=1, + groups=1, + deformable_groups=1, + ): + ctx.stride = stride + ctx.padding = padding + ctx.dilation = dilation + ctx.groups = groups + ctx.deformable_groups = deformable_groups + ctx.with_bias = bias is not None + if not ctx.with_bias: + bias = input.new_empty(1) # fake tensor + if not input.is_cuda: + raise NotImplementedError("Deformable Conv is not supported on CPUs!") + if ( + weight.requires_grad + or mask.requires_grad + or offset.requires_grad + or input.requires_grad + ): + ctx.save_for_backward(input, offset, mask, weight, bias) + output = input.new_empty(_ModulatedDeformConv._infer_shape(ctx, input, weight)) + ctx._bufs = [input.new_empty(0), input.new_empty(0)] + _C.modulated_deform_conv_forward( + input, + weight, + bias, + ctx._bufs[0], + offset, + mask, + output, + ctx._bufs[1], + weight.shape[2], + weight.shape[3], + ctx.stride, + ctx.stride, + ctx.padding, + ctx.padding, + ctx.dilation, + ctx.dilation, + ctx.groups, + ctx.deformable_groups, + ctx.with_bias, + ) + return output + + @staticmethod + @once_differentiable + def backward(ctx, grad_output): + if not grad_output.is_cuda: + raise NotImplementedError("Deformable Conv is not supported on CPUs!") + input, offset, mask, weight, bias = ctx.saved_tensors + grad_input = torch.zeros_like(input) + grad_offset = torch.zeros_like(offset) + grad_mask = torch.zeros_like(mask) + grad_weight = torch.zeros_like(weight) + grad_bias = torch.zeros_like(bias) + _C.modulated_deform_conv_backward( + input, + weight, + bias, + ctx._bufs[0], + offset, + mask, + ctx._bufs[1], + grad_input, + grad_weight, + grad_bias, + grad_offset, + grad_mask, + grad_output, + weight.shape[2], + weight.shape[3], + ctx.stride, + ctx.stride, + ctx.padding, + ctx.padding, + ctx.dilation, + ctx.dilation, + ctx.groups, + ctx.deformable_groups, + ctx.with_bias, + ) + if not ctx.with_bias: + grad_bias = None + + return ( + grad_input, + grad_offset, + grad_mask, + grad_weight, + grad_bias, + None, + None, + None, + None, + None, + ) + + @staticmethod + def _infer_shape(ctx, input, weight): + n = input.size(0) + channels_out = weight.size(0) + height, width = input.shape[2:4] + kernel_h, kernel_w = weight.shape[2:4] + height_out = ( + height + 2 * ctx.padding - (ctx.dilation * (kernel_h - 1) + 1) + ) // ctx.stride + 1 + width_out = ( + width + 2 * ctx.padding - (ctx.dilation * (kernel_w - 1) + 1) + ) // ctx.stride + 1 + return n, channels_out, height_out, width_out + + +deform_conv = _DeformConv.apply +modulated_deform_conv = _ModulatedDeformConv.apply + + +class DeformConv(nn.Module): + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + deformable_groups=1, + bias=False, + norm=None, + activation=None, + ): + """ + Deformable convolution from :paper:`deformconv`. + + Arguments are similar to :class:`Conv2D`. Extra arguments: + + Args: + deformable_groups (int): number of groups used in deformable convolution. + norm (nn.Module, optional): a normalization layer + activation (callable(Tensor) -> Tensor): a callable activation function + """ + super(DeformConv, self).__init__() + + assert not bias + assert in_channels % groups == 0, "in_channels {} cannot be divisible by groups {}".format( + in_channels, groups + ) + assert ( + out_channels % groups == 0 + ), "out_channels {} cannot be divisible by groups {}".format(out_channels, groups) + + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _pair(kernel_size) + self.stride = _pair(stride) + self.padding = _pair(padding) + self.dilation = _pair(dilation) + self.groups = groups + self.deformable_groups = deformable_groups + self.norm = norm + self.activation = activation + + self.weight = nn.Parameter( + torch.Tensor(out_channels, in_channels // self.groups, *self.kernel_size) + ) + self.bias = None + + nn.init.kaiming_uniform_(self.weight, nonlinearity="relu") + + def forward(self, x, offset): + if x.numel() == 0: + # When input is empty, we want to return a empty tensor with "correct" shape, + # So that the following operations will not panic + # if they check for the shape of the tensor. + # This computes the height and width of the output tensor + output_shape = [ + (i + 2 * p - (di * (k - 1) + 1)) // s + 1 + for i, p, di, k, s in zip( + x.shape[-2:], self.padding, self.dilation, self.kernel_size, self.stride + ) + ] + output_shape = [x.shape[0], self.weight.shape[0]] + output_shape + return _NewEmptyTensorOp.apply(x, output_shape) + + x = deform_conv( + x, + offset, + self.weight, + self.stride, + self.padding, + self.dilation, + self.groups, + self.deformable_groups, + ) + if self.norm is not None: + x = self.norm(x) + if self.activation is not None: + x = self.activation(x) + return x + + def extra_repr(self): + tmpstr = "in_channels=" + str(self.in_channels) + tmpstr += ", out_channels=" + str(self.out_channels) + tmpstr += ", kernel_size=" + str(self.kernel_size) + tmpstr += ", stride=" + str(self.stride) + tmpstr += ", padding=" + str(self.padding) + tmpstr += ", dilation=" + str(self.dilation) + tmpstr += ", groups=" + str(self.groups) + tmpstr += ", deformable_groups=" + str(self.deformable_groups) + tmpstr += ", bias=False" + return tmpstr + + +class ModulatedDeformConv(nn.Module): + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + padding=0, + dilation=1, + groups=1, + deformable_groups=1, + bias=True, + norm=None, + activation=None, + ): + """ + Modulated deformable convolution from :paper:`deformconv2`. + + Arguments are similar to :class:`Conv2D`. Extra arguments: + + Args: + deformable_groups (int): number of groups used in deformable convolution. + norm (nn.Module, optional): a normalization layer + activation (callable(Tensor) -> Tensor): a callable activation function + """ + super(ModulatedDeformConv, self).__init__() + self.in_channels = in_channels + self.out_channels = out_channels + self.kernel_size = _pair(kernel_size) + self.stride = stride + self.padding = padding + self.dilation = dilation + self.groups = groups + self.deformable_groups = deformable_groups + self.with_bias = bias + self.norm = norm + self.activation = activation + + self.weight = nn.Parameter( + torch.Tensor(out_channels, in_channels // groups, *self.kernel_size) + ) + if bias: + self.bias = nn.Parameter(torch.Tensor(out_channels)) + else: + self.bias = None + + nn.init.kaiming_uniform_(self.weight, nonlinearity="relu") + if self.bias is not None: + nn.init.constant_(self.bias, 0) + + def forward(self, x, offset, mask): + if x.numel() == 0: + output_shape = [ + (i + 2 * p - (di * (k - 1) + 1)) // s + 1 + for i, p, di, k, s in zip( + x.shape[-2:], self.padding, self.dilation, self.kernel_size, self.stride + ) + ] + output_shape = [x.shape[0], self.weight.shape[0]] + output_shape + return _NewEmptyTensorOp.apply(x, output_shape) + + x = modulated_deform_conv( + x, + offset, + mask, + self.weight, + self.bias, + self.stride, + self.padding, + self.dilation, + self.groups, + self.deformable_groups, + ) + if self.norm is not None: + x = self.norm(x) + if self.activation is not None: + x = self.activation(x) + return x + + def extra_repr(self): + tmpstr = "in_channels=" + str(self.in_channels) + tmpstr += ", out_channels=" + str(self.out_channels) + tmpstr += ", kernel_size=" + str(self.kernel_size) + tmpstr += ", stride=" + str(self.stride) + tmpstr += ", padding=" + str(self.padding) + tmpstr += ", dilation=" + str(self.dilation) + tmpstr += ", groups=" + str(self.groups) + tmpstr += ", deformable_groups=" + str(self.deformable_groups) + tmpstr += ", bias=" + str(self.with_bias) + return tmpstr diff --git a/detectron2/layers/mask_ops.py b/detectron2/layers/mask_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..c698a03c4d3faf30c08da97169f010b64c0d1058 --- /dev/null +++ b/detectron2/layers/mask_ops.py @@ -0,0 +1,260 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import numpy as np +from typing import Tuple +import torch +from PIL import Image +from torch.nn import functional as F + +from detectron2.structures import Boxes + +__all__ = ["paste_masks_in_image"] + + +BYTES_PER_FLOAT = 4 +# TODO: This memory limit may be too much or too little. It would be better to +# determine it based on available resources. +GPU_MEM_LIMIT = 1024 ** 3 # 1 GB memory limit + + +def _do_paste_mask(masks, boxes, img_h: int, img_w: int, skip_empty: bool = True): + """ + Args: + masks: N, 1, H, W + boxes: N, 4 + img_h, img_w (int): + skip_empty (bool): only paste masks within the region that + tightly bound all boxes, and returns the results this region only. + An important optimization for CPU. + + Returns: + if skip_empty == False, a mask of shape (N, img_h, img_w) + if skip_empty == True, a mask of shape (N, h', w'), and the slice + object for the corresponding region. + """ + # On GPU, paste all masks together (up to chunk size) + # by using the entire image to sample the masks + # Compared to pasting them one by one, + # this has more operations but is faster on COCO-scale dataset. + device = masks.device + + if skip_empty and not torch.jit.is_scripting(): + x0_int, y0_int = torch.clamp(boxes.min(dim=0).values.floor()[:2] - 1, min=0).to( + dtype=torch.int32 + ) + x1_int = torch.clamp(boxes[:, 2].max().ceil() + 1, max=img_w).to(dtype=torch.int32) + y1_int = torch.clamp(boxes[:, 3].max().ceil() + 1, max=img_h).to(dtype=torch.int32) + else: + x0_int, y0_int = 0, 0 + x1_int, y1_int = img_w, img_h + x0, y0, x1, y1 = torch.split(boxes, 1, dim=1) # each is Nx1 + + N = masks.shape[0] + + img_y = torch.arange(y0_int, y1_int, device=device, dtype=torch.float32) + 0.5 + img_x = torch.arange(x0_int, x1_int, device=device, dtype=torch.float32) + 0.5 + img_y = (img_y - y0) / (y1 - y0) * 2 - 1 + img_x = (img_x - x0) / (x1 - x0) * 2 - 1 + # img_x, img_y have shapes (N, w), (N, h) + + gx = img_x[:, None, :].expand(N, img_y.size(1), img_x.size(1)) + gy = img_y[:, :, None].expand(N, img_y.size(1), img_x.size(1)) + grid = torch.stack([gx, gy], dim=3) + + if not torch.jit.is_scripting(): + if not masks.dtype.is_floating_point: + masks = masks.float() + img_masks = F.grid_sample(masks, grid.to(masks.dtype), align_corners=False) + + if skip_empty and not torch.jit.is_scripting(): + return img_masks[:, 0], (slice(y0_int, y1_int), slice(x0_int, x1_int)) + else: + return img_masks[:, 0], () + + +def paste_masks_in_image( + masks: torch.Tensor, boxes: Boxes, image_shape: Tuple[int, int], threshold: float = 0.5 +): + """ + Paste a set of masks that are of a fixed resolution (e.g., 28 x 28) into an image. + The location, height, and width for pasting each mask is determined by their + corresponding bounding boxes in boxes. + + Note: + This is a complicated but more accurate implementation. In actual deployment, it is + often enough to use a faster but less accurate implementation. + See :func:`paste_mask_in_image_old` in this file for an alternative implementation. + + Args: + masks (tensor): Tensor of shape (Bimg, Hmask, Wmask), where Bimg is the number of + detected object instances in the image and Hmask, Wmask are the mask width and mask + height of the predicted mask (e.g., Hmask = Wmask = 28). Values are in [0, 1]. + boxes (Boxes or Tensor): A Boxes of length Bimg or Tensor of shape (Bimg, 4). + boxes[i] and masks[i] correspond to the same object instance. + image_shape (tuple): height, width + threshold (float): A threshold in [0, 1] for converting the (soft) masks to + binary masks. + + Returns: + img_masks (Tensor): A tensor of shape (Bimg, Himage, Wimage), where Bimg is the + number of detected object instances and Himage, Wimage are the image width + and height. img_masks[i] is a binary mask for object instance i. + """ + + assert masks.shape[-1] == masks.shape[-2], "Only square mask predictions are supported" + N = len(masks) + if N == 0: + return masks.new_empty((0,) + image_shape, dtype=torch.uint8) + if not isinstance(boxes, torch.Tensor): + boxes = boxes.tensor + device = boxes.device + assert len(boxes) == N, boxes.shape + + img_h, img_w = image_shape + + # The actual implementation split the input into chunks, + # and paste them chunk by chunk. + if device.type == "cpu" or torch.jit.is_scripting(): + # CPU is most efficient when they are pasted one by one with skip_empty=True + # so that it performs minimal number of operations. + num_chunks = N + else: + # GPU benefits from parallelism for larger chunks, but may have memory issue + # int(img_h) because shape may be tensors in tracing + num_chunks = int(np.ceil(N * int(img_h) * int(img_w) * BYTES_PER_FLOAT / GPU_MEM_LIMIT)) + assert ( + num_chunks <= N + ), "Default GPU_MEM_LIMIT in mask_ops.py is too small; try increasing it" + chunks = torch.chunk(torch.arange(N, device=device), num_chunks) + + img_masks = torch.zeros( + N, img_h, img_w, device=device, dtype=torch.bool if threshold >= 0 else torch.uint8 + ) + for inds in chunks: + masks_chunk, spatial_inds = _do_paste_mask( + masks[inds, None, :, :], boxes[inds], img_h, img_w, skip_empty=device.type == "cpu" + ) + + if threshold >= 0: + masks_chunk = (masks_chunk >= threshold).to(dtype=torch.bool) + else: + # for visualization and debugging + masks_chunk = (masks_chunk * 255).to(dtype=torch.uint8) + + if torch.jit.is_scripting(): # Scripting does not use the optimized codepath + img_masks[inds] = masks_chunk + else: + img_masks[(inds,) + spatial_inds] = masks_chunk + return img_masks + + +# The below are the original paste function (from Detectron1) which has +# larger quantization error. +# It is faster on CPU, while the aligned one is faster on GPU thanks to grid_sample. + + +def paste_mask_in_image_old(mask, box, img_h, img_w, threshold): + """ + Paste a single mask in an image. + This is a per-box implementation of :func:`paste_masks_in_image`. + This function has larger quantization error due to incorrect pixel + modeling and is not used any more. + + Args: + mask (Tensor): A tensor of shape (Hmask, Wmask) storing the mask of a single + object instance. Values are in [0, 1]. + box (Tensor): A tensor of shape (4, ) storing the x0, y0, x1, y1 box corners + of the object instance. + img_h, img_w (int): Image height and width. + threshold (float): Mask binarization threshold in [0, 1]. + + Returns: + im_mask (Tensor): + The resized and binarized object mask pasted into the original + image plane (a tensor of shape (img_h, img_w)). + """ + # Conversion from continuous box coordinates to discrete pixel coordinates + # via truncation (cast to int32). This determines which pixels to paste the + # mask onto. + box = box.to(dtype=torch.int32) # Continuous to discrete coordinate conversion + # An example (1D) box with continuous coordinates (x0=0.7, x1=4.3) will map to + # a discrete coordinates (x0=0, x1=4). Note that box is mapped to 5 = x1 - x0 + 1 + # pixels (not x1 - x0 pixels). + samples_w = box[2] - box[0] + 1 # Number of pixel samples, *not* geometric width + samples_h = box[3] - box[1] + 1 # Number of pixel samples, *not* geometric height + + # Resample the mask from it's original grid to the new samples_w x samples_h grid + mask = Image.fromarray(mask.cpu().numpy()) + mask = mask.resize((samples_w, samples_h), resample=Image.BILINEAR) + mask = np.array(mask, copy=False) + + if threshold >= 0: + mask = np.array(mask > threshold, dtype=np.uint8) + mask = torch.from_numpy(mask) + else: + # for visualization and debugging, we also + # allow it to return an unmodified mask + mask = torch.from_numpy(mask * 255).to(torch.uint8) + + im_mask = torch.zeros((img_h, img_w), dtype=torch.uint8) + x_0 = max(box[0], 0) + x_1 = min(box[2] + 1, img_w) + y_0 = max(box[1], 0) + y_1 = min(box[3] + 1, img_h) + + im_mask[y_0:y_1, x_0:x_1] = mask[ + (y_0 - box[1]) : (y_1 - box[1]), (x_0 - box[0]) : (x_1 - box[0]) + ] + return im_mask + + +# Our pixel modeling requires extrapolation for any continuous +# coordinate < 0.5 or > length - 0.5. When sampling pixels on the masks, +# we would like this extrapolation to be an interpolation between boundary values and zero, +# instead of using absolute zero or boundary values. +# Therefore `paste_mask_in_image_old` is often used with zero padding around the masks like this: +# masks, scale = pad_masks(masks[:, 0, :, :], 1) +# boxes = scale_boxes(boxes.tensor, scale) + + +def pad_masks(masks, padding): + """ + Args: + masks (tensor): A tensor of shape (B, M, M) representing B masks. + padding (int): Number of cells to pad on all sides. + + Returns: + The padded masks and the scale factor of the padding size / original size. + """ + B = masks.shape[0] + M = masks.shape[-1] + pad2 = 2 * padding + scale = float(M + pad2) / M + padded_masks = masks.new_zeros((B, M + pad2, M + pad2)) + padded_masks[:, padding:-padding, padding:-padding] = masks + return padded_masks, scale + + +def scale_boxes(boxes, scale): + """ + Args: + boxes (tensor): A tensor of shape (B, 4) representing B boxes with 4 + coords representing the corners x0, y0, x1, y1, + scale (float): The box scaling factor. + + Returns: + Scaled boxes. + """ + w_half = (boxes[:, 2] - boxes[:, 0]) * 0.5 + h_half = (boxes[:, 3] - boxes[:, 1]) * 0.5 + x_c = (boxes[:, 2] + boxes[:, 0]) * 0.5 + y_c = (boxes[:, 3] + boxes[:, 1]) * 0.5 + + w_half *= scale + h_half *= scale + + scaled_boxes = torch.zeros_like(boxes) + scaled_boxes[:, 0] = x_c - w_half + scaled_boxes[:, 2] = x_c + w_half + scaled_boxes[:, 1] = y_c - h_half + scaled_boxes[:, 3] = y_c + h_half + return scaled_boxes diff --git a/detectron2/layers/nms.py b/detectron2/layers/nms.py new file mode 100644 index 0000000000000000000000000000000000000000..ac14d459259b19a1a145adff2817a0ca0441b7eb --- /dev/null +++ b/detectron2/layers/nms.py @@ -0,0 +1,158 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +from typing import List +import torch +from torchvision.ops import boxes as box_ops +from torchvision.ops import nms # BC-compat + +from detectron2.utils.env import TORCH_VERSION + +if TORCH_VERSION < (1, 7): + from detectron2 import _C + + nms_rotated_func = _C.nms_rotated +else: + nms_rotated_func = torch.ops.detectron2.nms_rotated + + +def batched_nms( + boxes: torch.Tensor, scores: torch.Tensor, idxs: torch.Tensor, iou_threshold: float +): + """ + Same as torchvision.ops.boxes.batched_nms, but safer. + """ + assert boxes.shape[-1] == 4 + # TODO may need better strategy. + # Investigate after having a fully-cuda NMS op. + if len(boxes) < 40000: + # fp16 does not have enough range for batched NMS + return box_ops.batched_nms(boxes.float(), scores, idxs, iou_threshold) + + result_mask = scores.new_zeros(scores.size(), dtype=torch.bool) + for id in torch.jit.annotate(List[int], torch.unique(idxs).cpu().tolist()): + mask = (idxs == id).nonzero().view(-1) + keep = nms(boxes[mask], scores[mask], iou_threshold) + result_mask[mask[keep]] = True + keep = result_mask.nonzero().view(-1) + keep = keep[scores[keep].argsort(descending=True)] + return keep + + +# Note: this function (nms_rotated) might be moved into +# torchvision/ops/boxes.py in the future +def nms_rotated(boxes, scores, iou_threshold): + """ + Performs non-maximum suppression (NMS) on the rotated boxes according + to their intersection-over-union (IoU). + + Rotated NMS iteratively removes lower scoring rotated boxes which have an + IoU greater than iou_threshold with another (higher scoring) rotated box. + + Note that RotatedBox (5, 3, 4, 2, -90) covers exactly the same region as + RotatedBox (5, 3, 4, 2, 90) does, and their IoU will be 1. However, they + can be representing completely different objects in certain tasks, e.g., OCR. + + As for the question of whether rotated-NMS should treat them as faraway boxes + even though their IOU is 1, it depends on the application and/or ground truth annotation. + + As an extreme example, consider a single character v and the square box around it. + + If the angle is 0 degree, the object (text) would be read as 'v'; + + If the angle is 90 degrees, the object (text) would become '>'; + + If the angle is 180 degrees, the object (text) would become '^'; + + If the angle is 270/-90 degrees, the object (text) would become '<' + + All of these cases have IoU of 1 to each other, and rotated NMS that only + uses IoU as criterion would only keep one of them with the highest score - + which, practically, still makes sense in most cases because typically + only one of theses orientations is the correct one. Also, it does not matter + as much if the box is only used to classify the object (instead of transcribing + them with a sequential OCR recognition model) later. + + On the other hand, when we use IoU to filter proposals that are close to the + ground truth during training, we should definitely take the angle into account if + we know the ground truth is labeled with the strictly correct orientation (as in, + upside-down words are annotated with -180 degrees even though they can be covered + with a 0/90/-90 degree box, etc.) + + The way the original dataset is annotated also matters. For example, if the dataset + is a 4-point polygon dataset that does not enforce ordering of vertices/orientation, + we can estimate a minimum rotated bounding box to this polygon, but there's no way + we can tell the correct angle with 100% confidence (as shown above, there could be 4 different + rotated boxes, with angles differed by 90 degrees to each other, covering the exactly + same region). In that case we have to just use IoU to determine the box + proximity (as many detection benchmarks (even for text) do) unless there're other + assumptions we can make (like width is always larger than height, or the object is not + rotated by more than 90 degrees CCW/CW, etc.) + + In summary, not considering angles in rotated NMS seems to be a good option for now, + but we should be aware of its implications. + + Args: + boxes (Tensor[N, 5]): Rotated boxes to perform NMS on. They are expected to be in + (x_center, y_center, width, height, angle_degrees) format. + scores (Tensor[N]): Scores for each one of the rotated boxes + iou_threshold (float): Discards all overlapping rotated boxes with IoU < iou_threshold + + Returns: + keep (Tensor): int64 tensor with the indices of the elements that have been kept + by Rotated NMS, sorted in decreasing order of scores + """ + return nms_rotated_func(boxes, scores, iou_threshold) + + +# Note: this function (batched_nms_rotated) might be moved into +# torchvision/ops/boxes.py in the future +def batched_nms_rotated(boxes, scores, idxs, iou_threshold): + """ + Performs non-maximum suppression in a batched fashion. + + Each index value correspond to a category, and NMS + will not be applied between elements of different categories. + + Args: + boxes (Tensor[N, 5]): + boxes where NMS will be performed. They + are expected to be in (x_ctr, y_ctr, width, height, angle_degrees) format + scores (Tensor[N]): + scores for each one of the boxes + idxs (Tensor[N]): + indices of the categories for each one of the boxes. + iou_threshold (float): + discards all overlapping boxes + with IoU < iou_threshold + + Returns: + Tensor: + int64 tensor with the indices of the elements that have been kept + by NMS, sorted in decreasing order of scores + """ + assert boxes.shape[-1] == 5 + + if boxes.numel() == 0: + return torch.empty((0,), dtype=torch.int64, device=boxes.device) + boxes = boxes.float() # fp16 does not have enough range for batched NMS + # Strategy: in order to perform NMS independently per class, + # we add an offset to all the boxes. The offset is dependent + # only on the class idx, and is large enough so that boxes + # from different classes do not overlap + + # Note that batched_nms in torchvision/ops/boxes.py only uses max_coordinate, + # which won't handle negative coordinates correctly. + # Here by using min_coordinate we can make sure the negative coordinates are + # correctly handled. + max_coordinate = ( + torch.max(boxes[:, 0], boxes[:, 1]) + torch.max(boxes[:, 2], boxes[:, 3]) / 2 + ).max() + min_coordinate = ( + torch.min(boxes[:, 0], boxes[:, 1]) - torch.max(boxes[:, 2], boxes[:, 3]) / 2 + ).min() + offsets = idxs.to(boxes) * (max_coordinate - min_coordinate + 1) + boxes_for_nms = boxes.clone() # avoid modifying the original values in boxes + boxes_for_nms[:, :2] += offsets[:, None] + keep = nms_rotated(boxes_for_nms, scores, iou_threshold) + return keep diff --git a/detectron2/layers/roi_align.py b/detectron2/layers/roi_align.py new file mode 100644 index 0000000000000000000000000000000000000000..163462e1f194e1e4100da92d76d9516f7cc22e35 --- /dev/null +++ b/detectron2/layers/roi_align.py @@ -0,0 +1,74 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from torch import nn +from torchvision.ops import roi_align + + +# NOTE: torchvision's RoIAlign has a different default aligned=False +class ROIAlign(nn.Module): + def __init__(self, output_size, spatial_scale, sampling_ratio, aligned=True): + """ + Args: + output_size (tuple): h, w + spatial_scale (float): scale the input boxes by this number + sampling_ratio (int): number of inputs samples to take for each output + sample. 0 to take samples densely. + aligned (bool): if False, use the legacy implementation in + Detectron. If True, align the results more perfectly. + + Note: + The meaning of aligned=True: + + Given a continuous coordinate c, its two neighboring pixel indices (in our + pixel model) are computed by floor(c - 0.5) and ceil(c - 0.5). For example, + c=1.3 has pixel neighbors with discrete indices [0] and [1] (which are sampled + from the underlying signal at continuous coordinates 0.5 and 1.5). But the original + roi_align (aligned=False) does not subtract the 0.5 when computing neighboring + pixel indices and therefore it uses pixels with a slightly incorrect alignment + (relative to our pixel model) when performing bilinear interpolation. + + With `aligned=True`, + we first appropriately scale the ROI and then shift it by -0.5 + prior to calling roi_align. This produces the correct neighbors; see + detectron2/tests/test_roi_align.py for verification. + + The difference does not make a difference to the model's performance if + ROIAlign is used together with conv layers. + """ + super().__init__() + self.output_size = output_size + self.spatial_scale = spatial_scale + self.sampling_ratio = sampling_ratio + self.aligned = aligned + + from torchvision import __version__ + + version = tuple(int(x) for x in __version__.split(".")[:2]) + # https://github.com/pytorch/vision/pull/2438 + assert version >= (0, 7), "Require torchvision >= 0.7" + + def forward(self, input, rois): + """ + Args: + input: NCHW images + rois: Bx5 boxes. First column is the index into N. The other 4 columns are xyxy. + """ + assert rois.dim() == 2 and rois.size(1) == 5 + if input.is_quantized: + input = input.dequantize() + return roi_align( + input, + rois.to(dtype=input.dtype), + self.output_size, + self.spatial_scale, + self.sampling_ratio, + self.aligned, + ) + + def __repr__(self): + tmpstr = self.__class__.__name__ + "(" + tmpstr += "output_size=" + str(self.output_size) + tmpstr += ", spatial_scale=" + str(self.spatial_scale) + tmpstr += ", sampling_ratio=" + str(self.sampling_ratio) + tmpstr += ", aligned=" + str(self.aligned) + tmpstr += ")" + return tmpstr diff --git a/detectron2/layers/roi_align_rotated.py b/detectron2/layers/roi_align_rotated.py new file mode 100644 index 0000000000000000000000000000000000000000..e3775e08fc9b9172f73c8ec7025a51ef2edd0a1d --- /dev/null +++ b/detectron2/layers/roi_align_rotated.py @@ -0,0 +1,93 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import torch +from torch import nn +from torch.autograd import Function +from torch.autograd.function import once_differentiable +from torch.nn.modules.utils import _pair + +from detectron2 import _C + + +class _ROIAlignRotated(Function): + @staticmethod + def forward(ctx, input, roi, output_size, spatial_scale, sampling_ratio): + ctx.save_for_backward(roi) + ctx.output_size = _pair(output_size) + ctx.spatial_scale = spatial_scale + ctx.sampling_ratio = sampling_ratio + ctx.input_shape = input.size() + output = _C.roi_align_rotated_forward( + input, roi, spatial_scale, output_size[0], output_size[1], sampling_ratio + ) + return output + + @staticmethod + @once_differentiable + def backward(ctx, grad_output): + (rois,) = ctx.saved_tensors + output_size = ctx.output_size + spatial_scale = ctx.spatial_scale + sampling_ratio = ctx.sampling_ratio + bs, ch, h, w = ctx.input_shape + grad_input = _C.roi_align_rotated_backward( + grad_output, + rois, + spatial_scale, + output_size[0], + output_size[1], + bs, + ch, + h, + w, + sampling_ratio, + ) + return grad_input, None, None, None, None, None + + +roi_align_rotated = _ROIAlignRotated.apply + + +class ROIAlignRotated(nn.Module): + def __init__(self, output_size, spatial_scale, sampling_ratio): + """ + Args: + output_size (tuple): h, w + spatial_scale (float): scale the input boxes by this number + sampling_ratio (int): number of inputs samples to take for each output + sample. 0 to take samples densely. + + Note: + ROIAlignRotated supports continuous coordinate by default: + Given a continuous coordinate c, its two neighboring pixel indices (in our + pixel model) are computed by floor(c - 0.5) and ceil(c - 0.5). For example, + c=1.3 has pixel neighbors with discrete indices [0] and [1] (which are sampled + from the underlying signal at continuous coordinates 0.5 and 1.5). + """ + super(ROIAlignRotated, self).__init__() + self.output_size = output_size + self.spatial_scale = spatial_scale + self.sampling_ratio = sampling_ratio + + def forward(self, input, rois): + """ + Args: + input: NCHW images + rois: Bx6 boxes. First column is the index into N. + The other 5 columns are (x_ctr, y_ctr, width, height, angle_degrees). + """ + assert rois.dim() == 2 and rois.size(1) == 6 + orig_dtype = input.dtype + if orig_dtype == torch.float16: + input = input.float() + rois = rois.float() + return roi_align_rotated( + input, rois, self.output_size, self.spatial_scale, self.sampling_ratio + ).to(dtype=orig_dtype) + + def __repr__(self): + tmpstr = self.__class__.__name__ + "(" + tmpstr += "output_size=" + str(self.output_size) + tmpstr += ", spatial_scale=" + str(self.spatial_scale) + tmpstr += ", sampling_ratio=" + str(self.sampling_ratio) + tmpstr += ")" + return tmpstr diff --git a/detectron2/layers/rotated_boxes.py b/detectron2/layers/rotated_boxes.py new file mode 100644 index 0000000000000000000000000000000000000000..0004f765ef355ad47d92d26d3012be382e1b3eca --- /dev/null +++ b/detectron2/layers/rotated_boxes.py @@ -0,0 +1,22 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from __future__ import absolute_import, division, print_function, unicode_literals + +from detectron2 import _C + + +def pairwise_iou_rotated(boxes1, boxes2): + """ + Return intersection-over-union (Jaccard index) of boxes. + + Both sets of boxes are expected to be in + (x_center, y_center, width, height, angle) format. + + Arguments: + boxes1 (Tensor[N, 5]) + boxes2 (Tensor[M, 5]) + + Returns: + iou (Tensor[N, M]): the NxM matrix containing the pairwise + IoU values for every element in boxes1 and boxes2 + """ + return _C.box_iou_rotated(boxes1, boxes2) diff --git a/detectron2/layers/shape_spec.py b/detectron2/layers/shape_spec.py new file mode 100644 index 0000000000000000000000000000000000000000..fe7e8e261c1ab1bb1636bd7a245068d64e632306 --- /dev/null +++ b/detectron2/layers/shape_spec.py @@ -0,0 +1,20 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. +from collections import namedtuple + + +class ShapeSpec(namedtuple("_ShapeSpec", ["channels", "height", "width", "stride"])): + """ + A simple structure that contains basic shape specification about a tensor. + It is often used as the auxiliary inputs/outputs of models, + to complement the lack of shape inference ability among pytorch modules. + + Attributes: + channels: + height: + width: + stride: + """ + + def __new__(cls, channels=None, height=None, width=None, stride=None): + return super().__new__(cls, channels, height, width, stride) diff --git a/detectron2/layers/soft_nms.py b/detectron2/layers/soft_nms.py new file mode 100644 index 0000000000000000000000000000000000000000..576089db79278bbd62aa87b17fc1ed13ac2261c7 --- /dev/null +++ b/detectron2/layers/soft_nms.py @@ -0,0 +1,261 @@ +import torch + +from detectron2.structures import Boxes, RotatedBoxes, pairwise_iou, pairwise_iou_rotated + +""" Soft-NMS Pull request from: https://github.com/facebookresearch/detectron2/pull/1183/files +""" + +def soft_nms(boxes, scores, method, gaussian_sigma, linear_threshold, prune_threshold): + """ + Performs soft non-maximum suppression algorithm on axis aligned boxes + Args: + boxes (Tensor[N, 5]): + boxes where NMS will be performed. They + are expected to be in (x_ctr, y_ctr, width, height, angle_degrees) format + scores (Tensor[N]): + scores for each one of the boxes + method (str): + one of ['gaussian', 'linear', 'hard'] + see paper for details. users encouraged not to use "hard", as this is the + same nms available elsewhere in detectron2 + gaussian_sigma (float): + parameter for Gaussian penalty function + linear_threshold (float): + iou threshold for applying linear decay. Nt from the paper + re-used as threshold for standard "hard" nms + prune_threshold (float): + boxes with scores below this threshold are pruned at each iteration. + Dramatically reduces computation time. Authors use values in [10e-4, 10e-2] + Returns: + tuple(Tensor, Tensor): + [0]: int64 tensor with the indices of the elements that have been kept + by Soft NMS, sorted in decreasing order of scores + [1]: float tensor with the re-scored scores of the elements that were kept +""" + return _soft_nms( + Boxes, + pairwise_iou, + boxes, + scores, + method, + gaussian_sigma, + linear_threshold, + prune_threshold, + ) + + +def soft_nms_rotated(boxes, scores, method, gaussian_sigma, linear_threshold, prune_threshold): + """ + Performs soft non-maximum suppression algorithm on rotated boxes + Args: + boxes (Tensor[N, 5]): + boxes where NMS will be performed. They + are expected to be in (x_ctr, y_ctr, width, height, angle_degrees) format + scores (Tensor[N]): + scores for each one of the boxes + method (str): + one of ['gaussian', 'linear', 'hard'] + see paper for details. users encouraged not to use "hard", as this is the + same nms available elsewhere in detectron2 + gaussian_sigma (float): + parameter for Gaussian penalty function + linear_threshold (float): + iou threshold for applying linear decay. Nt from the paper + re-used as threshold for standard "hard" nms + prune_threshold (float): + boxes with scores below this threshold are pruned at each iteration. + Dramatically reduces computation time. Authors use values in [10e-4, 10e-2] + Returns: + tuple(Tensor, Tensor): + [0]: int64 tensor with the indices of the elements that have been kept + by Soft NMS, sorted in decreasing order of scores + [1]: float tensor with the re-scored scores of the elements that were kept """ + return _soft_nms( + RotatedBoxes, + pairwise_iou_rotated, + boxes, + scores, + method, + gaussian_sigma, + linear_threshold, + prune_threshold, + ) + + +def batched_soft_nms( + boxes, scores, idxs, method, gaussian_sigma, linear_threshold, prune_threshold +): + """ + Performs soft non-maximum suppression in a batched fashion. + Each index value correspond to a category, and NMS + will not be applied between elements of different categories. + Args: + boxes (Tensor[N, 4]): + boxes where NMS will be performed. They + are expected to be in (x1, y1, x2, y2) format + scores (Tensor[N]): + scores for each one of the boxes + idxs (Tensor[N]): + indices of the categories for each one of the boxes. + method (str): + one of ['gaussian', 'linear', 'hard'] + see paper for details. users encouraged not to use "hard", as this is the + same nms available elsewhere in detectron2 + gaussian_sigma (float): + parameter for Gaussian penalty function + linear_threshold (float): + iou threshold for applying linear decay. Nt from the paper + re-used as threshold for standard "hard" nms + prune_threshold (float): + boxes with scores below this threshold are pruned at each iteration. + Dramatically reduces computation time. Authors use values in [10e-4, 10e-2] + Returns: + tuple(Tensor, Tensor): + [0]: int64 tensor with the indices of the elements that have been kept + by Soft NMS, sorted in decreasing order of scores + [1]: float tensor with the re-scored scores of the elements that were kept + """ + if boxes.numel() == 0: + return ( + torch.empty((0,), dtype=torch.int64, device=boxes.device), + torch.empty((0,), dtype=torch.float32, device=scores.device), + ) + # strategy: in order to perform NMS independently per class. + # we add an offset to all the boxes. The offset is dependent + # only on the class idx, and is large enough so that boxes + # from different classes do not overlap + max_coordinate = boxes.max() + offsets = idxs.to(boxes) * (max_coordinate + 1) + boxes_for_nms = boxes + offsets[:, None] + return soft_nms( + boxes_for_nms, scores, method, gaussian_sigma, linear_threshold, prune_threshold + ) + + +def batched_soft_nms_rotated( + boxes, scores, idxs, method, gaussian_sigma, linear_threshold, prune_threshold +): + """ + Performs soft non-maximum suppression in a batched fashion on rotated bounding boxes. + Each index value correspond to a category, and NMS + will not be applied between elements of different categories. + Args: + boxes (Tensor[N, 5]): + boxes where NMS will be performed. They + are expected to be in (x_ctr, y_ctr, width, height, angle_degrees) format + scores (Tensor[N]): + scores for each one of the boxes + idxs (Tensor[N]): + indices of the categories for each one of the boxes. + method (str): + one of ['gaussian', 'linear', 'hard'] + see paper for details. users encouraged not to use "hard", as this is the + same nms available elsewhere in detectron2 + gaussian_sigma (float): + parameter for Gaussian penalty function + linear_threshold (float): + iou threshold for applying linear decay. Nt from the paper + re-used as threshold for standard "hard" nms + prune_threshold (float): + boxes with scores below this threshold are pruned at each iteration. + Dramatically reduces computation time. Authors use values in [10e-4, 10e-2] + Returns: + tuple(Tensor, Tensor): + [0]: int64 tensor with the indices of the elements that have been kept + by Soft NMS, sorted in decreasing order of scores + [1]: float tensor with the re-scored scores of the elements that were kept + """ + if boxes.numel() == 0: + return ( + torch.empty((0,), dtype=torch.int64, device=boxes.device), + torch.empty((0,), dtype=torch.float32, device=scores.device), + ) + # strategy: in order to perform NMS independently per class. + # we add an offset to all the boxes. The offset is dependent + # only on the class idx, and is large enough so that boxes + # from different classes do not overlap + max_coordinate = boxes[:, :2].max() + torch.norm(boxes[:, 2:4], 2, dim=1).max() + offsets = idxs.to(boxes) * (max_coordinate + 1) + boxes_for_nms = boxes.clone() + boxes_for_nms[:, :2] += offsets[:, None] + return soft_nms_rotated( + boxes_for_nms, scores, method, gaussian_sigma, linear_threshold, prune_threshold + ) + + +def _soft_nms( + box_class, + pairwise_iou_func, + boxes, + scores, + method, + gaussian_sigma, + linear_threshold, + prune_threshold, +): + """ + Soft non-max suppression algorithm. + Implementation of [Soft-NMS -- Improving Object Detection With One Line of Codec] + (https://arxiv.org/abs/1704.04503) + Args: + box_class (cls): one of Box, RotatedBoxes + pairwise_iou_func (func): one of pairwise_iou, pairwise_iou_rotated + boxes (Tensor[N, ?]): + boxes where NMS will be performed + if Boxes, in (x1, y1, x2, y2) format + if RotatedBoxes, in (x_ctr, y_ctr, width, height, angle_degrees) format + scores (Tensor[N]): + scores for each one of the boxes + method (str): + one of ['gaussian', 'linear', 'hard'] + see paper for details. users encouraged not to use "hard", as this is the + same nms available elsewhere in detectron2 + gaussian_sigma (float): + parameter for Gaussian penalty function + linear_threshold (float): + iou threshold for applying linear decay. Nt from the paper + re-used as threshold for standard "hard" nms + prune_threshold (float): + boxes with scores below this threshold are pruned at each iteration. + Dramatically reduces computation time. Authors use values in [10e-4, 10e-2] + Returns: + tuple(Tensor, Tensor): + [0]: int64 tensor with the indices of the elements that have been kept + by Soft NMS, sorted in decreasing order of scores + [1]: float tensor with the re-scored scores of the elements that were kept + """ + boxes = boxes.clone() + scores = scores.clone() + idxs = torch.arange(scores.size()[0]) + + idxs_out = [] + scores_out = [] + + while scores.numel() > 0: + top_idx = torch.argmax(scores) + idxs_out.append(idxs[top_idx].item()) + scores_out.append(scores[top_idx].item()) + + top_box = boxes[top_idx] + ious = pairwise_iou_func(box_class(top_box.unsqueeze(0)), box_class(boxes))[0] + + if method == "linear": + decay = torch.ones_like(ious) + decay_mask = ious > linear_threshold + decay[decay_mask] = 1 - ious[decay_mask] + elif method == "gaussian": + decay = torch.exp(-torch.pow(ious, 2) / gaussian_sigma) + elif method == "hard": # standard NMS + decay = (ious < linear_threshold).float() + else: + raise NotImplementedError("{} soft nms method not implemented.".format(method)) + + scores *= decay + keep = scores > prune_threshold + keep[top_idx] = False + + boxes = boxes[keep] + scores = scores[keep] + idxs = idxs[keep] + + return torch.tensor(idxs_out).to(boxes.device), torch.tensor(scores_out).to(scores.device) \ No newline at end of file diff --git a/detectron2/layers/wrappers.py b/detectron2/layers/wrappers.py new file mode 100644 index 0000000000000000000000000000000000000000..5bb4e7c1a1334c5501a6c492ddfa836dadf0beab --- /dev/null +++ b/detectron2/layers/wrappers.py @@ -0,0 +1,110 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +""" +Wrappers around on some nn functions, mainly to support empty tensors. + +Ideally, add support directly in PyTorch to empty tensors in those functions. + +These can be removed once https://github.com/pytorch/pytorch/issues/12013 +is implemented +""" + +from typing import List +import torch +from torch.nn import functional as F + + +def cat(tensors: List[torch.Tensor], dim: int = 0): + """ + Efficient version of torch.cat that avoids a copy if there is only a single element in a list + """ + assert isinstance(tensors, (list, tuple)) + if len(tensors) == 1: + return tensors[0] + return torch.cat(tensors, dim) + + +def cross_entropy(input, target, *, reduction="mean", **kwargs): + """ + Same as `torch.nn.functional.cross_entropy`, but returns 0 (instead of nan) + for empty inputs. + """ + if target.numel() == 0 and reduction == "mean": + return input.sum() * 0.0 # connect the gradient + return F.cross_entropy(input, target, **kwargs) + + +class _NewEmptyTensorOp(torch.autograd.Function): + @staticmethod + def forward(ctx, x, new_shape): + ctx.shape = x.shape + return x.new_empty(new_shape) + + @staticmethod + def backward(ctx, grad): + shape = ctx.shape + return _NewEmptyTensorOp.apply(grad, shape), None + + +class Conv2d(torch.nn.Conv2d): + """ + A wrapper around :class:`torch.nn.Conv2d` to support empty inputs and more features. + """ + + def __init__(self, *args, **kwargs): + """ + Extra keyword arguments supported in addition to those in `torch.nn.Conv2d`: + + Args: + norm (nn.Module, optional): a normalization layer + activation (callable(Tensor) -> Tensor): a callable activation function + + It assumes that norm layer is used before activation. + """ + norm = kwargs.pop("norm", None) + activation = kwargs.pop("activation", None) + super().__init__(*args, **kwargs) + + self.norm = norm + self.activation = activation + + def forward(self, x): + # torchscript does not support SyncBatchNorm yet + # https://github.com/pytorch/pytorch/issues/40507 + # and we skip these codes in torchscript since: + # 1. currently we only support torchscript in evaluation mode + # 2. features needed by exporting module to torchscript are added in PyTorch 1.6 or + # later version, `Conv2d` in these PyTorch versions has already supported empty inputs. + if not torch.jit.is_scripting(): + if x.numel() == 0 and self.training: + # https://github.com/pytorch/pytorch/issues/12013 + assert not isinstance( + self.norm, torch.nn.SyncBatchNorm + ), "SyncBatchNorm does not support empty inputs!" + + x = F.conv2d( + x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups + ) + if self.norm is not None: + x = self.norm(x) + if self.activation is not None: + x = self.activation(x) + return x + + +ConvTranspose2d = torch.nn.ConvTranspose2d +BatchNorm2d = torch.nn.BatchNorm2d +interpolate = F.interpolate +Linear = torch.nn.Linear + + +def nonzero_tuple(x): + """ + A 'as_tuple=True' version of torch.nonzero to support torchscript. + because of https://github.com/pytorch/pytorch/issues/38718 + """ + if torch.jit.is_scripting(): + if x.dim() == 0: + return x.unsqueeze(0).nonzero().unbind(1) + return x.nonzero().unbind(1) + else: + return x.nonzero(as_tuple=True) diff --git a/detectron2/model_zoo/__init__.py b/detectron2/model_zoo/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..fcae6e18502bab72d76e220b7144b8c262d80e1f --- /dev/null +++ b/detectron2/model_zoo/__init__.py @@ -0,0 +1,10 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +""" +Model Zoo API for Detectron2: a collection of functions to create common model architectures +listed in `MODEL_ZOO.md `_, +and optionally load their pre-trained weights. +""" + +from .model_zoo import get, get_config_file, get_checkpoint_url, get_config + +__all__ = ["get_checkpoint_url", "get", "get_config_file", "get_config"] diff --git a/detectron2/model_zoo/configs b/detectron2/model_zoo/configs new file mode 120000 index 0000000000000000000000000000000000000000..1047c95e90ee6a0b9fc2065d0092b2aa71529965 --- /dev/null +++ b/detectron2/model_zoo/configs @@ -0,0 +1 @@ +/datadrive/regionclip-demo/configs \ No newline at end of file diff --git a/detectron2/model_zoo/model_zoo.py b/detectron2/model_zoo/model_zoo.py new file mode 100644 index 0000000000000000000000000000000000000000..1ca234e0ec1f97be3e6cf761409a876bf2f05caf --- /dev/null +++ b/detectron2/model_zoo/model_zoo.py @@ -0,0 +1,200 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import os +from typing import Optional +import pkg_resources +import torch + +from detectron2.checkpoint import DetectionCheckpointer +from detectron2.config import CfgNode, LazyConfig, get_cfg, instantiate +from detectron2.modeling import build_model + + +class _ModelZooUrls(object): + """ + Mapping from names to officially released Detectron2 pre-trained models. + """ + + S3_PREFIX = "https://dl.fbaipublicfiles.com/detectron2/" + + # format: {config_path.yaml} -> model_id/model_final_{commit}.pkl + CONFIG_PATH_TO_URL_SUFFIX = { + # COCO Detection with Faster R-CNN + "COCO-Detection/faster_rcnn_R_50_C4_1x": "137257644/model_final_721ade.pkl", + "COCO-Detection/faster_rcnn_R_50_DC5_1x": "137847829/model_final_51d356.pkl", + "COCO-Detection/faster_rcnn_R_50_FPN_1x": "137257794/model_final_b275ba.pkl", + "COCO-Detection/faster_rcnn_R_50_C4_3x": "137849393/model_final_f97cb7.pkl", + "COCO-Detection/faster_rcnn_R_50_DC5_3x": "137849425/model_final_68d202.pkl", + "COCO-Detection/faster_rcnn_R_50_FPN_3x": "137849458/model_final_280758.pkl", + "COCO-Detection/faster_rcnn_R_101_C4_3x": "138204752/model_final_298dad.pkl", + "COCO-Detection/faster_rcnn_R_101_DC5_3x": "138204841/model_final_3e0943.pkl", + "COCO-Detection/faster_rcnn_R_101_FPN_3x": "137851257/model_final_f6e8b1.pkl", + "COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x": "139173657/model_final_68b088.pkl", + # COCO Detection with RetinaNet + "COCO-Detection/retinanet_R_50_FPN_1x": "190397773/model_final_bfca0b.pkl", + "COCO-Detection/retinanet_R_50_FPN_3x": "190397829/model_final_5bd44e.pkl", + "COCO-Detection/retinanet_R_101_FPN_3x": "190397697/model_final_971ab9.pkl", + # COCO Detection with RPN and Fast R-CNN + "COCO-Detection/rpn_R_50_C4_1x": "137258005/model_final_450694.pkl", + "COCO-Detection/rpn_R_50_FPN_1x": "137258492/model_final_02ce48.pkl", + "COCO-Detection/fast_rcnn_R_50_FPN_1x": "137635226/model_final_e5f7ce.pkl", + # COCO Instance Segmentation Baselines with Mask R-CNN + "COCO-InstanceSegmentation/mask_rcnn_R_50_C4_1x": "137259246/model_final_9243eb.pkl", + "COCO-InstanceSegmentation/mask_rcnn_R_50_DC5_1x": "137260150/model_final_4f86c3.pkl", + "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x": "137260431/model_final_a54504.pkl", + "COCO-InstanceSegmentation/mask_rcnn_R_50_C4_3x": "137849525/model_final_4ce675.pkl", + "COCO-InstanceSegmentation/mask_rcnn_R_50_DC5_3x": "137849551/model_final_84107b.pkl", + "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x": "137849600/model_final_f10217.pkl", + "COCO-InstanceSegmentation/mask_rcnn_R_101_C4_3x": "138363239/model_final_a2914c.pkl", + "COCO-InstanceSegmentation/mask_rcnn_R_101_DC5_3x": "138363294/model_final_0464b7.pkl", + "COCO-InstanceSegmentation/mask_rcnn_R_101_FPN_3x": "138205316/model_final_a3ec72.pkl", + "COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x": "139653917/model_final_2d9806.pkl", # noqa + # COCO Person Keypoint Detection Baselines with Keypoint R-CNN + "COCO-Keypoints/keypoint_rcnn_R_50_FPN_1x": "137261548/model_final_04e291.pkl", + "COCO-Keypoints/keypoint_rcnn_R_50_FPN_3x": "137849621/model_final_a6e10b.pkl", + "COCO-Keypoints/keypoint_rcnn_R_101_FPN_3x": "138363331/model_final_997cc7.pkl", + "COCO-Keypoints/keypoint_rcnn_X_101_32x8d_FPN_3x": "139686956/model_final_5ad38f.pkl", + # COCO Panoptic Segmentation Baselines with Panoptic FPN + "COCO-PanopticSegmentation/panoptic_fpn_R_50_1x": "139514544/model_final_dbfeb4.pkl", + "COCO-PanopticSegmentation/panoptic_fpn_R_50_3x": "139514569/model_final_c10459.pkl", + "COCO-PanopticSegmentation/panoptic_fpn_R_101_3x": "139514519/model_final_cafdb1.pkl", + # LVIS Instance Segmentation Baselines with Mask R-CNN + "LVISv0.5-InstanceSegmentation/mask_rcnn_R_50_FPN_1x": "144219072/model_final_571f7c.pkl", # noqa + "LVISv0.5-InstanceSegmentation/mask_rcnn_R_101_FPN_1x": "144219035/model_final_824ab5.pkl", # noqa + "LVISv0.5-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_1x": "144219108/model_final_5e3439.pkl", # noqa + # Cityscapes & Pascal VOC Baselines + "Cityscapes/mask_rcnn_R_50_FPN": "142423278/model_final_af9cf5.pkl", + "PascalVOC-Detection/faster_rcnn_R_50_C4": "142202221/model_final_b1acc2.pkl", + # Other Settings + "Misc/mask_rcnn_R_50_FPN_1x_dconv_c3-c5": "138602867/model_final_65c703.pkl", + "Misc/mask_rcnn_R_50_FPN_3x_dconv_c3-c5": "144998336/model_final_821d0b.pkl", + "Misc/cascade_mask_rcnn_R_50_FPN_1x": "138602847/model_final_e9d89b.pkl", + "Misc/cascade_mask_rcnn_R_50_FPN_3x": "144998488/model_final_480dd8.pkl", + "Misc/mask_rcnn_R_50_FPN_3x_syncbn": "169527823/model_final_3b3c51.pkl", + "Misc/mask_rcnn_R_50_FPN_3x_gn": "138602888/model_final_dc5d9e.pkl", + "Misc/scratch_mask_rcnn_R_50_FPN_3x_gn": "138602908/model_final_01ca85.pkl", + "Misc/scratch_mask_rcnn_R_50_FPN_9x_gn": "183808979/model_final_da7b4c.pkl", + "Misc/scratch_mask_rcnn_R_50_FPN_9x_syncbn": "184226666/model_final_5ce33e.pkl", + "Misc/panoptic_fpn_R_101_dconv_cascade_gn_3x": "139797668/model_final_be35db.pkl", + "Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv": "18131413/model_0039999_e76410.pkl", # noqa + # D1 Comparisons + "Detectron1-Comparisons/faster_rcnn_R_50_FPN_noaug_1x": "137781054/model_final_7ab50c.pkl", # noqa + "Detectron1-Comparisons/mask_rcnn_R_50_FPN_noaug_1x": "137781281/model_final_62ca52.pkl", # noqa + "Detectron1-Comparisons/keypoint_rcnn_R_50_FPN_1x": "137781195/model_final_cce136.pkl", + } + + @staticmethod + def query(config_path: str) -> Optional[str]: + """ + Args: + config_path: relative config filename + """ + name = config_path.replace(".yaml", "").replace(".py", "") + if name in _ModelZooUrls.CONFIG_PATH_TO_URL_SUFFIX: + suffix = _ModelZooUrls.CONFIG_PATH_TO_URL_SUFFIX[name] + return _ModelZooUrls.S3_PREFIX + name + "/" + suffix + return None + + +def get_checkpoint_url(config_path): + """ + Returns the URL to the model trained using the given config + + Args: + config_path (str): config file name relative to detectron2's "configs/" + directory, e.g., "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml" + + Returns: + str: a URL to the model + """ + url = _ModelZooUrls.query(config_path) + if url is None: + raise RuntimeError("Pretrained model for {} is not available!".format(config_path)) + return url + + +def get_config_file(config_path): + """ + Returns path to a builtin config file. + + Args: + config_path (str): config file name relative to detectron2's "configs/" + directory, e.g., "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml" + + Returns: + str: the real path to the config file. + """ + cfg_file = pkg_resources.resource_filename( + "detectron2.model_zoo", os.path.join("configs", config_path) + ) + if not os.path.exists(cfg_file): + raise RuntimeError("{} not available in Model Zoo!".format(config_path)) + return cfg_file + + +def get_config(config_path, trained: bool = False): + """ + Returns a config object for a model in model zoo. + + Args: + config_path (str): config file name relative to detectron2's "configs/" + directory, e.g., "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml" + trained (bool): If True, will set ``MODEL.WEIGHTS`` to trained model zoo weights. + If False, the checkpoint specified in the config file's ``MODEL.WEIGHTS`` is used + instead; this will typically (though not always) initialize a subset of weights using + an ImageNet pre-trained model, while randomly initializing the other weights. + + Returns: + CfgNode or omegaconf.DictConfig: a config object + """ + cfg_file = get_config_file(config_path) + if cfg_file.endswith(".yaml"): + cfg = get_cfg() + cfg.merge_from_file(cfg_file) + if trained: + cfg.MODEL.WEIGHTS = get_checkpoint_url(config_path) + return cfg + elif cfg_file.endswith(".py"): + cfg = LazyConfig.load(cfg_file) + if trained: + url = get_checkpoint_url(config_path) + if "train" in cfg and "init_checkpoint" in cfg.train: + cfg.train.init_checkpoint = url + else: + raise NotImplementedError + return cfg + + +def get(config_path, trained: bool = False, device: Optional[str] = None): + """ + Get a model specified by relative path under Detectron2's official ``configs/`` directory. + + Args: + config_path (str): config file name relative to detectron2's "configs/" + directory, e.g., "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml" + trained (bool): see :func:`get_config`. + device (str or None): overwrite the device in config, if given. + + Returns: + nn.Module: a detectron2 model. Will be in training mode. + + Example: + :: + from detectron2 import model_zoo + model = model_zoo.get("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml", trained=True) + """ + cfg = get_config(config_path, trained) + if device is None and not torch.cuda.is_available(): + device = "cpu" + if device is not None and isinstance(cfg, CfgNode): + cfg.MODEL.DEVICE = device + + if isinstance(cfg, CfgNode): + model = build_model(cfg) + DetectionCheckpointer(model).load(cfg.MODEL.WEIGHTS) + else: + model = instantiate(cfg.model) + if device is not None: + model = model.to(device) + if "train" in cfg and "init_checkpoint" in cfg.train: + DetectionCheckpointer(model).load(cfg.train.init_checkpoint) + return model diff --git a/detectron2/modeling/__init__.py b/detectron2/modeling/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..0655f96b4618d716f62290ce65e7ae82335ea61f --- /dev/null +++ b/detectron2/modeling/__init__.py @@ -0,0 +1,58 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from detectron2.layers import ShapeSpec + +from .anchor_generator import build_anchor_generator, ANCHOR_GENERATOR_REGISTRY +from .backbone import ( + BACKBONE_REGISTRY, + FPN, + Backbone, + ResNet, + ResNetBlockBase, + build_backbone, + build_resnet_backbone, + make_stage, +) +from .meta_arch import ( + META_ARCH_REGISTRY, + SEM_SEG_HEADS_REGISTRY, + GeneralizedRCNN, + PanopticFPN, + ProposalNetwork, + RetinaNet, + SemanticSegmentor, + build_model, + build_sem_seg_head, +) +from .postprocessing import detector_postprocess +from .proposal_generator import ( + PROPOSAL_GENERATOR_REGISTRY, + build_proposal_generator, + RPN_HEAD_REGISTRY, + build_rpn_head, +) +from .roi_heads import ( + ROI_BOX_HEAD_REGISTRY, + ROI_HEADS_REGISTRY, + ROI_KEYPOINT_HEAD_REGISTRY, + ROI_MASK_HEAD_REGISTRY, + ROIHeads, + StandardROIHeads, + BaseMaskRCNNHead, + BaseKeypointRCNNHead, + FastRCNNOutputLayers, + build_box_head, + build_keypoint_head, + build_mask_head, + build_roi_heads, +) +from .test_time_augmentation import DatasetMapperTTA, GeneralizedRCNNWithTTA +from .mmdet_wrapper import MMDetBackbone, MMDetDetector + +_EXCLUDE = {"ShapeSpec"} +__all__ = [k for k in globals().keys() if k not in _EXCLUDE and not k.startswith("_")] + + +from detectron2.utils.env import fixup_module_metadata + +fixup_module_metadata(__name__, globals(), __all__) +del fixup_module_metadata diff --git a/detectron2/modeling/__pycache__/__init__.cpython-39.pyc b/detectron2/modeling/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..eb78d891b31e96e440a68533f785bb0a70802397 Binary files /dev/null and b/detectron2/modeling/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/modeling/__pycache__/anchor_generator.cpython-39.pyc b/detectron2/modeling/__pycache__/anchor_generator.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fcd8750c915f5bfdb6d5845a9c12549429803647 Binary files /dev/null and b/detectron2/modeling/__pycache__/anchor_generator.cpython-39.pyc differ diff --git a/detectron2/modeling/__pycache__/box_regression.cpython-39.pyc b/detectron2/modeling/__pycache__/box_regression.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b9c1d6e85120b6665db0e1aab4fda9e0bfcc9957 Binary files /dev/null and b/detectron2/modeling/__pycache__/box_regression.cpython-39.pyc differ diff --git a/detectron2/modeling/__pycache__/matcher.cpython-39.pyc b/detectron2/modeling/__pycache__/matcher.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3d8f7849c019754fd704593232d34e7513b3560a Binary files /dev/null and b/detectron2/modeling/__pycache__/matcher.cpython-39.pyc differ diff --git a/detectron2/modeling/__pycache__/mmdet_wrapper.cpython-39.pyc b/detectron2/modeling/__pycache__/mmdet_wrapper.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..32bd05ab14f3fb3ca8cfd5d27f6a265419ed44eb Binary files /dev/null and b/detectron2/modeling/__pycache__/mmdet_wrapper.cpython-39.pyc differ diff --git a/detectron2/modeling/__pycache__/poolers.cpython-39.pyc b/detectron2/modeling/__pycache__/poolers.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..56acc46317616d75b7be630a14eb2c4464739e04 Binary files /dev/null and b/detectron2/modeling/__pycache__/poolers.cpython-39.pyc differ diff --git a/detectron2/modeling/__pycache__/postprocessing.cpython-39.pyc b/detectron2/modeling/__pycache__/postprocessing.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2acd29c8f067c325b754d9864f3e294bd94baabe Binary files /dev/null and b/detectron2/modeling/__pycache__/postprocessing.cpython-39.pyc differ diff --git a/detectron2/modeling/__pycache__/sampling.cpython-39.pyc b/detectron2/modeling/__pycache__/sampling.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..35f1bf993f48c3cc5150048b1c89561b19b47317 Binary files /dev/null and b/detectron2/modeling/__pycache__/sampling.cpython-39.pyc differ diff --git a/detectron2/modeling/__pycache__/test_time_augmentation.cpython-39.pyc b/detectron2/modeling/__pycache__/test_time_augmentation.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9638b1b1d13043819179e5e4a3ef6ebad9a34c4d Binary files /dev/null and b/detectron2/modeling/__pycache__/test_time_augmentation.cpython-39.pyc differ diff --git a/detectron2/modeling/anchor_generator.py b/detectron2/modeling/anchor_generator.py new file mode 100644 index 0000000000000000000000000000000000000000..ee4b98819445f95982ca89a72cdd3e27b39b367f --- /dev/null +++ b/detectron2/modeling/anchor_generator.py @@ -0,0 +1,382 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import collections +import math +from typing import List +import torch +from torch import nn + +from detectron2.config import configurable +from detectron2.layers import ShapeSpec +from detectron2.structures import Boxes, RotatedBoxes +from detectron2.utils.registry import Registry + +ANCHOR_GENERATOR_REGISTRY = Registry("ANCHOR_GENERATOR") +ANCHOR_GENERATOR_REGISTRY.__doc__ = """ +Registry for modules that creates object detection anchors for feature maps. + +The registered object will be called with `obj(cfg, input_shape)`. +""" + + +class BufferList(nn.Module): + """ + Similar to nn.ParameterList, but for buffers + """ + + def __init__(self, buffers): + super().__init__() + for i, buffer in enumerate(buffers): + # Use non-persistent buffer so the values are not saved in checkpoint + self.register_buffer(str(i), buffer, persistent=False) + + def __len__(self): + return len(self._buffers) + + def __iter__(self): + return iter(self._buffers.values()) + + +def _create_grid_offsets(size: List[int], stride: int, offset: float, device: torch.device): + grid_height, grid_width = size + shifts_x = torch.arange( + offset * stride, grid_width * stride, step=stride, dtype=torch.float32, device=device + ) + shifts_y = torch.arange( + offset * stride, grid_height * stride, step=stride, dtype=torch.float32, device=device + ) + + shift_y, shift_x = torch.meshgrid(shifts_y, shifts_x) + shift_x = shift_x.reshape(-1) + shift_y = shift_y.reshape(-1) + return shift_x, shift_y + + +def _broadcast_params(params, num_features, name): + """ + If one size (or aspect ratio) is specified and there are multiple feature + maps, we "broadcast" anchors of that single size (or aspect ratio) + over all feature maps. + + If params is list[float], or list[list[float]] with len(params) == 1, repeat + it num_features time. + + Returns: + list[list[float]]: param for each feature + """ + assert isinstance( + params, collections.abc.Sequence + ), f"{name} in anchor generator has to be a list! Got {params}." + assert len(params), f"{name} in anchor generator cannot be empty!" + if not isinstance(params[0], collections.abc.Sequence): # params is list[float] + return [params] * num_features + if len(params) == 1: + return list(params) * num_features + assert len(params) == num_features, ( + f"Got {name} of length {len(params)} in anchor generator, " + f"but the number of input features is {num_features}!" + ) + return params + + +@ANCHOR_GENERATOR_REGISTRY.register() +class DefaultAnchorGenerator(nn.Module): + """ + Compute anchors in the standard ways described in + "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks". + """ + + box_dim: torch.jit.Final[int] = 4 + """ + the dimension of each anchor box. + """ + + @configurable + def __init__(self, *, sizes, aspect_ratios, strides, offset=0.5): + """ + This interface is experimental. + + Args: + sizes (list[list[float]] or list[float]): + If ``sizes`` is list[list[float]], ``sizes[i]`` is the list of anchor sizes + (i.e. sqrt of anchor area) to use for the i-th feature map. + If ``sizes`` is list[float], ``sizes`` is used for all feature maps. + Anchor sizes are given in absolute lengths in units of + the input image; they do not dynamically scale if the input image size changes. + aspect_ratios (list[list[float]] or list[float]): list of aspect ratios + (i.e. height / width) to use for anchors. Same "broadcast" rule for `sizes` applies. + strides (list[int]): stride of each input feature. + offset (float): Relative offset between the center of the first anchor and the top-left + corner of the image. Value has to be in [0, 1). + Recommend to use 0.5, which means half stride. + """ + super().__init__() + + self.strides = strides + self.num_features = len(self.strides) + sizes = _broadcast_params(sizes, self.num_features, "sizes") + aspect_ratios = _broadcast_params(aspect_ratios, self.num_features, "aspect_ratios") + self.cell_anchors = self._calculate_anchors(sizes, aspect_ratios) + + self.offset = offset + assert 0.0 <= self.offset < 1.0, self.offset + + @classmethod + def from_config(cls, cfg, input_shape: List[ShapeSpec]): + return { + "sizes": cfg.MODEL.ANCHOR_GENERATOR.SIZES, + "aspect_ratios": cfg.MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS, + "strides": [x.stride for x in input_shape], + "offset": cfg.MODEL.ANCHOR_GENERATOR.OFFSET, + } + + def _calculate_anchors(self, sizes, aspect_ratios): + cell_anchors = [ + self.generate_cell_anchors(s, a).float() for s, a in zip(sizes, aspect_ratios) + ] + return BufferList(cell_anchors) + + @property + @torch.jit.unused + def num_cell_anchors(self): + """ + Alias of `num_anchors`. + """ + return self.num_anchors + + @property + @torch.jit.unused + def num_anchors(self): + """ + Returns: + list[int]: Each int is the number of anchors at every pixel + location, on that feature map. + For example, if at every pixel we use anchors of 3 aspect + ratios and 5 sizes, the number of anchors is 15. + (See also ANCHOR_GENERATOR.SIZES and ANCHOR_GENERATOR.ASPECT_RATIOS in config) + + In standard RPN models, `num_anchors` on every feature map is the same. + """ + return [len(cell_anchors) for cell_anchors in self.cell_anchors] + + def _grid_anchors(self, grid_sizes: List[List[int]]): + """ + Returns: + list[Tensor]: #featuremap tensors, each is (#locations x #cell_anchors) x 4 + """ + anchors = [] + # buffers() not supported by torchscript. use named_buffers() instead + buffers: List[torch.Tensor] = [x[1] for x in self.cell_anchors.named_buffers()] + for size, stride, base_anchors in zip(grid_sizes, self.strides, buffers): + shift_x, shift_y = _create_grid_offsets(size, stride, self.offset, base_anchors.device) + shifts = torch.stack((shift_x, shift_y, shift_x, shift_y), dim=1) + + anchors.append((shifts.view(-1, 1, 4) + base_anchors.view(1, -1, 4)).reshape(-1, 4)) + + return anchors + + def generate_cell_anchors(self, sizes=(32, 64, 128, 256, 512), aspect_ratios=(0.5, 1, 2)): + """ + Generate a tensor storing canonical anchor boxes, which are all anchor + boxes of different sizes and aspect_ratios centered at (0, 0). + We can later build the set of anchors for a full feature map by + shifting and tiling these tensors (see `meth:_grid_anchors`). + + Args: + sizes (tuple[float]): + aspect_ratios (tuple[float]]): + + Returns: + Tensor of shape (len(sizes) * len(aspect_ratios), 4) storing anchor boxes + in XYXY format. + """ + + # This is different from the anchor generator defined in the original Faster R-CNN + # code or Detectron. They yield the same AP, however the old version defines cell + # anchors in a less natural way with a shift relative to the feature grid and + # quantization that results in slightly different sizes for different aspect ratios. + # See also https://github.com/facebookresearch/Detectron/issues/227 + + anchors = [] + for size in sizes: + area = size ** 2.0 + for aspect_ratio in aspect_ratios: + # s * s = w * h + # a = h / w + # ... some algebra ... + # w = sqrt(s * s / a) + # h = a * w + w = math.sqrt(area / aspect_ratio) + h = aspect_ratio * w + x0, y0, x1, y1 = -w / 2.0, -h / 2.0, w / 2.0, h / 2.0 + anchors.append([x0, y0, x1, y1]) + return torch.tensor(anchors) + + def forward(self, features: List[torch.Tensor]): + """ + Args: + features (list[Tensor]): list of backbone feature maps on which to generate anchors. + + Returns: + list[Boxes]: a list of Boxes containing all the anchors for each feature map + (i.e. the cell anchors repeated over all locations in the feature map). + The number of anchors of each feature map is Hi x Wi x num_cell_anchors, + where Hi, Wi are resolution of the feature map divided by anchor stride. + """ + grid_sizes = [feature_map.shape[-2:] for feature_map in features] + anchors_over_all_feature_maps = self._grid_anchors(grid_sizes) + return [Boxes(x) for x in anchors_over_all_feature_maps] + + +@ANCHOR_GENERATOR_REGISTRY.register() +class RotatedAnchorGenerator(nn.Module): + """ + Compute rotated anchors used by Rotated RPN (RRPN), described in + "Arbitrary-Oriented Scene Text Detection via Rotation Proposals". + """ + + box_dim: int = 5 + """ + the dimension of each anchor box. + """ + + @configurable + def __init__(self, *, sizes, aspect_ratios, strides, angles, offset=0.5): + """ + This interface is experimental. + + Args: + sizes (list[list[float]] or list[float]): + If sizes is list[list[float]], sizes[i] is the list of anchor sizes + (i.e. sqrt of anchor area) to use for the i-th feature map. + If sizes is list[float], the sizes are used for all feature maps. + Anchor sizes are given in absolute lengths in units of + the input image; they do not dynamically scale if the input image size changes. + aspect_ratios (list[list[float]] or list[float]): list of aspect ratios + (i.e. height / width) to use for anchors. Same "broadcast" rule for `sizes` applies. + strides (list[int]): stride of each input feature. + angles (list[list[float]] or list[float]): list of angles (in degrees CCW) + to use for anchors. Same "broadcast" rule for `sizes` applies. + offset (float): Relative offset between the center of the first anchor and the top-left + corner of the image. Value has to be in [0, 1). + Recommend to use 0.5, which means half stride. + """ + super().__init__() + + self.strides = strides + self.num_features = len(self.strides) + sizes = _broadcast_params(sizes, self.num_features, "sizes") + aspect_ratios = _broadcast_params(aspect_ratios, self.num_features, "aspect_ratios") + angles = _broadcast_params(angles, self.num_features, "angles") + self.cell_anchors = self._calculate_anchors(sizes, aspect_ratios, angles) + + self.offset = offset + assert 0.0 <= self.offset < 1.0, self.offset + + @classmethod + def from_config(cls, cfg, input_shape: List[ShapeSpec]): + return { + "sizes": cfg.MODEL.ANCHOR_GENERATOR.SIZES, + "aspect_ratios": cfg.MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS, + "strides": [x.stride for x in input_shape], + "offset": cfg.MODEL.ANCHOR_GENERATOR.OFFSET, + "angles": cfg.MODEL.ANCHOR_GENERATOR.ANGLES, + } + + def _calculate_anchors(self, sizes, aspect_ratios, angles): + cell_anchors = [ + self.generate_cell_anchors(size, aspect_ratio, angle).float() + for size, aspect_ratio, angle in zip(sizes, aspect_ratios, angles) + ] + return BufferList(cell_anchors) + + @property + def num_cell_anchors(self): + """ + Alias of `num_anchors`. + """ + return self.num_anchors + + @property + def num_anchors(self): + """ + Returns: + list[int]: Each int is the number of anchors at every pixel + location, on that feature map. + For example, if at every pixel we use anchors of 3 aspect + ratios, 2 sizes and 5 angles, the number of anchors is 30. + (See also ANCHOR_GENERATOR.SIZES, ANCHOR_GENERATOR.ASPECT_RATIOS + and ANCHOR_GENERATOR.ANGLES in config) + + In standard RRPN models, `num_anchors` on every feature map is the same. + """ + return [len(cell_anchors) for cell_anchors in self.cell_anchors] + + def _grid_anchors(self, grid_sizes): + anchors = [] + for size, stride, base_anchors in zip(grid_sizes, self.strides, self.cell_anchors): + shift_x, shift_y = _create_grid_offsets(size, stride, self.offset, base_anchors.device) + zeros = torch.zeros_like(shift_x) + shifts = torch.stack((shift_x, shift_y, zeros, zeros, zeros), dim=1) + + anchors.append((shifts.view(-1, 1, 5) + base_anchors.view(1, -1, 5)).reshape(-1, 5)) + + return anchors + + def generate_cell_anchors( + self, + sizes=(32, 64, 128, 256, 512), + aspect_ratios=(0.5, 1, 2), + angles=(-90, -60, -30, 0, 30, 60, 90), + ): + """ + Generate a tensor storing canonical anchor boxes, which are all anchor + boxes of different sizes, aspect_ratios, angles centered at (0, 0). + We can later build the set of anchors for a full feature map by + shifting and tiling these tensors (see `meth:_grid_anchors`). + + Args: + sizes (tuple[float]): + aspect_ratios (tuple[float]]): + angles (tuple[float]]): + + Returns: + Tensor of shape (len(sizes) * len(aspect_ratios) * len(angles), 5) + storing anchor boxes in (x_ctr, y_ctr, w, h, angle) format. + """ + anchors = [] + for size in sizes: + area = size ** 2.0 + for aspect_ratio in aspect_ratios: + # s * s = w * h + # a = h / w + # ... some algebra ... + # w = sqrt(s * s / a) + # h = a * w + w = math.sqrt(area / aspect_ratio) + h = aspect_ratio * w + anchors.extend([0, 0, w, h, a] for a in angles) + + return torch.tensor(anchors) + + def forward(self, features): + """ + Args: + features (list[Tensor]): list of backbone feature maps on which to generate anchors. + + Returns: + list[RotatedBoxes]: a list of Boxes containing all the anchors for each feature map + (i.e. the cell anchors repeated over all locations in the feature map). + The number of anchors of each feature map is Hi x Wi x num_cell_anchors, + where Hi, Wi are resolution of the feature map divided by anchor stride. + """ + grid_sizes = [feature_map.shape[-2:] for feature_map in features] + anchors_over_all_feature_maps = self._grid_anchors(grid_sizes) + return [RotatedBoxes(x) for x in anchors_over_all_feature_maps] + + +def build_anchor_generator(cfg, input_shape): + """ + Built an anchor generator from `cfg.MODEL.ANCHOR_GENERATOR.NAME`. + """ + anchor_generator = cfg.MODEL.ANCHOR_GENERATOR.NAME + return ANCHOR_GENERATOR_REGISTRY.get(anchor_generator)(cfg, input_shape) diff --git a/detectron2/modeling/backbone/__init__.py b/detectron2/modeling/backbone/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b58e05c517a3adacd142cf1d68ce3a65f2f66447 --- /dev/null +++ b/detectron2/modeling/backbone/__init__.py @@ -0,0 +1,19 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .build import build_backbone, build_text_backbone, BACKBONE_REGISTRY # noqa F401 isort:skip + +from .backbone import Backbone +from .fpn import FPN, LastLevelMaxPool +from .regnet import RegNet +from .resnet import ( + BasicStem, + ResNet, + ResNetBlockBase, + build_resnet_backbone, + make_stage, + BottleneckBlock, +) +from .clip_backbone import ModifiedResNet, build_resnet_clip, build_clip_resnet_backbone, build_clip_language_encoder +from .clip_swin import build_clip_swin, build_clip_swin_backbone + +__all__ = [k for k in globals().keys() if not k.startswith("_")] +# TODO can expose more resnet blocks after careful consideration diff --git a/detectron2/modeling/backbone/__pycache__/__init__.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..9b604ef38c47654abe36da85a3675f21095c88c5 Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/__pycache__/backbone.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/backbone.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c874bc0b7bc1e29844f17c11456dcb19eb359fcc Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/backbone.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/__pycache__/build.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/build.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6e83e59541b8156f193b9e01512ff95131364672 Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/build.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/__pycache__/clip_backbone.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/clip_backbone.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ffbc04be9f3c86df6b198b8aa9c51eba55e77f8f Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/clip_backbone.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/__pycache__/clip_swin.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/clip_swin.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7a4b66408b2ac043a609e48fbf535fb88873bbbe Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/clip_swin.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/__pycache__/det_swin.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/det_swin.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..611587344ee24c85453db8e00b433ea03bb7691f Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/det_swin.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/__pycache__/fpn.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/fpn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..358d42d1f1652d88379b2e0a3b6521fde35c422f Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/fpn.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/__pycache__/regnet.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/regnet.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5ebd281ed288961273420d2ddf62a87681924dcb Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/regnet.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/__pycache__/resnet.cpython-39.pyc b/detectron2/modeling/backbone/__pycache__/resnet.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8b8b81f884d60a1c64c99616ed064cfc527eb1c0 Binary files /dev/null and b/detectron2/modeling/backbone/__pycache__/resnet.cpython-39.pyc differ diff --git a/detectron2/modeling/backbone/backbone.py b/detectron2/modeling/backbone/backbone.py new file mode 100644 index 0000000000000000000000000000000000000000..369fb884930c5dd82f94024c45303dafaab14d66 --- /dev/null +++ b/detectron2/modeling/backbone/backbone.py @@ -0,0 +1,53 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from abc import ABCMeta, abstractmethod +import torch.nn as nn + +from detectron2.layers import ShapeSpec + +__all__ = ["Backbone"] + + +class Backbone(nn.Module, metaclass=ABCMeta): + """ + Abstract base class for network backbones. + """ + + def __init__(self): + """ + The `__init__` method of any subclass can specify its own set of arguments. + """ + super().__init__() + + @abstractmethod + def forward(self): + """ + Subclasses must override this method, but adhere to the same return type. + + Returns: + dict[str->Tensor]: mapping from feature name (e.g., "res2") to tensor + """ + pass + + @property + def size_divisibility(self) -> int: + """ + Some backbones require the input height and width to be divisible by a + specific integer. This is typically true for encoder / decoder type networks + with lateral connection (e.g., FPN) for which feature maps need to match + dimension in the "bottom up" and "top down" paths. Set to 0 if no specific + input size divisibility is required. + """ + return 0 + + def output_shape(self): + """ + Returns: + dict[str->ShapeSpec] + """ + # this is a backward-compatible default + return { + name: ShapeSpec( + channels=self._out_feature_channels[name], stride=self._out_feature_strides[name] + ) + for name in self._out_features + } diff --git a/detectron2/modeling/backbone/build.py b/detectron2/modeling/backbone/build.py new file mode 100644 index 0000000000000000000000000000000000000000..3f407420c13b3c28878aaea98e6958093b9d2caf --- /dev/null +++ b/detectron2/modeling/backbone/build.py @@ -0,0 +1,48 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from detectron2.layers import ShapeSpec +from detectron2.utils.registry import Registry + +from .backbone import Backbone + +BACKBONE_REGISTRY = Registry("BACKBONE") +BACKBONE_REGISTRY.__doc__ = """ +Registry for backbones, which extract feature maps from images + +The registered object must be a callable that accepts two arguments: + +1. A :class:`detectron2.config.CfgNode` +2. A :class:`detectron2.layers.ShapeSpec`, which contains the input shape specification. + +Registered object must return instance of :class:`Backbone`. +""" + + +def build_backbone(cfg, input_shape=None): + """ + Build a backbone from `cfg.MODEL.BACKBONE.NAME`. + + Returns: + an instance of :class:`Backbone` + """ + if input_shape is None: + input_shape = ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN)) + + backbone_name = cfg.MODEL.BACKBONE.NAME + backbone = BACKBONE_REGISTRY.get(backbone_name)(cfg, input_shape) + assert isinstance(backbone, Backbone) + return backbone + +def build_text_backbone(cfg, input_shape=None): + """ + Build a backbone from `cfg.MODEL.BACKBONE.NAME`. + + Returns: + an instance of :class:`Backbone` + """ + if input_shape is None: + input_shape = ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN)) + + backbone_name = cfg.MODEL.BACKBONE.NAME + backbone = BACKBONE_REGISTRY.get(backbone_name)(cfg, input_shape) + assert isinstance(backbone, Backbone) + return backbone \ No newline at end of file diff --git a/detectron2/modeling/backbone/clip_backbone.py b/detectron2/modeling/backbone/clip_backbone.py new file mode 100644 index 0000000000000000000000000000000000000000..093886abf00ac853ae05c88a584eb1b9b4026d68 --- /dev/null +++ b/detectron2/modeling/backbone/clip_backbone.py @@ -0,0 +1,882 @@ +from collections import OrderedDict +from typing import Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from .backbone import Backbone +from .build import BACKBONE_REGISTRY +from detectron2.layers.blocks import FrozenBatchNorm2d +from detectron2.layers import ShapeSpec + +class Bottleneck(nn.Module): + expansion = 4 + + def __init__(self, inplanes, planes, stride=1, norm_type='FronzenBN'): + super().__init__() + + # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1 + self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False) + if norm_type == 'FronzenBN': + self.bn1 = FrozenBatchNorm2d(planes) # nn.BatchNorm2d(planes) + elif norm_type == 'SyncBN': + self.bn1 = nn.SyncBatchNorm(planes) + + self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False) + if norm_type == 'FronzenBN': + self.bn2 = FrozenBatchNorm2d(planes) # nn.BatchNorm2d(planes) + elif norm_type == 'SyncBN': + self.bn2 = nn.SyncBatchNorm(planes) + + self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity() + + self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False) + if norm_type == 'FronzenBN': + self.bn3 = FrozenBatchNorm2d(planes * self.expansion) # nn.BatchNorm2d(planes * self.expansion) + elif norm_type == 'SyncBN': + self.bn3 = nn.SyncBatchNorm(planes * self.expansion) + + self.relu = nn.ReLU(inplace=True) + self.downsample = None + self.stride = stride + + if stride > 1 or inplanes != planes * Bottleneck.expansion: + # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1 + if norm_type == 'FronzenBN': + this_norm = FrozenBatchNorm2d(planes * self.expansion) #("1", nn.BatchNorm2d(planes * self.expansion)) + elif norm_type == 'SyncBN': + this_norm = nn.SyncBatchNorm(planes * self.expansion) + self.downsample = nn.Sequential(OrderedDict([ + ("-1", nn.AvgPool2d(stride)), + ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)), + ("1", this_norm), #("1", nn.BatchNorm2d(planes * self.expansion)) + ])) + + def forward(self, x: torch.Tensor): + identity = x + + out = self.relu(self.bn1(self.conv1(x))) + out = self.relu(self.bn2(self.conv2(out))) + out = self.avgpool(out) + out = self.bn3(self.conv3(out)) + + if self.downsample is not None: + identity = self.downsample(x) + + out += identity + out = self.relu(out) + return out + + +class AttentionPool2d(nn.Module): + def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None): + super().__init__() + self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5) + self.k_proj = nn.Linear(embed_dim, embed_dim) + self.q_proj = nn.Linear(embed_dim, embed_dim) + self.v_proj = nn.Linear(embed_dim, embed_dim) + self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim) + self.num_heads = num_heads + + def forward(self, x): + x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(2, 0, 1) # NCHW -> (HW)NC + x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (HW+1)NC + x = x + self.positional_embedding[:, None, :].to(x.dtype) # (HW+1)NC + x, _ = F.multi_head_attention_forward( + query=x, key=x, value=x, + embed_dim_to_check=x.shape[-1], + num_heads=self.num_heads, + q_proj_weight=self.q_proj.weight, + k_proj_weight=self.k_proj.weight, + v_proj_weight=self.v_proj.weight, + in_proj_weight=None, + in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]), + bias_k=None, + bias_v=None, + add_zero_attn=False, + dropout_p=0, + out_proj_weight=self.c_proj.weight, + out_proj_bias=self.c_proj.bias, + use_separate_proj_weight=True, + training=self.training, + need_weights=False + ) + + return x[0] + + +class ModifiedResNet(Backbone): + """ + Extended from CLIP implementation. It contains following changes: + 1. change all nn.BatchNorm2d() to FrozenBatchNorm2d(), due to small batch size of detection training + 2. add self._out_feature_strides according to standard ResNet + 2. modify forward() to be compatible with Detectron2 + 3. add freeze() and output_shape() to be compatible with Detectron2 + 4. add build_clip_resnet_backbone() to build this ModifiedResNet + + A ResNet class that is similar to torchvision's but contains the following changes: + - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool. + - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1 + - The final pooling layer is a QKV attention instead of an average pool + """ + + def __init__(self, layers, output_dim, heads, input_resolution=224, width=64, + out_features=None, freeze_at=0, depth=None, pool_vec=True, create_att_pool=False, norm_type='FronzenBN'): + super().__init__() + self.output_dim = output_dim + self.input_resolution = input_resolution + self.norm_type = norm_type + + # the 3-layer stem + self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False) + if norm_type == 'FronzenBN': + self.bn1 = FrozenBatchNorm2d(width // 2) # nn.BatchNorm2d(width // 2) + elif norm_type == 'SyncBN': + self.bn1 = nn.SyncBatchNorm(width // 2) + self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False) + if norm_type == 'FronzenBN': + self.bn2 = FrozenBatchNorm2d(width // 2) # nn.BatchNorm2d(width // 2) + elif norm_type == 'SyncBN': + self.bn2 = nn.SyncBatchNorm(width // 2) + self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False) + if norm_type == 'FronzenBN': + self.bn3 = FrozenBatchNorm2d(width) # nn.BatchNorm2d(width) + elif norm_type == 'SyncBN': + self.bn3 = nn.SyncBatchNorm(width) + self.avgpool = nn.AvgPool2d(2) + self.relu = nn.ReLU(inplace=True) + + # residual layers + self._inplanes = width # this is a *mutable* variable used during construction + self.layer1 = self._make_layer(width, layers[0]) + self.layer2 = self._make_layer(width * 2, layers[1], stride=2) + self.layer3 = self._make_layer(width * 4, layers[2], stride=2) + if 'res5' in out_features: # FPN + self.layer4 = self._make_layer(width * 8, layers[3], stride=2) + else: # C4, layer4 created here won't be used in backbone, but used in roi_head + self.layer4 = self._make_layer(width * 8, layers[3], stride=2) # None + + self.pool_vec = pool_vec + if self.pool_vec or create_att_pool: # pool a vector representation for an image + embed_dim = width * 32 # the ResNet feature dimension + self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim) + # if create_att_pool: # freeze attnpool layer + # for p in self.attnpool.parameters(): p.requires_grad = False + + self._out_features = out_features if out_features else [] + if depth in [50,101]: # resnet50 or resnet 101 + # FPN: ["res2", "res3", "res4", "res5"]; C4: ["res4"] + self._out_feature_channels = {'stem': 64, 'res2': 256, 'res3': 512, 'res4': 1024, 'res5': 2048} if 'res5' in self._out_features \ + else {'stem': 64, 'res2': 256, 'res3': 512, 'res4': 1024} + self._out_feature_strides = {'stem': 4, 'res2': 4, 'res3': 8, 'res4': 16, 'res5': 32} if 'res5' in self._out_features \ + else {'stem': 4, 'res2': 4, 'res3': 8, 'res4': 16} # anti-aliasing strided conv??? + elif depth in [200]: # resnet50x4 + # FPN: ["res2", "res3", "res4", "res5"]; C4: ["res4"] + self._out_feature_channels = {'stem': 80, 'res2': 320, 'res3': 640, 'res4': 1280, 'res5': 2560} if 'res5' in self._out_features \ + else {'stem': 80, 'res2': 320, 'res3': 640, 'res4': 1280} + self._out_feature_strides = {'stem': 4, 'res2': 4, 'res3': 8, 'res4': 16, 'res5': 32} if 'res5' in self._out_features \ + else {'stem': 4, 'res2': 4, 'res3': 8, 'res4': 16} # anti-aliasing strided conv??? + self.freeze(freeze_at) + + + def _make_layer(self, planes, blocks, stride=1): + layers = [Bottleneck(self._inplanes, planes, stride, norm_type=self.norm_type)] + + self._inplanes = planes * Bottleneck.expansion + for _ in range(1, blocks): + layers.append(Bottleneck(self._inplanes, planes, norm_type=self.norm_type)) + + return nn.Sequential(*layers) + + def forward(self, x): + def stem(x): + for conv, bn in [(self.conv1, self.bn1), (self.conv2, self.bn2), (self.conv3, self.bn3)]: + x = self.relu(bn(conv(x))) + x = self.avgpool(x) + return x + + assert x.dim() == 4, f"ResNet takes an input of shape (N, C, H, W). Got {x.shape} instead!" + outputs = {} + x = x.type(self.conv1.weight.dtype) # det2 resnet50: [3, 800, 1216]; CLIP resnet50: [3, 224, 224] + x = stem(x) # det2 resnet50: [64, 200, 304]; CLIP resnet50: [64, 56, 56] + if "stem" in self._out_features: + outputs["stem"] = x + x = self.layer1(x) # det2 resnet50: [256, 200, 304]; CLIP resnet50: [256, 56, 56] + outputs['res2'] = x if "res2" in self._out_features else None + x = self.layer2(x) # det2 resnet50: [512, 100, 152]; CLIP resnet50: [512, 28, 28] + outputs['res3'] = x if "res3" in self._out_features else None + x = self.layer3(x) # det2 resnet50: [1024, 50, 76]; CLIP resnet50: [1024, 14, 14] + outputs['res4'] = x if "res4" in self._out_features else None + x = self.layer4(x) if "res5" in self._out_features else x # det2 resnet50: [2048, 25, 38]; CLIP resnet50: [2048, 7, 7] + outputs['res5'] = x if "res5" in self._out_features else None + + if self.pool_vec: # pool a vector representation for an image, for global image classification + x = self.attnpool(x) # CLIP resnet50: [1024] + return x + else: # for FPN + return outputs + + def freeze(self, freeze_at=0): + """ + Freeze the first several stages of the ResNet. Commonly used in + fine-tuning. + + Layers that produce the same feature map spatial size are defined as one + "stage" by :paper:`FPN`. + + Args: + freeze_at (int): number of stages to freeze. + `1` means freezing the stem. `2` means freezing the stem and + one residual stage, etc. + + Returns: + nn.Module: this ResNet itself + """ + def cnnblockbase_freeze(nn_module): + """ + Make this block not trainable. + This method sets all parameters to `requires_grad=False`, + and convert all BatchNorm layers to FrozenBatchNorm + + Returns: + the block itself + """ + for p in nn_module.parameters(): + p.requires_grad = False + FrozenBatchNorm2d.convert_frozen_batchnorm(nn_module) + + if freeze_at >= 1: # stem + cnnblockbase_freeze(self.conv1) + cnnblockbase_freeze(self.bn1) + cnnblockbase_freeze(self.conv2) + cnnblockbase_freeze(self.bn2) + cnnblockbase_freeze(self.conv3) + cnnblockbase_freeze(self.bn3) + # each stage is a torch.nn.modules.container.Sequential + for idx, stage in enumerate([self.layer1, self.layer2, self.layer3, self.layer4], start=2): + if freeze_at >= idx: + for block in stage.children(): # each block is a Bottleneck + cnnblockbase_freeze(block) + return self + + def output_shape(self): + return { + name: ShapeSpec( + channels=self._out_feature_channels[name], stride=self._out_feature_strides[name] + ) + for name in self._out_features + } + + +class LayerNorm(nn.LayerNorm): + """Subclass torch's LayerNorm to handle fp16.""" + + def forward(self, x: torch.Tensor): + orig_type = x.dtype + ret = super().forward(x.type(torch.float32)) + return ret.type(orig_type) + + +class QuickGELU(nn.Module): + def forward(self, x: torch.Tensor): + return x * torch.sigmoid(1.702 * x) + + +class ResidualAttentionBlock(nn.Module): + def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None): + super().__init__() + + self.attn = nn.MultiheadAttention(d_model, n_head) + self.ln_1 = LayerNorm(d_model) + self.mlp = nn.Sequential(OrderedDict([ + ("c_fc", nn.Linear(d_model, d_model * 4)), + ("gelu", QuickGELU()), + ("c_proj", nn.Linear(d_model * 4, d_model)) + ])) + self.ln_2 = LayerNorm(d_model) + self.attn_mask = attn_mask + + def attention(self, x: torch.Tensor): + self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None + return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0] + + def forward(self, x: torch.Tensor): + x = x + self.attention(self.ln_1(x)) + x = x + self.mlp(self.ln_2(x)) + return x + + +class Transformer(nn.Module): + def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None): + super().__init__() + self.width = width + self.layers = layers + self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)]) + + def forward(self, x: torch.Tensor): + return self.resblocks(x) + + +class VisualTransformer(nn.Module): + def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int): + super().__init__() + self.input_resolution = input_resolution + self.output_dim = output_dim + self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False) + + scale = width ** -0.5 + self.class_embedding = nn.Parameter(scale * torch.randn(width)) + self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width)) + self.ln_pre = LayerNorm(width) + + self.transformer = Transformer(width, layers, heads) + + self.ln_post = LayerNorm(width) + self.proj = nn.Parameter(scale * torch.randn(width, output_dim)) + + def forward(self, x: torch.Tensor): + x = self.conv1(x) # shape = [*, width, grid, grid] + x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2] + x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width] + x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1) # shape = [*, grid ** 2 + 1, width] + x = x + self.positional_embedding.to(x.dtype) + x = self.ln_pre(x) + + x = x.permute(1, 0, 2) # NLD -> LND + x = self.transformer(x) + x = x.permute(1, 0, 2) # LND -> NLD + + x = self.ln_post(x[:, 0, :]) + + if self.proj is not None: + x = x @ self.proj + + return x + + +class CLIP(Backbone): + def __init__(self, + embed_dim: int, + # vision + image_resolution: int, + vision_layers: Union[Tuple[int, int, int, int], int], + vision_width: int, + vision_patch_size: int, + # text + context_length: int, + vocab_size: int, + transformer_width: int, + transformer_heads: int, + transformer_layers: int, + out_features, + freeze_at, + ): + super().__init__() + + self.context_length = context_length + + if isinstance(vision_layers, (tuple, list)): + vision_heads = vision_width * 32 // 64 + self.visual = ModifiedResNet( + layers=vision_layers, + output_dim=embed_dim, + heads=vision_heads, + input_resolution=image_resolution, + width=vision_width, + out_features=out_features, + freeze_at=freeze_at, + ) + else: + vision_heads = vision_width // 64 + self.visual = VisualTransformer( + input_resolution=image_resolution, + patch_size=vision_patch_size, + width=vision_width, + layers=vision_layers, + heads=vision_heads, + output_dim=embed_dim + ) + + self.transformer = Transformer( + width=transformer_width, + layers=transformer_layers, + heads=transformer_heads, + attn_mask=self.build_attention_mask() + ) + + self.vocab_size = vocab_size + self.token_embedding = nn.Embedding(vocab_size, transformer_width) + self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width)) + self.ln_final = LayerNorm(transformer_width) + + self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim)) + self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07)) + + self.initialize_parameters() + + def initialize_parameters(self): + nn.init.normal_(self.token_embedding.weight, std=0.02) + nn.init.normal_(self.positional_embedding, std=0.01) + + if isinstance(self.visual, ModifiedResNet): + if self.visual.attnpool is not None: + std = self.visual.attnpool.c_proj.in_features ** -0.5 + nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std) + nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std) + nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std) + nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std) + + for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]: + for name, param in resnet_block.named_parameters(): + if name.endswith("bn3.weight"): + nn.init.zeros_(param) + + proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5) + attn_std = self.transformer.width ** -0.5 + fc_std = (2 * self.transformer.width) ** -0.5 + for block in self.transformer.resblocks: + nn.init.normal_(block.attn.in_proj_weight, std=attn_std) + nn.init.normal_(block.attn.out_proj.weight, std=proj_std) + nn.init.normal_(block.mlp.c_fc.weight, std=fc_std) + nn.init.normal_(block.mlp.c_proj.weight, std=proj_std) + + if self.text_projection is not None: + nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5) + + def build_attention_mask(self): + # lazily create causal attention mask, with full attention between the vision tokens + # pytorch uses additive attention mask; fill with -inf + mask = torch.empty(self.context_length, self.context_length) + mask.fill_(float("-inf")) + mask.triu_(1) # zero out the lower diagonal + return mask + + @property + def dtype(self): + return self.visual.conv1.weight.dtype + + def encode_image(self, image): + return self.visual(image.type(self.dtype)) + + def encode_text(self, text, norm=True): + x = self.token_embedding(text).type(self.dtype) # [batch_size, n_ctx, d_model] + + x = x + self.positional_embedding.type(self.dtype) + x = x.permute(1, 0, 2) # NLD -> LND + x = self.transformer(x) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.ln_final(x).type(self.dtype) + + # x.shape = [batch_size, n_ctx, transformer.width] + # take features from the eot embedding (eot_token is the highest number in each sequence) + x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection + if norm: + x = x / x.norm(dim=-1, keepdim=True) + return x + + def forward(self, image, text): + image_features = self.encode_image(image) + text_features = self.encode_text(text) + + # normalized features + image_features = image_features / image_features.norm(dim=-1, keepdim=True) + text_features = text_features / text_features.norm(dim=-1, keepdim=True) + + # cosine similarity as logits + logit_scale = self.logit_scale.exp() + logits_per_image = logit_scale * image_features @ text_features.t() + logits_per_text = logit_scale * text_features @ image_features.t() + + # shape = [global_batch_size, global_batch_size] + return logits_per_image, logits_per_text + + +def convert_weights(model: nn.Module): + """Convert applicable model parameters to fp16""" + + def _convert_weights_to_fp16(l): + if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)): + l.weight.data = l.weight.data.half() + if l.bias is not None: + l.bias.data = l.bias.data.half() + + if isinstance(l, nn.MultiheadAttention): + for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]: + tensor = getattr(l, attr) + if tensor is not None: + tensor.data = tensor.data.half() + + for name in ["text_projection", "proj"]: + if hasattr(l, name): + attr = getattr(l, name) + if attr is not None: + attr.data = attr.data.half() + + model.apply(_convert_weights_to_fp16) + + +def build_model(state_dict: dict): + vit = "visual.proj" in state_dict + + if vit: + vision_width = state_dict["visual.conv1.weight"].shape[0] + vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")]) + vision_patch_size = state_dict["visual.conv1.weight"].shape[-1] + grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5) + image_resolution = vision_patch_size * grid_size + else: + counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]] + vision_layers = tuple(counts) + vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0] + output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5) + vision_patch_size = None + assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0] + image_resolution = output_width * 32 + + embed_dim = state_dict["text_projection"].shape[1] + context_length = state_dict["positional_embedding"].shape[0] + vocab_size = state_dict["token_embedding.weight"].shape[0] + transformer_width = state_dict["ln_final.weight"].shape[0] + transformer_heads = transformer_width // 64 + transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith(f"transformer.resblocks"))) + + model = CLIP( + embed_dim, + image_resolution, vision_layers, vision_width, vision_patch_size, + context_length, vocab_size, transformer_width, transformer_heads, transformer_layers + ) + + for key in ["input_resolution", "context_length", "vocab_size"]: + if key in state_dict: + del state_dict[key] + + convert_weights(model) + model.load_state_dict(state_dict) + return model.eval() + + +@BACKBONE_REGISTRY.register() +def build_vit_clip(cfg, input_shape): + """ + Create the whole CLIP instance from config. + + Returns: + CLIP: a :class:`CLIP` instance. + """ + # port standard ResNet config to CLIP ModifiedResNet + freeze_at = cfg.MODEL.BACKBONE.FREEZE_AT + out_features = ['res5'] # includes the whole ResNet # cfg.MODEL.RESNETS.OUT_FEATURES + depth = cfg.MODEL.RESNETS.DEPTH + + # num_blocks_per_stage = { + # 18: [2, 2, 2, 2], + # 34: [3, 4, 6, 3], + # 50: [3, 4, 6, 3], + # 101: [3, 4, 23, 3], + # 152: [3, 8, 36, 3], + # }[depth] + vision_layers = 12 # num_blocks_per_stage + vision_width = 768 # cfg.MODEL.RESNETS.STEM_OUT_CHANNELS + + # default configs of CLIP + embed_dim = 512 # 1024 + image_resolution = 224 + vision_patch_size = 32 # None + context_length = 77 + vocab_size = 49408 + transformer_width = 512 + transformer_heads = 8 + transformer_layers = 12 + + model = CLIP( + embed_dim, + image_resolution, vision_layers, vision_width, vision_patch_size, + context_length, vocab_size, transformer_width, transformer_heads, transformer_layers, + out_features, freeze_at + ) + return model + +@BACKBONE_REGISTRY.register() +def build_resnet_clip(cfg, input_shape): + """ + Create the whole CLIP instance from config. + + Returns: + CLIP: a :class:`CLIP` instance. + """ + # port standard ResNet config to CLIP ModifiedResNet + freeze_at = cfg.MODEL.BACKBONE.FREEZE_AT + out_features = ['res5'] # includes the whole ResNet # cfg.MODEL.RESNETS.OUT_FEATURES + depth = cfg.MODEL.RESNETS.DEPTH + + num_blocks_per_stage = { + 18: [2, 2, 2, 2], + 34: [3, 4, 6, 3], + 50: [3, 4, 6, 3], + 101: [3, 4, 23, 3], + 152: [3, 8, 36, 3], + 200: [4, 6, 10, 6], # flag for ResNet50x4 + }[depth] + vision_layers = num_blocks_per_stage + vision_width = { + 50: 64, + 101: 64, + 200: 80, # flag for ResNet50x4 + }[depth] # cfg.MODEL.RESNETS.STEM_OUT_CHANNELS + + # default configs of CLIP + embed_dim = { + 50: 1024, + 101: 512, + 200: 640, # flag for ResNet50x4 + }[depth] + vision_heads = vision_width * 32 // 64 + image_resolution = { + 50: 224, + 101: 224, + 200: 288, # flag for ResNet50x4 + }[depth] + vision_patch_size = None + context_length = 77 + vocab_size = 49408 + transformer_width = { + 50: 512, + 101: 512, + 200: 640, # flag for ResNet50x4 + }[depth] + transformer_heads = { + 50: 8, + 101: 8, + 200: 10, # flag for ResNet50x4 + }[depth] + transformer_layers = 12 + + model = CLIP( + embed_dim, + image_resolution, vision_layers, vision_width, vision_patch_size, + context_length, vocab_size, transformer_width, transformer_heads, transformer_layers, + out_features, freeze_at + ) + return model + + +@BACKBONE_REGISTRY.register() +def build_clip_resnet_backbone(cfg, input_shape): + """ + Create a CLIP ResNet instance from config. + + Returns: + ModifiedResNet: a :class:`ModifiedResNet` instance. + """ + # port standard ResNet config to CLIP ModifiedResNet + freeze_at = cfg.MODEL.BACKBONE.FREEZE_AT + out_features = cfg.MODEL.RESNETS.OUT_FEATURES + depth = cfg.MODEL.RESNETS.DEPTH + # num_groups = cfg.MODEL.RESNETS.NUM_GROUPS + # width_per_group = cfg.MODEL.RESNETS.WIDTH_PER_GROUP + # bottleneck_channels = num_groups * width_per_group + # in_channels = cfg.MODEL.RESNETS.STEM_OUT_CHANNELS + # out_channels = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS + # stride_in_1x1 = cfg.MODEL.RESNETS.STRIDE_IN_1X1 + # res5_dilation = cfg.MODEL.RESNETS.RES5_DILATION + # deform_on_per_stage = cfg.MODEL.RESNETS.DEFORM_ON_PER_STAGE + # deform_modulated = cfg.MODEL.RESNETS.DEFORM_MODULATED + # deform_num_groups = cfg.MODEL.RESNETS.DEFORM_NUM_GROUPS + + num_blocks_per_stage = { + 18: [2, 2, 2, 2], + 34: [3, 4, 6, 3], + 50: [3, 4, 6, 3], + 101: [3, 4, 23, 3], + 152: [3, 8, 36, 3], + 200: [4, 6, 10, 6], # flag for ResNet50x4 + }[depth] + vision_layers = num_blocks_per_stage + vision_width = { + 50: 64, + 101: 64, + 200: 80, # flag for ResNet50x4 + }[depth] # cfg.MODEL.RESNETS.STEM_OUT_CHANNELS + + # default configs of CLIP ModifiedResNet, but not used if only building ModifiedResNet as backbone + embed_dim = { + 50: 1024, + 101: 512, + 200: 640, # flag for ResNet50x4 + }[depth] + vision_heads = vision_width * 32 // 64 + image_resolution = { + 50: 224, + 101: 224, + 200: 288, # flag for ResNet50x4 + }[depth] + + # if combine {ModifiedResNet of CLIP, C4, text emb as classifier}, then has to use att_pool to match dimension + create_att_pool = True if (cfg.MODEL.ROI_HEADS.NAME in ['CLIPRes5ROIHeads', 'CLIPStandardROIHeads'] and cfg.MODEL.CLIP.USE_TEXT_EMB_CLASSIFIER)\ + or cfg.MODEL.ROI_HEADS.NAME == 'PretrainRes5ROIHeads' else False + + return ModifiedResNet(layers=vision_layers, + output_dim=embed_dim, + heads=vision_heads, + input_resolution=image_resolution, + width=vision_width, + out_features=out_features, + freeze_at=freeze_at, + depth=depth, + pool_vec=False, + create_att_pool=create_att_pool, + ) + + +class CLIPLangEncoder(nn.Module): + def __init__(self, + embed_dim: int, + # vision + image_resolution: int, + vision_layers: Union[Tuple[int, int, int, int], int], + vision_width: int, + vision_patch_size: int, + # text + context_length: int, + vocab_size: int, + transformer_width: int, + transformer_heads: int, + transformer_layers: int, + out_features, + freeze_at, + ): + super().__init__() + + self.context_length = context_length + + self.transformer = Transformer( + width=transformer_width, + layers=transformer_layers, + heads=transformer_heads, + attn_mask=self.build_attention_mask() + ) + + self.vocab_size = vocab_size + self.token_embedding = nn.Embedding(vocab_size, transformer_width) + self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width)) + self.ln_final = LayerNorm(transformer_width) + + self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim)) + #self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07)) + + self.initialize_parameters() + + def initialize_parameters(self): + nn.init.normal_(self.token_embedding.weight, std=0.02) + nn.init.normal_(self.positional_embedding, std=0.01) + + proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5) + attn_std = self.transformer.width ** -0.5 + fc_std = (2 * self.transformer.width) ** -0.5 + for block in self.transformer.resblocks: + nn.init.normal_(block.attn.in_proj_weight, std=attn_std) + nn.init.normal_(block.attn.out_proj.weight, std=proj_std) + nn.init.normal_(block.mlp.c_fc.weight, std=fc_std) + nn.init.normal_(block.mlp.c_proj.weight, std=proj_std) + + if self.text_projection is not None: + nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5) + + def build_attention_mask(self): + # lazily create causal attention mask, with full attention between the vision tokens + # pytorch uses additive attention mask; fill with -inf + mask = torch.empty(self.context_length, self.context_length) + mask.fill_(float("-inf")) + mask.triu_(1) # zero out the lower diagonal + return mask + + @property + def dtype(self): + return self.transformer.resblocks[0].mlp[0].weight.dtype # torch.float32, not sure whether need to be fp16 in pretraining + + def encode_text(self, text, only_eot=True, norm=True): + x = self.token_embedding(text).type(self.dtype) # [batch_size, n_ctx, d_model] + + x = x + self.positional_embedding.type(self.dtype) + x = x.permute(1, 0, 2) # NLD -> LND + x = self.transformer(x) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.ln_final(x).type(self.dtype) + + if only_eot: + # x.shape = [batch_size, n_ctx, transformer.width] + # take features from the eot embedding (eot_token is the highest number in each sequence) + x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection + if norm: + x = x / x.norm(dim=-1, keepdim=True) + return x + else: + # return embeddings for all tokens, instead of the eot embedding as CLIP implementation below + x = x @ self.text_projection + if norm: + x = x / x.norm(dim=-1, keepdim=True) + return x + +def build_clip_language_encoder(cfg): + """ + Create the CLIP language encoder instance from config. + + Returns: + CLIP: a :class:`CLIP` instance. + """ + # port standard ResNet config to CLIP ModifiedResNet + freeze_at = cfg.MODEL.BACKBONE.FREEZE_AT + out_features = ['res5'] # includes the whole ResNet # cfg.MODEL.RESNETS.OUT_FEATURES + depth = cfg.MODEL.RESNETS.DEPTH + + num_blocks_per_stage = { + 18: [2, 2, 2, 2], + 34: [3, 4, 6, 3], + 50: [3, 4, 6, 3], + 101: [3, 4, 23, 3], + 152: [3, 8, 36, 3], + 200: [4, 6, 10, 6], # flag for ResNet50x4 + }[depth] + vision_layers = num_blocks_per_stage + vision_width = { + 50: 64, + 101: 64, + 200: 80, # flag for ResNet50x4 + }[depth] # cfg.MODEL.RESNETS.STEM_OUT_CHANNELS + + # default configs of CLIP + embed_dim = { + 50: 1024, + 101: 512, + 200: 640, # flag for ResNet50x4 + }[depth] + vision_heads = vision_width * 32 // 64 + image_resolution = { + 50: 224, + 101: 224, + 200: 288, # flag for ResNet50x4 + }[depth] + vision_patch_size = None + context_length = 77 + vocab_size = 49408 + transformer_width = { + 50: 512, + 101: 512, + 200: 640, # flag for ResNet50x4 + }[depth] + transformer_heads = { + 50: 8, + 101: 8, + 200: 10, # flag for ResNet50x4 + }[depth] + transformer_layers = 12 + + model = CLIPLangEncoder( + embed_dim, + image_resolution, vision_layers, vision_width, vision_patch_size, + context_length, vocab_size, transformer_width, transformer_heads, transformer_layers, + out_features, freeze_at + ) + return model \ No newline at end of file diff --git a/detectron2/modeling/backbone/clip_swin.py b/detectron2/modeling/backbone/clip_swin.py new file mode 100644 index 0000000000000000000000000000000000000000..bbd56ec3921cb61409f3b8757c32027059266da1 --- /dev/null +++ b/detectron2/modeling/backbone/clip_swin.py @@ -0,0 +1,289 @@ +from collections import OrderedDict +from typing import Tuple, Union +import logging +import os + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from timm.models.layers import DropPath, trunc_normal_ +from .backbone import Backbone +from .build import BACKBONE_REGISTRY +from .det_swin import SwinTransformer +from ..text_encoder import build_text_encoder +from ..text_encoder import build_tokenizer + +class LayerNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-12): + """Construct a layernorm module in the TF style (epsilon inside the square root). + """ + super(LayerNorm, self).__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.bias = nn.Parameter(torch.zeros(hidden_size)) + self.variance_epsilon = eps + + def forward(self, x): + pdtype = x.dtype + x = x.float() + u = x.mean(-1, keepdim=True) + s = (x - u).pow(2).mean(-1, keepdim=True) + x = (x - u) / torch.sqrt(s + self.variance_epsilon) + return self.weight * x.to(pdtype) + self.bias + + +class QuickGELU(nn.Module): + def forward(self, x: torch.Tensor): + return x * torch.sigmoid(1.702 * x) + + +class ResidualAttentionBlock(nn.Module): + def __init__(self, + d_model: int, + n_head: int, + attn_mask: torch.Tensor = None, + drop_path: float = 0.0): + super().__init__() + + self.attn = nn.MultiheadAttention(d_model, n_head) + self.ln_1 = LayerNorm(d_model) + self.mlp = nn.Sequential(OrderedDict([ + ("c_fc", nn.Linear(d_model, d_model * 4)), + ("gelu", QuickGELU()), + ("c_proj", nn.Linear(d_model * 4, d_model)) + ])) + self.ln_2 = LayerNorm(d_model) + self.attn_mask = attn_mask + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + + def attention(self, x: torch.Tensor): + self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) \ + if self.attn_mask is not None else None + return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0] + + def forward(self, x: torch.Tensor): + x = x + self.drop_path(self.attention(self.ln_1(x))) + x = x + self.drop_path(self.mlp(self.ln_2(x))) + return x + + +class Transformer(nn.Module): + def __init__(self, + context_length: int, + vocab_size: int, + width: int, + layers: int, + heads: int, + drop_path: float = 0.0): + super().__init__() + + self.token_embedding = nn.Embedding(vocab_size, width) + + self.context_length = context_length + self.positional_embedding = nn.Parameter( + torch.empty(self.context_length, width) + ) + + self.width = width + self.layers = layers + attn_mask = self.build_attention_mask() + dpr = [x.item() for x in torch.linspace(0, drop_path, layers)] # stochastic depth decay rule + self.resblocks = nn.Sequential( + *[ + ResidualAttentionBlock(width, heads, attn_mask, dpr[i]) + for i in range(layers) + ] + ) + + self.ln_final = LayerNorm(width) + + trunc_normal_(self.positional_embedding, std=.02) + # nn.init.normal_(self.token_embedding, std=.02) + trunc_normal_(self.token_embedding.weight, std=.02) + self.apply(self._init_weights) + + def build_attention_mask(self): + # lazily create causal attention mask, with full attention between the vision tokens + # pytorch uses additive attention mask; fill with -inf + mask = torch.empty(self.context_length, self.context_length) + mask.fill_(float("-inf")) + mask.triu_(1) # zero out the lower diagonal + return mask + + def _init_weights(self, m): + if isinstance(m, (nn.Linear, nn.Conv2d)): + trunc_normal_(m.weight, std=0.02) + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, (nn.LayerNorm, nn.BatchNorm2d)): + nn.init.constant_(m.bias, 0) + + @torch.jit.ignore + def no_weight_decay(self): + return { + 'positional_embedding', + 'token_embedding', + } + + def forward(self, text: torch.Tensor): + x = self.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.resblocks(x) + x = x.permute(1, 0, 2) # LND -> NLD + + x = self.ln_final(x) + + x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] + + return x + +class CLIP(Backbone): + def __init__(self, config: dict): + super().__init__() + spec_text = config['MODEL']['SPEC']['TEXT'] + assert spec_text['TOKENIZER'] == 'clip', 'Only support clip tokenizer' + self.tokenizer_style = spec_text['TOKENIZER'] + self.tokenizer = build_tokenizer(spec_text) + + self.text_encoder = build_text_encoder(spec_text, self.tokenizer, True) + + embed_dim = config['MODEL']['SPEC']['EMBED_DIM'] + self.text_projection = nn.Parameter( + torch.empty(spec_text['WIDTH'], embed_dim) + ) + + spec_vision = config['MODEL']['SPEC']['VISION'] + self.image_encoder = SwinTransformer( + patch_size=spec_vision['PATCH_SIZE'], + in_chans=spec_vision['IN_CHANS'], + embed_dim=spec_vision['EMBED_DIM'], + depths=spec_vision['DEPTHS'], + num_heads=spec_vision['NUM_HEADS'], + window_size=spec_vision['WINDOW_SIZE'], + mlp_ratio=spec_vision['MLP_RATIO'], + qkv_bias=spec_vision['QKV_BIAS'], + qk_scale=spec_vision.get('QK_SCALE', None), + drop_rate=spec_vision['DROP_RATE'], + attn_drop_rate=spec_vision['ATTN_DROP_RATE'], + drop_path_rate=spec_vision['DROP_PATH_RATE'], + ape=spec_vision['APE'], + patch_norm=spec_vision['PATCH_NORM'], + out_indices=(0, 1, 2, 3), + frozen_stages=-1, + use_checkpoint=False, + ) + + width = spec_vision['EMBED_DIM'] * 2 ** (len(spec_vision['DEPTHS']) - 1) + self.image_projection = nn.Parameter( + torch.empty(width, embed_dim) + ) + # self.logit_scale = nn.Parameter(torch.FloatTensor([np.log(1 / 0.07)])) + self.logit_scale = nn.Parameter(torch.ones([])) + + trunc_normal_(self.text_projection, std=.02) + trunc_normal_(self.image_projection, std=.02) + + def init_weights(self, pretrained='', pretrained_layers=[], verbose=True): + if os.path.isfile(pretrained): + pretrained_dict = torch.load(pretrained, map_location='cpu') + logger.info(f'=> loading pretrained model {pretrained}') + model_dict = self.state_dict() + pretrained_dict = { + k: v for k, v in pretrained_dict.items() + if k in model_dict.keys() + } + need_init_state_dict = {} + for k, v in pretrained_dict.items(): + need_init = ( + k.split('.')[0] in pretrained_layers + or pretrained_layers[0] is '*' + ) + if need_init: + if verbose: + logging.info(f'=> init {k} from {pretrained}') + need_init_state_dict[k] = v + self.load_state_dict(need_init_state_dict, strict=False) + + @torch.jit.ignore + def no_weight_decay(self): + no_weight_decay = {'logit_scale'} + for k in self.text_encoder.no_weight_decay(): + no_weight_decay.add('text.'+k) + + for k in self.image_encoder.no_weight_decay(): + no_weight_decay.add('visual.'+k) + + return no_weight_decay + + @torch.jit.ignore + def no_weight_decay_keywords(self): + return {'relative_position_bias_table'} + + @property + def dtype(self): + return self.image_encoder.conv1.weight.dtype + + def encode_image(self, image, norm=True): + x = self.image_encoder(image) + return x + + def encode_text(self, text, norm=True): + assert isinstance(text, str), "only support single query" + tokens = self.tokenizer( + text, padding='max_length', truncation=True, max_length=77, return_tensors='pt' + ) + tokens = {key:(val.cuda() if next(self.parameters()).is_cuda else val) for key,val in tokens.items()} + x = self.text_encoder(**tokens) + x = x['last_hidden_state'] + x = x[torch.arange(x.size(0)), tokens['input_ids'].argmax(dim=-1)] + + x = x @ self.text_projection + if norm: + x = x / x.norm(dim=-1, keepdim=True) + return x + + def forward(self, image): + features_image = self.image_encoder(image) + return features_image + + +@BACKBONE_REGISTRY.register() +def build_clip_swin_backbone(cfg, input_shape): + """ + Create a CLIP Swin instance from config. + + Returns: + SwinTransformer: a :class:`SwinTransformer` instance. + """ + spec_vision = cfg.MODEL.CLIP.VISION + return SwinTransformer( + patch_size=spec_vision['PATCH_SIZE'], + in_chans=spec_vision['IN_CHANS'], + embed_dim=spec_vision['EMBED_DIM'], + depths=spec_vision['DEPTHS'], + num_heads=spec_vision['NUM_HEADS'], + window_size=spec_vision['WINDOW_SIZE'], + mlp_ratio=spec_vision['MLP_RATIO'], + qkv_bias=spec_vision['QKV_BIAS'], + qk_scale=spec_vision.get('QK_SCALE', None), + drop_rate=spec_vision['DROP_RATE'], + attn_drop_rate=spec_vision['ATTN_DROP_RATE'], + drop_path_rate=spec_vision['DROP_PATH_RATE'], + ape=spec_vision['APE'], + patch_norm=spec_vision['PATCH_NORM'], + out_indices=(0, 1, 2, 3), + frozen_stages=-1, + use_checkpoint=False, + ) + +@BACKBONE_REGISTRY.register() +def build_clip_swin(cfg, input_shape): + """ + Create a CLIP Swin instance from config. + + Returns: + SwinTransformer: a :class:`SwinTransformer` instance. + """ + return CLIP(cfg) \ No newline at end of file diff --git a/detectron2/modeling/backbone/det_swin.py b/detectron2/modeling/backbone/det_swin.py new file mode 100644 index 0000000000000000000000000000000000000000..1ec74aafa2393832fbe1a32e25780aef64e8e667 --- /dev/null +++ b/detectron2/modeling/backbone/det_swin.py @@ -0,0 +1,717 @@ +# -------------------------------------------------------- +# Swin Transformer +# Copyright (c) 2021 Microsoft +# Licensed under The MIT License [see LICENSE for details] +# Written by Ze Liu, Yutong Lin, Yixuan Wei +# -------------------------------------------------------- +import logging +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.utils.checkpoint as checkpoint +import numpy as np +from timm.models.layers import DropPath, to_2tuple, trunc_normal_ +from detectron2.layers import ShapeSpec +from .backbone import Backbone + +logger = logging.getLogger(__name__) + +class Mlp(nn.Module): + """ Multilayer perceptron.""" + + def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = nn.Linear(in_features, hidden_features) + self.act = act_layer() + self.fc2 = nn.Linear(hidden_features, out_features) + self.drop = nn.Dropout(drop) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.drop(x) + x = self.fc2(x) + x = self.drop(x) + return x + + +def window_partition(x, window_size): + """ + Args: + x: (B, H, W, C) + window_size (int): window size + + Returns: + windows: (num_windows*B, window_size, window_size, C) + """ + B, H, W, C = x.shape + x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) + windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) + return windows + + +def window_reverse(windows, window_size, H, W): + """ + Args: + windows: (num_windows*B, window_size, window_size, C) + window_size (int): Window size + H (int): Height of image + W (int): Width of image + + Returns: + x: (B, H, W, C) + """ + B = int(windows.shape[0] / (H * W / window_size / window_size)) + x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) + x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) + return x + + +class WindowAttention(nn.Module): + r""" Window based multi-head self attention (W-MSA) module with relative position bias. + It supports both of shifted and non-shifted window. + + Args: + dim (int): Number of input channels. + window_size (tuple[int]): The height and width of the window. + num_heads (int): Number of attention heads. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set + attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 + proj_drop (float, optional): Dropout ratio of output. Default: 0.0 + """ + + def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): + + super().__init__() + self.dim = dim + self.window_size = window_size # Wh, Ww + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + # define a parameter table of relative position bias + self.relative_position_bias_table = nn.Parameter( + torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH + + # get pair-wise relative position index for each token inside the window + coords_h = torch.arange(self.window_size[0]) + coords_w = torch.arange(self.window_size[1]) + coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 + relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 + relative_coords[:, :, 1] += self.window_size[1] - 1 + relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 + relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww + self.register_buffer("relative_position_index", relative_position_index) + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + trunc_normal_(self.relative_position_bias_table, std=.02) + self.softmax = nn.Softmax(dim=-1) + + def forward(self, x, mask=None): + """ + Args: + x: input features with shape of (num_windows*B, N, C) + mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None + """ + B_, N, C = x.shape + qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) + + q = q * self.scale + attn = (q @ k.transpose(-2, -1)) + + relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view( + self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH + relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww + attn = attn + relative_position_bias.unsqueeze(0) + + if mask is not None: + nW = mask.shape[0] + attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) + attn = attn.view(-1, self.num_heads, N, N) + attn = self.softmax(attn) + else: + attn = self.softmax(attn) + + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B_, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class SwinTransformerBlock(nn.Module): + """ Swin Transformer Block. + + Args: + dim (int): Number of input channels. + num_heads (int): Number of attention heads. + window_size (int): Window size. + shift_size (int): Shift size for SW-MSA. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float, optional): Stochastic depth rate. Default: 0.0 + act_layer (nn.Module, optional): Activation layer. Default: nn.GELU + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + + def __init__(self, dim, num_heads, window_size=7, shift_size=0, + mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0., + act_layer=nn.GELU, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.num_heads = num_heads + self.window_size = window_size + self.shift_size = shift_size + self.mlp_ratio = mlp_ratio + assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size" + + self.norm1 = norm_layer(dim) + self.attn = WindowAttention( + dim, window_size=to_2tuple(self.window_size), num_heads=num_heads, + qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) + + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) + + self.H = None + self.W = None + + def forward(self, x, mask_matrix): + """ Forward function. + + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + mask_matrix: Attention mask for cyclic shift. + """ + B, L, C = x.shape + H, W = self.H, self.W + assert L == H * W, "input feature has wrong size" + + shortcut = x + x = self.norm1(x) + x = x.view(B, H, W, C) + + # pad feature maps to multiples of window size + pad_l = pad_t = 0 + pad_r = (self.window_size - W % self.window_size) % self.window_size + pad_b = (self.window_size - H % self.window_size) % self.window_size + x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) + _, Hp, Wp, _ = x.shape + + # cyclic shift + if self.shift_size > 0: + shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) + attn_mask = mask_matrix + else: + shifted_x = x + attn_mask = None + + # partition windows + x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C + x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C + + # W-MSA/SW-MSA + attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C + + # merge windows + attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) + shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C + + # reverse cyclic shift + if self.shift_size > 0: + x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) + else: + x = shifted_x + + if pad_r > 0 or pad_b > 0: + x = x[:, :H, :W, :].contiguous() + + x = x.view(B, H * W, C) + + # FFN + x = shortcut + self.drop_path(x) + x = x + self.drop_path(self.mlp(self.norm2(x))) + + return x + + +class PatchMerging(nn.Module): + """ Patch Merging Layer + + Args: + dim (int): Number of input channels. + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + def __init__(self, dim, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) + self.norm = norm_layer(4 * dim) + + def forward(self, x, H, W): + """ Forward function. + + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + """ + B, L, C = x.shape + assert L == H * W, "input feature has wrong size" + + x = x.view(B, H, W, C) + + # padding + pad_input = (H % 2 == 1) or (W % 2 == 1) + if pad_input: + x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2)) + + x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C + x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C + x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C + x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C + x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C + x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C + + x = self.norm(x) + x = self.reduction(x) + + return x + + +class BasicLayer(nn.Module): + """ A basic Swin Transformer layer for one stage. + + Args: + dim (int): Number of feature channels + depth (int): Depths of this stage. + num_heads (int): Number of attention head. + window_size (int): Local window size. Default: 7. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, + dim, + depth, + num_heads, + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + drop=0., + attn_drop=0., + drop_path=0., + norm_layer=nn.LayerNorm, + downsample=None, + use_checkpoint=False): + super().__init__() + self.window_size = window_size + self.shift_size = window_size // 2 + self.depth = depth + self.use_checkpoint = use_checkpoint + + # build blocks + self.blocks = nn.ModuleList([ + SwinTransformerBlock( + dim=dim, + num_heads=num_heads, + window_size=window_size, + shift_size=0 if (i % 2 == 0) else window_size // 2, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop, + attn_drop=attn_drop, + drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, + norm_layer=norm_layer) + for i in range(depth)]) + + # patch merging layer + if downsample is not None: + self.downsample = downsample(dim=dim, norm_layer=norm_layer) + else: + self.downsample = None + + def forward(self, x, H, W): + """ Forward function. + + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + """ + + # calculate attention mask for SW-MSA + Hp = int(np.ceil(H / self.window_size)) * self.window_size + Wp = int(np.ceil(W / self.window_size)) * self.window_size + img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1 + h_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + w_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + cnt = 0 + for h in h_slices: + for w in w_slices: + img_mask[:, h, w, :] = cnt + cnt += 1 + + mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1 + mask_windows = mask_windows.view(-1, self.window_size * self.window_size) + attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) + attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) + + for blk in self.blocks: + blk.H, blk.W = H, W + if self.use_checkpoint: + x = checkpoint.checkpoint(blk, x, attn_mask) + else: + x = blk(x, attn_mask) + if self.downsample is not None: + x_down = self.downsample(x, H, W) + Wh, Ww = (H + 1) // 2, (W + 1) // 2 + return x, H, W, x_down, Wh, Ww + else: + return x, H, W, x, H, W + + +class PatchEmbed(nn.Module): + """ Image to Patch Embedding + + Args: + patch_size (int): Patch token size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + norm_layer (nn.Module, optional): Normalization layer. Default: None + """ + + def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): + super().__init__() + patch_size = to_2tuple(patch_size) + self.patch_size = patch_size + + self.in_chans = in_chans + self.embed_dim = embed_dim + + self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + if norm_layer is not None: + self.norm = norm_layer(embed_dim) + else: + self.norm = None + + def forward(self, x): + """Forward function.""" + # padding + _, _, H, W = x.size() + if W % self.patch_size[1] != 0: + x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1])) + if H % self.patch_size[0] != 0: + x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0])) + + x = self.proj(x) # B C Wh Ww + if self.norm is not None: + Wh, Ww = x.size(2), x.size(3) + x = x.flatten(2).transpose(1, 2) + x = self.norm(x) + x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww) + + return x + + +class SwinTransformer(Backbone): + """ Swin Transformer backbone. + A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - + https://arxiv.org/pdf/2103.14030 + + Args: + pretrain_img_size (int): Input image size for training the pretrained model, + used in absolute postion embedding. Default 224. + patch_size (int | tuple(int)): Patch size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + depths (tuple[int]): Depths of each Swin Transformer stage. + num_heads (tuple[int]): Number of attention head of each stage. + window_size (int): Window size. Default: 7. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. + qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. + drop_rate (float): Dropout rate. + attn_drop_rate (float): Attention dropout rate. Default: 0. + drop_path_rate (float): Stochastic depth rate. Default: 0.2. + norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. + ape (bool): If True, add absolute position embedding to the patch embedding. Default: False. + patch_norm (bool): If True, add normalization after patch embedding. Default: True. + out_indices (Sequence[int]): Output from which stages. + frozen_stages (int): Stages to be frozen (stop grad and set eval mode). + -1 means not freezing any parameters. + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, + pretrain_img_size=224, + patch_size=4, + in_chans=3, + embed_dim=96, + depths=[2, 2, 6, 2], + num_heads=[3, 6, 12, 24], + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + drop_rate=0., + attn_drop_rate=0., + drop_path_rate=0.2, + norm_layer=nn.LayerNorm, + ape=False, + patch_norm=True, + out_indices=(0, 1, 2, 3), + frozen_stages=-1, + out_features=["stage2", "stage3", "stage4", "stage5"], + use_checkpoint=False): + super().__init__() + + self.pretrain_img_size = pretrain_img_size + self.num_layers = len(depths) + self.embed_dim = embed_dim + self.ape = ape + self.patch_norm = patch_norm + self.out_indices = out_indices + self.frozen_stages = frozen_stages + self.out_features = out_features + + # split image into non-overlapping patches + self.patch_embed = PatchEmbed( + patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, + norm_layer=norm_layer if self.patch_norm else None) + + # absolute position embedding + if self.ape: + pretrain_img_size = to_2tuple(pretrain_img_size) + patch_size = to_2tuple(patch_size) + patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]] + + self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1])) + trunc_normal_(self.absolute_pos_embed, std=.02) + + self.pos_drop = nn.Dropout(p=drop_rate) + + # stochastic depth + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule + + self._out_feature_strides = {} + self._out_feature_channels = {} + + # build layers + self.layers = nn.ModuleList() + for i_layer in range(self.num_layers): + layer = BasicLayer( + dim=int(embed_dim * 2 ** i_layer), + depth=depths[i_layer], + num_heads=num_heads[i_layer], + window_size=window_size, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop_rate, + attn_drop=attn_drop_rate, + drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], + norm_layer=norm_layer, + downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, + use_checkpoint=use_checkpoint) + self.layers.append(layer) + + stage = f'stage{i_layer + 2}' + if stage in self.out_features: + self._out_feature_channels[stage] = embed_dim * 2 ** i_layer + self._out_feature_strides[stage] = 4 * 2 ** i_layer + + num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)] + self.num_features = num_features + + self.norm = norm_layer(self.num_features[-1]) + + self._freeze_stages() + + def output_shape(self): + return { + name: ShapeSpec( + channels=self._out_feature_channels[name], stride=self._out_feature_strides[name] + ) + for name in self.out_features + } + + def _freeze_stages(self): + if self.frozen_stages >= 0: + self.patch_embed.eval() + for param in self.patch_embed.parameters(): + param.requires_grad = False + + if self.frozen_stages >= 1 and self.ape: + self.absolute_pos_embed.requires_grad = False + + if self.frozen_stages >= 2: + self.pos_drop.eval() + for i in range(0, self.frozen_stages - 1): + m = self.layers[i] + m.eval() + for param in m.parameters(): + param.requires_grad = False + + @torch.jit.ignore + def no_weight_decay(self): + return {'absolute_pos_embed'} + + @torch.jit.ignore + def no_weight_decay_keywords(self): + return {'relative_position_bias_table', 'norm'} + + # def init_weights(self, pretrained=None): + # """Initialize the weights in backbone. + + # Args: + # pretrained (str, optional): Path to pre-trained weights. + # Defaults to None. + # """ + + # def _init_weights(m): + # if isinstance(m, nn.Linear): + # trunc_normal_(m.weight, std=.02) + # if isinstance(m, nn.Linear) and m.bias is not None: + # nn.init.constant_(m.bias, 0) + # elif isinstance(m, nn.LayerNorm): + # nn.init.constant_(m.bias, 0) + # nn.init.constant_(m.weight, 1.0) + + # if isinstance(pretrained, str): + # self.apply(_init_weights) + # logger = get_root_logger() + # load_checkpoint(self, pretrained, strict=False, logger=logger) + # elif pretrained is None: + # self.apply(_init_weights) + # else: + # raise TypeError('pretrained must be a str or None') + + def init_weights(self, pretrained='', pretrained_layers=[], verbose=True): + if not os.path.isfile(pretrained): + logger.warning(f'=> Pretrained model ({pretrained}) is not a file, skip init weight') + return + + pretrained_dict = torch.load(pretrained, map_location='cpu') + logger.info(f'=> Loading pretrained model {pretrained}') + model_dict = self.state_dict() + pretrained_dict = { + k: v for k, v in pretrained_dict.items() + if k in model_dict.keys() + } + need_init_state_dict = {} + for k, v in pretrained_dict.items(): + need_init = ( + ( + k.split('.')[0] in pretrained_layers + or pretrained_layers[0] == '*' + ) + and 'relative_position_index' not in k + and 'attn_mask' not in k + ) + + if need_init: + if verbose: + logger.info(f'=> init {k} from {pretrained}') + + if 'relative_position_bias_table' in k and v.size() != model_dict[k].size(): + relative_position_bias_table_pretrained = v + relative_position_bias_table_current = model_dict[k] + L1, nH1 = relative_position_bias_table_pretrained.size() + L2, nH2 = relative_position_bias_table_current.size() + if nH1 != nH2: + logger.info(f"Error in loading {k}, passing") + else: + if L1 != L2: + logger.info( + '=> load_pretrained: resized variant: {} to {}' + .format((L1, nH1), (L2, nH2)) + ) + S1 = int(L1 ** 0.5) + S2 = int(L2 ** 0.5) + relative_position_bias_table_pretrained_resized = torch.nn.functional.interpolate( + relative_position_bias_table_pretrained.permute(1, 0).view(1, nH1, S1, S1), + size=(S2, S2), + mode='bicubic') + v = relative_position_bias_table_pretrained_resized.view(nH2, L2).permute(1, 0) + + if 'absolute_pos_embed' in k and v.size() != model_dict[k].size(): + absolute_pos_embed_pretrained = v + absolute_pos_embed_current = model_dict[k] + _, L1, C1 = absolute_pos_embed_pretrained.size() + _, L2, C2 = absolute_pos_embed_current.size() + if C1 != C1: + logger.info(f"Error in loading {k}, passing") + else: + if L1 != L2: + logger.info( + '=> load_pretrained: resized variant: {} to {}' + .format((1, L1, C1), (1, L2, C2)) + ) + S1 = int(L1 ** 0.5) + S2 = int(L2 ** 0.5) + absolute_pos_embed_pretrained = absolute_pos_embed_pretrained.reshape(-1, S1, S1, C1) + absolute_pos_embed_pretrained = absolute_pos_embed_pretrained.permute(0, 3, 1, 2) + absolute_pos_embed_pretrained_resized = torch.nn.functional.interpolate( + absolute_pos_embed_pretrained, size=(S2, S2), mode='bicubic') + v = absolute_pos_embed_pretrained_resized.permute(0, 2, 3, 1).flatten(1, 2) + + need_init_state_dict[k] = v + self.load_state_dict(need_init_state_dict, strict=False) + + def forward(self, x): + """Forward function.""" + x = self.patch_embed(x) + + Wh, Ww = x.size(2), x.size(3) + if self.ape: + # interpolate the position embedding to the corresponding size + absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic') + x = (x + absolute_pos_embed).flatten(2).transpose(1, 2) # B Wh*Ww C + else: + x = x.flatten(2).transpose(1, 2) + x = self.pos_drop(x) + + outs = {} + for i in range(self.num_layers): + layer = self.layers[i] + x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww) + name = f'stage{i + 2}' + if name in self.out_features: + out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous() + outs[name] = out + return outs + + def train(self, mode=True): + """Convert the model into training mode while keep layers freezed.""" + super(SwinTransformer, self).train(mode) + self._freeze_stages() \ No newline at end of file diff --git a/detectron2/modeling/backbone/fpn.py b/detectron2/modeling/backbone/fpn.py new file mode 100644 index 0000000000000000000000000000000000000000..532711d882b7baf109eef1fded128069e144d6ba --- /dev/null +++ b/detectron2/modeling/backbone/fpn.py @@ -0,0 +1,277 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import math +import fvcore.nn.weight_init as weight_init +import torch +import torch.nn.functional as F +from torch import nn + +from detectron2.layers import Conv2d, ShapeSpec, get_norm + +from .backbone import Backbone +from .build import BACKBONE_REGISTRY +from .resnet import build_resnet_backbone +from .clip_backbone import build_clip_resnet_backbone + +__all__ = ["build_clip_resnet_fpn_backbone", "build_resnet_fpn_backbone", "build_retinanet_resnet_fpn_backbone", "FPN"] + + +class FPN(Backbone): + """ + This module implements :paper:`FPN`. + It creates pyramid features built on top of some input feature maps. + """ + + _fuse_type: torch.jit.Final[str] + + def __init__( + self, bottom_up, in_features, out_channels, norm="", top_block=None, fuse_type="sum" + ): + """ + Args: + bottom_up (Backbone): module representing the bottom up subnetwork. + Must be a subclass of :class:`Backbone`. The multi-scale feature + maps generated by the bottom up network, and listed in `in_features`, + are used to generate FPN levels. + in_features (list[str]): names of the input feature maps coming + from the backbone to which FPN is attached. For example, if the + backbone produces ["res2", "res3", "res4"], any *contiguous* sublist + of these may be used; order must be from high to low resolution. + out_channels (int): number of channels in the output feature maps. + norm (str): the normalization to use. + top_block (nn.Module or None): if provided, an extra operation will + be performed on the output of the last (smallest resolution) + FPN output, and the result will extend the result list. The top_block + further downsamples the feature map. It must have an attribute + "num_levels", meaning the number of extra FPN levels added by + this block, and "in_feature", which is a string representing + its input feature (e.g., p5). + fuse_type (str): types for fusing the top down features and the lateral + ones. It can be "sum" (default), which sums up element-wise; or "avg", + which takes the element-wise mean of the two. + """ + super(FPN, self).__init__() + assert isinstance(bottom_up, Backbone) + assert in_features, in_features + + # Feature map strides and channels from the bottom up network (e.g. ResNet) + input_shapes = bottom_up.output_shape() + strides = [input_shapes[f].stride for f in in_features] + in_channels_per_feature = [input_shapes[f].channels for f in in_features] + + _assert_strides_are_log2_contiguous(strides) + lateral_convs = [] + output_convs = [] + + use_bias = norm == "" + for idx, in_channels in enumerate(in_channels_per_feature): + lateral_norm = get_norm(norm, out_channels) + output_norm = get_norm(norm, out_channels) + + lateral_conv = Conv2d( + in_channels, out_channels, kernel_size=1, bias=use_bias, norm=lateral_norm + ) + output_conv = Conv2d( + out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1, + bias=use_bias, + norm=output_norm, + ) + weight_init.c2_xavier_fill(lateral_conv) + weight_init.c2_xavier_fill(output_conv) + stage = int(math.log2(strides[idx])) + self.add_module("fpn_lateral{}".format(stage), lateral_conv) + self.add_module("fpn_output{}".format(stage), output_conv) + + lateral_convs.append(lateral_conv) + output_convs.append(output_conv) + # Place convs into top-down order (from low to high resolution) + # to make the top-down computation in forward clearer. + self.lateral_convs = lateral_convs[::-1] + self.output_convs = output_convs[::-1] + self.top_block = top_block + self.in_features = tuple(in_features) + self.bottom_up = bottom_up + # Return feature names are "p", like ["p2", "p3", ..., "p6"] + self._out_feature_strides = {"p{}".format(int(math.log2(s))): s for s in strides} + # top block output feature maps. + if self.top_block is not None: + for s in range(stage, stage + self.top_block.num_levels): + self._out_feature_strides["p{}".format(s + 1)] = 2 ** (s + 1) + + self._out_features = list(self._out_feature_strides.keys()) + self._out_feature_channels = {k: out_channels for k in self._out_features} + self._size_divisibility = strides[-1] + assert fuse_type in {"avg", "sum"} + self._fuse_type = fuse_type + + @property + def size_divisibility(self): + return self._size_divisibility + + def forward(self, x): + """ + Args: + input (dict[str->Tensor]): mapping feature map name (e.g., "res5") to + feature map tensor for each feature level in high to low resolution order. + + Returns: + dict[str->Tensor]: + mapping from feature map name to FPN feature map tensor + in high to low resolution order. Returned feature names follow the FPN + paper convention: "p", where stage has stride = 2 ** stage e.g., + ["p2", "p3", ..., "p6"]. + """ + bottom_up_features = self.bottom_up(x) + results = [] + prev_features = self.lateral_convs[0](bottom_up_features[self.in_features[-1]]) + results.append(self.output_convs[0](prev_features)) + + # Reverse feature maps into top-down order (from low to high resolution) + for idx, (lateral_conv, output_conv) in enumerate( + zip(self.lateral_convs, self.output_convs) + ): + # Slicing of ModuleList is not supported https://github.com/pytorch/pytorch/issues/47336 + # Therefore we loop over all modules but skip the first one + if idx > 0: + features = self.in_features[-idx - 1] + features = bottom_up_features[features] + top_down_features = F.interpolate(prev_features, scale_factor=2.0, mode="nearest") + lateral_features = lateral_conv(features) + prev_features = lateral_features + top_down_features + if self._fuse_type == "avg": + prev_features /= 2 + results.insert(0, output_conv(prev_features)) + + if self.top_block is not None: + if self.top_block.in_feature in bottom_up_features: + top_block_in_feature = bottom_up_features[self.top_block.in_feature] + else: + top_block_in_feature = results[self._out_features.index(self.top_block.in_feature)] + results.extend(self.top_block(top_block_in_feature)) + assert len(self._out_features) == len(results) + return {f: res for f, res in zip(self._out_features, results)} + + def output_shape(self): + return { + name: ShapeSpec( + channels=self._out_feature_channels[name], stride=self._out_feature_strides[name] + ) + for name in self._out_features + } + + +def _assert_strides_are_log2_contiguous(strides): + """ + Assert that each stride is 2x times its preceding stride, i.e. "contiguous in log2". + """ + for i, stride in enumerate(strides[1:], 1): + assert stride == 2 * strides[i - 1], "Strides {} {} are not log2 contiguous".format( + stride, strides[i - 1] + ) + + +class LastLevelMaxPool(nn.Module): + """ + This module is used in the original FPN to generate a downsampled + P6 feature from P5. + """ + + def __init__(self): + super().__init__() + self.num_levels = 1 + self.in_feature = "p5" + + def forward(self, x): + return [F.max_pool2d(x, kernel_size=1, stride=2, padding=0)] + + +class LastLevelP6P7(nn.Module): + """ + This module is used in RetinaNet to generate extra layers, P6 and P7 from + C5 feature. + """ + + def __init__(self, in_channels, out_channels, in_feature="res5"): + super().__init__() + self.num_levels = 2 + self.in_feature = in_feature + self.p6 = nn.Conv2d(in_channels, out_channels, 3, 2, 1) + self.p7 = nn.Conv2d(out_channels, out_channels, 3, 2, 1) + for module in [self.p6, self.p7]: + weight_init.c2_xavier_fill(module) + + def forward(self, c5): + p6 = self.p6(c5) + p7 = self.p7(F.relu(p6)) + return [p6, p7] + + +@BACKBONE_REGISTRY.register() +def build_resnet_fpn_backbone(cfg, input_shape: ShapeSpec): + """ + Args: + cfg: a detectron2 CfgNode + + Returns: + backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`. + """ + bottom_up = build_resnet_backbone(cfg, input_shape) + in_features = cfg.MODEL.FPN.IN_FEATURES + out_channels = cfg.MODEL.FPN.OUT_CHANNELS + backbone = FPN( + bottom_up=bottom_up, + in_features=in_features, + out_channels=out_channels, + norm=cfg.MODEL.FPN.NORM, + top_block=LastLevelMaxPool(), + fuse_type=cfg.MODEL.FPN.FUSE_TYPE, + ) + return backbone + +@BACKBONE_REGISTRY.register() +def build_clip_resnet_fpn_backbone(cfg, input_shape: ShapeSpec): + """ + Args: + cfg: a detectron2 CfgNode + + Returns: + backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`. + """ + bottom_up = build_clip_resnet_backbone(cfg, input_shape) + in_features = cfg.MODEL.FPN.IN_FEATURES + out_channels = cfg.MODEL.FPN.OUT_CHANNELS + backbone = FPN( + bottom_up=bottom_up, + in_features=in_features, + out_channels=out_channels, + norm=cfg.MODEL.FPN.NORM, + top_block=LastLevelMaxPool(), + fuse_type=cfg.MODEL.FPN.FUSE_TYPE, + ) + return backbone + +@BACKBONE_REGISTRY.register() +def build_retinanet_resnet_fpn_backbone(cfg, input_shape: ShapeSpec): + """ + Args: + cfg: a detectron2 CfgNode + + Returns: + backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`. + """ + bottom_up = build_resnet_backbone(cfg, input_shape) + in_features = cfg.MODEL.FPN.IN_FEATURES + out_channels = cfg.MODEL.FPN.OUT_CHANNELS + in_channels_p6p7 = bottom_up.output_shape()["res5"].channels + backbone = FPN( + bottom_up=bottom_up, + in_features=in_features, + out_channels=out_channels, + norm=cfg.MODEL.FPN.NORM, + top_block=LastLevelP6P7(in_channels_p6p7, out_channels), + fuse_type=cfg.MODEL.FPN.FUSE_TYPE, + ) + return backbone diff --git a/detectron2/modeling/backbone/regnet.py b/detectron2/modeling/backbone/regnet.py new file mode 100644 index 0000000000000000000000000000000000000000..3533d63385d1324cfc1559eae9576b3fa52585af --- /dev/null +++ b/detectron2/modeling/backbone/regnet.py @@ -0,0 +1,452 @@ +# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved +""" +Implementation of RegNet models from :paper:`dds` and :paper:`scaling`. + +This code is adapted from https://github.com/facebookresearch/pycls with minimal modifications. +Some code duplication exists between RegNet and ResNets (e.g., ResStem) in order to simplify +model loading. +""" + +import numpy as np +from torch import nn + +from detectron2.layers import CNNBlockBase, ShapeSpec, get_norm + +from .backbone import Backbone + +__all__ = [ + "AnyNet", + "RegNet", + "ResStem", + "SimpleStem", + "VanillaBlock", + "ResBasicBlock", + "ResBottleneckBlock", +] + + +def conv2d(w_in, w_out, k, *, stride=1, groups=1, bias=False): + """Helper for building a conv2d layer.""" + assert k % 2 == 1, "Only odd size kernels supported to avoid padding issues." + s, p, g, b = stride, (k - 1) // 2, groups, bias + return nn.Conv2d(w_in, w_out, k, stride=s, padding=p, groups=g, bias=b) + + +def gap2d(): + """Helper for building a global average pooling layer.""" + return nn.AdaptiveAvgPool2d((1, 1)) + + +def pool2d(k, *, stride=1): + """Helper for building a pool2d layer.""" + assert k % 2 == 1, "Only odd size kernels supported to avoid padding issues." + return nn.MaxPool2d(k, stride=stride, padding=(k - 1) // 2) + + +def init_weights(m): + """Performs ResNet-style weight initialization.""" + if isinstance(m, nn.Conv2d): + # Note that there is no bias due to BN + fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels + m.weight.data.normal_(mean=0.0, std=np.sqrt(2.0 / fan_out)) + elif isinstance(m, nn.BatchNorm2d): + m.weight.data.fill_(1.0) + m.bias.data.zero_() + elif isinstance(m, nn.Linear): + m.weight.data.normal_(mean=0.0, std=0.01) + m.bias.data.zero_() + + +class ResStem(CNNBlockBase): + """ResNet stem for ImageNet: 7x7, BN, AF, MaxPool.""" + + def __init__(self, w_in, w_out, norm, activation_class): + super().__init__(w_in, w_out, 4) + self.conv = conv2d(w_in, w_out, 7, stride=2) + self.bn = get_norm(norm, w_out) + self.af = activation_class() + self.pool = pool2d(3, stride=2) + + def forward(self, x): + for layer in self.children(): + x = layer(x) + return x + + +class SimpleStem(CNNBlockBase): + """Simple stem for ImageNet: 3x3, BN, AF.""" + + def __init__(self, w_in, w_out, norm, activation_class): + super().__init__(w_in, w_out, 2) + self.conv = conv2d(w_in, w_out, 3, stride=2) + self.bn = get_norm(norm, w_out) + self.af = activation_class() + + def forward(self, x): + for layer in self.children(): + x = layer(x) + return x + + +class SE(nn.Module): + """Squeeze-and-Excitation (SE) block: AvgPool, FC, Act, FC, Sigmoid.""" + + def __init__(self, w_in, w_se, activation_class): + super().__init__() + self.avg_pool = gap2d() + self.f_ex = nn.Sequential( + conv2d(w_in, w_se, 1, bias=True), + activation_class(), + conv2d(w_se, w_in, 1, bias=True), + nn.Sigmoid(), + ) + + def forward(self, x): + return x * self.f_ex(self.avg_pool(x)) + + +class VanillaBlock(CNNBlockBase): + """Vanilla block: [3x3 conv, BN, Relu] x2.""" + + def __init__(self, w_in, w_out, stride, norm, activation_class, _params): + super().__init__(w_in, w_out, stride) + self.a = conv2d(w_in, w_out, 3, stride=stride) + self.a_bn = get_norm(norm, w_out) + self.a_af = activation_class() + self.b = conv2d(w_out, w_out, 3) + self.b_bn = get_norm(norm, w_out) + self.b_af = activation_class() + + def forward(self, x): + for layer in self.children(): + x = layer(x) + return x + + +class BasicTransform(nn.Module): + """Basic transformation: [3x3 conv, BN, Relu] x2.""" + + def __init__(self, w_in, w_out, stride, norm, activation_class, _params): + super().__init__() + self.a = conv2d(w_in, w_out, 3, stride=stride) + self.a_bn = get_norm(norm, w_out) + self.a_af = activation_class() + self.b = conv2d(w_out, w_out, 3) + self.b_bn = get_norm(norm, w_out) + self.b_bn.final_bn = True + + def forward(self, x): + for layer in self.children(): + x = layer(x) + return x + + +class ResBasicBlock(CNNBlockBase): + """Residual basic block: x + f(x), f = basic transform.""" + + def __init__(self, w_in, w_out, stride, norm, activation_class, params): + super().__init__(w_in, w_out, stride) + self.proj, self.bn = None, None + if (w_in != w_out) or (stride != 1): + self.proj = conv2d(w_in, w_out, 1, stride=stride) + self.bn = get_norm(norm, w_out) + self.f = BasicTransform(w_in, w_out, stride, norm, activation_class, params) + self.af = activation_class() + + def forward(self, x): + x_p = self.bn(self.proj(x)) if self.proj else x + return self.af(x_p + self.f(x)) + + +class BottleneckTransform(nn.Module): + """Bottleneck transformation: 1x1, 3x3 [+SE], 1x1.""" + + def __init__(self, w_in, w_out, stride, norm, activation_class, params): + super().__init__() + w_b = int(round(w_out * params["bot_mul"])) + w_se = int(round(w_in * params["se_r"])) + groups = w_b // params["group_w"] + self.a = conv2d(w_in, w_b, 1) + self.a_bn = get_norm(norm, w_b) + self.a_af = activation_class() + self.b = conv2d(w_b, w_b, 3, stride=stride, groups=groups) + self.b_bn = get_norm(norm, w_b) + self.b_af = activation_class() + self.se = SE(w_b, w_se, activation_class) if w_se else None + self.c = conv2d(w_b, w_out, 1) + self.c_bn = get_norm(norm, w_out) + self.c_bn.final_bn = True + + def forward(self, x): + for layer in self.children(): + x = layer(x) + return x + + +class ResBottleneckBlock(CNNBlockBase): + """Residual bottleneck block: x + f(x), f = bottleneck transform.""" + + def __init__(self, w_in, w_out, stride, norm, activation_class, params): + super().__init__(w_in, w_out, stride) + self.proj, self.bn = None, None + if (w_in != w_out) or (stride != 1): + self.proj = conv2d(w_in, w_out, 1, stride=stride) + self.bn = get_norm(norm, w_out) + self.f = BottleneckTransform(w_in, w_out, stride, norm, activation_class, params) + self.af = activation_class() + + def forward(self, x): + x_p = self.bn(self.proj(x)) if self.proj else x + return self.af(x_p + self.f(x)) + + +class AnyStage(nn.Module): + """AnyNet stage (sequence of blocks w/ the same output shape).""" + + def __init__(self, w_in, w_out, stride, d, block_class, norm, activation_class, params): + super().__init__() + for i in range(d): + block = block_class(w_in, w_out, stride, norm, activation_class, params) + self.add_module("b{}".format(i + 1), block) + stride, w_in = 1, w_out + + def forward(self, x): + for block in self.children(): + x = block(x) + return x + + +class AnyNet(Backbone): + """AnyNet model. See :paper:`dds`.""" + + def __init__( + self, + *, + stem_class, + stem_width, + block_class, + depths, + widths, + group_widths, + strides, + bottleneck_ratios, + se_ratio, + activation_class, + freeze_at=0, + norm="BN", + out_features=None, + ): + """ + Args: + stem_class (callable): A callable taking 4 arguments (channels in, channels out, + normalization, callable returning an activation function) that returns another + callable implementing the stem module. + stem_width (int): The number of output channels that the stem produces. + block_class (callable): A callable taking 6 arguments (channels in, channels out, + stride, normalization, callable returning an activation function, a dict of + block-specific parameters) that returns another callable implementing the repeated + block module. + depths (list[int]): Number of blocks in each stage. + widths (list[int]): For each stage, the number of output channels of each block. + group_widths (list[int]): For each stage, the number of channels per group in group + convolution, if the block uses group convolution. + strides (list[int]): The stride that each network stage applies to its input. + bottleneck_ratios (list[float]): For each stage, the ratio of the number of bottleneck + channels to the number of block input channels (or, equivalently, output channels), + if the block uses a bottleneck. + se_ratio (float): The ratio of the number of channels used inside the squeeze-excitation + (SE) module to it number of input channels, if SE the block uses SE. + activation_class (callable): A callable taking no arguments that returns another + callable implementing an activation function. + freeze_at (int): The number of stages at the beginning to freeze. + see :meth:`freeze` for detailed explanation. + norm (str or callable): normalization for all conv layers. + See :func:`layers.get_norm` for supported format. + out_features (list[str]): name of the layers whose outputs should + be returned in forward. RegNet's use "stem" and "s1", "s2", etc for the stages after + the stem. If None, will return the output of the last layer. + """ + super().__init__() + self.stem = stem_class(3, stem_width, norm, activation_class) + + current_stride = self.stem.stride + self._out_feature_strides = {"stem": current_stride} + self._out_feature_channels = {"stem": self.stem.out_channels} + self.stages_and_names = [] + prev_w = stem_width + + for i, (d, w, s, b, g) in enumerate( + zip(depths, widths, strides, bottleneck_ratios, group_widths) + ): + params = {"bot_mul": b, "group_w": g, "se_r": se_ratio} + stage = AnyStage(prev_w, w, s, d, block_class, norm, activation_class, params) + name = "s{}".format(i + 1) + self.add_module(name, stage) + self.stages_and_names.append((stage, name)) + self._out_feature_strides[name] = current_stride = int( + current_stride * np.prod([k.stride for k in stage.children()]) + ) + self._out_feature_channels[name] = list(stage.children())[-1].out_channels + prev_w = w + + self.apply(init_weights) + + if out_features is None: + out_features = [name] + self._out_features = out_features + assert len(self._out_features) + children = [x[0] for x in self.named_children()] + for out_feature in self._out_features: + assert out_feature in children, "Available children: {} does not include {}".format( + ", ".join(children), out_feature + ) + self.freeze(freeze_at) + + def forward(self, x): + """ + Args: + x: Tensor of shape (N,C,H,W). H, W must be a multiple of ``self.size_divisibility``. + + Returns: + dict[str->Tensor]: names and the corresponding features + """ + assert x.dim() == 4, f"Model takes an input of shape (N, C, H, W). Got {x.shape} instead!" + outputs = {} + x = self.stem(x) + if "stem" in self._out_features: + outputs["stem"] = x + for stage, name in self.stages_and_names: + x = stage(x) + if name in self._out_features: + outputs[name] = x + return outputs + + def output_shape(self): + return { + name: ShapeSpec( + channels=self._out_feature_channels[name], stride=self._out_feature_strides[name] + ) + for name in self._out_features + } + + def freeze(self, freeze_at=0): + """ + Freeze the first several stages of the model. Commonly used in fine-tuning. + + Layers that produce the same feature map spatial size are defined as one + "stage" by :paper:`FPN`. + + Args: + freeze_at (int): number of stages to freeze. + `1` means freezing the stem. `2` means freezing the stem and + one residual stage, etc. + + Returns: + nn.Module: this model itself + """ + if freeze_at >= 1: + self.stem.freeze() + for idx, (stage, _) in enumerate(self.stages_and_names, start=2): + if freeze_at >= idx: + for block in stage.children(): + block.freeze() + return self + + +def adjust_block_compatibility(ws, bs, gs): + """Adjusts the compatibility of widths, bottlenecks, and groups.""" + assert len(ws) == len(bs) == len(gs) + assert all(w > 0 and b > 0 and g > 0 for w, b, g in zip(ws, bs, gs)) + vs = [int(max(1, w * b)) for w, b in zip(ws, bs)] + gs = [int(min(g, v)) for g, v in zip(gs, vs)] + ms = [np.lcm(g, b) if b > 1 else g for g, b in zip(gs, bs)] + vs = [max(m, int(round(v / m) * m)) for v, m in zip(vs, ms)] + ws = [int(v / b) for v, b in zip(vs, bs)] + assert all(w * b % g == 0 for w, b, g in zip(ws, bs, gs)) + return ws, bs, gs + + +def generate_regnet_parameters(w_a, w_0, w_m, d, q=8): + """Generates per stage widths and depths from RegNet parameters.""" + assert w_a >= 0 and w_0 > 0 and w_m > 1 and w_0 % q == 0 + # Generate continuous per-block ws + ws_cont = np.arange(d) * w_a + w_0 + # Generate quantized per-block ws + ks = np.round(np.log(ws_cont / w_0) / np.log(w_m)) + ws_all = w_0 * np.power(w_m, ks) + ws_all = np.round(np.divide(ws_all, q)).astype(int) * q + # Generate per stage ws and ds (assumes ws_all are sorted) + ws, ds = np.unique(ws_all, return_counts=True) + # Compute number of actual stages and total possible stages + num_stages, total_stages = len(ws), ks.max() + 1 + # Convert numpy arrays to lists and return + ws, ds, ws_all, ws_cont = (x.tolist() for x in (ws, ds, ws_all, ws_cont)) + return ws, ds, num_stages, total_stages, ws_all, ws_cont + + +class RegNet(AnyNet): + """RegNet model. See :paper:`dds`.""" + + def __init__( + self, + *, + stem_class, + stem_width, + block_class, + depth, + w_a, + w_0, + w_m, + group_width, + stride=2, + bottleneck_ratio=1.0, + se_ratio=0.0, + activation_class=None, + freeze_at=0, + norm="BN", + out_features=None, + ): + """ + Build a RegNet from the parameterization described in :paper:`dds` Section 3.3. + + Args: + See :class:`AnyNet` for arguments that are not listed here. + depth (int): Total number of blocks in the RegNet. + w_a (float): Factor by which block width would increase prior to quantizing block widths + by stage. See :paper:`dds` Section 3.3. + w_0 (int): Initial block width. See :paper:`dds` Section 3.3. + w_m (float): Parameter controlling block width quantization. + See :paper:`dds` Section 3.3. + group_width (int): Number of channels per group in group convolution, if the block uses + group convolution. + bottleneck_ratio (float): The ratio of the number of bottleneck channels to the number + of block input channels (or, equivalently, output channels), if the block uses a + bottleneck. + stride (int): The stride that each network stage applies to its input. + """ + ws, ds = generate_regnet_parameters(w_a, w_0, w_m, depth)[0:2] + ss = [stride for _ in ws] + bs = [bottleneck_ratio for _ in ws] + gs = [group_width for _ in ws] + ws, bs, gs = adjust_block_compatibility(ws, bs, gs) + + def default_activation_class(): + return nn.ReLU(inplace=True) + + super().__init__( + stem_class=stem_class, + stem_width=stem_width, + block_class=block_class, + depths=ds, + widths=ws, + strides=ss, + group_widths=gs, + bottleneck_ratios=bs, + se_ratio=se_ratio, + activation_class=default_activation_class + if activation_class is None + else activation_class, + freeze_at=freeze_at, + norm=norm, + out_features=out_features, + ) diff --git a/detectron2/modeling/backbone/resnet.py b/detectron2/modeling/backbone/resnet.py new file mode 100644 index 0000000000000000000000000000000000000000..5b8e842c585a81b5345ade4ca1da62a4904a122a --- /dev/null +++ b/detectron2/modeling/backbone/resnet.py @@ -0,0 +1,694 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import numpy as np +import fvcore.nn.weight_init as weight_init +import torch +import torch.nn.functional as F +from torch import nn + +from detectron2.layers import ( + CNNBlockBase, + Conv2d, + DeformConv, + ModulatedDeformConv, + ShapeSpec, + get_norm, +) + +from .backbone import Backbone +from .build import BACKBONE_REGISTRY + +__all__ = [ + "ResNetBlockBase", + "BasicBlock", + "BottleneckBlock", + "DeformBottleneckBlock", + "BasicStem", + "ResNet", + "make_stage", + "build_resnet_backbone", +] + + +class BasicBlock(CNNBlockBase): + """ + The basic residual block for ResNet-18 and ResNet-34 defined in :paper:`ResNet`, + with two 3x3 conv layers and a projection shortcut if needed. + """ + + def __init__(self, in_channels, out_channels, *, stride=1, norm="BN"): + """ + Args: + in_channels (int): Number of input channels. + out_channels (int): Number of output channels. + stride (int): Stride for the first conv. + norm (str or callable): normalization for all conv layers. + See :func:`layers.get_norm` for supported format. + """ + super().__init__(in_channels, out_channels, stride) + + if in_channels != out_channels: + self.shortcut = Conv2d( + in_channels, + out_channels, + kernel_size=1, + stride=stride, + bias=False, + norm=get_norm(norm, out_channels), + ) + else: + self.shortcut = None + + self.conv1 = Conv2d( + in_channels, + out_channels, + kernel_size=3, + stride=stride, + padding=1, + bias=False, + norm=get_norm(norm, out_channels), + ) + + self.conv2 = Conv2d( + out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1, + bias=False, + norm=get_norm(norm, out_channels), + ) + + for layer in [self.conv1, self.conv2, self.shortcut]: + if layer is not None: # shortcut can be None + weight_init.c2_msra_fill(layer) + + def forward(self, x): + out = self.conv1(x) + out = F.relu_(out) + out = self.conv2(out) + + if self.shortcut is not None: + shortcut = self.shortcut(x) + else: + shortcut = x + + out += shortcut + out = F.relu_(out) + return out + + +class BottleneckBlock(CNNBlockBase): + """ + The standard bottleneck residual block used by ResNet-50, 101 and 152 + defined in :paper:`ResNet`. It contains 3 conv layers with kernels + 1x1, 3x3, 1x1, and a projection shortcut if needed. + """ + + def __init__( + self, + in_channels, + out_channels, + *, + bottleneck_channels, + stride=1, + num_groups=1, + norm="BN", + stride_in_1x1=False, + dilation=1, + ): + """ + Args: + bottleneck_channels (int): number of output channels for the 3x3 + "bottleneck" conv layers. + num_groups (int): number of groups for the 3x3 conv layer. + norm (str or callable): normalization for all conv layers. + See :func:`layers.get_norm` for supported format. + stride_in_1x1 (bool): when stride>1, whether to put stride in the + first 1x1 convolution or the bottleneck 3x3 convolution. + dilation (int): the dilation rate of the 3x3 conv layer. + """ + super().__init__(in_channels, out_channels, stride) + + if in_channels != out_channels: + self.shortcut = Conv2d( + in_channels, + out_channels, + kernel_size=1, + stride=stride, + bias=False, + norm=get_norm(norm, out_channels), + ) + else: + self.shortcut = None + + # The original MSRA ResNet models have stride in the first 1x1 conv + # The subsequent fb.torch.resnet and Caffe2 ResNe[X]t implementations have + # stride in the 3x3 conv + stride_1x1, stride_3x3 = (stride, 1) if stride_in_1x1 else (1, stride) + + self.conv1 = Conv2d( + in_channels, + bottleneck_channels, + kernel_size=1, + stride=stride_1x1, + bias=False, + norm=get_norm(norm, bottleneck_channels), + ) + + self.conv2 = Conv2d( + bottleneck_channels, + bottleneck_channels, + kernel_size=3, + stride=stride_3x3, + padding=1 * dilation, + bias=False, + groups=num_groups, + dilation=dilation, + norm=get_norm(norm, bottleneck_channels), + ) + + self.conv3 = Conv2d( + bottleneck_channels, + out_channels, + kernel_size=1, + bias=False, + norm=get_norm(norm, out_channels), + ) + + for layer in [self.conv1, self.conv2, self.conv3, self.shortcut]: + if layer is not None: # shortcut can be None + weight_init.c2_msra_fill(layer) + + # Zero-initialize the last normalization in each residual branch, + # so that at the beginning, the residual branch starts with zeros, + # and each residual block behaves like an identity. + # See Sec 5.1 in "Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour": + # "For BN layers, the learnable scaling coefficient γ is initialized + # to be 1, except for each residual block's last BN + # where γ is initialized to be 0." + + # nn.init.constant_(self.conv3.norm.weight, 0) + # TODO this somehow hurts performance when training GN models from scratch. + # Add it as an option when we need to use this code to train a backbone. + + def forward(self, x): + out = self.conv1(x) + out = F.relu_(out) + + out = self.conv2(out) + out = F.relu_(out) + + out = self.conv3(out) + + if self.shortcut is not None: + shortcut = self.shortcut(x) + else: + shortcut = x + + out += shortcut + out = F.relu_(out) + return out + + +class DeformBottleneckBlock(CNNBlockBase): + """ + Similar to :class:`BottleneckBlock`, but with :paper:`deformable conv ` + in the 3x3 convolution. + """ + + def __init__( + self, + in_channels, + out_channels, + *, + bottleneck_channels, + stride=1, + num_groups=1, + norm="BN", + stride_in_1x1=False, + dilation=1, + deform_modulated=False, + deform_num_groups=1, + ): + super().__init__(in_channels, out_channels, stride) + self.deform_modulated = deform_modulated + + if in_channels != out_channels: + self.shortcut = Conv2d( + in_channels, + out_channels, + kernel_size=1, + stride=stride, + bias=False, + norm=get_norm(norm, out_channels), + ) + else: + self.shortcut = None + + stride_1x1, stride_3x3 = (stride, 1) if stride_in_1x1 else (1, stride) + + self.conv1 = Conv2d( + in_channels, + bottleneck_channels, + kernel_size=1, + stride=stride_1x1, + bias=False, + norm=get_norm(norm, bottleneck_channels), + ) + + if deform_modulated: + deform_conv_op = ModulatedDeformConv + # offset channels are 2 or 3 (if with modulated) * kernel_size * kernel_size + offset_channels = 27 + else: + deform_conv_op = DeformConv + offset_channels = 18 + + self.conv2_offset = Conv2d( + bottleneck_channels, + offset_channels * deform_num_groups, + kernel_size=3, + stride=stride_3x3, + padding=1 * dilation, + dilation=dilation, + ) + self.conv2 = deform_conv_op( + bottleneck_channels, + bottleneck_channels, + kernel_size=3, + stride=stride_3x3, + padding=1 * dilation, + bias=False, + groups=num_groups, + dilation=dilation, + deformable_groups=deform_num_groups, + norm=get_norm(norm, bottleneck_channels), + ) + + self.conv3 = Conv2d( + bottleneck_channels, + out_channels, + kernel_size=1, + bias=False, + norm=get_norm(norm, out_channels), + ) + + for layer in [self.conv1, self.conv2, self.conv3, self.shortcut]: + if layer is not None: # shortcut can be None + weight_init.c2_msra_fill(layer) + + nn.init.constant_(self.conv2_offset.weight, 0) + nn.init.constant_(self.conv2_offset.bias, 0) + + def forward(self, x): + out = self.conv1(x) + out = F.relu_(out) + + if self.deform_modulated: + offset_mask = self.conv2_offset(out) + offset_x, offset_y, mask = torch.chunk(offset_mask, 3, dim=1) + offset = torch.cat((offset_x, offset_y), dim=1) + mask = mask.sigmoid() + out = self.conv2(out, offset, mask) + else: + offset = self.conv2_offset(out) + out = self.conv2(out, offset) + out = F.relu_(out) + + out = self.conv3(out) + + if self.shortcut is not None: + shortcut = self.shortcut(x) + else: + shortcut = x + + out += shortcut + out = F.relu_(out) + return out + + +class BasicStem(CNNBlockBase): + """ + The standard ResNet stem (layers before the first residual block), + with a conv, relu and max_pool. + """ + + def __init__(self, in_channels=3, out_channels=64, norm="BN"): + """ + Args: + norm (str or callable): norm after the first conv layer. + See :func:`layers.get_norm` for supported format. + """ + super().__init__(in_channels, out_channels, 4) + self.in_channels = in_channels + self.conv1 = Conv2d( + in_channels, + out_channels, + kernel_size=7, + stride=2, + padding=3, + bias=False, + norm=get_norm(norm, out_channels), + ) + weight_init.c2_msra_fill(self.conv1) + + def forward(self, x): + x = self.conv1(x) + x = F.relu_(x) + x = F.max_pool2d(x, kernel_size=3, stride=2, padding=1) + return x + + +class ResNet(Backbone): + """ + Implement :paper:`ResNet`. + """ + + def __init__(self, stem, stages, num_classes=None, out_features=None, freeze_at=0): + """ + Args: + stem (nn.Module): a stem module + stages (list[list[CNNBlockBase]]): several (typically 4) stages, + each contains multiple :class:`CNNBlockBase`. + num_classes (None or int): if None, will not perform classification. + Otherwise, will create a linear layer. + out_features (list[str]): name of the layers whose outputs should + be returned in forward. Can be anything in "stem", "linear", or "res2" ... + If None, will return the output of the last layer. + freeze_at (int): The number of stages at the beginning to freeze. + see :meth:`freeze` for detailed explanation. + """ + super().__init__() + self.stem = stem + self.num_classes = num_classes + + current_stride = self.stem.stride + self._out_feature_strides = {"stem": current_stride} + self._out_feature_channels = {"stem": self.stem.out_channels} + + self.stage_names, self.stages = [], [] + + if out_features is not None: + # Avoid keeping unused layers in this module. They consume extra memory + # and may cause allreduce to fail + num_stages = max( + [{"res2": 1, "res3": 2, "res4": 3, "res5": 4}.get(f, 0) for f in out_features] + ) + stages = stages[:num_stages] + for i, blocks in enumerate(stages): + assert len(blocks) > 0, len(blocks) + for block in blocks: + assert isinstance(block, CNNBlockBase), block + + name = "res" + str(i + 2) + stage = nn.Sequential(*blocks) + + self.add_module(name, stage) + self.stage_names.append(name) + self.stages.append(stage) + + self._out_feature_strides[name] = current_stride = int( + current_stride * np.prod([k.stride for k in blocks]) + ) + self._out_feature_channels[name] = curr_channels = blocks[-1].out_channels + self.stage_names = tuple(self.stage_names) # Make it static for scripting + + if num_classes is not None: + self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) + self.linear = nn.Linear(curr_channels, num_classes) + + # Sec 5.1 in "Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour": + # "The 1000-way fully-connected layer is initialized by + # drawing weights from a zero-mean Gaussian with standard deviation of 0.01." + nn.init.normal_(self.linear.weight, std=0.01) + name = "linear" + + if out_features is None: + out_features = [name] + self._out_features = out_features + assert len(self._out_features) + children = [x[0] for x in self.named_children()] + for out_feature in self._out_features: + assert out_feature in children, "Available children: {}".format(", ".join(children)) + self.freeze(freeze_at) + + def forward(self, x): + """ + Args: + x: Tensor of shape (N,C,H,W). H, W must be a multiple of ``self.size_divisibility``. + + Returns: + dict[str->Tensor]: names and the corresponding features + """ + assert x.dim() == 4, f"ResNet takes an input of shape (N, C, H, W). Got {x.shape} instead!" + outputs = {} + x = self.stem(x) + if "stem" in self._out_features: + outputs["stem"] = x + for name, stage in zip(self.stage_names, self.stages): + x = stage(x) + if name in self._out_features: + outputs[name] = x + if self.num_classes is not None: + x = self.avgpool(x) + x = torch.flatten(x, 1) + x = self.linear(x) + if "linear" in self._out_features: + outputs["linear"] = x + return outputs + + def output_shape(self): + return { + name: ShapeSpec( + channels=self._out_feature_channels[name], stride=self._out_feature_strides[name] + ) + for name in self._out_features + } + + def freeze(self, freeze_at=0): + """ + Freeze the first several stages of the ResNet. Commonly used in + fine-tuning. + + Layers that produce the same feature map spatial size are defined as one + "stage" by :paper:`FPN`. + + Args: + freeze_at (int): number of stages to freeze. + `1` means freezing the stem. `2` means freezing the stem and + one residual stage, etc. + + Returns: + nn.Module: this ResNet itself + """ + if freeze_at >= 1: + self.stem.freeze() + for idx, stage in enumerate(self.stages, start=2): + if freeze_at >= idx: + for block in stage.children(): + block.freeze() + return self + + @staticmethod + def make_stage(block_class, num_blocks, *, in_channels, out_channels, **kwargs): + """ + Create a list of blocks of the same type that forms one ResNet stage. + + Args: + block_class (type): a subclass of CNNBlockBase that's used to create all blocks in this + stage. A module of this type must not change spatial resolution of inputs unless its + stride != 1. + num_blocks (int): number of blocks in this stage + in_channels (int): input channels of the entire stage. + out_channels (int): output channels of **every block** in the stage. + kwargs: other arguments passed to the constructor of + `block_class`. If the argument name is "xx_per_block", the + argument is a list of values to be passed to each block in the + stage. Otherwise, the same argument is passed to every block + in the stage. + + Returns: + list[CNNBlockBase]: a list of block module. + + Examples: + :: + stage = ResNet.make_stage( + BottleneckBlock, 3, in_channels=16, out_channels=64, + bottleneck_channels=16, num_groups=1, + stride_per_block=[2, 1, 1], + dilations_per_block=[1, 1, 2] + ) + + Usually, layers that produce the same feature map spatial size are defined as one + "stage" (in :paper:`FPN`). Under such definition, ``stride_per_block[1:]`` should + all be 1. + """ + blocks = [] + for i in range(num_blocks): + curr_kwargs = {} + for k, v in kwargs.items(): + if k.endswith("_per_block"): + assert len(v) == num_blocks, ( + f"Argument '{k}' of make_stage should have the " + f"same length as num_blocks={num_blocks}." + ) + newk = k[: -len("_per_block")] + assert newk not in kwargs, f"Cannot call make_stage with both {k} and {newk}!" + curr_kwargs[newk] = v[i] + else: + curr_kwargs[k] = v + + blocks.append( + block_class(in_channels=in_channels, out_channels=out_channels, **curr_kwargs) + ) + in_channels = out_channels + return blocks + + @staticmethod + def make_default_stages(depth, block_class=None, **kwargs): + """ + Created list of ResNet stages from pre-defined depth (one of 18, 34, 50, 101, 152). + If it doesn't create the ResNet variant you need, please use :meth:`make_stage` + instead for fine-grained customization. + + Args: + depth (int): depth of ResNet + block_class (type): the CNN block class. Has to accept + `bottleneck_channels` argument for depth > 50. + By default it is BasicBlock or BottleneckBlock, based on the + depth. + kwargs: + other arguments to pass to `make_stage`. Should not contain + stride and channels, as they are predefined for each depth. + + Returns: + list[list[CNNBlockBase]]: modules in all stages; see arguments of + :class:`ResNet.__init__`. + """ + num_blocks_per_stage = { + 18: [2, 2, 2, 2], + 34: [3, 4, 6, 3], + 50: [3, 4, 6, 3], + 101: [3, 4, 23, 3], + 152: [3, 8, 36, 3], + }[depth] + if block_class is None: + block_class = BasicBlock if depth < 50 else BottleneckBlock + if depth < 50: + in_channels = [64, 64, 128, 256] + out_channels = [64, 128, 256, 512] + else: + in_channels = [64, 256, 512, 1024] + out_channels = [256, 512, 1024, 2048] + ret = [] + for (n, s, i, o) in zip(num_blocks_per_stage, [1, 2, 2, 2], in_channels, out_channels): + if depth >= 50: + kwargs["bottleneck_channels"] = o // 4 + ret.append( + ResNet.make_stage( + block_class=block_class, + num_blocks=n, + stride_per_block=[s] + [1] * (n - 1), + in_channels=i, + out_channels=o, + **kwargs, + ) + ) + return ret + + +ResNetBlockBase = CNNBlockBase +""" +Alias for backward compatibiltiy. +""" + + +def make_stage(*args, **kwargs): + """ + Deprecated alias for backward compatibiltiy. + """ + return ResNet.make_stage(*args, **kwargs) + + +@BACKBONE_REGISTRY.register() +def build_resnet_backbone(cfg, input_shape): + """ + Create a ResNet instance from config. + + Returns: + ResNet: a :class:`ResNet` instance. + """ + # need registration of new blocks/stems? + norm = cfg.MODEL.RESNETS.NORM + stem = BasicStem( + in_channels=input_shape.channels, + out_channels=cfg.MODEL.RESNETS.STEM_OUT_CHANNELS, + norm=norm, + ) + + # fmt: off + freeze_at = cfg.MODEL.BACKBONE.FREEZE_AT + out_features = cfg.MODEL.RESNETS.OUT_FEATURES + depth = cfg.MODEL.RESNETS.DEPTH + num_groups = cfg.MODEL.RESNETS.NUM_GROUPS + width_per_group = cfg.MODEL.RESNETS.WIDTH_PER_GROUP + bottleneck_channels = num_groups * width_per_group + in_channels = cfg.MODEL.RESNETS.STEM_OUT_CHANNELS + out_channels = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS + stride_in_1x1 = cfg.MODEL.RESNETS.STRIDE_IN_1X1 + res5_dilation = cfg.MODEL.RESNETS.RES5_DILATION + deform_on_per_stage = cfg.MODEL.RESNETS.DEFORM_ON_PER_STAGE + deform_modulated = cfg.MODEL.RESNETS.DEFORM_MODULATED + deform_num_groups = cfg.MODEL.RESNETS.DEFORM_NUM_GROUPS + # fmt: on + assert res5_dilation in {1, 2}, "res5_dilation cannot be {}.".format(res5_dilation) + + num_blocks_per_stage = { + 18: [2, 2, 2, 2], + 34: [3, 4, 6, 3], + 50: [3, 4, 6, 3], + 101: [3, 4, 23, 3], + 152: [3, 8, 36, 3], + }[depth] + + if depth in [18, 34]: + assert out_channels == 64, "Must set MODEL.RESNETS.RES2_OUT_CHANNELS = 64 for R18/R34" + assert not any( + deform_on_per_stage + ), "MODEL.RESNETS.DEFORM_ON_PER_STAGE unsupported for R18/R34" + assert res5_dilation == 1, "Must set MODEL.RESNETS.RES5_DILATION = 1 for R18/R34" + assert num_groups == 1, "Must set MODEL.RESNETS.NUM_GROUPS = 1 for R18/R34" + + stages = [] + + for idx, stage_idx in enumerate(range(2, 6)): + # res5_dilation is used this way as a convention in R-FCN & Deformable Conv paper + dilation = res5_dilation if stage_idx == 5 else 1 + first_stride = 1 if idx == 0 or (stage_idx == 5 and dilation == 2) else 2 + stage_kargs = { + "num_blocks": num_blocks_per_stage[idx], + "stride_per_block": [first_stride] + [1] * (num_blocks_per_stage[idx] - 1), + "in_channels": in_channels, + "out_channels": out_channels, + "norm": norm, + } + # Use BasicBlock for R18 and R34. + if depth in [18, 34]: + stage_kargs["block_class"] = BasicBlock + else: + stage_kargs["bottleneck_channels"] = bottleneck_channels + stage_kargs["stride_in_1x1"] = stride_in_1x1 + stage_kargs["dilation"] = dilation + stage_kargs["num_groups"] = num_groups + if deform_on_per_stage[idx]: + stage_kargs["block_class"] = DeformBottleneckBlock + stage_kargs["deform_modulated"] = deform_modulated + stage_kargs["deform_num_groups"] = deform_num_groups + else: + stage_kargs["block_class"] = BottleneckBlock + blocks = ResNet.make_stage(**stage_kargs) + in_channels = out_channels + out_channels *= 2 + bottleneck_channels *= 2 + stages.append(blocks) + return ResNet(stem, stages, out_features=out_features, freeze_at=freeze_at) diff --git a/detectron2/modeling/box_regression.py b/detectron2/modeling/box_regression.py new file mode 100644 index 0000000000000000000000000000000000000000..12be0008b66bd4954a5139aeb6e07d71f8159caa --- /dev/null +++ b/detectron2/modeling/box_regression.py @@ -0,0 +1,270 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import math +from typing import List, Tuple +import torch +from fvcore.nn import giou_loss, smooth_l1_loss + +from detectron2.layers import cat +from detectron2.structures import Boxes + +# Value for clamping large dw and dh predictions. The heuristic is that we clamp +# such that dw and dh are no larger than what would transform a 16px box into a +# 1000px box (based on a small anchor, 16px, and a typical image size, 1000px). +_DEFAULT_SCALE_CLAMP = math.log(1000.0 / 16) + + +__all__ = ["Box2BoxTransform", "Box2BoxTransformRotated"] + + +@torch.jit.script +class Box2BoxTransform(object): + """ + The box-to-box transform defined in R-CNN. The transformation is parameterized + by 4 deltas: (dx, dy, dw, dh). The transformation scales the box's width and height + by exp(dw), exp(dh) and shifts a box's center by the offset (dx * width, dy * height). + """ + + def __init__( + self, weights: Tuple[float, float, float, float], scale_clamp: float = _DEFAULT_SCALE_CLAMP + ): + """ + Args: + weights (4-element tuple): Scaling factors that are applied to the + (dx, dy, dw, dh) deltas. In Fast R-CNN, these were originally set + such that the deltas have unit variance; now they are treated as + hyperparameters of the system. + scale_clamp (float): When predicting deltas, the predicted box scaling + factors (dw and dh) are clamped such that they are <= scale_clamp. + """ + self.weights = weights + self.scale_clamp = scale_clamp + + def get_deltas(self, src_boxes, target_boxes): + """ + Get box regression transformation deltas (dx, dy, dw, dh) that can be used + to transform the `src_boxes` into the `target_boxes`. That is, the relation + ``target_boxes == self.apply_deltas(deltas, src_boxes)`` is true (unless + any delta is too large and is clamped). + + Args: + src_boxes (Tensor): source boxes, e.g., object proposals + target_boxes (Tensor): target of the transformation, e.g., ground-truth + boxes. + """ + assert isinstance(src_boxes, torch.Tensor), type(src_boxes) + assert isinstance(target_boxes, torch.Tensor), type(target_boxes) + + src_widths = src_boxes[:, 2] - src_boxes[:, 0] + src_heights = src_boxes[:, 3] - src_boxes[:, 1] + src_ctr_x = src_boxes[:, 0] + 0.5 * src_widths + src_ctr_y = src_boxes[:, 1] + 0.5 * src_heights + + target_widths = target_boxes[:, 2] - target_boxes[:, 0] + target_heights = target_boxes[:, 3] - target_boxes[:, 1] + target_ctr_x = target_boxes[:, 0] + 0.5 * target_widths + target_ctr_y = target_boxes[:, 1] + 0.5 * target_heights + + wx, wy, ww, wh = self.weights + dx = wx * (target_ctr_x - src_ctr_x) / src_widths + dy = wy * (target_ctr_y - src_ctr_y) / src_heights + dw = ww * torch.log(target_widths / src_widths) + dh = wh * torch.log(target_heights / src_heights) + + deltas = torch.stack((dx, dy, dw, dh), dim=1) + assert (src_widths > 0).all().item(), "Input boxes to Box2BoxTransform are not valid!" + return deltas + + def apply_deltas(self, deltas, boxes): + """ + Apply transformation `deltas` (dx, dy, dw, dh) to `boxes`. + + Args: + deltas (Tensor): transformation deltas of shape (N, k*4), where k >= 1. + deltas[i] represents k potentially different class-specific + box transformations for the single box boxes[i]. + boxes (Tensor): boxes to transform, of shape (N, 4) + """ + deltas = deltas.float() # ensure fp32 for decoding precision + boxes = boxes.to(deltas.dtype) + + widths = boxes[:, 2] - boxes[:, 0] + heights = boxes[:, 3] - boxes[:, 1] + ctr_x = boxes[:, 0] + 0.5 * widths + ctr_y = boxes[:, 1] + 0.5 * heights + + wx, wy, ww, wh = self.weights + dx = deltas[:, 0::4] / wx + dy = deltas[:, 1::4] / wy + dw = deltas[:, 2::4] / ww + dh = deltas[:, 3::4] / wh + + # Prevent sending too large values into torch.exp() + dw = torch.clamp(dw, max=self.scale_clamp) + dh = torch.clamp(dh, max=self.scale_clamp) + + pred_ctr_x = dx * widths[:, None] + ctr_x[:, None] + pred_ctr_y = dy * heights[:, None] + ctr_y[:, None] + pred_w = torch.exp(dw) * widths[:, None] + pred_h = torch.exp(dh) * heights[:, None] + + x1 = pred_ctr_x - 0.5 * pred_w + y1 = pred_ctr_y - 0.5 * pred_h + x2 = pred_ctr_x + 0.5 * pred_w + y2 = pred_ctr_y + 0.5 * pred_h + pred_boxes = torch.stack((x1, y1, x2, y2), dim=-1) + return pred_boxes.reshape(deltas.shape) + + +@torch.jit.script +class Box2BoxTransformRotated(object): + """ + The box-to-box transform defined in Rotated R-CNN. The transformation is parameterized + by 5 deltas: (dx, dy, dw, dh, da). The transformation scales the box's width and height + by exp(dw), exp(dh), shifts a box's center by the offset (dx * width, dy * height), + and rotate a box's angle by da (radians). + Note: angles of deltas are in radians while angles of boxes are in degrees. + """ + + def __init__( + self, + weights: Tuple[float, float, float, float, float], + scale_clamp: float = _DEFAULT_SCALE_CLAMP, + ): + """ + Args: + weights (5-element tuple): Scaling factors that are applied to the + (dx, dy, dw, dh, da) deltas. These are treated as + hyperparameters of the system. + scale_clamp (float): When predicting deltas, the predicted box scaling + factors (dw and dh) are clamped such that they are <= scale_clamp. + """ + self.weights = weights + self.scale_clamp = scale_clamp + + def get_deltas(self, src_boxes, target_boxes): + """ + Get box regression transformation deltas (dx, dy, dw, dh, da) that can be used + to transform the `src_boxes` into the `target_boxes`. That is, the relation + ``target_boxes == self.apply_deltas(deltas, src_boxes)`` is true (unless + any delta is too large and is clamped). + + Args: + src_boxes (Tensor): Nx5 source boxes, e.g., object proposals + target_boxes (Tensor): Nx5 target of the transformation, e.g., ground-truth + boxes. + """ + assert isinstance(src_boxes, torch.Tensor), type(src_boxes) + assert isinstance(target_boxes, torch.Tensor), type(target_boxes) + + src_ctr_x, src_ctr_y, src_widths, src_heights, src_angles = torch.unbind(src_boxes, dim=1) + + target_ctr_x, target_ctr_y, target_widths, target_heights, target_angles = torch.unbind( + target_boxes, dim=1 + ) + + wx, wy, ww, wh, wa = self.weights + dx = wx * (target_ctr_x - src_ctr_x) / src_widths + dy = wy * (target_ctr_y - src_ctr_y) / src_heights + dw = ww * torch.log(target_widths / src_widths) + dh = wh * torch.log(target_heights / src_heights) + # Angles of deltas are in radians while angles of boxes are in degrees. + # the conversion to radians serve as a way to normalize the values + da = target_angles - src_angles + da = (da + 180.0) % 360.0 - 180.0 # make it in [-180, 180) + da *= wa * math.pi / 180.0 + + deltas = torch.stack((dx, dy, dw, dh, da), dim=1) + assert ( + (src_widths > 0).all().item() + ), "Input boxes to Box2BoxTransformRotated are not valid!" + return deltas + + def apply_deltas(self, deltas, boxes): + """ + Apply transformation `deltas` (dx, dy, dw, dh, da) to `boxes`. + + Args: + deltas (Tensor): transformation deltas of shape (N, k*5). + deltas[i] represents box transformation for the single box boxes[i]. + boxes (Tensor): boxes to transform, of shape (N, 5) + """ + assert deltas.shape[1] % 5 == 0 and boxes.shape[1] == 5 + + boxes = boxes.to(deltas.dtype).unsqueeze(2) + + ctr_x = boxes[:, 0] + ctr_y = boxes[:, 1] + widths = boxes[:, 2] + heights = boxes[:, 3] + angles = boxes[:, 4] + + wx, wy, ww, wh, wa = self.weights + + dx = deltas[:, 0::5] / wx + dy = deltas[:, 1::5] / wy + dw = deltas[:, 2::5] / ww + dh = deltas[:, 3::5] / wh + da = deltas[:, 4::5] / wa + + # Prevent sending too large values into torch.exp() + dw = torch.clamp(dw, max=self.scale_clamp) + dh = torch.clamp(dh, max=self.scale_clamp) + + pred_boxes = torch.zeros_like(deltas) + pred_boxes[:, 0::5] = dx * widths + ctr_x # x_ctr + pred_boxes[:, 1::5] = dy * heights + ctr_y # y_ctr + pred_boxes[:, 2::5] = torch.exp(dw) * widths # width + pred_boxes[:, 3::5] = torch.exp(dh) * heights # height + + # Following original RRPN implementation, + # angles of deltas are in radians while angles of boxes are in degrees. + pred_angle = da * 180.0 / math.pi + angles + pred_angle = (pred_angle + 180.0) % 360.0 - 180.0 # make it in [-180, 180) + + pred_boxes[:, 4::5] = pred_angle + + return pred_boxes + + +def _dense_box_regression_loss( + anchors: List[Boxes], + box2box_transform: Box2BoxTransform, + pred_anchor_deltas: List[torch.Tensor], + gt_boxes: List[torch.Tensor], + fg_mask: torch.Tensor, + box_reg_loss_type="smooth_l1", + smooth_l1_beta=0.0, +): + """ + Compute loss for dense multi-level box regression. + Loss is accumulated over ``fg_mask``. + + Args: + anchors: #lvl anchor boxes, each is (HixWixA, 4) + pred_anchor_deltas: #lvl predictions, each is (N, HixWixA, 4) + gt_boxes: N ground truth boxes, each has shape (R, 4) (R = sum(Hi * Wi * A)) + fg_mask: the foreground boolean mask of shape (N, R) to compute loss on + box_reg_loss_type (str): Loss type to use. Supported losses: "smooth_l1", "giou". + smooth_l1_beta (float): beta parameter for the smooth L1 regression loss. Default to + use L1 loss. Only used when `box_reg_loss_type` is "smooth_l1" + """ + anchors = type(anchors[0]).cat(anchors).tensor # (R, 4) + if box_reg_loss_type == "smooth_l1": + gt_anchor_deltas = [box2box_transform.get_deltas(anchors, k) for k in gt_boxes] + gt_anchor_deltas = torch.stack(gt_anchor_deltas) # (N, R, 4) + loss_box_reg = smooth_l1_loss( + cat(pred_anchor_deltas, dim=1)[fg_mask], + gt_anchor_deltas[fg_mask], + beta=smooth_l1_beta, + reduction="sum", + ) + elif box_reg_loss_type == "giou": + pred_boxes = [ + box2box_transform.apply_deltas(k, anchors) for k in cat(pred_anchor_deltas, dim=1) + ] + loss_box_reg = giou_loss( + torch.stack(pred_boxes)[fg_mask], torch.stack(gt_boxes)[fg_mask], reduction="sum" + ) + else: + raise ValueError(f"Invalid dense box regression loss type '{box_reg_loss_type}'") + return loss_box_reg diff --git a/detectron2/modeling/matcher.py b/detectron2/modeling/matcher.py new file mode 100644 index 0000000000000000000000000000000000000000..666913f76fb0b9d8a277541716f91872d8246250 --- /dev/null +++ b/detectron2/modeling/matcher.py @@ -0,0 +1,126 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from typing import List +import torch + +from detectron2.layers import nonzero_tuple + + +class Matcher(object): + """ + This class assigns to each predicted "element" (e.g., a box) a ground-truth + element. Each predicted element will have exactly zero or one matches; each + ground-truth element may be matched to zero or more predicted elements. + + The matching is determined by the MxN match_quality_matrix, that characterizes + how well each (ground-truth, prediction)-pair match each other. For example, + if the elements are boxes, this matrix may contain box intersection-over-union + overlap values. + + The matcher returns (a) a vector of length N containing the index of the + ground-truth element m in [0, M) that matches to prediction n in [0, N). + (b) a vector of length N containing the labels for each prediction. + """ + + def __init__( + self, thresholds: List[float], labels: List[int], allow_low_quality_matches: bool = False + ): + """ + Args: + thresholds (list): a list of thresholds used to stratify predictions + into levels. + labels (list): a list of values to label predictions belonging at + each level. A label can be one of {-1, 0, 1} signifying + {ignore, negative class, positive class}, respectively. + allow_low_quality_matches (bool): if True, produce additional matches + for predictions with maximum match quality lower than high_threshold. + See set_low_quality_matches_ for more details. + + For example, + thresholds = [0.3, 0.5] + labels = [0, -1, 1] + All predictions with iou < 0.3 will be marked with 0 and + thus will be considered as false positives while training. + All predictions with 0.3 <= iou < 0.5 will be marked with -1 and + thus will be ignored. + All predictions with 0.5 <= iou will be marked with 1 and + thus will be considered as true positives. + """ + # Add -inf and +inf to first and last position in thresholds + thresholds = thresholds[:] + assert thresholds[0] > 0 + thresholds.insert(0, -float("inf")) + thresholds.append(float("inf")) + # Currently torchscript does not support all + generator + assert all([low <= high for (low, high) in zip(thresholds[:-1], thresholds[1:])]) + assert all([l in [-1, 0, 1] for l in labels]) + assert len(labels) == len(thresholds) - 1 + self.thresholds = thresholds + self.labels = labels + self.allow_low_quality_matches = allow_low_quality_matches + + def __call__(self, match_quality_matrix): + """ + Args: + match_quality_matrix (Tensor[float]): an MxN tensor, containing the + pairwise quality between M ground-truth elements and N predicted + elements. All elements must be >= 0 (due to the us of `torch.nonzero` + for selecting indices in :meth:`set_low_quality_matches_`). + + Returns: + matches (Tensor[int64]): a vector of length N, where matches[i] is a matched + ground-truth index in [0, M) + match_labels (Tensor[int8]): a vector of length N, where pred_labels[i] indicates + whether a prediction is a true or false positive or ignored + """ + assert match_quality_matrix.dim() == 2 + if match_quality_matrix.numel() == 0: + default_matches = match_quality_matrix.new_full( + (match_quality_matrix.size(1),), 0, dtype=torch.int64 + ) + # When no gt boxes exist, we define IOU = 0 and therefore set labels + # to `self.labels[0]`, which usually defaults to background class 0 + # To choose to ignore instead, can make labels=[-1,0,-1,1] + set appropriate thresholds + default_match_labels = match_quality_matrix.new_full( + (match_quality_matrix.size(1),), self.labels[0], dtype=torch.int8 + ) + return default_matches, default_match_labels + + assert torch.all(match_quality_matrix >= 0) + + # match_quality_matrix is M (gt) x N (predicted) + # Max over gt elements (dim 0) to find best gt candidate for each prediction + matched_vals, matches = match_quality_matrix.max(dim=0) + + match_labels = matches.new_full(matches.size(), 1, dtype=torch.int8) + + for (l, low, high) in zip(self.labels, self.thresholds[:-1], self.thresholds[1:]): + low_high = (matched_vals >= low) & (matched_vals < high) + match_labels[low_high] = l + + if self.allow_low_quality_matches: + self.set_low_quality_matches_(match_labels, match_quality_matrix) + + return matches, match_labels + + def set_low_quality_matches_(self, match_labels, match_quality_matrix): + """ + Produce additional matches for predictions that have only low-quality matches. + Specifically, for each ground-truth G find the set of predictions that have + maximum overlap with it (including ties); for each prediction in that set, if + it is unmatched, then match it to the ground-truth G. + + This function implements the RPN assignment case (i) in Sec. 3.1.2 of + :paper:`Faster R-CNN`. + """ + # For each gt, find the prediction with which it has highest quality + highest_quality_foreach_gt, _ = match_quality_matrix.max(dim=1) + # Find the highest quality match available, even if it is low, including ties. + # Note that the matches qualities must be positive due to the use of + # `torch.nonzero`. + _, pred_inds_with_highest_quality = nonzero_tuple( + match_quality_matrix == highest_quality_foreach_gt[:, None] + ) + # If an anchor was labeled positive only due to a low-quality match + # with gt_A, but it has larger overlap with gt_B, it's matched index will still be gt_B. + # This follows the implementation in Detectron, and is found to have no significant impact. + match_labels[pred_inds_with_highest_quality] = 1 diff --git a/detectron2/modeling/meta_arch/__init__.py b/detectron2/modeling/meta_arch/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..04caae8693a51e59f1f31d1daac18df484842e93 --- /dev/null +++ b/detectron2/modeling/meta_arch/__init__.py @@ -0,0 +1,15 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +from .build import META_ARCH_REGISTRY, build_model # isort:skip + +from .panoptic_fpn import PanopticFPN + +# import all the meta_arch, so they will be registered +from .rcnn import GeneralizedRCNN, ProposalNetwork +from .retinanet import RetinaNet +from .semantic_seg import SEM_SEG_HEADS_REGISTRY, SemanticSegmentor, build_sem_seg_head +from .clip_rcnn import CLIPRCNN, CLIPFastRCNN, PretrainFastRCNN + + +__all__ = list(globals().keys()) diff --git a/detectron2/modeling/meta_arch/__pycache__/__init__.cpython-39.pyc b/detectron2/modeling/meta_arch/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7c7e7d90901f756c2a892fc935f9e8f35dfc4f7f Binary files /dev/null and b/detectron2/modeling/meta_arch/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/modeling/meta_arch/__pycache__/build.cpython-39.pyc b/detectron2/modeling/meta_arch/__pycache__/build.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b82e0763afcc38b2e73e0a87572ea8df93e7fef1 Binary files /dev/null and b/detectron2/modeling/meta_arch/__pycache__/build.cpython-39.pyc differ diff --git a/detectron2/modeling/meta_arch/__pycache__/clip_rcnn.cpython-39.pyc b/detectron2/modeling/meta_arch/__pycache__/clip_rcnn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6fc45c29eb8a7419c243dfa963e6c7729af66f14 Binary files /dev/null and b/detectron2/modeling/meta_arch/__pycache__/clip_rcnn.cpython-39.pyc differ diff --git a/detectron2/modeling/meta_arch/__pycache__/panoptic_fpn.cpython-39.pyc b/detectron2/modeling/meta_arch/__pycache__/panoptic_fpn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8ce8e770f29e33e04223c0b94581f21cb590641d Binary files /dev/null and b/detectron2/modeling/meta_arch/__pycache__/panoptic_fpn.cpython-39.pyc differ diff --git a/detectron2/modeling/meta_arch/__pycache__/rcnn.cpython-39.pyc b/detectron2/modeling/meta_arch/__pycache__/rcnn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0b09c1c662b8b8479896188bc4676b4974218623 Binary files /dev/null and b/detectron2/modeling/meta_arch/__pycache__/rcnn.cpython-39.pyc differ diff --git a/detectron2/modeling/meta_arch/__pycache__/retinanet.cpython-39.pyc b/detectron2/modeling/meta_arch/__pycache__/retinanet.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0a762985e6c6081dc84bea1e6e8c6f1f221acd36 Binary files /dev/null and b/detectron2/modeling/meta_arch/__pycache__/retinanet.cpython-39.pyc differ diff --git a/detectron2/modeling/meta_arch/__pycache__/semantic_seg.cpython-39.pyc b/detectron2/modeling/meta_arch/__pycache__/semantic_seg.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..95ce23238827ae9dd2004bb36d4d192e214ba90f Binary files /dev/null and b/detectron2/modeling/meta_arch/__pycache__/semantic_seg.cpython-39.pyc differ diff --git a/detectron2/modeling/meta_arch/build.py b/detectron2/modeling/meta_arch/build.py new file mode 100644 index 0000000000000000000000000000000000000000..3427215746c9a146bd902f22ea9b26d121c36b27 --- /dev/null +++ b/detectron2/modeling/meta_arch/build.py @@ -0,0 +1,25 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import torch + +from detectron2.utils.logger import _log_api_usage +from detectron2.utils.registry import Registry + +META_ARCH_REGISTRY = Registry("META_ARCH") # noqa F401 isort:skip +META_ARCH_REGISTRY.__doc__ = """ +Registry for meta-architectures, i.e. the whole model. + +The registered object will be called with `obj(cfg)` +and expected to return a `nn.Module` object. +""" + + +def build_model(cfg): + """ + Build the whole model architecture, defined by ``cfg.MODEL.META_ARCHITECTURE``. + Note that it does not load any weights from ``cfg``. + """ + meta_arch = cfg.MODEL.META_ARCHITECTURE + model = META_ARCH_REGISTRY.get(meta_arch)(cfg) + model.to(torch.device(cfg.MODEL.DEVICE)) + _log_api_usage("modeling.meta_arch." + meta_arch) + return model diff --git a/detectron2/modeling/meta_arch/clip_rcnn.py b/detectron2/modeling/meta_arch/clip_rcnn.py new file mode 100644 index 0000000000000000000000000000000000000000..f7749ccb26e26d9750367161a9785fa6b9b1171a --- /dev/null +++ b/detectron2/modeling/meta_arch/clip_rcnn.py @@ -0,0 +1,1557 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import numpy as np +from typing import Dict, List, Optional, Tuple +from numpy.lib import pad +import torch +from torch import nn +from torch.nn import functional as F +from random import randint + +from detectron2.config import configurable +from detectron2.data.detection_utils import convert_image_to_rgb +from detectron2.structures import ImageList, Instances, Boxes +from detectron2.utils.events import get_event_storage +from detectron2.utils.logger import log_first_n + +from ..backbone import Backbone, build_backbone, build_text_backbone +from ..postprocessing import detector_postprocess +from ..proposal_generator import build_proposal_generator +from ..roi_heads import build_roi_heads +from .build import META_ARCH_REGISTRY + +from PIL import Image +import torchvision +from torchvision.transforms import Resize, CenterCrop +from detectron2.data.datasets.clip_prompt_utils import get_cls_names, pre_tokenize +import copy +from ..backbone.fpn import build_resnet_fpn_backbone +from ..roi_heads.fast_rcnn import fast_rcnn_inference +from detectron2.layers import ShapeSpec +from ..backbone.clip_backbone import build_clip_language_encoder +from detectron2.utils.comm import gather_tensors, MILCrossEntropy, SoftTargetCrossEntropy + +__all__ = ["CLIPRCNN", "CLIPFastRCNN", "PretrainFastRCNN"] + +@META_ARCH_REGISTRY.register() +class CLIPRCNN(nn.Module): + """ + CLIP in R-CNN format. + It takes the image regions as inputs and classifies each image. + It contains the following two components: + 1. Per-image feature extraction (visual encoder) + 2. Per-image prediction (text-based classifier) + """ + @configurable + def __init__( + self, + *, + clip: Backbone, + offline_backbone: Backbone, + offline_proposal_generator: nn.Module, + roi_heads: nn.Module, + pixel_mean: Tuple[float], + pixel_std: Tuple[float], + input_format: Optional[str] = None, + vis_period: int = 0, + clip_crop_region_type: str = 'GT', + test_score_thresh: float = 0.0001, + test_nms_thresh: float = 0.5, + test_topk_per_image: float = 300, + ): + """ + Args: + backbone: a backbone module, must follow detectron2's backbone interface + proposal_generator: a module that generates proposals using backbone features + roi_heads: a ROI head that performs per-region computation + pixel_mean, pixel_std: list or tuple with #channels element, representing + the per-channel mean and std to be used to normalize the input image + input_format: describe the meaning of channels of input. Needed by visualization + vis_period: the period to run visualization. Set to 0 to disable. + """ + super().__init__() + self.clip_backbone = clip + self.offline_backbone = offline_backbone + self.offline_proposal_generator = offline_proposal_generator + self.roi_heads = roi_heads + + self.input_format = input_format + self.vis_period = vis_period + if vis_period > 0: + assert input_format is not None, "input_format is required for visualization!" + + self.register_buffer("pixel_mean", torch.tensor(pixel_mean).view(-1, 1, 1), False) + self.register_buffer("pixel_std", torch.tensor(pixel_std).view(-1, 1, 1), False) + assert ( + self.pixel_mean.shape == self.pixel_std.shape + ), f"{self.pixel_mean} and {self.pixel_std} have different shapes!" + # Detectron2 default pixel mean and std + self.register_buffer("detectron_pixel_mean", torch.tensor([103.530, 116.280, 123.675]).view(-1, 1, 1), False) + self.register_buffer("detectron_pixel_std", torch.tensor([1.0, 1.0, 1.0]).view(-1, 1, 1), False) + + # CLIP image loading + if np.sum(pixel_mean) < 3.0: # converrt pixel value to range [0.0, 1.0] by dividing 255.0 + assert input_format == 'RGB' + self.div_pixel = True + else: # default setting + self.div_pixel = False + n_px = 224 + self.clip_resize = Resize(n_px, interpolation=Image.BICUBIC) # shorter side becomes n_px + self.clip_center_crop = CenterCrop(n_px) # crop image into n_px * n_px at the center + self.region_crop_scales = (1.0, 1.5) # (1.0, 2.0) # (1.0, 1.2) # (1.0,) # + + # CLIP text prompt loading + print("Working on pre_tokenize...") + cls_names = get_cls_names(filter_novel=False, from_file='/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/trained_models/concept_pool/googlecc_nouns_filtered_100.txt') # filter_novel=True; coco='all', coco='base', coco='target'; from_file: a file path for concept pool + # from_file='/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/trained_models/concept_pool/googlecc_nouns_triplet_parser_filtered_100.txt' + print("Got {} class names: {}\n {} class names in total.".format(len(cls_names), cls_names, len(cls_names))) + input_ids = pre_tokenize(cls_names) + self.num_cls = input_ids.size(0) + self.num_prompt = input_ids.size(1) + self.input_ids_flat = input_ids.view(-1, input_ids.size(2)) # [#cls*#prompts, #context_length] + self.clss_emb_all = None + + # CLIP crop image configs + self.clip_crop_region_type = clip_crop_region_type + self.test_score_thresh = test_score_thresh + self.test_nms_thresh = test_nms_thresh + self.test_topk_per_image = test_topk_per_image + + @classmethod + def from_config(cls, cfg): + if cfg.MODEL.CLIP.CROP_REGION_TYPE == "RPN": + offline_backbone = build_resnet_fpn_backbone(cfg, ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN))) # build_backbone(cfg) + offline_rpn = build_proposal_generator(cfg, offline_backbone.output_shape()) + roi_heads = None # build_roi_heads(cfg, backbone.output_shape()), + elif cfg.MODEL.CLIP.CROP_REGION_TYPE == "GT": + offline_backbone = None + offline_rpn = None + roi_heads = None + clip = build_backbone(cfg) + return { + "clip": clip, + "offline_backbone": offline_backbone, + "offline_proposal_generator": offline_rpn, + "roi_heads": roi_heads, + "input_format": cfg.INPUT.FORMAT, + "vis_period": cfg.VIS_PERIOD, + "pixel_mean": cfg.MODEL.PIXEL_MEAN, + "pixel_std": cfg.MODEL.PIXEL_STD, + "clip_crop_region_type" : cfg.MODEL.CLIP.CROP_REGION_TYPE, + "test_score_thresh" : cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST, + "test_nms_thresh" : cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST, + "test_topk_per_image" : cfg.TEST.DETECTIONS_PER_IMAGE, + } + + @property + def device(self): + return self.pixel_mean.device + + def forward(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Args: + batched_inputs: a list, batched outputs of :class:`DatasetMapper` . + Each item in the list contains the inputs for one image. + For now, each item in the list is a dict that contains: + + * image: Tensor, image in (C, H, W) format. + * instances (optional): groundtruth :class:`Instances` + * proposals (optional): :class:`Instances`, precomputed proposals. + + Other information that's included in the original dicts, such as: + + * "height", "width" (int): the output resolution of the model, used in inference. + See :meth:`postprocess` for details. + + Returns: + list[dict]: + Each dict is the output for one input image. + The dict contains one key "instances" whose value is a :class:`Instances`. + The :class:`Instances` object has the following keys: + "pred_boxes", "pred_classes", "scores", "pred_masks", "pred_keypoints" + """ + if not self.training: + return self.inference(batched_inputs) + # No training mode for this arch + + def inference( + self, + batched_inputs: List[Dict[str, torch.Tensor]], + detected_instances: Optional[List[Instances]] = None, + do_postprocess: bool = True, + ): + """ + Run inference on the given inputs. + + Args: + batched_inputs (list[dict]): same as in :meth:`forward` + detected_instances (None or list[Instances]): if not None, it + contains an `Instances` object per image. The `Instances` + object contains "pred_boxes" and "pred_classes" which are + known boxes in the image. + The inference will then skip the detection of bounding boxes, + and only predict other per-ROI outputs. + do_postprocess (bool): whether to apply post-processing on the outputs. + + Returns: + When do_postprocess=True, same as in :meth:`forward`. + Otherwise, a list[Instances] containing raw network outputs. + """ + assert not self.training + + # get the label prompt, and use CLIP.encode_text() to compute text emb only once + if self.clss_emb_all is None: # compute only once + num_instances = self.input_ids_flat.size(0) + per_split = 1000 + num_splits = num_instances // per_split + input_ids_flat = self.input_ids_flat.to(self.device) + #self.clss_emb_all = torch.ones((1203, 512)).to(self.device) + clss_emb_all = [] + for i in range(num_splits+1): + if i < num_splits: + clss_emb_i = self.clip_backbone.encode_text(input_ids_flat[per_split*i:per_split*(i+1)]) # per_split x D + else: + clss_emb_i = self.clip_backbone.encode_text(input_ids_flat[per_split*i:]) # per_split x D + # clss_emb_i = clip_model.encode_label(torch.arange(0, 1000).view(-1, 1).long().to(device)) # per_split x D + clss_emb_all.append(clss_emb_i) + self.clss_emb_all = torch.cat(clss_emb_all, 0).view(self.num_cls, self.num_prompt, -1) # [#cls, #prompts, D] + self.clss_emb_all = self.clss_emb_all.mean(1) # ensemble different prompts for each class + # torch.save(self.clss_emb_all.cpu(), "/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/trained_models/lvis_cls_emb/coco_17_target_cls_emb_notnorm_rn50x4.pth") + self.clss_emb_all = F.normalize(self.clss_emb_all, p=2.0, dim=1) # [#cls, emb_dim] + else: + assert self.clss_emb_all.device == self.device + + # get the region proposals, from the backbone & RPN of standard Mask-RCNN, trained on base classes + if self.clip_crop_region_type == "GT": + proposals = None + elif self.clip_crop_region_type == "RPN": + images = self.preprocess_image(batched_inputs) + features = self.offline_backbone(images.tensor) + if detected_instances is None: + if self.offline_proposal_generator is not None: + proposals, _ = self.offline_proposal_generator(images, features, None) + + # crop image regions, and use CLIP.encode_image() to get the visual feature + images, bbs, num_bbs = self.preprocess_image_crop(batched_inputs, rpn_proposals=proposals) + img_emb = self.clip_backbone.encode_image(images.tensor) + img_emb = img_emb.view(-1, len(self.region_crop_scales), img_emb.size(1)) + img_emb = torch.sum(img_emb, dim=1) # ensemble different scales for each region + img_emb = F.normalize(img_emb, p=2.0, dim=1) + + # cosine similarity as logits + all_scores = torch.mm(img_emb, self.clss_emb_all.T) + all_scores = F.softmax(all_scores, dim=-1) + scores, pred_cls = torch.max(all_scores, dim=-1) # Note: [0, #cls-1] representing the categories. The value #cls represents "background". + + # convert model outputs into regular output result format + scores_per_img = scores.split(num_bbs) + pred_cls_per_img = pred_cls.split(num_bbs) + all_scores_per_img = all_scores.split(num_bbs) + + # per-class NMS + if self.clip_crop_region_type == "GT": + image_shapes = [x['instances']._image_size for x in batched_inputs] + bbs = [bb.to(self.device) for bb in bbs] + pred_instances, _ = fast_rcnn_inference(bbs, all_scores_per_img, image_shapes, \ + self.test_score_thresh, self.test_nms_thresh, self.test_topk_per_image) + results = pred_instances + + # results = [] + # for r_i, (b_input, bb, sc, prd) in enumerate(zip(batched_inputs, bbs, scores_per_img, pred_cls_per_img)): + # this_result = copy.deepcopy(b_input["instances"]) # Instance + # if self.clip_crop_region_type == "GT": + # result_boxes = this_result._fields['gt_boxes'].to(self.device) + # elif self.clip_crop_region_type == "RPN": # directly use RPN boxes without per-class NMS + # result_boxes = bb # result_boxes = Boxes(bb) + # this_result._fields = {'pred_boxes': result_boxes, 'scores': sc, 'pred_classes': prd} + # results.append(this_result) + + # sanity check: GT boxes + GT classes + # results = [] + # for b_input in batched_inputs: + # this_result = copy.deepcopy(b_input["instances"]) # Instance + # gt_boxes = this_result._fields['gt_boxes'].to(self.device) + # gt_cls = this_result._fields['gt_classes'].to(self.device) + # this_result._fields = {'pred_boxes': gt_boxes, 'scores': torch.ones(gt_cls.size(0)).to(self.device), 'pred_classes': gt_cls} + # #this_result._fields = {'pred_boxes': gt_boxes, 'scores': sc, 'pred_classes': prd} + # results.append(this_result) + elif self.clip_crop_region_type == "RPN": + image_shapes = [x.image_size for x in proposals] + pred_instances, _ = fast_rcnn_inference(bbs, all_scores_per_img, image_shapes, \ + self.test_score_thresh, self.test_nms_thresh, self.test_topk_per_image) + results = pred_instances + + if do_postprocess: + assert not torch.jit.is_scripting(), "Scripting is not supported for postprocess." + return CLIPRCNN._postprocess(results, batched_inputs) + else: + return results + + def preprocess_image(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Normalize, pad and batch the input images. Use detectron2 default processing (pixel mean & std). + Note: Due to FPN size_divisibility, images are padded by right/bottom border. So FPN is consistent with C4 and GT boxes. + """ + images = [x["image"].to(self.device) for x in batched_inputs] + images = [(x - self.detectron_pixel_mean) / self.detectron_pixel_std for x in images] + images = ImageList.from_tensors(images, self.offline_backbone.size_divisibility) + return images + + def preprocess_image_crop(self, batched_inputs: List[Dict[str, torch.Tensor]], rpn_proposals=None, max_num_rpn=1000): + """ + Crop image regions based on GT or RPN boxes with different scales. + Then apply CLIP tranformation: resizing / cropping the regions into square shape (224 * 224). + Followed by the default preprocessing in Detectron2 as follows. + Normalize, pad and batch the input images. + """ + def clip_crop_region(image, box, scales=(1.0, 1.5)): + """Crop image regions based on given boxes. Return different scales of region crops. (3 hrs)""" + img_h, img_w = image.size(1), image.size(2) + x1, y1, x2, y2 = list(box) + assert x1 < x2 and y1 < y2 and x2 < (img_w + 1) and y2 < (img_h + 1) + x_center = (x1 + x2) / 2.0 + y_center = (y1 + y2) / 2.0 + half_w = x_center - x1 + half_h = y_center - y1 + regions = [] + for scale in scales: # get region coordinates + r_y1 = int(max(0, (y_center - half_h * scale).item())) + r_y2 = int(min(img_h, (y_center + half_h * scale).item())) + r_x1 = int(max(0, (x_center - half_w * scale).item())) + r_x2 = int(min(img_w, (x_center + half_w * scale).item())) + # sanity check + if r_y2 - r_y1 <= 1: + r_y2 = int(min(img_h, r_y2 + 2)) + if r_y2 - r_y1 <= 1: + r_y1 = int(max(0, r_y1 - 2)) + if r_x2 - r_x1 <= 1: + r_x2 = int(min(img_w, r_x2 + 2)) + if r_x2 - r_x1 <= 1: + r_x1 = int(max(0, r_x1 - 2)) + regions.append(image[:, r_y1:r_y2, r_x1:r_x2]) + return regions + + def clip_square_crop(image, box, scales=(1.0,)): + """Crop image regions based on given boxes. Ensure square region as much as possible. (1.75 hrs)""" + img_h, img_w = image.size(1), image.size(2) + x1, y1, x2, y2 = list(box) + assert x1 < x2 and y1 < y2 and x2 < (img_w + 1) and y2 < (img_h + 1) + x_center = (x1 + x2) / 2.0 + y_center = (y1 + y2) / 2.0 + half_w = x_center - x1 + half_h = y_center - y1 + square_side = max(half_w, half_h) + half_w = square_side + half_h = square_side + regions = [] + for scale in scales: # get region coordinates + if square_side * square_side < 2500: # crop larger context area for tiny objects + scale = 1.5 if scale == 1.0 else 4.0 + # elif square_side * square_side > 90000: # crop exact area for large objects + # scale = 1.0 if scale == 1.0 else 1.1 + r_y1 = int(max(0, (y_center - half_h * scale).item())) + r_y2 = int(min(img_h, (y_center + half_h * scale).item())) + r_x1 = int(max(0, (x_center - half_w * scale).item())) + r_x2 = int(min(img_w, (x_center + half_w * scale).item())) + # sanity check + if r_y2 - r_y1 <= 1: + r_y2 = int(min(img_h, r_y2 + 2)) + if r_y2 - r_y1 <= 1: + r_y1 = int(max(0, r_y1 - 2)) + if r_x2 - r_x1 <= 1: + r_x2 = int(min(img_w, r_x2 + 2)) + if r_x2 - r_x1 <= 1: + r_x1 = int(max(0, r_x1 - 2)) + #regions.append(image[:, r_y1:r_y2, r_x1:r_x2]) + # if the cropped image isn't square (due to image boundaries), pad the cropped region + crop_image = image[:, r_y1:r_y2, r_x1:r_x2] + r_h, r_w = crop_image.size(1), crop_image.size(2) + pad_image = torch.zeros((3, int(2 * half_h.item() * scale) + 4 , int(2 * half_w.item() * scale) + 4)) #.fill_(torch.mean(crop_image.float())) + p_h, p_w = pad_image.size(1), pad_image.size(2) + pad_image[:, int(((p_h - r_h) / 2)):int(((p_h - r_h) / 2 + r_h)), int(((p_w - r_w) / 2)):int(((p_w - r_w) / 2 + r_w))] = crop_image + regions.append(pad_image.type(torch.uint8)) + return regions + + def vis_crop(f_n, images): + """visualize the crop regions to diagnose the accuracy.""" + if f_n not in ['datasets/coco/train2017/000000008691.jpg']: + for p_i, pad_image in enumerate(images): + to_save = pad_image.permute(1, 2, 0).numpy() + to_save = Image.fromarray(np.array(to_save, np.uint8)) + to_save.save("output/regions/" + f_n.split("/")[-1].split(".")[0] + "-{}.png".format(p_i)) + pass + + # crop image region + images = [] + bbs = [] + num_bbs = [] + for img_i, b_input in enumerate(batched_inputs): + this_img = b_input["image"] + if self.clip_crop_region_type == "GT": + this_boxes = b_input["instances"]._fields['gt_boxes'].tensor # variant #bbox (eg, max 759), might lead to OOM + elif self.clip_crop_region_type == "RPN": + this_boxes = rpn_proposals[img_i]._fields['proposal_boxes'].tensor[:max_num_rpn] + + bbs.append(this_boxes) + num_bbs.append(this_boxes.size(0)) + for this_box in this_boxes: + #images.extend(clip_crop_region(this_img, this_box, self.region_crop_scales)) + images.extend(clip_square_crop(this_img, this_box, self.region_crop_scales)) + #vis_crop(batched_inputs[0]['file_name'], images) + images = [self.clip_resize(x) for x in images] + images = [self.clip_center_crop(x) for x in images] + images = [x.to(self.device) for x in images] + if self.div_pixel: + images = [((x / 255.0) - self.pixel_mean) / self.pixel_std for x in images] + else: + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + images = ImageList.from_tensors(images, self.clip_backbone.size_divisibility) # batch images into single tensor by padding to same size + return images, bbs, num_bbs + + @staticmethod + def _postprocess(instances, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Rescale the output instances to the target size. + """ + # note: private function; subject to changes + processed_results = [] + for results_per_image, input_per_image in zip( + instances, batched_inputs): + height = input_per_image["height"] # original image size, before resizing + width = input_per_image["width"] # original image size, before resizing + r = detector_postprocess(results_per_image, height, width) + processed_results.append({"instances": r}) + return processed_results + + def inference_on_cifar(self, pseudo_input): + """ Evaluate recoginition accuracy on CIFAR-10 for sanity check """ + # get the label prompt, and use CLIP.encode_text() to compute text emb only once + cifar_cls_names = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck'] + input_ids = pre_tokenize(cifar_cls_names) + num_cls = input_ids.size(0) + input_ids_flat = input_ids.view(-1, input_ids.size(2)) + input_ids_flat = input_ids_flat.to(self.device) + + clss_emb_all = self.clip_backbone.encode_text(input_ids_flat) + clss_emb_all = clss_emb_all.view(num_cls, self.num_prompt, -1) + clss_emb_all = clss_emb_all.mean(1) + clss_emb_all = F.normalize(clss_emb_all, p=2.0, dim=1) # [#cls, emb_dim] + + # dataset loads images and labels + testset = torchvision.datasets.CIFAR10(root='./datasets', train=False, + download=False, transform=None) + # testloader = torch.utils.data.DataLoader(testset, batch_size=4, + # shuffle=False, num_workers=0) + + # inference on each image and calculate accuracy + correct = 0 + wrong = 0 + for idx, inputs in enumerate(testset): + if idx % 1000 == 0: + print(idx) + # preprocess images + raw_image, label = inputs + image = np.array(raw_image) # [h, w, 3] + image = torch.from_numpy(image) + image = image.permute(2, 0, 1) # [3, h, w] + images = [image] + images = [self.clip_resize(x) for x in images] + images = [self.clip_center_crop(x) for x in images] + images = [x.to(self.device) for x in images] + if self.div_pixel: + images = [((x / 255.0) - self.pixel_mean) / self.pixel_std for x in images] + else: + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + + # get image embedding + img_emb = self.clip_backbone.encode_image(images[0].unsqueeze(0)) + img_emb = img_emb.view(-1, 1, img_emb.size(1)) + img_emb = torch.sum(img_emb, dim=1) # ensemble different scales for each region + img_emb = F.normalize(img_emb, p=2.0, dim=1) + + # cosine similarity as logits + all_scores = torch.mm(img_emb, clss_emb_all.T) + scores, pred_cls = torch.max(all_scores, dim=1) # Note: [0, #cls-1] representing the categories. The value #cls represents "background". + pred_cls = pred_cls.item() + if pred_cls == label: + correct += 1 + else: + wrong += 1 + + print("\n\nGot correct {} and wrong {}. Accuracy is {} / {} = {}\n\n".format(correct,wrong,correct,correct+wrong,correct/(correct+wrong))) + return + +@META_ARCH_REGISTRY.register() +class CLIPFastRCNN(nn.Module): + """ + CLIP in Fast R-CNN format, where the cropping is conducted on feature maps instead of raw images. + It contains the following two components: + 1. Localization modules: pretrained backbone+RPN or equivalent modules and is able to output object proposals + 2. Recognition branch: initialized by CLIP and is able to recognize zero-shot regions + """ + @configurable + def __init__( + self, + *, + offline_backbone: Backbone, + backbone: Backbone, + backbone_type: str = "resnet", + text_backbone: Backbone, + offline_proposal_generator: nn.Module, + roi_heads: nn.Module, + pixel_mean: Tuple[float], + pixel_std: Tuple[float], + input_format: Optional[str] = None, + vis_period: int = 0, + clip_crop_region_type: str = 'GT', + use_clip_c4: False, + use_clip_attpool: False, + offline_input_format: Optional[str] = None, + offline_pixel_mean: Tuple[float], + offline_pixel_std: Tuple[float], + ): + """ + Args: + backbone: a backbone module, must follow detectron2's backbone interface + proposal_generator: a module that generates proposals using backbone features + roi_heads: a ROI head that performs per-region computation + pixel_mean, pixel_std: list or tuple with #channels element, representing + the per-channel mean and std to be used to normalize the input image + input_format: describe the meaning of channels of input. Needed by visualization + vis_period: the period to run visualization. Set to 0 to disable. + """ + super().__init__() + self.offline_backbone = offline_backbone + self.backbone = backbone + self.backbone_type = backbone_type + self.offline_proposal_generator = offline_proposal_generator + self.roi_heads = roi_heads + self.lang_encoder = text_backbone + + self.input_format = input_format + self.vis_period = vis_period + if vis_period > 0: + assert input_format is not None, "input_format is required for visualization!" + + self.register_buffer("pixel_mean", torch.tensor(pixel_mean).view(-1, 1, 1), False) + self.register_buffer("pixel_std", torch.tensor(pixel_std).view(-1, 1, 1), False) + assert ( + self.pixel_mean.shape == self.pixel_std.shape + ), f"{self.pixel_mean} and {self.pixel_std} have different shapes!" + if np.sum(pixel_mean) < 3.0: # converrt pixel value to range [0.0, 1.0] by dividing 255.0 + assert input_format == 'RGB' + self.div_pixel = True + else: # default setting + self.div_pixel = False + + # input format, pixel mean and std for offline modules + if offline_input_format and offline_pixel_mean and offline_pixel_std: + self.offline_input_format = offline_input_format + self.register_buffer("offline_pixel_mean", torch.tensor(offline_pixel_mean).view(-1, 1, 1), False) + self.register_buffer("offline_pixel_std", torch.tensor(offline_pixel_std).view(-1, 1, 1), False) + if np.sum(offline_pixel_mean) < 3.0: # converrt pixel value to range [0.0, 1.0] by dividing 255.0 + assert offline_input_format == 'RGB' + self.offline_div_pixel = True + else: # default setting + self.offline_div_pixel = False + + self.clip_crop_region_type = clip_crop_region_type + self.use_clip_c4 = use_clip_c4 # if True, use C4 mode where roi_head uses the last resnet layer from backbone + self.use_clip_attpool = use_clip_attpool # if True (C4+text_emb_as_classifier), use att_pool to replace default mean pool + + + @classmethod + def from_config(cls, cfg): + if cfg.MODEL.CLIP.CROP_REGION_TYPE == "RPN": # create isolated backbone & RPN + # create offline cfg for the pretrained backbone & RPN + from detectron2.config import get_cfg + offline_cfg = get_cfg() + offline_cfg.merge_from_file(cfg.MODEL.CLIP.OFFLINE_RPN_CONFIG) + if cfg.MODEL.CLIP.OFFLINE_RPN_LSJ_PRETRAINED: # large-scale jittering (LSJ) pretrained RPN + offline_cfg.MODEL.BACKBONE.FREEZE_AT = 0 # make all fronzon layers to "SyncBN" + offline_cfg.MODEL.RESNETS.NORM = "SyncBN" # 5 resnet layers + offline_cfg.MODEL.FPN.NORM = "SyncBN" # fpn layers + offline_cfg.MODEL.RPN.CONV_DIMS = [-1, -1] # rpn layers + if cfg.MODEL.CLIP.OFFLINE_RPN_NMS_THRESH: + offline_cfg.MODEL.RPN.NMS_THRESH = cfg.MODEL.CLIP.OFFLINE_RPN_NMS_THRESH # 0.9 + + # create offline backbone and RPN + offline_backbone = build_backbone(offline_cfg) # build_resnet_fpn_backbone(cfg, ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN))) + offline_rpn = build_proposal_generator(offline_cfg, offline_backbone.output_shape()) + # convert to evaluation mode + for p in offline_backbone.parameters(): p.requires_grad = False + for p in offline_rpn.parameters(): p.requires_grad = False + offline_backbone.eval() + offline_rpn.eval() + elif cfg.MODEL.CLIP.CROP_REGION_TYPE == "GT": + offline_backbone = None + offline_rpn = None + offline_cfg = None + + backbone = build_backbone(cfg) + text_backbone = build_clip_language_encoder(cfg) + + backbone_type = "swin" if "swin" in cfg.MODEL.BACKBONE.NAME else "resnet" + + if backbone_type == "swin": + roi_heads = build_roi_heads(cfg, backbone.image_encoder.output_shape()) + else: + roi_heads = build_roi_heads(cfg, backbone.output_shape()) + + return { + "offline_backbone": offline_backbone, + "offline_proposal_generator": offline_rpn, + "backbone": backbone, + "backbone_type": backbone_type, + "text_backbone": text_backbone, + "roi_heads": roi_heads, + "input_format": cfg.INPUT.FORMAT, + "vis_period": cfg.VIS_PERIOD, + "pixel_mean": cfg.MODEL.PIXEL_MEAN, + "pixel_std": cfg.MODEL.PIXEL_STD, + "clip_crop_region_type" : cfg.MODEL.CLIP.CROP_REGION_TYPE, + "use_clip_c4": 'FPN' not in cfg.MODEL.BACKBONE.NAME, + "use_clip_attpool": cfg.MODEL.ROI_HEADS.NAME in ['CLIPRes5ROIHeads', 'CLIPStandardROIHeads'] and cfg.MODEL.CLIP.USE_TEXT_EMB_CLASSIFIER, + "offline_input_format": offline_cfg.INPUT.FORMAT if offline_cfg else None, + "offline_pixel_mean": offline_cfg.MODEL.PIXEL_MEAN if offline_cfg else None, + "offline_pixel_std": offline_cfg.MODEL.PIXEL_STD if offline_cfg else None, + } + + @property + def device(self): + return self.pixel_mean.device + + def forward(self, queries, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Args: + batched_inputs: a list, batched outputs of :class:`DatasetMapper` . + Each item in the list contains the inputs for one image. + For now, each item in the list is a dict that contains: + + * image: Tensor, image in (C, H, W) format. + * instances (optional): groundtruth :class:`Instances` + * proposals (optional): :class:`Instances`, precomputed proposals. + + Other information that's included in the original dicts, such as: + + * "height", "width" (int): the output resolution of the model, used in inference. + See :meth:`postprocess` for details. + + Returns: + list[dict]: + Each dict is the output for one input image. + The dict contains one key "instances" whose value is a :class:`Instances`. + The :class:`Instances` object has the following keys: + "pred_boxes", "pred_classes", "scores", "pred_masks", "pred_keypoints" + """ + if not self.training: + return self.inference(queries, batched_inputs) + if "instances" in batched_inputs[0]: + gt_instances = [x["instances"].to(self.device) for x in batched_inputs] + else: + gt_instances = None + + # localization branch: offline modules to get the region proposals + with torch.no_grad(): + if self.clip_crop_region_type == "GT": # from ground-truth + proposals = [] + for r_i, b_input in enumerate(batched_inputs): + this_gt = copy.deepcopy(b_input["instances"]) # Instance + gt_boxes = this_gt._fields['gt_boxes'].to(self.device) + this_gt._fields = {'proposal_boxes': gt_boxes, 'objectness_logits': torch.ones(gt_boxes.tensor.size(0)).to(self.device)} + proposals.append(this_gt) + elif self.clip_crop_region_type == "RPN": # from the backbone & RPN of standard Mask-RCNN, trained on base classes + if self.offline_backbone.training or self.offline_proposal_generator.training: # was set to True in training script + self.offline_backbone.eval() + self.offline_proposal_generator.eval() + images = self.offline_preprocess_image(batched_inputs) + features = self.offline_backbone(images.tensor) + if self.offline_proposal_generator is not None: + proposals, _ = self.offline_proposal_generator(images, features, None) + + # recognition branch: get 2D feature maps using the backbone of recognition branch + images = self.preprocess_image(batched_inputs) + features = self.backbone(images.tensor) + + if self.backbone_type == "resnet": + head = self.backbone.layer4 + elif self.backbone_type == "swin": + head = self.backbone.layers[-1] + + # Given the proposals, crop region features from 2D image features and classify the regions + if self.use_clip_c4: # use C4 + resnet weights from CLIP + if self.use_clip_attpool: # use att_pool from CLIP to match dimension + _, detector_losses = self.roi_heads(images, features, proposals, gt_instances, res5=head, attnpool=self.backbone.attnpool) + else: # use default mean pool + _, detector_losses = self.roi_heads(images, features, proposals, gt_instances, res5=head) + else: # default setting + if self.use_clip_attpool: # use att_pool from CLIP to match dimension + _, detector_losses = self.roi_heads(images, features, proposals, gt_instances, attnpool=self.backbone.bottom_up.attnpool) + else: # use default mean pool + _, detector_losses = self.roi_heads(images, features, proposals, gt_instances) + if self.vis_period > 0: + storage = get_event_storage() + if storage.iter % self.vis_period == 0: + self.visualize_training(batched_inputs, proposals) + #visualize_proposals(batched_inputs, proposals, self.input_format) + + losses = {} + losses.update(detector_losses) + return losses + + def inference( + self, + queries, + batched_inputs: List[Dict[str, torch.Tensor]], + detected_instances: Optional[List[Instances]] = None, + do_postprocess: bool = True, + ): + """ + Run inference on the given inputs. + + Args: + batched_inputs (list[dict]): same as in :meth:`forward` + detected_instances (None or list[Instances]): if not None, it + contains an `Instances` object per image. The `Instances` + object contains "pred_boxes" and "pred_classes" which are + known boxes in the image. + The inference will then skip the detection of bounding boxes, + and only predict other per-ROI outputs. + do_postprocess (bool): whether to apply post-processing on the outputs. + + Returns: + When do_postprocess=True, same as in :meth:`forward`. + Otherwise, a list[Instances] containing raw network outputs. + """ + assert not self.training + + # localization branch: offline modules to get the region proposals + if self.clip_crop_region_type == "GT": # from ground-truth + proposals = [] + for r_i, b_input in enumerate(batched_inputs): + this_gt = copy.deepcopy(b_input["instances"]) # Instance + gt_boxes = this_gt._fields['gt_boxes'].to(self.device) + this_gt._fields = {'proposal_boxes': gt_boxes} #, 'objectness_logits': None} + proposals.append(this_gt) + elif self.clip_crop_region_type == "RPN": # from the backbone & RPN of standard Mask-RCNN, trained on base classes + images = self.offline_preprocess_image(batched_inputs) + features = self.offline_backbone(images.tensor) + if detected_instances is None: + if self.offline_proposal_generator is not None: + proposals, _ = self.offline_proposal_generator(images, features, None) + + # recognition branch: get 2D feature maps using the backbone of recognition branch + print(batched_inputs[0]['image'][0][:10, :10]) + print(batched_inputs[0]['image'].shape) + images = self.preprocess_image(batched_inputs) + + if self.backbone_type == "swin": + features = self.backbone.encode_image(images.tensor) + text_features = self.backbone.encode_text(queries) + else: + features = self.backbone(images.tensor) + token_embeddings = pre_tokenize([queries])[:, 0].to(images.tensor.device) + text_features = self.lang_encoder.encode_text(token_embeddings) + + if self.backbone_type == "resnet": + head = self.backbone.layer4 + downsampler = None + norm = None + vision_projection = None + elif self.backbone_type == "swin": + downsampler = self.backbone.image_encoder.layers[-2].downsample + head = self.backbone.image_encoder.layers[-1] + norm = self.backbone.image_encoder.norm + vision_projection = self.backbone.image_projection + + # Given the proposals, crop region features from 2D image features and classify the regions + if self.use_clip_c4: # use C4 + resnet weights from CLIP + if self.use_clip_attpool: # use att_pool from CLIP to match dimension + results, _ = self.roi_heads(images, features, proposals, text_features, None, + res5=head, ds=downsampler, norm=norm, vision_projection=vision_projection, attnpool=self.backbone.attnpool) + else: # use default mean pool + results, _ = self.roi_heads(images, features, proposals, text_features, None, + res5=head, ds=downsampler, norm=norm, vision_projection=vision_projection) + else: # default setting + if self.use_clip_attpool: # use att_pool from CLIP to match dimension + results, _ = self.roi_heads(images, features, proposals, text_features, None, + attnpool=self.backbone.bottom_up.attnpool) + else: + results, _ = self.roi_heads(images, features, proposals, text_features, None) + + visualize_proposals(batched_inputs, proposals, self.input_format) + vis = visualize_results(batched_inputs, results, self.input_format) + return vis + + def offline_preprocess_image(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Normalize, pad and batch the input images. Use detectron2 default processing (pixel mean & std). + Note: Due to FPN size_divisibility, images are padded by right/bottom border. So FPN is consistent with C4 and GT boxes. + """ + images = [x["image"].to(self.device) for x in batched_inputs] + if (self.input_format == 'RGB' and self.offline_input_format == 'BGR') or \ + (self.input_format == 'BGR' and self.offline_input_format == 'RGB'): # the input image follows the main config format ('RGB' or 'BGR') + images = [x[[2,1,0],:,:] for x in images] + if self.offline_div_pixel: + images = [((x / 255.0) - self.offline_pixel_mean) / self.offline_pixel_std for x in images] + else: + images = [(x - self.offline_pixel_mean) / self.offline_pixel_std for x in images] + images = ImageList.from_tensors(images, self.offline_backbone.size_divisibility) + return images + + def preprocess_image(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Normalize, pad and batch the input images. Use CLIP default processing (pixel mean & std). + Note: Due to FPN size_divisibility, images are padded by right/bottom border. So FPN is consistent with C4 and GT boxes. + """ + images = [x["image"].to(self.device) for x in batched_inputs] + if self.div_pixel: + images = [((x / 255.0) - self.pixel_mean) / self.pixel_std for x in images] + else: + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + images = ImageList.from_tensors(images, self.backbone.size_divisibility) + return images + + @staticmethod + def _postprocess(instances, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Rescale the output instances to the target size. + """ + # note: private function; subject to changes + processed_results = [] + for results_per_image, input_per_image in zip( + instances, batched_inputs): + height = input_per_image["height"] # original image size, before resizing + width = input_per_image["width"] # original image size, before resizing + r = detector_postprocess(results_per_image, height, width) + processed_results.append({"instances": r}) + return processed_results + +@META_ARCH_REGISTRY.register() +class PretrainFastRCNN(nn.Module): + """ + Open-vocabulary region representation via vision-language pretraining from image-text pairs + 1. image-text level matching: weakly supervised grounding task with contrastive learning based on region-token representation + 2. region-token level matching: use pseudo text to train model, provided by teacher model + """ + @configurable + def __init__( + self, + *, + offline_backbone: Backbone, + backbone: Backbone, + offline_proposal_generator: nn.Module, + roi_heads: nn.Module, + teacher_backbone: nn.Module, + teacher_roi_heads: nn.Module, + pixel_mean: Tuple[float], + pixel_std: Tuple[float], + input_format: Optional[str] = None, + vis_period: int = 0, + clip_crop_region_type: str = 'GT', + use_clip_c4: False, + use_clip_attpool: False, + offline_input_format: Optional[str] = None, + offline_pixel_mean: Tuple[float], + offline_pixel_std: Tuple[float], + language_encoder: nn.Module, + matching_temp: None, + num_regions_per_img: int = 0, + img_txt_level: None, + gather_gpus: False, + grid_regions: False, + concept_emb: None, + ): + """ + Args: + backbone: a backbone module, must follow detectron2's backbone interface + proposal_generator: a module that generates proposals using backbone features + roi_heads: a ROI head that performs per-region computation + pixel_mean, pixel_std: list or tuple with #channels element, representing + the per-channel mean and std to be used to normalize the input image + input_format: describe the meaning of channels of input. Needed by visualization + vis_period: the period to run visualization. Set to 0 to disable. + """ + super().__init__() + self.offline_backbone = offline_backbone + self.backbone = backbone + self.offline_proposal_generator = offline_proposal_generator + self.roi_heads = roi_heads + + self.input_format = input_format + self.vis_period = vis_period + if vis_period > 0: + assert input_format is not None, "input_format is required for visualization!" + + self.register_buffer("pixel_mean", torch.tensor(pixel_mean).view(-1, 1, 1), False) + self.register_buffer("pixel_std", torch.tensor(pixel_std).view(-1, 1, 1), False) + assert ( + self.pixel_mean.shape == self.pixel_std.shape + ), f"{self.pixel_mean} and {self.pixel_std} have different shapes!" + if np.sum(pixel_mean) < 3.0: # converrt pixel value to range [0.0, 1.0] by dividing 255.0 + assert input_format == 'RGB' + self.div_pixel = True + else: # default setting + self.div_pixel = False + + # input format, pixel mean and std for offline modules + if offline_input_format and offline_pixel_mean and offline_pixel_std: + self.offline_input_format = offline_input_format + self.register_buffer("offline_pixel_mean", torch.tensor(offline_pixel_mean).view(-1, 1, 1), False) + self.register_buffer("offline_pixel_std", torch.tensor(offline_pixel_std).view(-1, 1, 1), False) + if np.sum(offline_pixel_mean) < 3.0: # converrt pixel value to range [0.0, 1.0] by dividing 255.0 + assert offline_input_format == 'RGB' + self.offline_div_pixel = True + else: # default setting + self.offline_div_pixel = False + + self.clip_crop_region_type = clip_crop_region_type + self.use_clip_c4 = use_clip_c4 # if True, use C4 mode where roi_head uses the last resnet layer from backbone + self.use_clip_attpool = use_clip_attpool # if True (C4+text_emb_as_classifier), use att_pool to replace default mean pool + + # image-text level pretraining + self.img_txt_level = img_txt_level[0] + self.only_eot = img_txt_level[1] + if self.img_txt_level: + self.lang_encoder = language_encoder + for p in self.lang_encoder.parameters(): # freeze language encoder + p.requires_grad = False + if matching_temp > 0.0: # fixed temp + self.matching_temp = matching_temp + else: # leanable temp + self.matching_temp = nn.Parameter(torch.ones([]) * 4.6052) # nn.Parameter(torch.ones([]) * np.log(1 / 0.07)) + self.context_length = 77 # defined in clip_img_txt_pair_tsv class + self.num_regions_per_img = num_regions_per_img + self.gather_gpus = gather_gpus + self.grid_regions = grid_regions + + # region-token level pretraining + if concept_emb[0]: + self.register_buffer("concept_emb", torch.load(concept_emb[0]), False) # [#concepts, 1024] + self.concept_thres = concept_emb[1] + self.teacher_backbone = teacher_backbone # None + # when resume, create teacher model in advance to load ckpt + # self.teacher_backbone = copy.deepcopy(self.backbone) + # # # oai_clip = torch.load("/mnt/output_storage/trained_models/oai_clip_weights/RN50_OAI_CLIP.pth") #("/home/v-yiwuzhong/projects/azureblobs/vyiwuzhong_phillytools/trained_models/oai_clip_weights/RN50_OAI_CLIP.pth") + # # # oai_clip_visual = {} + # # # for key in oai_clip['model']: + # # # if 'visual' in key and 'num_batches_tracked' not in key: + # # # oai_clip_visual[key.replace('visual.','')] = oai_clip['model'][key] + # # # self.teacher_backbone.load_state_dict(oai_clip_visual) + for p in self.teacher_backbone.parameters(): # freeze visual encoder of teacher model + p.requires_grad = False + if concept_emb[2] is None: # teacher model uses the same concept embedding as student model + self.register_buffer("teacher_concept_emb", torch.load(concept_emb[0]), False) + else: # teacher model uses a seperate concept embedding + self.register_buffer("teacher_concept_emb", torch.load(concept_emb[2]), False) + self.teacher_roi_heads = teacher_roi_heads + else: + self.concept_emb = None + + @classmethod + def from_config(cls, cfg): + if cfg.MODEL.CLIP.CROP_REGION_TYPE == "RPN": # create isolated backbone & RPN + # create offline cfg for the pretrained backbone & RPN + from detectron2.config import get_cfg + offline_cfg = get_cfg() + offline_cfg.merge_from_file(cfg.MODEL.CLIP.OFFLINE_RPN_CONFIG) + if cfg.MODEL.CLIP.OFFLINE_RPN_LSJ_PRETRAINED: # large-scale jittering (LSJ) pretrained RPN + offline_cfg.MODEL.BACKBONE.FREEZE_AT = 0 # make all fronzon layers to "SyncBN" + offline_cfg.MODEL.RESNETS.NORM = "SyncBN" # 5 resnet layers + offline_cfg.MODEL.FPN.NORM = "SyncBN" # fpn layers + offline_cfg.MODEL.RPN.CONV_DIMS = [-1, -1] # rpn layers + if cfg.MODEL.CLIP.PRETRAIN_RPN_REGIONS: + offline_cfg.MODEL.RPN.POST_NMS_TOPK_TEST = cfg.MODEL.CLIP.PRETRAIN_RPN_REGIONS + if cfg.MODEL.CLIP.OFFLINE_RPN_NMS_THRESH: + offline_cfg.MODEL.RPN.NMS_THRESH = cfg.MODEL.CLIP.OFFLINE_RPN_NMS_THRESH # 0.9 + # offline_cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST = 0.6 + # print("\n\n Set offline RPN.NMS_THRESH to {} and ROI_HEADS.NMS_THRESH_TEST to {}.\n\n".format(offline_cfg.MODEL.RPN.NMS_THRESH, offline_cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST)) + # create offline backbone and RPN + offline_backbone = build_backbone(offline_cfg) # build_resnet_fpn_backbone(cfg, ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN))) + offline_rpn = build_proposal_generator(offline_cfg, offline_backbone.output_shape()) + # convert to evaluation mode + for p in offline_backbone.parameters(): p.requires_grad = False + for p in offline_rpn.parameters(): p.requires_grad = False + offline_backbone.eval() + offline_rpn.eval() + elif cfg.MODEL.CLIP.CROP_REGION_TYPE in ["GLOBAL", "GRID", "RANDOM"]: + offline_backbone = None + offline_rpn = None + offline_cfg = None + # visual encoder and roi_heads of student model + backbone = build_backbone(cfg) + + if "swin" in cfg.MODEL.BACKBONE.NAME: + roi_heads = build_roi_heads(cfg, backbone.image_encoder.output_shape()) + else: + roi_heads = build_roi_heads(cfg, backbone.output_shape()) + + # language encoder of student model + language_encoder = build_clip_language_encoder(cfg) + # visual encoder of teacher model + teacher_cfg = copy.deepcopy(cfg) + teacher_cfg.defrost() + teacher_cfg.MODEL.RESNETS.DEPTH = teacher_cfg.MODEL.CLIP.TEACHER_RESNETS_DEPTH + teacher_backbone = build_backbone(teacher_cfg) + teacher_cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION = teacher_cfg.MODEL.CLIP.TEACHER_POOLER_RESOLUTION + teacher_roi_heads = build_roi_heads(teacher_cfg, teacher_backbone.output_shape()) + return { + "offline_backbone": offline_backbone, + "offline_proposal_generator": offline_rpn, + "backbone": backbone, + "roi_heads": roi_heads, + "teacher_backbone": teacher_backbone, + "teacher_roi_heads": teacher_roi_heads, + "input_format": cfg.INPUT.FORMAT, + "vis_period": cfg.VIS_PERIOD, + "pixel_mean": cfg.MODEL.PIXEL_MEAN, + "pixel_std": cfg.MODEL.PIXEL_STD, + "clip_crop_region_type" : cfg.MODEL.CLIP.CROP_REGION_TYPE, + "use_clip_c4": 'FPN' not in cfg.MODEL.BACKBONE.NAME, + "use_clip_attpool": cfg.MODEL.ROI_HEADS.NAME == 'PretrainRes5ROIHeads', + "offline_input_format": offline_cfg.INPUT.FORMAT if offline_cfg else None, + "offline_pixel_mean": offline_cfg.MODEL.PIXEL_MEAN if offline_cfg else None, + "offline_pixel_std": offline_cfg.MODEL.PIXEL_STD if offline_cfg else None, + "language_encoder": language_encoder, + "matching_temp": cfg.MODEL.CLIP.CLSS_TEMP, + "num_regions_per_img": cfg.MODEL.CLIP.PRETRAIN_SAMPLE_REGIONS, + "img_txt_level": (cfg.MODEL.CLIP.PRETRAIN_IMG_TXT_LEVEL, cfg.MODEL.CLIP.PRETRAIN_ONLY_EOT), + "gather_gpus": cfg.MODEL.CLIP.GATHER_GPUS, + "grid_regions": cfg.MODEL.CLIP.GRID_REGIONS, + "concept_emb": (cfg.MODEL.CLIP.CONCEPT_POOL_EMB, cfg.MODEL.CLIP.CONCEPT_THRES, cfg.MODEL.CLIP.TEACHER_CONCEPT_POOL_EMB), + } + + @property + def device(self): + return self.pixel_mean.device + + def forward(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Args: + batched_inputs: a list, batched outputs of :class:`DatasetMapper` . + Each item in the list contains the inputs for one image. + For now, each item in the list is a dict that contains: + + * image: Tensor, image in (C, H, W) format. + * instances (optional): groundtruth :class:`Instances` + * proposals (optional): :class:`Instances`, precomputed proposals. + + Other information that's included in the original dicts, such as: + + * "height", "width" (int): the output resolution of the model, used in inference. + See :meth:`postprocess` for details. + + Returns: + list[dict]: + Each dict is the output for one input image. + The dict contains one key "instances" whose value is a :class:`Instances`. + The :class:`Instances` object has the following keys: + "pred_boxes", "pred_classes", "scores", "pred_masks", "pred_keypoints" + """ + if not self.training: + return self.inference(batched_inputs) + if self.concept_emb is not None and self.teacher_backbone is None: # create a teacher model from an initialized student model; if resume, simply comment out this section + self.teacher_backbone = copy.deepcopy(self.backbone) + for p in self.teacher_backbone.parameters(): # freeze visual encoder of teacher model + p.requires_grad = False + gt_instances = None + losses = {} + + # localization branch: offline modules to get the region proposals + proposals = self.get_region_proposals(batched_inputs) + global_proposals = self.create_global_proposals(batched_inputs) + # for prop, g_prop in zip(proposals, global_proposals): # append global proposal into each image + # prop.proposal_boxes.tensor = torch.cat((prop.proposal_boxes.tensor, g_prop.tensor), dim=0) + + # recognition branch: get 2D feature maps using the backbone of recognition branch + images = self.preprocess_image(batched_inputs) + features = self.backbone(images.tensor) + region_feats = self.get_region_features(images, features, proposals, gt_instances) + global_feats = self.get_region_features(images, features, global_proposals, gt_instances) + + # image-text level matching + if self.img_txt_level: + self.image_text_matching(batched_inputs, proposals, region_feats, losses, global_feats=global_feats, only_global=True) + + # region-phrase level matching + if len(batched_inputs[0]) > 6: # controlled by dataset loading + phrase_text_embs = self.encode_phrase_text(batched_inputs) + else: + phrase_text_embs = None + + # region-concept level matching + if self.concept_emb is not None: + self.region_concept_matching(images, proposals, gt_instances, region_feats, losses, phrase_embs=phrase_text_embs) + + return losses + + def encode_phrase_text(self, batched_inputs): + text = [x[6].view(-1,self.context_length).to(self.device) for i, x in enumerate(batched_inputs)] + text = torch.cat(text, dim=0) + text_embs = self.lang_encoder.encode_text(text, only_eot=True) # [#phrases, transformer.width] + return text_embs + + def region_concept_matching(self, images, proposals, gt_instances, region_feats, losses, phrase_embs=None): + use_distill = True + use_contrastive = True + # get psuedo concept labels from teacher model + concept_scores, target_inds, keep_regions, target_embs, label_mtx, phrase_label_mtx, phrase_target_regions \ + = self.get_psuedo_concept_labels(images, proposals, gt_instances, phrase_embs=phrase_embs) + + # prepare region features for the kept regions + keep_region_feats = region_feats[keep_regions] + keep_region_feats = keep_region_feats / keep_region_feats.norm(dim=-1, keepdim=True) + + if use_distill: + # distillation learning: learns from the predictions of teacher model + concept_emb = self.concept_emb / self.concept_emb.norm(dim=-1, keepdim=True) + cls_scores = keep_region_feats @ concept_emb.t() # [#kept_regions, #concepts] + if isinstance(self.matching_temp, float): # Typical good values are 100.0 for euclidean, 10.0 for dot, 0.01 for cosine + cls_scores_temp = cls_scores / self.matching_temp + else: + cls_scores_temp = cls_scores * self.matching_temp.exp() + + # loss weights + #rpn_weights = torch.cat([torch.sigmoid(p.objectness_logits) for p in proposals])[keep_regions] + #focal_weights = self.focal_scaling(cls_scores_temp, target_inds) + + # calculate loss + cls_loss = F.kl_div(F.softmax(cls_scores_temp, dim=1).log(), concept_scores, reduction='batchmean') # input is log-probabilities, target is probabilities + #cls_loss = SoftTargetCrossEntropy()(cls_scores_temp, concept_scores) + #cls_loss = F.cross_entropy(cls_scores_temp, target_inds) + #cls_loss = (F.cross_entropy(cls_scores_temp, target_inds, reduction="none") * focal_weights).mean() + losses.update({"loss_region_distill": cls_loss}) # * 0.8}) + + if use_contrastive: + # contrastive learning: matching student visual features with target teacher concept embs + target_embs = target_embs / target_embs.norm(dim=-1, keepdim=True) + match_scores = keep_region_feats @ target_embs.t() # [#kept_regions, #kept_regions] + if isinstance(self.matching_temp, float): # Typical good values are 100.0 for euclidean, 10.0 for dot, 0.01 for cosine + match_scores_temp = match_scores / self.matching_temp + else: + match_scores_temp = match_scores * self.matching_temp.exp() + + # loss weights + #rpn_weights = torch.cat([torch.sigmoid(p.objectness_logits) for p in proposals])[keep_regions] + #focal_weights = (1 - torch.sigmoid(torch.diag(match_scores_temp))) ** 0.8 # 1.0 # 2.0 # + + # calculate loss given matching scores and label matrix + contrastive_loss = MILCrossEntropy()(match_scores_temp, label_mtx, weights=None, avg_positives=False) # SoftTargetCrossEntropy()(match_scores_temp, label_mtx) + #contrastive_loss = (MILCrossEntropy()(match_scores, label_mtx) + MILCrossEntropy()(match_scores.t(), label_mtx)) / 2.0 + losses.update({"loss_concept_contrastive": contrastive_loss}) + + if phrase_embs is not None: + phrase_embs = phrase_embs / phrase_embs.norm(dim=-1, keepdim=True) + phrase_scores = phrase_embs @ phrase_target_regions.t() + if isinstance(self.matching_temp, float): # Typical good values are 100.0 for euclidean, 10.0 for dot, 0.01 for cosine + phrase_scores_temp = phrase_scores / self.matching_temp + else: + phrase_scores_temp = phrase_scores * self.matching_temp.exp() + contrastive_loss = MILCrossEntropy()(phrase_scores_temp, phrase_label_mtx, weights=None, avg_positives=False) + #contrastive_loss = SoftTargetCrossEntropy()(phrase_scores_temp, phrase_label_mtx) + losses.update({"loss_phrase_contrastive": contrastive_loss}) + + def image_text_matching(self, batched_inputs, proposals, region_feats, losses, global_feats=None, only_global=False): + # encode text + num_cap = int(batched_inputs[0][1].size(0) / self.context_length) + if num_cap == 1: # one caption per image + text = [x[1].view(1,-1).to(self.device) for x in batched_inputs] + else: # multiple caption pers image, then randomly pick one + rand_ind = [randint(0, num_cap-1) for _ in range(len(batched_inputs))] + text = [x[1].view(-1,self.context_length)[rand_ind[i]:rand_ind[i]+1].to(self.device) for i, x in enumerate(batched_inputs)] + text = torch.cat(text, dim=0) + text_embs = self.lang_encoder.encode_text(text, only_eot=self.only_eot) # [img_batch, n_ctx, transformer.width] or [img_batch, transformer.width] + eot_pos = text.argmax(dim=-1) + + # prepare region features and text embeddings + if isinstance(proposals[0], Boxes): + num_bbs = [len(prop) for prop in proposals] + else: + num_bbs = [len(prop.proposal_boxes) for prop in proposals] + if global_feats is not None and only_global: # only global feature + assert self.only_eot + region_feats = global_feats + region_feats = region_feats / region_feats.norm(dim=-1, keepdim=True) + text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True) + num_bbs = [1 for _ in num_bbs] + elif global_feats is not None and not only_global: # combine both global and region features + assert self.only_eot + keep_num = 20 + region_feats = region_feats.split(num_bbs) + region_feats = [torch.mean(rg_f, dim=0, keepdim=True) for rg_f in region_feats] + region_g_feats = [torch.cat((r_f[:keep_num], global_feats[i:i+1]), dim=0) for i, r_f in enumerate(region_feats)] + region_g_feats = [torch.mean(rg_f, dim=0, keepdim=True) for rg_f in region_g_feats] + region_g_feats = [rg_f / rg_f.norm(dim=-1, keepdim=True) for rg_f in region_g_feats] + region_feats = torch.cat(region_g_feats) + text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True) + num_bbs = [1 for _ in num_bbs] + else: # only region features + num_bbs = torch.tensor(num_bbs).long().to(self.device) + + region_feats_full, min_bs = gather_tensors(region_feats) if self.gather_gpus else (region_feats, None) # gather across GPUs + text_embs_full, min_bs = gather_tensors(text_embs) if self.gather_gpus else (text_embs, None) # gather across GPUs + + # matching visual features with text embs + match_scores = region_feats_full @ text_embs_full.view(-1, text_embs_full.size(-1)).t() # [#regions, img_batch * n_ctx] + if global_feats is not None: # only global feature or combine both global and region features + img_b = int(region_feats_full.size(0)) + pooled_score = match_scores + else: # only region features + eot_pos_full, min_bs = gather_tensors(eot_pos) if self.gather_gpus else (eot_pos, None) # gather across GPUs + num_bbs_full, min_bs = gather_tensors(num_bbs) if self.gather_gpus else (num_bbs, None) # gather across GPUs + pooled_score = [] + token_b = self.context_length + # region_b = self.num_regions_per_img if global_feats is None else 1 + # img_b = int(region_feats_full.size(0) / region_b) + img_b = num_bbs_full.size(0) + rb_start = 0 # the starting index of regions + for i in range(img_b): # for each image + region_b = num_bbs_full[i].item() + for j in range(img_b): # for each text + if self.only_eot: # sentence level embs + # max pool over regions + this_s = torch.max(match_scores[rb_start:(rb_start+region_b), j:(j+1)], dim=0)[0] + else: # token level embs + # 3. softmax over regions as soft attention, then multiply attention with original logits, finally sum over matrix and divided by #tokens + # this_matrix = match_scores[rb_start:(rb_start+region_b), j*token_b:(j*token_b+eot_pos_full[j]+1)] + # this_att = F.softmax(this_matrix, dim=0) + # this_s = torch.sum(this_matrix * this_att) / (eot_pos_full[j]+1) + # 2. max pool over regions, and then avg over text tokens + # this_s = torch.sum(torch.max(match_scores[rb_start:(rb_start+region_b), j*token_b:(j*token_b+eot_pos_full[j]+1)], dim=0)[0]) / (eot_pos_full[j]+1) + # 1. max pool over regions, and then sum over text tokens + this_s = torch.sum(torch.max(match_scores[rb_start:(rb_start+region_b), j*token_b:(j*token_b+eot_pos_full[j]+1)], dim=0)[0]) + pooled_score.append(this_s.view(1,1)) + rb_start += region_b + assert rb_start == match_scores.size(0) + pooled_score = torch.cat(pooled_score).view(img_b, img_b) # diagnal elements are positive pairs and the others are negative pairs + + if isinstance(self.matching_temp,float): # Typical good values are 100.0 for euclidean, 10.0 for dot, 0.01 for cosine + pooled_score = pooled_score / self.matching_temp + else: + pooled_score = pooled_score * self.matching_temp.exp() + contrast_target = torch.arange(img_b).to(self.device) + row_loss = F.cross_entropy(pooled_score, contrast_target) + col_loss = F.cross_entropy(pooled_score.t(), contrast_target) + losses.update({"loss_img_txt_level": (row_loss + col_loss) / 2.0}) # losses.update({"loss_img_txt_level": (row_loss + col_loss) / 4.0}) # + + def focal_scaling(self, logits, targets, gamma=1.0): + p = F.softmax(logits, dim=1) + p_t = p[torch.arange(p.size(0)).to(p.device), targets] # get prob of target class + weights = (1 - p_t) ** gamma + return weights + + def get_psuedo_concept_labels(self, images, proposals, gt_instances, s_temp=0.01, norm=True, phrase_embs=None): + """ Input images and region proposals, return matching results from teacher model + """ + with torch.no_grad(): + # extract visual features from teacher model + features = self.teacher_backbone(images.tensor) + teacher_region_feats = self.teacher_roi_heads(images, features, proposals, gt_instances, res5=self.teacher_backbone.layer4, attnpool=self.teacher_backbone.attnpool) + # match teacher visual features with teacher concept embs to create pseudo labels + if norm: + teacher_region_feats = teacher_region_feats / teacher_region_feats.norm(dim=-1, keepdim=True) + teacher_concept_emb = self.teacher_concept_emb / self.teacher_concept_emb.norm(dim=-1, keepdim=True) + else: + teacher_concept_emb = self.teacher_concept_emb + concept_scores = teacher_region_feats @ teacher_concept_emb.t() # [#regions, #concepts] + concept_scores = F.softmax(concept_scores / s_temp, dim=1) + max_scores, max_inds = torch.max(concept_scores, dim=1) + keep_regions = max_scores > self.concept_thres # only keep the regions that have high matching score with a concept + if keep_regions.nonzero().size(0) == 0: # if all regions can't match to any concept + print("all regions can't match to any concept!") + keep_regions = max_scores > 0.0 + target_inds = max_inds[keep_regions] + target_embs = self.concept_emb[target_inds] # the target embedding of student model + label_mtx = (target_inds.view(-1, 1) == target_inds.view(1, -1)).type_as(teacher_region_feats) + concept_scores = concept_scores[keep_regions] + # matching kept regions with phrase-text to create labels + if phrase_embs is None: + phrase_label_mtx = None + phrase_target_regions = None + else: + if norm: + phrase_embs = phrase_embs / phrase_embs.norm(dim=-1, keepdim=True) + teacher_kept_feats = teacher_region_feats[keep_regions] + phrase_scores = phrase_embs @ teacher_kept_feats.t() # [#phrases, #keep regions] + phrase_scores = F.softmax(phrase_scores / s_temp, dim=1) + _, max_region_inds = torch.max(phrase_scores, dim=1) + phrase_label_mtx = (max_region_inds.view(-1, 1) == max_region_inds.view(1, -1)).type_as(teacher_region_feats) + phrase_target_regions = teacher_kept_feats[max_region_inds] + + return concept_scores, target_inds, keep_regions, target_embs, label_mtx, phrase_label_mtx, phrase_target_regions + + def get_region_features(self, images, features, proposals, gt_instances): + """ Input images and region proposals, return region features + """ + # Given the proposals, crop region features from 2D image features + if self.use_clip_c4: # use C4 + resnet weights from CLIP + if self.use_clip_attpool: # use att_pool from CLIP to match dimension + region_feats = self.roi_heads(images, features, proposals, gt_instances, res5=self.backbone.layer4, attnpool=self.backbone.attnpool) + else: # use default mean pool + region_feats = self.roi_heads(images, features, proposals, gt_instances, res5=self.backbone.layer4) + else: # default setting + region_feats = self.roi_heads(images, features, proposals, gt_instances) + return region_feats + + def get_region_proposals(self, batched_inputs): + """ Given image, return object proposals + """ + if self.grid_regions: # use grid boxes + proposals = self.create_grid_boxes(batched_inputs) + else: # use object proposals + with torch.no_grad(): + if self.clip_crop_region_type == "GLOBAL": # from a global box per image + proposals = self.create_global_proposals(batched_inputs) + elif self.clip_crop_region_type == "GRID": # from grid proposals + proposals = self.create_grid_boxes(batched_inputs) + elif self.clip_crop_region_type == "RANDOM": # from random proposals + proposals = self.create_rand_boxes(batched_inputs) + elif self.clip_crop_region_type == "RPN": # from the backbone & RPN of standard Mask-RCNN, trained on base classes + if self.offline_backbone.training or self.offline_proposal_generator.training: # was set to True in training script + self.offline_backbone.eval() + self.offline_proposal_generator.eval() + images = self.offline_preprocess_image(batched_inputs) + features = self.offline_backbone(images.tensor) + if self.offline_proposal_generator is not None: + proposals, _ = self.offline_proposal_generator(images, features, None) + #visualize_proposals(batched_inputs, proposals, self.input_format, vis_pretrain=True) + # randomly select proposals to avoid overfitting + if self.training: + #rand_inds = [torch.arange(len(p))[:self.num_regions_per_img].to(self.device) for p in proposals] + rand_inds = [torch.randperm(len(p))[:self.num_regions_per_img].to(self.device) for p in proposals] + proposals = [p[rand_inds[i]] for i, p in enumerate(proposals)] + return proposals + + def offline_preprocess_image(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + NOTE: the image tsv in pretraining are already normalized pixel values and thus opposite to Detectron2 default input. + Normalize, pad and batch the input images. Use detectron2 default processing (pixel mean & std). + Note: Due to FPN size_divisibility, images are padded by right/bottom border. So FPN is consistent with C4 and GT boxes. + """ + images = [x[0].to(self.device) for x in batched_inputs] + if (self.input_format == 'RGB' and self.offline_input_format == 'BGR') or \ + (self.input_format == 'BGR' and self.offline_input_format == 'RGB'): # the input image follows the main config format ('RGB' or 'BGR') + images = [x[[2,1,0],:,:] for x in images] + if self.offline_div_pixel: + images = [(x - self.offline_pixel_mean) / self.offline_pixel_std for x in images] + else: + images = [((x * 255.0) - self.offline_pixel_mean) / self.offline_pixel_std for x in images] + images = ImageList.from_tensors(images, self.offline_backbone.size_divisibility) + return images + + def preprocess_image(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + NOTE: the image tsv in pretraining are already normalized pixel values and thus opposite to Detectron2 default input. + Normalize, pad and batch the input images. Use CLIP default processing (pixel mean & std). + Note: Due to FPN size_divisibility, images are padded by right/bottom border. So FPN is consistent with C4 and GT boxes. + """ + images = [x[0].to(self.device) for x in batched_inputs] + if self.div_pixel: + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + else: + images = [((x * 255.0) - self.pixel_mean) / self.pixel_std for x in images] + images = ImageList.from_tensors(images, self.backbone.size_divisibility) + return images + + def create_rand_boxes(self, batched_inputs, grid_length=8): + """ create random boxes within an image, output random self.num_regions_per_img boxes + return a list of Boxes + """ + images = self.preprocess_image(batched_inputs) + image_height = images.tensor.size(2) + image_width = images.tensor.size(3) + + left_top_x = torch.tensor([i*(grid_length) for i in range(image_width // grid_length)]) + left_top_y = torch.tensor([i*(grid_length) for i in range(image_height // grid_length)]) + right_bot_x = torch.tensor([(i+1)*(grid_length) for i in range(image_width // grid_length)]) + right_bot_y = torch.tensor([(i+1)*(grid_length) for i in range(image_height // grid_length)]) + x_inds = torch.randint(0, left_top_x.size(0), (self.num_regions_per_img,)) + y_inds = torch.randint(0, left_top_y.size(0), (self.num_regions_per_img,)) + + proposals = [] + for i in range(self.num_regions_per_img): + rb_x_candidates = right_bot_x[x_inds[i]:] + rb_x = rb_x_candidates[torch.randperm(rb_x_candidates.size(0))[0]] + rb_y_candidates = right_bot_y[y_inds[i]:] + rb_y = rb_y_candidates[torch.randperm(rb_y_candidates.size(0))[0]] + this_box = torch.cat((left_top_x[x_inds[i]].view(1,1), left_top_y[y_inds[i]].view(1,1), rb_x.view(1,1), rb_y.view(1,1)),dim=-1) + proposals.append(this_box) + proposals = torch.cat(proposals).float().to(self.device) + proposals = [Boxes(proposals) for i in range(len(batched_inputs))] # a list of Boxes + return proposals + + def create_grid_boxes(self, batched_inputs, grid_length=32): + """ create (image_height/32) * (image_width/32) pseudo grid boxes, and randomly sample self.num_regions_per_img boxes + return a list of Boxes + """ + images = self.preprocess_image(batched_inputs) + image_height = images.tensor.size(2) + image_width = images.tensor.size(3) + + left_top_x = torch.tensor([i*(grid_length) for i in range(image_width // grid_length)]) + left_top_y = torch.tensor([i*(grid_length) for i in range(image_height // grid_length)]) + right_bot_x = torch.tensor([(i+1)*(grid_length) for i in range(image_width // grid_length)]) + right_bot_y = torch.tensor([(i+1)*(grid_length) for i in range(image_height // grid_length)]) + left_top_x, left_top_y = torch.meshgrid(left_top_x, left_top_y) + right_bot_x, right_bot_y = torch.meshgrid(right_bot_x, right_bot_y) + grid_boxes = torch.cat((left_top_x.flatten().view(-1,1), left_top_y.flatten().view(-1,1),\ + right_bot_x.flatten().view(-1,1), right_bot_y.flatten().view(-1,1),), dim=1) + sample_ind = torch.randperm(grid_boxes.size(0))[:self.num_regions_per_img] + grid_boxes = grid_boxes[sample_ind] + grid_boxes = grid_boxes.float().to(self.device) + proposals = [Boxes(grid_boxes) for i in range(len(batched_inputs))] # a list of Boxes + return proposals + + def create_global_proposals(self, batched_inputs): + """ create a single global box for an image, so as to extract global image features with RoIAlign on high-resolution images. + """ + images = self.preprocess_image(batched_inputs) + image_height = images.tensor.size(2) + image_width = images.tensor.size(3) + + global_box = torch.tensor([0, 0, image_width, image_height]).view(1,4).float().to(self.device) + proposals = [Boxes(global_box) for i in range(len(batched_inputs))] # a list of Boxes + return proposals + + def inference(self, batched_inputs, detected_instances=None, do_postprocess=True): + """ + Grounding inference: map region features with sentence tokens + return: matching scores between region features and tokenized texts, region boxes in raw image resolution, image id & raw string texts & tokenized texts + """ + assert len(batched_inputs) == 1 # only one instance per image during inference + gt_instances = None + losses = {} + + # localization branch: offline modules to get the region proposals + proposals = self.get_region_proposals(batched_inputs) + + # recognition branch: get 2D feature maps using the backbone of recognition branch + images = self.preprocess_image(batched_inputs) + features = self.backbone(images.tensor) + region_feats = self.get_region_features(images, features, proposals, gt_instances) + + # encode text + num_cap = int(batched_inputs[0][1].size(0) / self.context_length) + text = batched_inputs[0][1].view(num_cap, -1).to(self.device) # [num_cap, context_length] + text_embs = self.lang_encoder.encode_text(text, only_eot=False) # [img_batch, n_ctx, transformer.width] or [img_batch, transformer.width] + + # matching visual features with text embs + region_feats = region_feats / region_feats.norm(dim=-1, keepdim=True) + text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True) + match_scores = region_feats @ text_embs.view(-1, text_embs.size(-1)).t() # [#regions, img_batch * n_ctx] + # visualize_proposals(batched_inputs, proposals, self.input_format, vis_pretrain=True) + + # multiply RPN logits + rpn_scores = [p.get('objectness_logits') for p in proposals][0] + match_scores = (match_scores * rpn_scores[:, None]) ** 0.5 + + # scale the object proposals back to raw image resolution + if do_postprocess: + assert not torch.jit.is_scripting(), "Scripting is not supported for postprocess." + processed_results = PretrainFastRCNN._postprocess(proposals, batched_inputs) + return match_scores, processed_results + + @staticmethod + def _postprocess(instances, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Rescale the output instances to the target size. + """ + # note: private function; subject to changes + processed_results = [] + for results_per_image, input_per_image in zip(instances, batched_inputs): + height, width = input_per_image[-1][2] # original image size, before resizing + r = detector_postprocess(results_per_image, height, width) + processed_results.append({"instances": r}) + return processed_results + + +def visualize_proposals(batched_inputs, proposals, input_format, vis_pretrain=False): + """ + A function used to visualize images and proposals. It shows ground truth + bounding boxes on the original image and up to 20 top-scoring predicted + object proposals on the original image. Users can implement different + visualization functions for different models. + + Args: + batched_inputs (list): a list that contains input to the model. + proposals (list): a list that contains predicted proposals. Both + batched_inputs and proposals should have the same length. + """ + from detectron2.utils.visualizer import Visualizer + + max_vis_prop = 50 + if vis_pretrain: + for i, (input, prop) in enumerate(zip(batched_inputs, proposals)): + img = input[0] * 255.0 + img = convert_image_to_rgb(img.permute(1, 2, 0), input_format) + box_size = min(len(prop.proposal_boxes), max_vis_prop) + v_pred = Visualizer(img, None) + v_pred = v_pred.overlay_instances( + boxes=prop.proposal_boxes[0:box_size].tensor.cpu().numpy() + ) + prop_img = v_pred.get_image() + vis_img = prop_img + to_save = Image.fromarray(np.array(vis_img, np.uint8)) + to_save.save("output/regions/" + str(i) + ".png") + #break # only visualize one image in a batch + else: + for input, prop in zip(batched_inputs, proposals): + img = input["image"] + img = convert_image_to_rgb(img.permute(1, 2, 0), input_format) + box_size = min(len(prop.proposal_boxes), max_vis_prop) + v_pred = Visualizer(img, None) + v_pred = v_pred.overlay_instances( + boxes=prop.proposal_boxes[0:box_size].tensor.cpu().numpy() + ) + prop_img = v_pred.get_image() + vis_img = prop_img + # f_n = input['file_name'] + to_save = Image.fromarray(np.array(vis_img, np.uint8)) + to_save.save("output/regions/" + "proposals.png") + #break # only visualize one image in a batch + +def visualize_results(batched_inputs, results, input_format, vis_pretrain=False): + """ + A function used to visualize images and results. It shows ground truth + bounding boxes on the original image and up to 20 top-scoring predicted + object results on the original image. Users can implement different + visualization functions for different models. + + Args: + batched_inputs (list): a list that contains input to the model. + results (list): a list that contains predicted results. Both + batched_inputs and results should have the same length. + """ + from detectron2.utils.visualizer import Visualizer + + max_vis_prop = 1 + if vis_pretrain: + for i, (input, prop) in enumerate(zip(batched_inputs, results)): + img = input[0] * 255.0 + img = convert_image_to_rgb(img.permute(1, 2, 0), input_format) + box_size = min(len(prop.proposal_boxes), max_vis_prop) + v_pred = Visualizer(img, None) + v_pred = v_pred.overlay_instances( + boxes=prop.proposal_boxes[0:box_size].tensor.cpu().numpy() + ) + prop_img = v_pred.get_image() + vis_img = prop_img + to_save = Image.fromarray(np.array(vis_img, np.uint8)) + # to_save.save("output/regions/" + str(i) + ".png") + #break # only visualize one image in a batch + else: + for input, prop in zip(batched_inputs, results): + img = input["image"] + img = convert_image_to_rgb(img.permute(1, 2, 0), input_format) + box_size = min(len(prop.pred_boxes), max_vis_prop) + v_pred = Visualizer(img, None) + v_pred = v_pred.overlay_instances( + boxes=prop.pred_boxes[0:box_size].tensor.cpu().numpy() + ) + prop_img = v_pred.get_image() + vis_img = prop_img + # f_n = input['file_name'] + to_save = Image.fromarray(np.array(vis_img, np.uint8)) + to_save.save("output/regions/" + "results.png") + #break # only visualize one image in a batch + return to_save \ No newline at end of file diff --git a/detectron2/modeling/meta_arch/panoptic_fpn.py b/detectron2/modeling/meta_arch/panoptic_fpn.py new file mode 100644 index 0000000000000000000000000000000000000000..13aeabce162f4114109efe2c7fb4770b89087ab0 --- /dev/null +++ b/detectron2/modeling/meta_arch/panoptic_fpn.py @@ -0,0 +1,266 @@ +# -*- coding: utf-8 -*- +# Copyright (c) Facebook, Inc. and its affiliates. + +import logging +from typing import Dict, List +import torch +from torch import nn + +from detectron2.config import configurable +from detectron2.structures import ImageList + +from ..postprocessing import detector_postprocess, sem_seg_postprocess +from .build import META_ARCH_REGISTRY +from .rcnn import GeneralizedRCNN +from .semantic_seg import build_sem_seg_head + +__all__ = ["PanopticFPN"] + + +@META_ARCH_REGISTRY.register() +class PanopticFPN(GeneralizedRCNN): + """ + Implement the paper :paper:`PanopticFPN`. + """ + + @configurable + def __init__( + self, + *, + sem_seg_head: nn.Module, + combine_overlap_thresh: float = 0.5, + combine_stuff_area_thresh: float = 4096, + combine_instances_score_thresh: float = 0.5, + **kwargs, + ): + """ + NOTE: this interface is experimental. + + Args: + sem_seg_head: a module for the semantic segmentation head. + combine_overlap_thresh: combine masks into one instances if + they have enough overlap + combine_stuff_area_thresh: ignore stuff areas smaller than this threshold + combine_instances_score_thresh: ignore instances whose score is + smaller than this threshold + + Other arguments are the same as :class:`GeneralizedRCNN`. + """ + super().__init__(**kwargs) + self.sem_seg_head = sem_seg_head + # options when combining instance & semantic outputs + self.combine_overlap_thresh = combine_overlap_thresh + self.combine_stuff_area_thresh = combine_stuff_area_thresh + self.combine_instances_score_thresh = combine_instances_score_thresh + + @classmethod + def from_config(cls, cfg): + ret = super().from_config(cfg) + ret.update( + { + "combine_overlap_thresh": cfg.MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH, + "combine_stuff_area_thresh": cfg.MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT, + "combine_instances_score_thresh": cfg.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH, # noqa + } + ) + ret["sem_seg_head"] = build_sem_seg_head(cfg, ret["backbone"].output_shape()) + logger = logging.getLogger(__name__) + if not cfg.MODEL.PANOPTIC_FPN.COMBINE.ENABLED: + logger.warning( + "PANOPTIC_FPN.COMBINED.ENABLED is no longer used. " + " model.inference(do_postprocess=) should be used to toggle postprocessing." + ) + if cfg.MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT != 1.0: + w = cfg.MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT + logger.warning( + "PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT should be replaced by weights on each ROI head." + ) + + def update_weight(x): + if isinstance(x, dict): + return {k: v * w for k, v in x.items()} + else: + return x * w + + roi_heads = ret["roi_heads"] + roi_heads.box_predictor.loss_weight = update_weight(roi_heads.box_predictor.loss_weight) + roi_heads.mask_head.loss_weight = update_weight(roi_heads.mask_head.loss_weight) + return ret + + def forward(self, batched_inputs): + """ + Args: + batched_inputs: a list, batched outputs of :class:`DatasetMapper`. + Each item in the list contains the inputs for one image. + + For now, each item in the list is a dict that contains: + + * "image": Tensor, image in (C, H, W) format. + * "instances": Instances + * "sem_seg": semantic segmentation ground truth. + * Other information that's included in the original dicts, such as: + "height", "width" (int): the output resolution of the model, used in inference. + See :meth:`postprocess` for details. + + Returns: + list[dict]: + each dict has the results for one image. The dict contains the following keys: + + * "instances": see :meth:`GeneralizedRCNN.forward` for its format. + * "sem_seg": see :meth:`SemanticSegmentor.forward` for its format. + * "panoptic_seg": See the return value of + :func:`combine_semantic_and_instance_outputs` for its format. + """ + if not self.training: + return self.inference(batched_inputs) + images = self.preprocess_image(batched_inputs) + features = self.backbone(images.tensor) + + assert "sem_seg" in batched_inputs[0] + gt_sem_seg = [x["sem_seg"].to(self.device) for x in batched_inputs] + gt_sem_seg = ImageList.from_tensors( + gt_sem_seg, self.backbone.size_divisibility, self.sem_seg_head.ignore_value + ).tensor + sem_seg_results, sem_seg_losses = self.sem_seg_head(features, gt_sem_seg) + + gt_instances = [x["instances"].to(self.device) for x in batched_inputs] + proposals, proposal_losses = self.proposal_generator(images, features, gt_instances) + detector_results, detector_losses = self.roi_heads( + images, features, proposals, gt_instances + ) + + losses = sem_seg_losses + losses.update(proposal_losses) + losses.update(detector_losses) + return losses + + def inference(self, batched_inputs: List[Dict[str, torch.Tensor]], do_postprocess: bool = True): + """ + Run inference on the given inputs. + + Args: + batched_inputs (list[dict]): same as in :meth:`forward` + do_postprocess (bool): whether to apply post-processing on the outputs. + + Returns: + When do_postprocess=True, see docs in :meth:`forward`. + Otherwise, returns a (list[Instances], list[Tensor]) that contains + the raw detector outputs, and raw semantic segmentation outputs. + """ + images = self.preprocess_image(batched_inputs) + features = self.backbone(images.tensor) + sem_seg_results, sem_seg_losses = self.sem_seg_head(features, None) + proposals, _ = self.proposal_generator(images, features, None) + detector_results, _ = self.roi_heads(images, features, proposals, None) + + if do_postprocess: + processed_results = [] + for sem_seg_result, detector_result, input_per_image, image_size in zip( + sem_seg_results, detector_results, batched_inputs, images.image_sizes + ): + height = input_per_image.get("height", image_size[0]) + width = input_per_image.get("width", image_size[1]) + sem_seg_r = sem_seg_postprocess(sem_seg_result, image_size, height, width) + detector_r = detector_postprocess(detector_result, height, width) + + processed_results.append({"sem_seg": sem_seg_r, "instances": detector_r}) + + panoptic_r = combine_semantic_and_instance_outputs( + detector_r, + sem_seg_r.argmax(dim=0), + self.combine_overlap_thresh, + self.combine_stuff_area_thresh, + self.combine_instances_score_thresh, + ) + processed_results[-1]["panoptic_seg"] = panoptic_r + return processed_results + else: + return detector_results, sem_seg_results + + +def combine_semantic_and_instance_outputs( + instance_results, + semantic_results, + overlap_threshold, + stuff_area_thresh, + instances_score_thresh, +): + """ + Implement a simple combining logic following + "combine_semantic_and_instance_predictions.py" in panopticapi + to produce panoptic segmentation outputs. + + Args: + instance_results: output of :func:`detector_postprocess`. + semantic_results: an (H, W) tensor, each element is the contiguous semantic + category id + + Returns: + panoptic_seg (Tensor): of shape (height, width) where the values are ids for each segment. + segments_info (list[dict]): Describe each segment in `panoptic_seg`. + Each dict contains keys "id", "category_id", "isthing". + """ + panoptic_seg = torch.zeros_like(semantic_results, dtype=torch.int32) + + # sort instance outputs by scores + sorted_inds = torch.argsort(-instance_results.scores) + + current_segment_id = 0 + segments_info = [] + + instance_masks = instance_results.pred_masks.to(dtype=torch.bool, device=panoptic_seg.device) + + # Add instances one-by-one, check for overlaps with existing ones + for inst_id in sorted_inds: + score = instance_results.scores[inst_id].item() + if score < instances_score_thresh: + break + mask = instance_masks[inst_id] # H,W + mask_area = mask.sum().item() + + if mask_area == 0: + continue + + intersect = (mask > 0) & (panoptic_seg > 0) + intersect_area = intersect.sum().item() + + if intersect_area * 1.0 / mask_area > overlap_threshold: + continue + + if intersect_area > 0: + mask = mask & (panoptic_seg == 0) + + current_segment_id += 1 + panoptic_seg[mask] = current_segment_id + segments_info.append( + { + "id": current_segment_id, + "isthing": True, + "score": score, + "category_id": instance_results.pred_classes[inst_id].item(), + "instance_id": inst_id.item(), + } + ) + + # Add semantic results to remaining empty areas + semantic_labels = torch.unique(semantic_results).cpu().tolist() + for semantic_label in semantic_labels: + if semantic_label == 0: # 0 is a special "thing" class + continue + mask = (semantic_results == semantic_label) & (panoptic_seg == 0) + mask_area = mask.sum().item() + if mask_area < stuff_area_thresh: + continue + + current_segment_id += 1 + panoptic_seg[mask] = current_segment_id + segments_info.append( + { + "id": current_segment_id, + "isthing": False, + "category_id": semantic_label, + "area": mask_area, + } + ) + + return panoptic_seg, segments_info diff --git a/detectron2/modeling/meta_arch/rcnn.py b/detectron2/modeling/meta_arch/rcnn.py new file mode 100644 index 0000000000000000000000000000000000000000..ce61a45cc73bd57506b90b938a92df51e03100b5 --- /dev/null +++ b/detectron2/modeling/meta_arch/rcnn.py @@ -0,0 +1,373 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import numpy as np +from typing import Dict, List, Optional, Tuple +from numpy.lib import pad +import torch +from torch import nn +from torch.nn import functional as F +from random import randint + +from detectron2.config import configurable +from detectron2.data.detection_utils import convert_image_to_rgb +from detectron2.structures import ImageList, Instances, Boxes +from detectron2.utils.events import get_event_storage +from detectron2.utils.logger import log_first_n + +from ..backbone import Backbone, build_backbone +from ..postprocessing import detector_postprocess +from ..proposal_generator import build_proposal_generator +from ..roi_heads import build_roi_heads +from .build import META_ARCH_REGISTRY + +__all__ = ["GeneralizedRCNN", "ProposalNetwork"] + +@META_ARCH_REGISTRY.register() +class GeneralizedRCNN(nn.Module): + """ + Generalized R-CNN. Any models that contains the following three components: + 1. Per-image feature extraction (aka backbone) + 2. Region proposal generation + 3. Per-region feature extraction and prediction + """ + + @configurable + def __init__( + self, + *, + backbone: Backbone, + proposal_generator: nn.Module, + roi_heads: nn.Module, + pixel_mean: Tuple[float], + pixel_std: Tuple[float], + input_format: Optional[str] = None, + vis_period: int = 0, + use_clip_c4: False, + use_clip_attpool: False, + ): + """ + Args: + backbone: a backbone module, must follow detectron2's backbone interface + proposal_generator: a module that generates proposals using backbone features + roi_heads: a ROI head that performs per-region computation + pixel_mean, pixel_std: list or tuple with #channels element, representing + the per-channel mean and std to be used to normalize the input image + input_format: describe the meaning of channels of input. Needed by visualization + vis_period: the period to run visualization. Set to 0 to disable. + """ + super().__init__() + self.backbone = backbone + self.proposal_generator = proposal_generator + self.roi_heads = roi_heads + + self.input_format = input_format + self.vis_period = vis_period + if vis_period > 0: + assert input_format is not None, "input_format is required for visualization!" + + self.register_buffer("pixel_mean", torch.tensor(pixel_mean).view(-1, 1, 1), False) + self.register_buffer("pixel_std", torch.tensor(pixel_std).view(-1, 1, 1), False) + assert ( + self.pixel_mean.shape == self.pixel_std.shape + ), f"{self.pixel_mean} and {self.pixel_std} have different shapes!" + if np.sum(pixel_mean) < 3.0: # converrt pixel value to range [0.0, 1.0] by dividing 255.0 + assert input_format == 'RGB' + self.div_pixel = True + else: # default setting + self.div_pixel = False + self.use_clip_c4 = use_clip_c4 # if True, use C4 mode where roi_head uses the last resnet layer from backbone + self.use_clip_attpool = use_clip_attpool # if True (C4+text_emb_as_classifier), use att_pool to replace default mean pool + + @classmethod + def from_config(cls, cfg): + backbone = build_backbone(cfg) + return { + "backbone": backbone, + "proposal_generator": build_proposal_generator(cfg, backbone.output_shape()), + "roi_heads": build_roi_heads(cfg, backbone.output_shape()), + "input_format": cfg.INPUT.FORMAT, + "vis_period": cfg.VIS_PERIOD, + "pixel_mean": cfg.MODEL.PIXEL_MEAN, + "pixel_std": cfg.MODEL.PIXEL_STD, + "use_clip_c4": cfg.MODEL.BACKBONE.NAME == "build_clip_resnet_backbone", + "use_clip_attpool": cfg.MODEL.ROI_HEADS.NAME == 'CLIPRes5ROIHeads' and cfg.MODEL.CLIP.USE_TEXT_EMB_CLASSIFIER, + } + + @property + def device(self): + return self.pixel_mean.device + + def visualize_training(self, batched_inputs, proposals): + """ + A function used to visualize images and proposals. It shows ground truth + bounding boxes on the original image and up to 20 top-scoring predicted + object proposals on the original image. Users can implement different + visualization functions for different models. + + Args: + batched_inputs (list): a list that contains input to the model. + proposals (list): a list that contains predicted proposals. Both + batched_inputs and proposals should have the same length. + """ + from detectron2.utils.visualizer import Visualizer + + storage = get_event_storage() + max_vis_prop = 20 + + for input, prop in zip(batched_inputs, proposals): + img = input["image"] + img = convert_image_to_rgb(img.permute(1, 2, 0), self.input_format) + v_gt = Visualizer(img, None) + v_gt = v_gt.overlay_instances(boxes=input["instances"].gt_boxes) + anno_img = v_gt.get_image() + box_size = min(len(prop.proposal_boxes), max_vis_prop) + v_pred = Visualizer(img, None) + v_pred = v_pred.overlay_instances( + boxes=prop.proposal_boxes[0:box_size].tensor.cpu().numpy() + ) + prop_img = v_pred.get_image() + vis_img = np.concatenate((anno_img, prop_img), axis=1) + vis_img = vis_img.transpose(2, 0, 1) + vis_name = "Left: GT bounding boxes; Right: Predicted proposals" + storage.put_image(vis_name, vis_img) + break # only visualize one image in a batch + + def forward(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Args: + batched_inputs: a list, batched outputs of :class:`DatasetMapper` . + Each item in the list contains the inputs for one image. + For now, each item in the list is a dict that contains: + + * image: Tensor, image in (C, H, W) format. + * instances (optional): groundtruth :class:`Instances` + * proposals (optional): :class:`Instances`, precomputed proposals. + + Other information that's included in the original dicts, such as: + + * "height", "width" (int): the output resolution of the model, used in inference. + See :meth:`postprocess` for details. + + Returns: + list[dict]: + Each dict is the output for one input image. + The dict contains one key "instances" whose value is a :class:`Instances`. + The :class:`Instances` object has the following keys: + "pred_boxes", "pred_classes", "scores", "pred_masks", "pred_keypoints" + """ + if not self.training: + return self.inference(batched_inputs) + + images = self.preprocess_image(batched_inputs) + if "instances" in batched_inputs[0]: + gt_instances = [x["instances"].to(self.device) for x in batched_inputs] + else: + gt_instances = None + # eg: {'p2': torch.Size([b, c, 200, 304]), 'p3': torch.Size([b, c, 100, 152]), 'p4': torch.Size([b, c, 50, 76]), 'p5': torch.Size([b, c, 25, 38]), 'p6': torch.Size([b, c, 13, 19])} + features = self.backbone(images.tensor) + + if self.proposal_generator is not None: + proposals, proposal_losses = self.proposal_generator(images, features, gt_instances) + else: + assert "proposals" in batched_inputs[0] + proposals = [x["proposals"].to(self.device) for x in batched_inputs] + proposal_losses = {} + + if self.use_clip_c4: # use C4 + resnet weights from CLIP + if self.use_clip_attpool: # use att_pool from CLIP to match dimension + _, detector_losses = self.roi_heads(images, features, proposals, gt_instances, res5=self.backbone.layer4, attnpool=self.backbone.attnpool) + else: # use default mean pool + _, detector_losses = self.roi_heads(images, features, proposals, gt_instances, res5=self.backbone.layer4) + else: # default setting + _, detector_losses = self.roi_heads(images, features, proposals, gt_instances) + if self.vis_period > 0: + storage = get_event_storage() + if storage.iter % self.vis_period == 0: + self.visualize_training(batched_inputs, proposals) + + losses = {} + losses.update(detector_losses) + losses.update(proposal_losses) + return losses + + def inference( + self, + batched_inputs: List[Dict[str, torch.Tensor]], + detected_instances: Optional[List[Instances]] = None, + do_postprocess: bool = True, + ): + """ + Run inference on the given inputs. + + Args: + batched_inputs (list[dict]): same as in :meth:`forward` + detected_instances (None or list[Instances]): if not None, it + contains an `Instances` object per image. The `Instances` + object contains "pred_boxes" and "pred_classes" which are + known boxes in the image. + The inference will then skip the detection of bounding boxes, + and only predict other per-ROI outputs. + do_postprocess (bool): whether to apply post-processing on the outputs. + + Returns: + When do_postprocess=True, same as in :meth:`forward`. + Otherwise, a list[Instances] containing raw network outputs. + """ + assert not self.training + + images = self.preprocess_image(batched_inputs) + features = self.backbone(images.tensor) + + if detected_instances is None: + if self.proposal_generator is not None: + proposals, _ = self.proposal_generator(images, features, None) + else: + assert "proposals" in batched_inputs[0] + proposals = [x["proposals"].to(self.device) for x in batched_inputs] + + if self.use_clip_c4: # use C4 + resnet weights from CLIP + if self.use_clip_attpool: # use att_pool from CLIP to match dimension + results, _ = self.roi_heads(images, features, proposals, None, res5=self.backbone.layer4, attnpool=self.backbone.attnpool) + else: # use default mean pool + results, _ = self.roi_heads(images, features, proposals, None, res5=self.backbone.layer4) + else: # default setting + results, _ = self.roi_heads(images, features, proposals, None) + else: + detected_instances = [x.to(self.device) for x in detected_instances] + + if self.use_clip_c4: # use C4 + resnet weights from CLIP + if self.use_clip_attpool: # use att_pool from CLIP to match dimension + results = self.roi_heads.forward_with_given_boxes(features, detected_instances, res5=self.backbone.layer4, attnpool=self.backbone.attnpool) + else: # use default mean pool + results = self.roi_heads.forward_with_given_boxes(features, detected_instances, res5=self.backbone.layer4) + else: # default setting + results = self.roi_heads.forward_with_given_boxes(features, detected_instances) + + #visualize_proposals(batched_inputs, proposals, self.input_format) + if do_postprocess: + assert not torch.jit.is_scripting(), "Scripting is not supported for postprocess." + return GeneralizedRCNN._postprocess(results, batched_inputs, images.image_sizes) + else: + return results + + def preprocess_image(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Normalize, pad and batch the input images. + """ + images = [x["image"].to(self.device) for x in batched_inputs] + if self.div_pixel: + images = [((x / 255.0) - self.pixel_mean) / self.pixel_std for x in images] + else: + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + images = ImageList.from_tensors(images, self.backbone.size_divisibility) + return images + + @staticmethod + def _postprocess(instances, batched_inputs: List[Dict[str, torch.Tensor]], image_sizes): + """ + Rescale the output instances to the target size. + """ + # note: private function; subject to changes + processed_results = [] + for results_per_image, input_per_image, image_size in zip( + instances, batched_inputs, image_sizes + ): + height = input_per_image.get("height", image_size[0]) + width = input_per_image.get("width", image_size[1]) + r = detector_postprocess(results_per_image, height, width) + processed_results.append({"instances": r}) + return processed_results + + +@META_ARCH_REGISTRY.register() +class ProposalNetwork(nn.Module): + """ + A meta architecture that only predicts object proposals. + """ + + @configurable + def __init__( + self, + *, + backbone: Backbone, + proposal_generator: nn.Module, + pixel_mean: Tuple[float], + pixel_std: Tuple[float], + input_format: Optional[str] = None, + ): + """ + Args: + backbone: a backbone module, must follow detectron2's backbone interface + proposal_generator: a module that generates proposals using backbone features + pixel_mean, pixel_std: list or tuple with #channels element, representing + the per-channel mean and std to be used to normalize the input image + """ + super().__init__() + self.backbone = backbone + self.proposal_generator = proposal_generator + self.register_buffer("pixel_mean", torch.tensor(pixel_mean).view(-1, 1, 1), False) + self.register_buffer("pixel_std", torch.tensor(pixel_std).view(-1, 1, 1), False) + if np.sum(pixel_mean) < 3.0: # converrt pixel value to range [0.0, 1.0] by dividing 255.0 + assert input_format == 'RGB' + self.div_pixel = True + else: # default setting + self.div_pixel = False + + @classmethod + def from_config(cls, cfg): + backbone = build_backbone(cfg) + return { + "backbone": backbone, + "proposal_generator": build_proposal_generator(cfg, backbone.output_shape()), + "input_format": cfg.INPUT.FORMAT, + "pixel_mean": cfg.MODEL.PIXEL_MEAN, + "pixel_std": cfg.MODEL.PIXEL_STD, + } + + @property + def device(self): + return self.pixel_mean.device + + def forward(self, batched_inputs): + """ + Args: + Same as in :class:`GeneralizedRCNN.forward` + + Returns: + list[dict]: + Each dict is the output for one input image. + The dict contains one key "proposals" whose value is a + :class:`Instances` with keys "proposal_boxes" and "objectness_logits". + """ + images = [x["image"].to(self.device) for x in batched_inputs] + if self.div_pixel: + images = [((x / 255.0) - self.pixel_mean) / self.pixel_std for x in images] + else: + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + images = ImageList.from_tensors(images, self.backbone.size_divisibility) + features = self.backbone(images.tensor) + + if "instances" in batched_inputs[0]: + gt_instances = [x["instances"].to(self.device) for x in batched_inputs] + elif "targets" in batched_inputs[0]: + log_first_n( + logging.WARN, "'targets' in the model inputs is now renamed to 'instances'!", n=10 + ) + gt_instances = [x["targets"].to(self.device) for x in batched_inputs] + else: + gt_instances = None + proposals, proposal_losses = self.proposal_generator(images, features, gt_instances) + # In training, the proposals are not useful at all but we generate them anyway. + # This makes RPN-only models about 5% slower. + if self.training: + return proposal_losses + + processed_results = [] + for results_per_image, input_per_image, image_size in zip( + proposals, batched_inputs, images.image_sizes + ): + height = input_per_image.get("height", image_size[0]) + width = input_per_image.get("width", image_size[1]) + r = detector_postprocess(results_per_image, height, width) + processed_results.append({"proposals": r}) + return processed_results diff --git a/detectron2/modeling/meta_arch/retinanet.py b/detectron2/modeling/meta_arch/retinanet.py new file mode 100644 index 0000000000000000000000000000000000000000..81992a3bc6d7f17ab64eb88a157901e69d3f0e16 --- /dev/null +++ b/detectron2/modeling/meta_arch/retinanet.py @@ -0,0 +1,609 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import math +import numpy as np +from typing import Dict, List, Tuple +import torch +from fvcore.nn import sigmoid_focal_loss_jit +from torch import Tensor, nn +from torch.nn import functional as F + +from detectron2.config import configurable +from detectron2.data.detection_utils import convert_image_to_rgb +from detectron2.layers import ShapeSpec, batched_nms, cat, get_norm, nonzero_tuple +from detectron2.structures import Boxes, ImageList, Instances, pairwise_iou +from detectron2.utils.events import get_event_storage + +from ..anchor_generator import build_anchor_generator +from ..backbone import Backbone, build_backbone +from ..box_regression import Box2BoxTransform, _dense_box_regression_loss +from ..matcher import Matcher +from ..postprocessing import detector_postprocess +from .build import META_ARCH_REGISTRY + +__all__ = ["RetinaNet"] + + +logger = logging.getLogger(__name__) + + +def permute_to_N_HWA_K(tensor, K: int): + """ + Transpose/reshape a tensor from (N, (Ai x K), H, W) to (N, (HxWxAi), K) + """ + assert tensor.dim() == 4, tensor.shape + N, _, H, W = tensor.shape + tensor = tensor.view(N, -1, K, H, W) + tensor = tensor.permute(0, 3, 4, 1, 2) + tensor = tensor.reshape(N, -1, K) # Size=(N,HWA,K) + return tensor + + +@META_ARCH_REGISTRY.register() +class RetinaNet(nn.Module): + """ + Implement RetinaNet in :paper:`RetinaNet`. + """ + + @configurable + def __init__( + self, + *, + backbone: Backbone, + head: nn.Module, + head_in_features, + anchor_generator, + box2box_transform, + anchor_matcher, + num_classes, + focal_loss_alpha=0.25, + focal_loss_gamma=2.0, + smooth_l1_beta=0.0, + box_reg_loss_type="smooth_l1", + test_score_thresh=0.05, + test_topk_candidates=1000, + test_nms_thresh=0.5, + max_detections_per_image=100, + pixel_mean, + pixel_std, + vis_period=0, + input_format="BGR", + ): + """ + NOTE: this interface is experimental. + + Args: + backbone: a backbone module, must follow detectron2's backbone interface + head (nn.Module): a module that predicts logits and regression deltas + for each level from a list of per-level features + head_in_features (Tuple[str]): Names of the input feature maps to be used in head + anchor_generator (nn.Module): a module that creates anchors from a + list of features. Usually an instance of :class:`AnchorGenerator` + box2box_transform (Box2BoxTransform): defines the transform from anchors boxes to + instance boxes + anchor_matcher (Matcher): label the anchors by matching them with ground truth. + num_classes (int): number of classes. Used to label background proposals. + + # Loss parameters: + focal_loss_alpha (float): focal_loss_alpha + focal_loss_gamma (float): focal_loss_gamma + smooth_l1_beta (float): smooth_l1_beta + box_reg_loss_type (str): Options are "smooth_l1", "giou" + + # Inference parameters: + test_score_thresh (float): Inference cls score threshold, only anchors with + score > INFERENCE_TH are considered for inference (to improve speed) + test_topk_candidates (int): Select topk candidates before NMS + test_nms_thresh (float): Overlap threshold used for non-maximum suppression + (suppress boxes with IoU >= this threshold) + max_detections_per_image (int): + Maximum number of detections to return per image during inference + (100 is based on the limit established for the COCO dataset). + + # Input parameters + pixel_mean (Tuple[float]): + Values to be used for image normalization (BGR order). + To train on images of different number of channels, set different mean & std. + Default values are the mean pixel value from ImageNet: [103.53, 116.28, 123.675] + pixel_std (Tuple[float]): + When using pre-trained models in Detectron1 or any MSRA models, + std has been absorbed into its conv1 weights, so the std needs to be set 1. + Otherwise, you can use [57.375, 57.120, 58.395] (ImageNet std) + vis_period (int): + The period (in terms of steps) for minibatch visualization at train time. + Set to 0 to disable. + input_format (str): Whether the model needs RGB, YUV, HSV etc. + """ + super().__init__() + + self.backbone = backbone + self.head = head + self.head_in_features = head_in_features + if len(self.backbone.output_shape()) != len(self.head_in_features): + logger.warning("[RetinaNet] Backbone produces unused features.") + + # Anchors + self.anchor_generator = anchor_generator + self.box2box_transform = box2box_transform + self.anchor_matcher = anchor_matcher + + self.num_classes = num_classes + # Loss parameters: + self.focal_loss_alpha = focal_loss_alpha + self.focal_loss_gamma = focal_loss_gamma + self.smooth_l1_beta = smooth_l1_beta + self.box_reg_loss_type = box_reg_loss_type + # Inference parameters: + self.test_score_thresh = test_score_thresh + self.test_topk_candidates = test_topk_candidates + self.test_nms_thresh = test_nms_thresh + self.max_detections_per_image = max_detections_per_image + # Vis parameters + self.vis_period = vis_period + self.input_format = input_format + + self.register_buffer("pixel_mean", torch.tensor(pixel_mean).view(-1, 1, 1), False) + self.register_buffer("pixel_std", torch.tensor(pixel_std).view(-1, 1, 1), False) + + """ + In Detectron1, loss is normalized by number of foreground samples in the batch. + When batch size is 1 per GPU, #foreground has a large variance and + using it lead to lower performance. Here we maintain an EMA of #foreground to + stabilize the normalizer. + """ + self.loss_normalizer = 100 # initialize with any reasonable #fg that's not too small + self.loss_normalizer_momentum = 0.9 + + @classmethod + def from_config(cls, cfg): + backbone = build_backbone(cfg) + backbone_shape = backbone.output_shape() + feature_shapes = [backbone_shape[f] for f in cfg.MODEL.RETINANET.IN_FEATURES] + head = RetinaNetHead(cfg, feature_shapes) + anchor_generator = build_anchor_generator(cfg, feature_shapes) + return { + "backbone": backbone, + "head": head, + "anchor_generator": anchor_generator, + "box2box_transform": Box2BoxTransform(weights=cfg.MODEL.RETINANET.BBOX_REG_WEIGHTS), + "anchor_matcher": Matcher( + cfg.MODEL.RETINANET.IOU_THRESHOLDS, + cfg.MODEL.RETINANET.IOU_LABELS, + allow_low_quality_matches=True, + ), + "pixel_mean": cfg.MODEL.PIXEL_MEAN, + "pixel_std": cfg.MODEL.PIXEL_STD, + "num_classes": cfg.MODEL.RETINANET.NUM_CLASSES, + "head_in_features": cfg.MODEL.RETINANET.IN_FEATURES, + # Loss parameters: + "focal_loss_alpha": cfg.MODEL.RETINANET.FOCAL_LOSS_ALPHA, + "focal_loss_gamma": cfg.MODEL.RETINANET.FOCAL_LOSS_GAMMA, + "smooth_l1_beta": cfg.MODEL.RETINANET.SMOOTH_L1_LOSS_BETA, + "box_reg_loss_type": cfg.MODEL.RETINANET.BBOX_REG_LOSS_TYPE, + # Inference parameters: + "test_score_thresh": cfg.MODEL.RETINANET.SCORE_THRESH_TEST, + "test_topk_candidates": cfg.MODEL.RETINANET.TOPK_CANDIDATES_TEST, + "test_nms_thresh": cfg.MODEL.RETINANET.NMS_THRESH_TEST, + "max_detections_per_image": cfg.TEST.DETECTIONS_PER_IMAGE, + # Vis parameters + "vis_period": cfg.VIS_PERIOD, + "input_format": cfg.INPUT.FORMAT, + } + + @property + def device(self): + return self.pixel_mean.device + + def visualize_training(self, batched_inputs, results): + """ + A function used to visualize ground truth images and final network predictions. + It shows ground truth bounding boxes on the original image and up to 20 + predicted object bounding boxes on the original image. + + Args: + batched_inputs (list): a list that contains input to the model. + results (List[Instances]): a list of #images elements. + """ + from detectron2.utils.visualizer import Visualizer + + assert len(batched_inputs) == len( + results + ), "Cannot visualize inputs and results of different sizes" + storage = get_event_storage() + max_boxes = 20 + + image_index = 0 # only visualize a single image + img = batched_inputs[image_index]["image"] + img = convert_image_to_rgb(img.permute(1, 2, 0), self.input_format) + v_gt = Visualizer(img, None) + v_gt = v_gt.overlay_instances(boxes=batched_inputs[image_index]["instances"].gt_boxes) + anno_img = v_gt.get_image() + processed_results = detector_postprocess(results[image_index], img.shape[0], img.shape[1]) + predicted_boxes = processed_results.pred_boxes.tensor.detach().cpu().numpy() + + v_pred = Visualizer(img, None) + v_pred = v_pred.overlay_instances(boxes=predicted_boxes[0:max_boxes]) + prop_img = v_pred.get_image() + vis_img = np.vstack((anno_img, prop_img)) + vis_img = vis_img.transpose(2, 0, 1) + vis_name = f"Top: GT bounding boxes; Bottom: {max_boxes} Highest Scoring Results" + storage.put_image(vis_name, vis_img) + + def forward(self, batched_inputs: List[Dict[str, Tensor]]): + """ + Args: + batched_inputs: a list, batched outputs of :class:`DatasetMapper` . + Each item in the list contains the inputs for one image. + For now, each item in the list is a dict that contains: + + * image: Tensor, image in (C, H, W) format. + * instances: Instances + + Other information that's included in the original dicts, such as: + + * "height", "width" (int): the output resolution of the model, used in inference. + See :meth:`postprocess` for details. + Returns: + In training, dict[str, Tensor]: mapping from a named loss to a tensor storing the + loss. Used during training only. In inference, the standard output format, described + in :doc:`/tutorials/models`. + """ + images = self.preprocess_image(batched_inputs) + features = self.backbone(images.tensor) + features = [features[f] for f in self.head_in_features] + + anchors = self.anchor_generator(features) + pred_logits, pred_anchor_deltas = self.head(features) + # Transpose the Hi*Wi*A dimension to the middle: + pred_logits = [permute_to_N_HWA_K(x, self.num_classes) for x in pred_logits] + pred_anchor_deltas = [permute_to_N_HWA_K(x, 4) for x in pred_anchor_deltas] + + if self.training: + assert not torch.jit.is_scripting(), "Not supported" + assert "instances" in batched_inputs[0], "Instance annotations are missing in training!" + gt_instances = [x["instances"].to(self.device) for x in batched_inputs] + + gt_labels, gt_boxes = self.label_anchors(anchors, gt_instances) + losses = self.losses(anchors, pred_logits, gt_labels, pred_anchor_deltas, gt_boxes) + + if self.vis_period > 0: + storage = get_event_storage() + if storage.iter % self.vis_period == 0: + results = self.inference( + anchors, pred_logits, pred_anchor_deltas, images.image_sizes + ) + self.visualize_training(batched_inputs, results) + + return losses + else: + results = self.inference(anchors, pred_logits, pred_anchor_deltas, images.image_sizes) + if torch.jit.is_scripting(): + return results + processed_results = [] + for results_per_image, input_per_image, image_size in zip( + results, batched_inputs, images.image_sizes + ): + height = input_per_image.get("height", image_size[0]) + width = input_per_image.get("width", image_size[1]) + r = detector_postprocess(results_per_image, height, width) + processed_results.append({"instances": r}) + return processed_results + + def losses(self, anchors, pred_logits, gt_labels, pred_anchor_deltas, gt_boxes): + """ + Args: + anchors (list[Boxes]): a list of #feature level Boxes + gt_labels, gt_boxes: see output of :meth:`RetinaNet.label_anchors`. + Their shapes are (N, R) and (N, R, 4), respectively, where R is + the total number of anchors across levels, i.e. sum(Hi x Wi x Ai) + pred_logits, pred_anchor_deltas: both are list[Tensor]. Each element in the + list corresponds to one level and has shape (N, Hi * Wi * Ai, K or 4). + Where K is the number of classes used in `pred_logits`. + + Returns: + dict[str, Tensor]: + mapping from a named loss to a scalar tensor + storing the loss. Used during training only. The dict keys are: + "loss_cls" and "loss_box_reg" + """ + num_images = len(gt_labels) + gt_labels = torch.stack(gt_labels) # (N, R) + + valid_mask = gt_labels >= 0 + pos_mask = (gt_labels >= 0) & (gt_labels != self.num_classes) + num_pos_anchors = pos_mask.sum().item() + get_event_storage().put_scalar("num_pos_anchors", num_pos_anchors / num_images) + self.loss_normalizer = self.loss_normalizer_momentum * self.loss_normalizer + ( + 1 - self.loss_normalizer_momentum + ) * max(num_pos_anchors, 1) + + # classification and regression loss + gt_labels_target = F.one_hot(gt_labels[valid_mask], num_classes=self.num_classes + 1)[ + :, :-1 + ] # no loss for the last (background) class + loss_cls = sigmoid_focal_loss_jit( + cat(pred_logits, dim=1)[valid_mask], + gt_labels_target.to(pred_logits[0].dtype), + alpha=self.focal_loss_alpha, + gamma=self.focal_loss_gamma, + reduction="sum", + ) + + loss_box_reg = _dense_box_regression_loss( + anchors, + self.box2box_transform, + pred_anchor_deltas, + gt_boxes, + pos_mask, + box_reg_loss_type=self.box_reg_loss_type, + smooth_l1_beta=self.smooth_l1_beta, + ) + + return { + "loss_cls": loss_cls / self.loss_normalizer, + "loss_box_reg": loss_box_reg / self.loss_normalizer, + } + + @torch.no_grad() + def label_anchors(self, anchors, gt_instances): + """ + Args: + anchors (list[Boxes]): A list of #feature level Boxes. + The Boxes contains anchors of this image on the specific feature level. + gt_instances (list[Instances]): a list of N `Instances`s. The i-th + `Instances` contains the ground-truth per-instance annotations + for the i-th input image. + + Returns: + list[Tensor]: List of #img tensors. i-th element is a vector of labels whose length is + the total number of anchors across all feature maps (sum(Hi * Wi * A)). + Label values are in {-1, 0, ..., K}, with -1 means ignore, and K means background. + + list[Tensor]: i-th element is a Rx4 tensor, where R is the total number of anchors + across feature maps. The values are the matched gt boxes for each anchor. + Values are undefined for those anchors not labeled as foreground. + """ + anchors = Boxes.cat(anchors) # Rx4 + + gt_labels = [] + matched_gt_boxes = [] + for gt_per_image in gt_instances: + match_quality_matrix = pairwise_iou(gt_per_image.gt_boxes, anchors) + matched_idxs, anchor_labels = self.anchor_matcher(match_quality_matrix) + del match_quality_matrix + + if len(gt_per_image) > 0: + matched_gt_boxes_i = gt_per_image.gt_boxes.tensor[matched_idxs] + + gt_labels_i = gt_per_image.gt_classes[matched_idxs] + # Anchors with label 0 are treated as background. + gt_labels_i[anchor_labels == 0] = self.num_classes + # Anchors with label -1 are ignored. + gt_labels_i[anchor_labels == -1] = -1 + else: + matched_gt_boxes_i = torch.zeros_like(anchors.tensor) + gt_labels_i = torch.zeros_like(matched_idxs) + self.num_classes + + gt_labels.append(gt_labels_i) + matched_gt_boxes.append(matched_gt_boxes_i) + + return gt_labels, matched_gt_boxes + + def inference( + self, + anchors: List[Boxes], + pred_logits: List[Tensor], + pred_anchor_deltas: List[Tensor], + image_sizes: List[Tuple[int, int]], + ): + """ + Arguments: + anchors (list[Boxes]): A list of #feature level Boxes. + The Boxes contain anchors of this image on the specific feature level. + pred_logits, pred_anchor_deltas: list[Tensor], one per level. Each + has shape (N, Hi * Wi * Ai, K or 4) + image_sizes (List[(h, w)]): the input image sizes + + Returns: + results (List[Instances]): a list of #images elements. + """ + results: List[Instances] = [] + for img_idx, image_size in enumerate(image_sizes): + pred_logits_per_image = [x[img_idx] for x in pred_logits] + deltas_per_image = [x[img_idx] for x in pred_anchor_deltas] + results_per_image = self.inference_single_image( + anchors, pred_logits_per_image, deltas_per_image, image_size + ) + results.append(results_per_image) + return results + + def inference_single_image( + self, + anchors: List[Boxes], + box_cls: List[Tensor], + box_delta: List[Tensor], + image_size: Tuple[int, int], + ): + """ + Single-image inference. Return bounding-box detection results by thresholding + on scores and applying non-maximum suppression (NMS). + + Arguments: + anchors (list[Boxes]): list of #feature levels. Each entry contains + a Boxes object, which contains all the anchors in that feature level. + box_cls (list[Tensor]): list of #feature levels. Each entry contains + tensor of size (H x W x A, K) + box_delta (list[Tensor]): Same shape as 'box_cls' except that K becomes 4. + image_size (tuple(H, W)): a tuple of the image height and width. + + Returns: + Same as `inference`, but for only one image. + """ + boxes_all = [] + scores_all = [] + class_idxs_all = [] + + # Iterate over every feature level + for box_cls_i, box_reg_i, anchors_i in zip(box_cls, box_delta, anchors): + # (HxWxAxK,) + predicted_prob = box_cls_i.flatten().sigmoid_() + + # Apply two filtering below to make NMS faster. + # 1. Keep boxes with confidence score higher than threshold + keep_idxs = predicted_prob > self.test_score_thresh + predicted_prob = predicted_prob[keep_idxs] + topk_idxs = nonzero_tuple(keep_idxs)[0] + + # 2. Keep top k top scoring boxes only + num_topk = min(self.test_topk_candidates, topk_idxs.size(0)) + # torch.sort is actually faster than .topk (at least on GPUs) + predicted_prob, idxs = predicted_prob.sort(descending=True) + predicted_prob = predicted_prob[:num_topk] + topk_idxs = topk_idxs[idxs[:num_topk]] + + anchor_idxs = topk_idxs // self.num_classes + classes_idxs = topk_idxs % self.num_classes + + box_reg_i = box_reg_i[anchor_idxs] + anchors_i = anchors_i[anchor_idxs] + # predict boxes + predicted_boxes = self.box2box_transform.apply_deltas(box_reg_i, anchors_i.tensor) + + boxes_all.append(predicted_boxes) + scores_all.append(predicted_prob) + class_idxs_all.append(classes_idxs) + + boxes_all, scores_all, class_idxs_all = [ + cat(x) for x in [boxes_all, scores_all, class_idxs_all] + ] + keep = batched_nms(boxes_all, scores_all, class_idxs_all, self.test_nms_thresh) + keep = keep[: self.max_detections_per_image] + + result = Instances(image_size) + result.pred_boxes = Boxes(boxes_all[keep]) + result.scores = scores_all[keep] + result.pred_classes = class_idxs_all[keep] + return result + + def preprocess_image(self, batched_inputs: List[Dict[str, Tensor]]): + """ + Normalize, pad and batch the input images. + """ + images = [x["image"].to(self.device) for x in batched_inputs] + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + images = ImageList.from_tensors(images, self.backbone.size_divisibility) + return images + + +class RetinaNetHead(nn.Module): + """ + The head used in RetinaNet for object classification and box regression. + It has two subnets for the two tasks, with a common structure but separate parameters. + """ + + @configurable + def __init__( + self, + *, + input_shape: List[ShapeSpec], + num_classes, + num_anchors, + conv_dims: List[int], + norm="", + prior_prob=0.01, + ): + """ + NOTE: this interface is experimental. + + Args: + input_shape (List[ShapeSpec]): input shape + num_classes (int): number of classes. Used to label background proposals. + num_anchors (int): number of generated anchors + conv_dims (List[int]): dimensions for each convolution layer + norm (str or callable): + Normalization for conv layers except for the two output layers. + See :func:`detectron2.layers.get_norm` for supported types. + prior_prob (float): Prior weight for computing bias + """ + super().__init__() + + if norm == "BN" or norm == "SyncBN": + logger.warning("Shared norm does not work well for BN, SyncBN, expect poor results") + + cls_subnet = [] + bbox_subnet = [] + for in_channels, out_channels in zip( + [input_shape[0].channels] + list(conv_dims), conv_dims + ): + cls_subnet.append( + nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) + ) + if norm: + cls_subnet.append(get_norm(norm, out_channels)) + cls_subnet.append(nn.ReLU()) + bbox_subnet.append( + nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) + ) + if norm: + bbox_subnet.append(get_norm(norm, out_channels)) + bbox_subnet.append(nn.ReLU()) + + self.cls_subnet = nn.Sequential(*cls_subnet) + self.bbox_subnet = nn.Sequential(*bbox_subnet) + self.cls_score = nn.Conv2d( + conv_dims[-1], num_anchors * num_classes, kernel_size=3, stride=1, padding=1 + ) + self.bbox_pred = nn.Conv2d( + conv_dims[-1], num_anchors * 4, kernel_size=3, stride=1, padding=1 + ) + + # Initialization + for modules in [self.cls_subnet, self.bbox_subnet, self.cls_score, self.bbox_pred]: + for layer in modules.modules(): + if isinstance(layer, nn.Conv2d): + torch.nn.init.normal_(layer.weight, mean=0, std=0.01) + torch.nn.init.constant_(layer.bias, 0) + + # Use prior in model initialization to improve stability + bias_value = -(math.log((1 - prior_prob) / prior_prob)) + torch.nn.init.constant_(self.cls_score.bias, bias_value) + + @classmethod + def from_config(cls, cfg, input_shape: List[ShapeSpec]): + num_anchors = build_anchor_generator(cfg, input_shape).num_cell_anchors + assert ( + len(set(num_anchors)) == 1 + ), "Using different number of anchors between levels is not currently supported!" + num_anchors = num_anchors[0] + + return { + "input_shape": input_shape, + "num_classes": cfg.MODEL.RETINANET.NUM_CLASSES, + "conv_dims": [input_shape[0].channels] * cfg.MODEL.RETINANET.NUM_CONVS, + "prior_prob": cfg.MODEL.RETINANET.PRIOR_PROB, + "norm": cfg.MODEL.RETINANET.NORM, + "num_anchors": num_anchors, + } + + def forward(self, features: List[Tensor]): + """ + Arguments: + features (list[Tensor]): FPN feature map tensors in high to low resolution. + Each tensor in the list correspond to different feature levels. + + Returns: + logits (list[Tensor]): #lvl tensors, each has shape (N, AxK, Hi, Wi). + The tensor predicts the classification probability + at each spatial position for each of the A anchors and K object + classes. + bbox_reg (list[Tensor]): #lvl tensors, each has shape (N, Ax4, Hi, Wi). + The tensor predicts 4-vector (dx,dy,dw,dh) box + regression values for every anchor. These values are the + relative offset between the anchor and the ground truth box. + """ + logits = [] + bbox_reg = [] + for feature in features: + logits.append(self.cls_score(self.cls_subnet(feature))) + bbox_reg.append(self.bbox_pred(self.bbox_subnet(feature))) + return logits, bbox_reg diff --git a/detectron2/modeling/meta_arch/semantic_seg.py b/detectron2/modeling/meta_arch/semantic_seg.py new file mode 100644 index 0000000000000000000000000000000000000000..7db8410c26c9809b5f13e4681ca5eca64afc8dca --- /dev/null +++ b/detectron2/modeling/meta_arch/semantic_seg.py @@ -0,0 +1,250 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import numpy as np +from typing import Callable, Dict, Optional, Tuple, Union +import fvcore.nn.weight_init as weight_init +import torch +from torch import nn +from torch.nn import functional as F + +from detectron2.config import configurable +from detectron2.layers import Conv2d, ShapeSpec, get_norm +from detectron2.structures import ImageList +from detectron2.utils.registry import Registry + +from ..backbone import Backbone, build_backbone +from ..postprocessing import sem_seg_postprocess +from .build import META_ARCH_REGISTRY + +__all__ = ["SemanticSegmentor", "SEM_SEG_HEADS_REGISTRY", "SemSegFPNHead", "build_sem_seg_head"] + + +SEM_SEG_HEADS_REGISTRY = Registry("SEM_SEG_HEADS") +SEM_SEG_HEADS_REGISTRY.__doc__ = """ +Registry for semantic segmentation heads, which make semantic segmentation predictions +from feature maps. +""" + + +@META_ARCH_REGISTRY.register() +class SemanticSegmentor(nn.Module): + """ + Main class for semantic segmentation architectures. + """ + + @configurable + def __init__( + self, + *, + backbone: Backbone, + sem_seg_head: nn.Module, + pixel_mean: Tuple[float], + pixel_std: Tuple[float], + ): + """ + Args: + backbone: a backbone module, must follow detectron2's backbone interface + sem_seg_head: a module that predicts semantic segmentation from backbone features + pixel_mean, pixel_std: list or tuple with #channels element, representing + the per-channel mean and std to be used to normalize the input image + """ + super().__init__() + self.backbone = backbone + self.sem_seg_head = sem_seg_head + self.register_buffer("pixel_mean", torch.tensor(pixel_mean).view(-1, 1, 1), False) + self.register_buffer("pixel_std", torch.tensor(pixel_std).view(-1, 1, 1), False) + + @classmethod + def from_config(cls, cfg): + backbone = build_backbone(cfg) + sem_seg_head = build_sem_seg_head(cfg, backbone.output_shape()) + return { + "backbone": backbone, + "sem_seg_head": sem_seg_head, + "pixel_mean": cfg.MODEL.PIXEL_MEAN, + "pixel_std": cfg.MODEL.PIXEL_STD, + } + + @property + def device(self): + return self.pixel_mean.device + + def forward(self, batched_inputs): + """ + Args: + batched_inputs: a list, batched outputs of :class:`DatasetMapper`. + Each item in the list contains the inputs for one image. + + For now, each item in the list is a dict that contains: + + * "image": Tensor, image in (C, H, W) format. + * "sem_seg": semantic segmentation ground truth + * Other information that's included in the original dicts, such as: + "height", "width" (int): the output resolution of the model (may be different + from input resolution), used in inference. + + + Returns: + list[dict]: + Each dict is the output for one input image. + The dict contains one key "sem_seg" whose value is a + Tensor that represents the + per-pixel segmentation prediced by the head. + The prediction has shape KxHxW that represents the logits of + each class for each pixel. + """ + images = [x["image"].to(self.device) for x in batched_inputs] + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + images = ImageList.from_tensors(images, self.backbone.size_divisibility) + + features = self.backbone(images.tensor) + + if "sem_seg" in batched_inputs[0]: + targets = [x["sem_seg"].to(self.device) for x in batched_inputs] + targets = ImageList.from_tensors( + targets, self.backbone.size_divisibility, self.sem_seg_head.ignore_value + ).tensor + else: + targets = None + results, losses = self.sem_seg_head(features, targets) + + if self.training: + return losses + + processed_results = [] + for result, input_per_image, image_size in zip(results, batched_inputs, images.image_sizes): + height = input_per_image.get("height") + width = input_per_image.get("width") + r = sem_seg_postprocess(result, image_size, height, width) + processed_results.append({"sem_seg": r}) + return processed_results + + +def build_sem_seg_head(cfg, input_shape): + """ + Build a semantic segmentation head from `cfg.MODEL.SEM_SEG_HEAD.NAME`. + """ + name = cfg.MODEL.SEM_SEG_HEAD.NAME + return SEM_SEG_HEADS_REGISTRY.get(name)(cfg, input_shape) + + +@SEM_SEG_HEADS_REGISTRY.register() +class SemSegFPNHead(nn.Module): + """ + A semantic segmentation head described in :paper:`PanopticFPN`. + It takes a list of FPN features as input, and applies a sequence of + 3x3 convs and upsampling to scale all of them to the stride defined by + ``common_stride``. Then these features are added and used to make final + predictions by another 1x1 conv layer. + """ + + @configurable + def __init__( + self, + input_shape: Dict[str, ShapeSpec], + *, + num_classes: int, + conv_dims: int, + common_stride: int, + loss_weight: float = 1.0, + norm: Optional[Union[str, Callable]] = None, + ignore_value: int = -1, + ): + """ + NOTE: this interface is experimental. + + Args: + input_shape: shapes (channels and stride) of the input features + num_classes: number of classes to predict + conv_dims: number of output channels for the intermediate conv layers. + common_stride: the common stride that all features will be upscaled to + loss_weight: loss weight + norm (str or callable): normalization for all conv layers + ignore_value: category id to be ignored during training. + """ + super().__init__() + input_shape = sorted(input_shape.items(), key=lambda x: x[1].stride) + self.in_features = [k for k, v in input_shape] + feature_strides = [v.stride for k, v in input_shape] + feature_channels = [v.channels for k, v in input_shape] + + self.ignore_value = ignore_value + self.common_stride = common_stride + self.loss_weight = loss_weight + + self.scale_heads = [] + for in_feature, stride, channels in zip( + self.in_features, feature_strides, feature_channels + ): + head_ops = [] + head_length = max(1, int(np.log2(stride) - np.log2(self.common_stride))) + for k in range(head_length): + norm_module = get_norm(norm, conv_dims) + conv = Conv2d( + channels if k == 0 else conv_dims, + conv_dims, + kernel_size=3, + stride=1, + padding=1, + bias=not norm, + norm=norm_module, + activation=F.relu, + ) + weight_init.c2_msra_fill(conv) + head_ops.append(conv) + if stride != self.common_stride: + head_ops.append( + nn.Upsample(scale_factor=2, mode="bilinear", align_corners=False) + ) + self.scale_heads.append(nn.Sequential(*head_ops)) + self.add_module(in_feature, self.scale_heads[-1]) + self.predictor = Conv2d(conv_dims, num_classes, kernel_size=1, stride=1, padding=0) + weight_init.c2_msra_fill(self.predictor) + + @classmethod + def from_config(cls, cfg, input_shape: Dict[str, ShapeSpec]): + return { + "input_shape": { + k: v for k, v in input_shape.items() if k in cfg.MODEL.SEM_SEG_HEAD.IN_FEATURES + }, + "ignore_value": cfg.MODEL.SEM_SEG_HEAD.IGNORE_VALUE, + "num_classes": cfg.MODEL.SEM_SEG_HEAD.NUM_CLASSES, + "conv_dims": cfg.MODEL.SEM_SEG_HEAD.CONVS_DIM, + "common_stride": cfg.MODEL.SEM_SEG_HEAD.COMMON_STRIDE, + "norm": cfg.MODEL.SEM_SEG_HEAD.NORM, + "loss_weight": cfg.MODEL.SEM_SEG_HEAD.LOSS_WEIGHT, + } + + def forward(self, features, targets=None): + """ + Returns: + In training, returns (None, dict of losses) + In inference, returns (CxHxW logits, {}) + """ + x = self.layers(features) + if self.training: + return None, self.losses(x, targets) + else: + x = F.interpolate( + x, scale_factor=self.common_stride, mode="bilinear", align_corners=False + ) + return x, {} + + def layers(self, features): + for i, f in enumerate(self.in_features): + if i == 0: + x = self.scale_heads[i](features[f]) + else: + x = x + self.scale_heads[i](features[f]) + x = self.predictor(x) + return x + + def losses(self, predictions, targets): + predictions = predictions.float() # https://github.com/pytorch/pytorch/issues/48163 + predictions = F.interpolate( + predictions, scale_factor=self.common_stride, mode="bilinear", align_corners=False + ) + loss = F.cross_entropy( + predictions, targets, reduction="mean", ignore_index=self.ignore_value + ) + losses = {"loss_sem_seg": loss * self.loss_weight} + return losses diff --git a/detectron2/modeling/mmdet_wrapper.py b/detectron2/modeling/mmdet_wrapper.py new file mode 100644 index 0000000000000000000000000000000000000000..df0fe2c6f070e7a63c1b9a0464ce81e9160c12b3 --- /dev/null +++ b/detectron2/modeling/mmdet_wrapper.py @@ -0,0 +1,277 @@ +# -*- coding: utf-8 -*- + +import itertools +import logging +import numpy as np +from collections import OrderedDict +from collections.abc import Mapping +from typing import Dict, List, Optional, Tuple, Union +import torch +from omegaconf import DictConfig, OmegaConf +from torch import Tensor, nn + +from detectron2.layers import ShapeSpec +from detectron2.structures import BitMasks, Boxes, ImageList, Instances +from detectron2.utils.events import get_event_storage + +from .backbone import Backbone + +logger = logging.getLogger(__name__) + + +def _to_container(cfg): + """ + mmdet will assert the type of dict/list. + So convert omegaconf objects to dict/list. + """ + if isinstance(cfg, DictConfig): + cfg = OmegaConf.to_container(cfg, resolve=True) + from mmcv.utils import ConfigDict + + return ConfigDict(cfg) + + +class MMDetBackbone(Backbone): + """ + Wrapper of mmdetection backbones to use in detectron2. + + mmdet backbones produce list/tuple of tensors, while detectron2 backbones + produce a dict of tensors. This class wraps the given backbone to produce + output in detectron2's convention, so it can be used in place of detectron2 + backbones. + """ + + def __init__( + self, + backbone: Union[nn.Module, Mapping], + neck: Union[nn.Module, Mapping, None] = None, + *, + pretrained_backbone: Optional[str] = None, + output_shapes: List[ShapeSpec], + output_names: Optional[List[str]] = None, + ): + """ + Args: + backbone: either a backbone module or a mmdet config dict that defines a + backbone. The backbone takes a 4D image tensor and returns a + sequence of tensors. + neck: either a backbone module or a mmdet config dict that defines a + neck. The neck takes outputs of backbone and returns a + sequence of tensors. If None, no neck is used. + pretrained_backbone: defines the backbone weights that can be loaded by + mmdet, such as "torchvision://resnet50". + output_shapes: shape for every output of the backbone (or neck, if given). + stride and channels are often needed. + output_names: names for every output of the backbone (or neck, if given). + By default, will use "out0", "out1", ... + """ + super().__init__() + if isinstance(backbone, Mapping): + from mmdet.models import build_backbone + + backbone = build_backbone(_to_container(backbone)) + self.backbone = backbone + + if isinstance(neck, Mapping): + from mmdet.models import build_neck + + neck = build_neck(_to_container(neck)) + self.neck = neck + + # It's confusing that backbone weights are given as a separate argument, + # but "neck" weights, if any, are part of neck itself. This is the interface + # of mmdet so we follow it. Reference: + # https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/detectors/two_stage.py + logger.info(f"Initializing mmdet backbone weights: {pretrained_backbone} ...") + self.backbone.init_weights(pretrained_backbone) + # train() in mmdet modules is non-trivial, and has to be explicitly + # called. Reference: + # https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/backbones/resnet.py + self.backbone.train() + if self.neck is not None: + logger.info("Initializing mmdet neck weights ...") + if isinstance(self.neck, nn.Sequential): + for m in self.neck: + m.init_weights() + else: + self.neck.init_weights() + self.neck.train() + + self._output_shapes = output_shapes + if not output_names: + output_names = [f"out{i}" for i in range(len(output_shapes))] + self._output_names = output_names + + def forward(self, x) -> Dict[str, Tensor]: + outs = self.backbone(x) + if self.neck is not None: + outs = self.neck(outs) + assert isinstance( + outs, (list, tuple) + ), "mmdet backbone should return a list/tuple of tensors!" + if len(outs) != len(self._output_shapes): + raise ValueError( + "Length of output_shapes does not match outputs from the mmdet backbone: " + f"{len(outs)} != {len(self._output_shapes)}" + ) + return {k: v for k, v in zip(self._output_names, outs)} + + def output_shape(self) -> Dict[str, ShapeSpec]: + return {k: v for k, v in zip(self._output_names, self._output_shapes)} + + +class MMDetDetector(nn.Module): + """ + Wrapper of a mmdetection detector model, for detection and instance segmentation. + Input/output formats of this class follow detectron2's convention, so a + mmdetection model can be trained and evaluated in detectron2. + """ + + def __init__( + self, + detector: Union[nn.Module, Mapping], + *, + # Default is 32 regardless of model: + # https://github.com/open-mmlab/mmdetection/tree/master/configs/_base_/datasets + size_divisibility=32, + pixel_mean: Tuple[float], + pixel_std: Tuple[float], + ): + """ + Args: + detector: a mmdet detector, or a mmdet config dict that defines a detector. + size_divisibility: pad input images to multiple of this number + pixel_mean: per-channel mean to normalize input image + pixel_std: per-channel stddev to normalize input image + """ + super().__init__() + if isinstance(detector, Mapping): + from mmdet.models import build_detector + + detector = build_detector(_to_container(detector)) + self.detector = detector + self.size_divisibility = size_divisibility + + self.register_buffer("pixel_mean", torch.tensor(pixel_mean).view(-1, 1, 1), False) + self.register_buffer("pixel_std", torch.tensor(pixel_std).view(-1, 1, 1), False) + assert ( + self.pixel_mean.shape == self.pixel_std.shape + ), f"{self.pixel_mean} and {self.pixel_std} have different shapes!" + + def forward(self, batched_inputs: List[Dict[str, torch.Tensor]]): + images = [x["image"].to(self.device) for x in batched_inputs] + images = [(x - self.pixel_mean) / self.pixel_std for x in images] + images = ImageList.from_tensors(images, size_divisibility=self.size_divisibility).tensor + metas = [] + rescale = {"height" in x for x in batched_inputs} + if len(rescale) != 1: + raise ValueError("Some inputs have original height/width, but some don't!") + rescale = list(rescale)[0] + output_shapes = [] + for input in batched_inputs: + meta = {} + c, h, w = input["image"].shape + meta["img_shape"] = meta["ori_shape"] = (h, w, c) + if rescale: + scale_factor = np.array( + [w / input["width"], h / input["height"]] * 2, dtype="float32" + ) + ori_shape = (input["height"], input["width"]) + output_shapes.append(ori_shape) + meta["ori_shape"] = ori_shape + (c,) + else: + scale_factor = 1.0 + output_shapes.append((h, w)) + meta["scale_factor"] = scale_factor + meta["flip"] = False + padh, padw = images.shape[-2:] + meta["pad_shape"] = (padh, padw, c) + metas.append(meta) + + if self.training: + gt_instances = [x["instances"].to(self.device) for x in batched_inputs] + if gt_instances[0].has("gt_masks"): + from mmdet.core import PolygonMasks as mm_PolygonMasks, BitmapMasks as mm_BitMasks + + def convert_mask(m, shape): + # mmdet mask format + if isinstance(m, BitMasks): + return mm_BitMasks(m.tensor.cpu().numpy(), shape[0], shape[1]) + else: + return mm_PolygonMasks(m.polygons, shape[0], shape[1]) + + gt_masks = [convert_mask(x.gt_masks, x.image_size) for x in gt_instances] + losses_and_metrics = self.detector.forward_train( + images, + metas, + [x.gt_boxes.tensor for x in gt_instances], + [x.gt_classes for x in gt_instances], + gt_masks=gt_masks, + ) + else: + losses_and_metrics = self.detector.forward_train( + images, + metas, + [x.gt_boxes.tensor for x in gt_instances], + [x.gt_classes for x in gt_instances], + ) + return _parse_losses(losses_and_metrics) + else: + results = self.detector.simple_test(images, metas, rescale=rescale) + results = [ + {"instances": _convert_mmdet_result(r, shape)} + for r, shape in zip(results, output_shapes) + ] + return results + + @property + def device(self): + return self.pixel_mean.device + + +# Reference: show_result() in +# https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/detectors/base.py +def _convert_mmdet_result(result, shape: Tuple[int, int]) -> Instances: + if isinstance(result, tuple): + bbox_result, segm_result = result + if isinstance(segm_result, tuple): + segm_result = segm_result[0] + else: + bbox_result, segm_result = result, None + + bboxes = torch.from_numpy(np.vstack(bbox_result)) # Nx5 + bboxes, scores = bboxes[:, :4], bboxes[:, -1] + labels = [ + torch.full((bbox.shape[0],), i, dtype=torch.int32) for i, bbox in enumerate(bbox_result) + ] + labels = torch.cat(labels) + inst = Instances(shape) + inst.pred_boxes = Boxes(bboxes) + inst.scores = scores + inst.pred_classes = labels + + if segm_result is not None and len(labels) > 0: + segm_result = list(itertools.chain(*segm_result)) + segm_result = [torch.from_numpy(x) if isinstance(x, np.ndarray) else x for x in segm_result] + segm_result = torch.stack(segm_result, dim=0) + inst.pred_masks = segm_result + return inst + + +# reference: https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/detectors/base.py +def _parse_losses(losses: Dict[str, Tensor]) -> Dict[str, Tensor]: + log_vars = OrderedDict() + for loss_name, loss_value in losses.items(): + if isinstance(loss_value, torch.Tensor): + log_vars[loss_name] = loss_value.mean() + elif isinstance(loss_value, list): + log_vars[loss_name] = sum(_loss.mean() for _loss in loss_value) + else: + raise TypeError(f"{loss_name} is not a tensor or list of tensors") + + if "loss" not in loss_name: + # put metrics to storage; don't return them + storage = get_event_storage() + value = log_vars.pop(loss_name).cpu().item() + storage.put_scalar(loss_name, value) + return log_vars diff --git a/detectron2/modeling/poolers.py b/detectron2/modeling/poolers.py new file mode 100644 index 0000000000000000000000000000000000000000..e5d72abf462ebc9c2ac9ad9dd7c6cc39eac4054c --- /dev/null +++ b/detectron2/modeling/poolers.py @@ -0,0 +1,250 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import math +from typing import List +import torch +from torch import nn +from torchvision.ops import RoIPool + +from detectron2.layers import ROIAlign, ROIAlignRotated, cat, nonzero_tuple +from detectron2.structures import Boxes + +""" +To export ROIPooler to torchscript, in this file, variables that should be annotated with +`Union[List[Boxes], List[RotatedBoxes]]` are only annotated with `List[Boxes]`. + +TODO: Correct these annotations when torchscript support `Union`. +https://github.com/pytorch/pytorch/issues/41412 +""" + +__all__ = ["ROIPooler"] + + +def assign_boxes_to_levels( + box_lists: List[Boxes], + min_level: int, + max_level: int, + canonical_box_size: int, + canonical_level: int, +): + """ + Map each box in `box_lists` to a feature map level index and return the assignment + vector. + + Args: + box_lists (list[Boxes] | list[RotatedBoxes]): A list of N Boxes or N RotatedBoxes, + where N is the number of images in the batch. + min_level (int): Smallest feature map level index. The input is considered index 0, + the output of stage 1 is index 1, and so. + max_level (int): Largest feature map level index. + canonical_box_size (int): A canonical box size in pixels (sqrt(box area)). + canonical_level (int): The feature map level index on which a canonically-sized box + should be placed. + + Returns: + A tensor of length M, where M is the total number of boxes aggregated over all + N batch images. The memory layout corresponds to the concatenation of boxes + from all images. Each element is the feature map index, as an offset from + `self.min_level`, for the corresponding box (so value i means the box is at + `self.min_level + i`). + """ + box_sizes = torch.sqrt(cat([boxes.area() for boxes in box_lists])) + # Eqn.(1) in FPN paper + level_assignments = torch.floor( + canonical_level + torch.log2(box_sizes / canonical_box_size + 1e-8) + ) + # clamp level to (min, max), in case the box size is too large or too small + # for the available feature maps + level_assignments = torch.clamp(level_assignments, min=min_level, max=max_level) + return level_assignments.to(torch.int64) - min_level + + +def _fmt_box_list(box_tensor, batch_index: int): + repeated_index = torch.full_like( + box_tensor[:, :1], batch_index, dtype=box_tensor.dtype, device=box_tensor.device + ) + return cat((repeated_index, box_tensor), dim=1) + + +def convert_boxes_to_pooler_format(box_lists: List[Boxes]): + """ + Convert all boxes in `box_lists` to the low-level format used by ROI pooling ops + (see description under Returns). + + Args: + box_lists (list[Boxes] | list[RotatedBoxes]): + A list of N Boxes or N RotatedBoxes, where N is the number of images in the batch. + + Returns: + When input is list[Boxes]: + A tensor of shape (M, 5), where M is the total number of boxes aggregated over all + N batch images. + The 5 columns are (batch index, x0, y0, x1, y1), where batch index + is the index in [0, N) identifying which batch image the box with corners at + (x0, y0, x1, y1) comes from. + When input is list[RotatedBoxes]: + A tensor of shape (M, 6), where M is the total number of boxes aggregated over all + N batch images. + The 6 columns are (batch index, x_ctr, y_ctr, width, height, angle_degrees), + where batch index is the index in [0, N) identifying which batch image the + rotated box (x_ctr, y_ctr, width, height, angle_degrees) comes from. + """ + pooler_fmt_boxes = cat( + [_fmt_box_list(box_list.tensor, i) for i, box_list in enumerate(box_lists)], dim=0 + ) + + return pooler_fmt_boxes + + +class ROIPooler(nn.Module): + """ + Region of interest feature map pooler that supports pooling from one or more + feature maps. + """ + + def __init__( + self, + output_size, + scales, + sampling_ratio, + pooler_type, + canonical_box_size=224, + canonical_level=4, + ): + """ + Args: + output_size (int, tuple[int] or list[int]): output size of the pooled region, + e.g., 14 x 14. If tuple or list is given, the length must be 2. + scales (list[float]): The scale for each low-level pooling op relative to + the input image. For a feature map with stride s relative to the input + image, scale is defined as 1/s. The stride must be power of 2. + When there are multiple scales, they must form a pyramid, i.e. they must be + a monotically decreasing geometric sequence with a factor of 1/2. + sampling_ratio (int): The `sampling_ratio` parameter for the ROIAlign op. + pooler_type (string): Name of the type of pooling operation that should be applied. + For instance, "ROIPool" or "ROIAlignV2". + canonical_box_size (int): A canonical box size in pixels (sqrt(box area)). The default + is heuristically defined as 224 pixels in the FPN paper (based on ImageNet + pre-training). + canonical_level (int): The feature map level index from which a canonically-sized box + should be placed. The default is defined as level 4 (stride=16) in the FPN paper, + i.e., a box of size 224x224 will be placed on the feature with stride=16. + The box placement for all boxes will be determined from their sizes w.r.t + canonical_box_size. For example, a box whose area is 4x that of a canonical box + should be used to pool features from feature level ``canonical_level+1``. + + Note that the actual input feature maps given to this module may not have + sufficiently many levels for the input boxes. If the boxes are too large or too + small for the input feature maps, the closest level will be used. + """ + super().__init__() + + if isinstance(output_size, int): + output_size = (output_size, output_size) + assert len(output_size) == 2 + assert isinstance(output_size[0], int) and isinstance(output_size[1], int) + self.output_size = output_size + + if pooler_type == "ROIAlign": + self.level_poolers = nn.ModuleList( + ROIAlign( + output_size, spatial_scale=scale, sampling_ratio=sampling_ratio, aligned=False + ) + for scale in scales + ) + elif pooler_type == "ROIAlignV2": + self.level_poolers = nn.ModuleList( + ROIAlign( + output_size, spatial_scale=scale, sampling_ratio=sampling_ratio, aligned=True + ) + for scale in scales + ) + elif pooler_type == "ROIPool": + self.level_poolers = nn.ModuleList( + RoIPool(output_size, spatial_scale=scale) for scale in scales + ) + elif pooler_type == "ROIAlignRotated": + self.level_poolers = nn.ModuleList( + ROIAlignRotated(output_size, spatial_scale=scale, sampling_ratio=sampling_ratio) + for scale in scales + ) + else: + raise ValueError("Unknown pooler type: {}".format(pooler_type)) + + # Map scale (defined as 1 / stride) to its feature map level under the + # assumption that stride is a power of 2. + min_level = -(math.log2(scales[0])) + max_level = -(math.log2(scales[-1])) + assert math.isclose(min_level, int(min_level)) and math.isclose( + max_level, int(max_level) + ), "Featuremap stride is not power of 2!" + self.min_level = int(min_level) + self.max_level = int(max_level) + assert ( + len(scales) == self.max_level - self.min_level + 1 + ), "[ROIPooler] Sizes of input featuremaps do not form a pyramid!" + assert 0 <= self.min_level and self.min_level <= self.max_level + self.canonical_level = canonical_level + assert canonical_box_size > 0 + self.canonical_box_size = canonical_box_size + + def forward(self, x: List[torch.Tensor], box_lists: List[Boxes]): + """ + Args: + x (list[Tensor]): A list of feature maps of NCHW shape, with scales matching those + used to construct this module. + box_lists (list[Boxes] | list[RotatedBoxes]): + A list of N Boxes or N RotatedBoxes, where N is the number of images in the batch. + The box coordinates are defined on the original image and + will be scaled by the `scales` argument of :class:`ROIPooler`. + + Returns: + Tensor: + A tensor of shape (M, C, output_size, output_size) where M is the total number of + boxes aggregated over all N batch images and C is the number of channels in `x`. + """ + num_level_assignments = len(self.level_poolers) + + assert isinstance(x, list) and isinstance( + box_lists, list + ), "Arguments to pooler must be lists" + assert ( + len(x) == num_level_assignments + ), "unequal value, num_level_assignments={}, but x is list of {} Tensors".format( + num_level_assignments, len(x) + ) + + assert len(box_lists) == x[0].size( + 0 + ), "unequal value, x[0] batch dim 0 is {}, but box_list has length {}".format( + x[0].size(0), len(box_lists) + ) + if len(box_lists) == 0: + return torch.zeros( + (0, x[0].shape[1]) + self.output_size, device=x[0].device, dtype=x[0].dtype + ) + + pooler_fmt_boxes = convert_boxes_to_pooler_format(box_lists) + + if num_level_assignments == 1: + return self.level_poolers[0](x[0], pooler_fmt_boxes) + + level_assignments = assign_boxes_to_levels( + box_lists, self.min_level, self.max_level, self.canonical_box_size, self.canonical_level + ) + + num_boxes = pooler_fmt_boxes.size(0) + num_channels = x[0].shape[1] + output_size = self.output_size[0] + + dtype, device = x[0].dtype, x[0].device + output = torch.zeros( + (num_boxes, num_channels, output_size, output_size), dtype=dtype, device=device + ) + + for level, pooler in enumerate(self.level_poolers): + inds = nonzero_tuple(level_assignments == level)[0] + pooler_fmt_boxes_level = pooler_fmt_boxes[inds] + # Use index_put_ instead of advance indexing, to avoid pytorch/issues/49852 + output.index_put_((inds,), pooler(x[level], pooler_fmt_boxes_level)) + + return output diff --git a/detectron2/modeling/postprocessing.py b/detectron2/modeling/postprocessing.py new file mode 100644 index 0000000000000000000000000000000000000000..1a3d287eeb6c2cb3070f1aa7157b006e9aa857f5 --- /dev/null +++ b/detectron2/modeling/postprocessing.py @@ -0,0 +1,101 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import torch +from torch.nn import functional as F + +from detectron2.structures import Instances, ROIMasks + + +# perhaps should rename to "resize_instance" +def detector_postprocess( + results: Instances, output_height: int, output_width: int, mask_threshold: float = 0.5 +): + """ + Resize the output instances. + The input images are often resized when entering an object detector. + As a result, we often need the outputs of the detector in a different + resolution from its inputs. + + This function will resize the raw outputs of an R-CNN detector + to produce outputs according to the desired output resolution. + + Args: + results (Instances): the raw outputs from the detector. + `results.image_size` contains the input image resolution the detector sees. + This object might be modified in-place. + output_height, output_width: the desired output resolution. + + Returns: + Instances: the resized output from the model, based on the output resolution + """ + # Change to 'if is_tracing' after PT1.7 + if isinstance(output_height, torch.Tensor): + # Converts integer tensors to float temporaries to ensure true + # division is performed when computing scale_x and scale_y. + output_width_tmp = output_width.float() + output_height_tmp = output_height.float() + new_size = torch.stack([output_height, output_width]) + else: + new_size = (output_height, output_width) + output_width_tmp = output_width + output_height_tmp = output_height + + scale_x, scale_y = ( + output_width_tmp / results.image_size[1], + output_height_tmp / results.image_size[0], + ) + results = Instances(new_size, **results.get_fields()) + + if results.has("pred_boxes"): + output_boxes = results.pred_boxes + elif results.has("proposal_boxes"): + output_boxes = results.proposal_boxes + else: + output_boxes = None + assert output_boxes is not None, "Predictions must contain boxes!" + + output_boxes.scale(scale_x, scale_y) + output_boxes.clip(results.image_size) + + results = results[output_boxes.nonempty()] + + if results.has("pred_masks"): + if isinstance(results.pred_masks, ROIMasks): + roi_masks = results.pred_masks + else: + # pred_masks is a tensor of shape (N, 1, M, M) + roi_masks = ROIMasks(results.pred_masks[:, 0, :, :]) + results.pred_masks = roi_masks.to_bitmasks( + results.pred_boxes, output_height, output_width, mask_threshold + ).tensor # TODO return ROIMasks/BitMask object in the future + + if results.has("pred_keypoints"): + results.pred_keypoints[:, :, 0] *= scale_x + results.pred_keypoints[:, :, 1] *= scale_y + + return results + + +def sem_seg_postprocess(result, img_size, output_height, output_width): + """ + Return semantic segmentation predictions in the original resolution. + + The input images are often resized when entering semantic segmentor. Moreover, in same + cases, they also padded inside segmentor to be divisible by maximum network stride. + As a result, we often need the predictions of the segmentor in a different + resolution from its inputs. + + Args: + result (Tensor): semantic segmentation prediction logits. A tensor of shape (C, H, W), + where C is the number of classes, and H, W are the height and width of the prediction. + img_size (tuple): image size that segmentor is taking as input. + output_height, output_width: the desired output resolution. + + Returns: + semantic segmentation prediction (Tensor): A tensor of the shape + (C, output_height, output_width) that contains per-pixel soft predictions. + """ + result = result[:, : img_size[0], : img_size[1]].expand(1, -1, -1, -1) + result = F.interpolate( + result, size=(output_height, output_width), mode="bilinear", align_corners=False + )[0] + return result diff --git a/detectron2/modeling/proposal_generator/__init__.py b/detectron2/modeling/proposal_generator/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3f4e4df7645c67b7a013295207b98fe70b2e574c --- /dev/null +++ b/detectron2/modeling/proposal_generator/__init__.py @@ -0,0 +1,5 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .build import PROPOSAL_GENERATOR_REGISTRY, build_proposal_generator +from .rpn import RPN_HEAD_REGISTRY, build_rpn_head, RPN, StandardRPNHead + +__all__ = list(globals().keys()) diff --git a/detectron2/modeling/proposal_generator/__pycache__/__init__.cpython-39.pyc b/detectron2/modeling/proposal_generator/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..34b63a5612c5523b2c005bc034c677ff91bbc217 Binary files /dev/null and b/detectron2/modeling/proposal_generator/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/modeling/proposal_generator/__pycache__/build.cpython-39.pyc b/detectron2/modeling/proposal_generator/__pycache__/build.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8c683e08fde71d4c2115dc7fdfdf1fc539917830 Binary files /dev/null and b/detectron2/modeling/proposal_generator/__pycache__/build.cpython-39.pyc differ diff --git a/detectron2/modeling/proposal_generator/__pycache__/proposal_utils.cpython-39.pyc b/detectron2/modeling/proposal_generator/__pycache__/proposal_utils.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..3463798512b52613adc4d85eee2f9f41a0ff65f3 Binary files /dev/null and b/detectron2/modeling/proposal_generator/__pycache__/proposal_utils.cpython-39.pyc differ diff --git a/detectron2/modeling/proposal_generator/__pycache__/rpn.cpython-39.pyc b/detectron2/modeling/proposal_generator/__pycache__/rpn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6a8fcc1cb20831f19102a936d0125af802911dd7 Binary files /dev/null and b/detectron2/modeling/proposal_generator/__pycache__/rpn.cpython-39.pyc differ diff --git a/detectron2/modeling/proposal_generator/__pycache__/rrpn.cpython-39.pyc b/detectron2/modeling/proposal_generator/__pycache__/rrpn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..675496fbc6e7b6ebc607f1fc81b07e46abbf0406 Binary files /dev/null and b/detectron2/modeling/proposal_generator/__pycache__/rrpn.cpython-39.pyc differ diff --git a/detectron2/modeling/proposal_generator/build.py b/detectron2/modeling/proposal_generator/build.py new file mode 100644 index 0000000000000000000000000000000000000000..34eb12d00d94ff905b796e75e2c4c5845257c8e9 --- /dev/null +++ b/detectron2/modeling/proposal_generator/build.py @@ -0,0 +1,24 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from detectron2.utils.registry import Registry + +PROPOSAL_GENERATOR_REGISTRY = Registry("PROPOSAL_GENERATOR") +PROPOSAL_GENERATOR_REGISTRY.__doc__ = """ +Registry for proposal generator, which produces object proposals from feature maps. + +The registered object will be called with `obj(cfg, input_shape)`. +The call should return a `nn.Module` object. +""" + +from . import rpn, rrpn # noqa F401 isort:skip + + +def build_proposal_generator(cfg, input_shape): + """ + Build a proposal generator from `cfg.MODEL.PROPOSAL_GENERATOR.NAME`. + The name can be "PrecomputedProposals" to use no proposal generator. + """ + name = cfg.MODEL.PROPOSAL_GENERATOR.NAME + if name == "PrecomputedProposals": + return None + + return PROPOSAL_GENERATOR_REGISTRY.get(name)(cfg, input_shape) diff --git a/detectron2/modeling/proposal_generator/proposal_utils.py b/detectron2/modeling/proposal_generator/proposal_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..48485a420a8afafc4097bef982c8d23b91b95269 --- /dev/null +++ b/detectron2/modeling/proposal_generator/proposal_utils.py @@ -0,0 +1,200 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import math +from typing import List, Tuple, Union +import torch + +from detectron2.layers import batched_nms, cat +from detectron2.structures import Boxes, Instances +from detectron2.utils.env import TORCH_VERSION + +logger = logging.getLogger(__name__) + + +def _is_tracing(): + if torch.jit.is_scripting(): + # https://github.com/pytorch/pytorch/issues/47379 + return False + else: + return TORCH_VERSION >= (1, 7) and torch.jit.is_tracing() + + +def find_top_rpn_proposals( + proposals: List[torch.Tensor], + pred_objectness_logits: List[torch.Tensor], + image_sizes: List[Tuple[int, int]], + nms_thresh: float, + pre_nms_topk: int, + post_nms_topk: int, + min_box_size: float, + training: bool, +): + """ + For each feature map, select the `pre_nms_topk` highest scoring proposals, + apply NMS, clip proposals, and remove small boxes. Return the `post_nms_topk` + highest scoring proposals among all the feature maps for each image. + + Args: + proposals (list[Tensor]): A list of L tensors. Tensor i has shape (N, Hi*Wi*A, 4). + All proposal predictions on the feature maps. + pred_objectness_logits (list[Tensor]): A list of L tensors. Tensor i has shape (N, Hi*Wi*A). + image_sizes (list[tuple]): sizes (h, w) for each image + nms_thresh (float): IoU threshold to use for NMS + pre_nms_topk (int): number of top k scoring proposals to keep before applying NMS. + When RPN is run on multiple feature maps (as in FPN) this number is per + feature map. + post_nms_topk (int): number of top k scoring proposals to keep after applying NMS. + When RPN is run on multiple feature maps (as in FPN) this number is total, + over all feature maps. + min_box_size (float): minimum proposal box side length in pixels (absolute units + wrt input images). + training (bool): True if proposals are to be used in training, otherwise False. + This arg exists only to support a legacy bug; look for the "NB: Legacy bug ..." + comment. + + Returns: + list[Instances]: list of N Instances. The i-th Instances + stores post_nms_topk object proposals for image i, sorted by their + objectness score in descending order. + """ + num_images = len(image_sizes) + device = proposals[0].device + + # 1. Select top-k anchor for every level and every image + topk_scores = [] # #lvl Tensor, each of shape N x topk + topk_proposals = [] + level_ids = [] # #lvl Tensor, each of shape (topk,) + batch_idx = torch.arange(num_images, device=device) + for level_id, (proposals_i, logits_i) in enumerate(zip(proposals, pred_objectness_logits)): + Hi_Wi_A = logits_i.shape[1] + if isinstance(Hi_Wi_A, torch.Tensor): # it's a tensor in tracing + num_proposals_i = torch.clamp(Hi_Wi_A, max=pre_nms_topk) + else: + num_proposals_i = min(Hi_Wi_A, pre_nms_topk) + + # sort is faster than topk: https://github.com/pytorch/pytorch/issues/22812 + # topk_scores_i, topk_idx = logits_i.topk(num_proposals_i, dim=1) + logits_i, idx = logits_i.sort(descending=True, dim=1) + topk_scores_i = logits_i.narrow(1, 0, num_proposals_i) + topk_idx = idx.narrow(1, 0, num_proposals_i) + + # each is N x topk + topk_proposals_i = proposals_i[batch_idx[:, None], topk_idx] # N x topk x 4 + + topk_proposals.append(topk_proposals_i) + topk_scores.append(topk_scores_i) + level_ids.append(torch.full((num_proposals_i,), level_id, dtype=torch.int64, device=device)) + + # 2. Concat all levels together + topk_scores = cat(topk_scores, dim=1) + topk_proposals = cat(topk_proposals, dim=1) + level_ids = cat(level_ids, dim=0) + + # 3. For each image, run a per-level NMS, and choose topk results. + results: List[Instances] = [] + for n, image_size in enumerate(image_sizes): + boxes = Boxes(topk_proposals[n]) + scores_per_img = topk_scores[n] + lvl = level_ids + + valid_mask = torch.isfinite(boxes.tensor).all(dim=1) & torch.isfinite(scores_per_img) + if not valid_mask.all(): + if training: + raise FloatingPointError( + "Predicted boxes or scores contain Inf/NaN. Training has diverged." + ) + boxes = boxes[valid_mask] + scores_per_img = scores_per_img[valid_mask] + lvl = lvl[valid_mask] + boxes.clip(image_size) + + # filter empty boxes + keep = boxes.nonempty(threshold=min_box_size) + if _is_tracing() or keep.sum().item() != len(boxes): + boxes, scores_per_img, lvl = boxes[keep], scores_per_img[keep], lvl[keep] + + keep = batched_nms(boxes.tensor, scores_per_img, lvl, nms_thresh) + # In Detectron1, there was different behavior during training vs. testing. + # (https://github.com/facebookresearch/Detectron/issues/459) + # During training, topk is over the proposals from *all* images in the training batch. + # During testing, it is over the proposals for each image separately. + # As a result, the training behavior becomes batch-dependent, + # and the configuration "POST_NMS_TOPK_TRAIN" end up relying on the batch size. + # This bug is addressed in Detectron2 to make the behavior independent of batch size. + keep = keep[:post_nms_topk] # keep is already sorted + + res = Instances(image_size) + res.proposal_boxes = boxes[keep] + res.objectness_logits = scores_per_img[keep] + results.append(res) + return results + + +def add_ground_truth_to_proposals( + gt: Union[List[Instances], List[Boxes]], proposals: List[Instances] +) -> List[Instances]: + """ + Call `add_ground_truth_to_proposals_single_image` for all images. + + Args: + gt(Union[List[Instances], List[Boxes]): list of N elements. Element i is a Instances + representing the ground-truth for image i. + proposals (list[Instances]): list of N elements. Element i is a Instances + representing the proposals for image i. + + Returns: + list[Instances]: list of N Instances. Each is the proposals for the image, + with field "proposal_boxes" and "objectness_logits". + """ + assert gt is not None + + if len(proposals) != len(gt): + raise ValueError("proposals and gt should have the same length as the number of images!") + if len(proposals) == 0: + return proposals + + return [ + add_ground_truth_to_proposals_single_image(gt_i, proposals_i) + for gt_i, proposals_i in zip(gt, proposals) + ] + + +def add_ground_truth_to_proposals_single_image( + gt: Union[Instances, Boxes], proposals: Instances +) -> Instances: + """ + Augment `proposals` with `gt`. + + Args: + Same as `add_ground_truth_to_proposals`, but with gt and proposals + per image. + + Returns: + Same as `add_ground_truth_to_proposals`, but for only one image. + """ + if isinstance(gt, Boxes): + # convert Boxes to Instances + gt = Instances(proposals.image_size, gt_boxes=gt) + + gt_boxes = gt.gt_boxes + device = proposals.objectness_logits.device + # Assign all ground-truth boxes an objectness logit corresponding to + # P(object) = sigmoid(logit) =~ 1. + gt_logit_value = math.log((1.0 - 1e-10) / (1 - (1.0 - 1e-10))) + gt_logits = gt_logit_value * torch.ones(len(gt_boxes), device=device) + + # Concatenating gt_boxes with proposals requires them to have the same fields + gt_proposal = Instances(proposals.image_size, **gt.get_fields()) + gt_proposal.proposal_boxes = gt_boxes + gt_proposal.objectness_logits = gt_logits + + for key in proposals.get_fields().keys(): + assert gt_proposal.has( + key + ), "The attribute '{}' in `proposals` does not exist in `gt`".format(key) + + # NOTE: Instances.cat only use fields from the first item. Extra fields in latter items + # will be thrown away. + new_proposals = Instances.cat([proposals, gt_proposal]) + + return new_proposals diff --git a/detectron2/modeling/proposal_generator/rpn.py b/detectron2/modeling/proposal_generator/rpn.py new file mode 100644 index 0000000000000000000000000000000000000000..99cd536d2f9880d2049390c45f73eb22335e1b82 --- /dev/null +++ b/detectron2/modeling/proposal_generator/rpn.py @@ -0,0 +1,533 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from typing import Dict, List, Optional, Tuple, Union +import torch +import torch.nn.functional as F +from torch import nn + +from detectron2.config import configurable +from detectron2.layers import Conv2d, ShapeSpec, cat +from detectron2.structures import Boxes, ImageList, Instances, pairwise_iou +from detectron2.utils.events import get_event_storage +from detectron2.utils.memory import retry_if_cuda_oom +from detectron2.utils.registry import Registry + +from ..anchor_generator import build_anchor_generator +from ..box_regression import Box2BoxTransform, _dense_box_regression_loss +from ..matcher import Matcher +from ..sampling import subsample_labels +from .build import PROPOSAL_GENERATOR_REGISTRY +from .proposal_utils import find_top_rpn_proposals + +RPN_HEAD_REGISTRY = Registry("RPN_HEAD") +RPN_HEAD_REGISTRY.__doc__ = """ +Registry for RPN heads, which take feature maps and perform +objectness classification and bounding box regression for anchors. + +The registered object will be called with `obj(cfg, input_shape)`. +The call should return a `nn.Module` object. +""" + + +""" +Shape shorthand in this module: + + N: number of images in the minibatch + L: number of feature maps per image on which RPN is run + A: number of cell anchors (must be the same for all feature maps) + Hi, Wi: height and width of the i-th feature map + B: size of the box parameterization + +Naming convention: + + objectness: refers to the binary classification of an anchor as object vs. not object. + + deltas: refers to the 4-d (dx, dy, dw, dh) deltas that parameterize the box2box + transform (see :class:`box_regression.Box2BoxTransform`), or 5d for rotated boxes. + + pred_objectness_logits: predicted objectness scores in [-inf, +inf]; use + sigmoid(pred_objectness_logits) to estimate P(object). + + gt_labels: ground-truth binary classification labels for objectness + + pred_anchor_deltas: predicted box2box transform deltas + + gt_anchor_deltas: ground-truth box2box transform deltas +""" + + +def build_rpn_head(cfg, input_shape): + """ + Build an RPN head defined by `cfg.MODEL.RPN.HEAD_NAME`. + """ + name = cfg.MODEL.RPN.HEAD_NAME + return RPN_HEAD_REGISTRY.get(name)(cfg, input_shape) + + +@RPN_HEAD_REGISTRY.register() +class StandardRPNHead(nn.Module): + """ + Standard RPN classification and regression heads described in :paper:`Faster R-CNN`. + Uses a 3x3 conv to produce a shared hidden state from which one 1x1 conv predicts + objectness logits for each anchor and a second 1x1 conv predicts bounding-box deltas + specifying how to deform each anchor into an object proposal. + """ + + @configurable + def __init__( + self, *, in_channels: int, num_anchors: int, box_dim: int = 4, conv_dims: List[int] = (-1,) + ): + """ + NOTE: this interface is experimental. + + Args: + in_channels (int): number of input feature channels. When using multiple + input features, they must have the same number of channels. + num_anchors (int): number of anchors to predict for *each spatial position* + on the feature map. The total number of anchors for each + feature map will be `num_anchors * H * W`. + box_dim (int): dimension of a box, which is also the number of box regression + predictions to make for each anchor. An axis aligned box has + box_dim=4, while a rotated box has box_dim=5. + conv_dims (list[int]): a list of integers representing the output channels + of N conv layers. Set it to -1 to use the same number of output channels + as input channels. + """ + super().__init__() + cur_channels = in_channels + # Keeping the old variable names and structure for backwards compatiblity. + # Otherwise the old checkpoints will fail to load. + if len(conv_dims) == 1: + out_channels = cur_channels if conv_dims[0] == -1 else conv_dims[0] + # 3x3 conv for the hidden representation + self.conv = self._get_rpn_conv(cur_channels, out_channels) + cur_channels = out_channels + else: + self.conv = nn.Sequential() + for k, conv_dim in enumerate(conv_dims): + out_channels = cur_channels if conv_dim == -1 else conv_dim + if out_channels <= 0: + raise ValueError( + f"Conv output channels should be greater than 0. Got {out_channels}" + ) + conv = self._get_rpn_conv(cur_channels, out_channels) + self.conv.add_module(f"conv{k}", conv) + cur_channels = out_channels + # 1x1 conv for predicting objectness logits + self.objectness_logits = nn.Conv2d(cur_channels, num_anchors, kernel_size=1, stride=1) + # 1x1 conv for predicting box2box transform deltas + self.anchor_deltas = nn.Conv2d(cur_channels, num_anchors * box_dim, kernel_size=1, stride=1) + + # Keeping the order of weights initialization same for backwards compatiblility. + for layer in self.modules(): + if isinstance(layer, nn.Conv2d): + nn.init.normal_(layer.weight, std=0.01) + nn.init.constant_(layer.bias, 0) + + def _get_rpn_conv(self, in_channels, out_channels): + return Conv2d( + in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1, + activation=nn.ReLU(), + ) + + @classmethod + def from_config(cls, cfg, input_shape): + # Standard RPN is shared across levels: + in_channels = [s.channels for s in input_shape] + assert len(set(in_channels)) == 1, "Each level must have the same channel!" + in_channels = in_channels[0] + + # RPNHead should take the same input as anchor generator + # NOTE: it assumes that creating an anchor generator does not have unwanted side effect. + anchor_generator = build_anchor_generator(cfg, input_shape) + num_anchors = anchor_generator.num_anchors + box_dim = anchor_generator.box_dim + assert ( + len(set(num_anchors)) == 1 + ), "Each level must have the same number of anchors per spatial position" + return { + "in_channels": in_channels, + "num_anchors": num_anchors[0], + "box_dim": box_dim, + "conv_dims": cfg.MODEL.RPN.CONV_DIMS, + } + + def forward(self, features: List[torch.Tensor]): + """ + Args: + features (list[Tensor]): list of feature maps + + Returns: + list[Tensor]: A list of L elements. + Element i is a tensor of shape (N, A, Hi, Wi) representing + the predicted objectness logits for all anchors. A is the number of cell anchors. + list[Tensor]: A list of L elements. Element i is a tensor of shape + (N, A*box_dim, Hi, Wi) representing the predicted "deltas" used to transform anchors + to proposals. + """ + pred_objectness_logits = [] + pred_anchor_deltas = [] + for x in features: + t = self.conv(x) + pred_objectness_logits.append(self.objectness_logits(t)) + pred_anchor_deltas.append(self.anchor_deltas(t)) + return pred_objectness_logits, pred_anchor_deltas + + +@PROPOSAL_GENERATOR_REGISTRY.register() +class RPN(nn.Module): + """ + Region Proposal Network, introduced by :paper:`Faster R-CNN`. + """ + + @configurable + def __init__( + self, + *, + in_features: List[str], + head: nn.Module, + anchor_generator: nn.Module, + anchor_matcher: Matcher, + box2box_transform: Box2BoxTransform, + batch_size_per_image: int, + positive_fraction: float, + pre_nms_topk: Tuple[float, float], + post_nms_topk: Tuple[float, float], + nms_thresh: float = 0.7, + min_box_size: float = 0.0, + anchor_boundary_thresh: float = -1.0, + loss_weight: Union[float, Dict[str, float]] = 1.0, + box_reg_loss_type: str = "smooth_l1", + smooth_l1_beta: float = 0.0, + ): + """ + NOTE: this interface is experimental. + + Args: + in_features (list[str]): list of names of input features to use + head (nn.Module): a module that predicts logits and regression deltas + for each level from a list of per-level features + anchor_generator (nn.Module): a module that creates anchors from a + list of features. Usually an instance of :class:`AnchorGenerator` + anchor_matcher (Matcher): label the anchors by matching them with ground truth. + box2box_transform (Box2BoxTransform): defines the transform from anchors boxes to + instance boxes + batch_size_per_image (int): number of anchors per image to sample for training + positive_fraction (float): fraction of foreground anchors to sample for training + pre_nms_topk (tuple[float]): (train, test) that represents the + number of top k proposals to select before NMS, in + training and testing. + post_nms_topk (tuple[float]): (train, test) that represents the + number of top k proposals to select after NMS, in + training and testing. + nms_thresh (float): NMS threshold used to de-duplicate the predicted proposals + min_box_size (float): remove proposal boxes with any side smaller than this threshold, + in the unit of input image pixels + anchor_boundary_thresh (float): legacy option + loss_weight (float|dict): weights to use for losses. Can be single float for weighting + all rpn losses together, or a dict of individual weightings. Valid dict keys are: + "loss_rpn_cls" - applied to classification loss + "loss_rpn_loc" - applied to box regression loss + box_reg_loss_type (str): Loss type to use. Supported losses: "smooth_l1", "giou". + smooth_l1_beta (float): beta parameter for the smooth L1 regression loss. Default to + use L1 loss. Only used when `box_reg_loss_type` is "smooth_l1" + """ + super().__init__() + self.in_features = in_features + self.rpn_head = head + self.anchor_generator = anchor_generator + self.anchor_matcher = anchor_matcher + self.box2box_transform = box2box_transform + self.batch_size_per_image = batch_size_per_image + self.positive_fraction = positive_fraction + # Map from self.training state to train/test settings + self.pre_nms_topk = {True: pre_nms_topk[0], False: pre_nms_topk[1]} + self.post_nms_topk = {True: post_nms_topk[0], False: post_nms_topk[1]} + self.nms_thresh = nms_thresh + self.min_box_size = float(min_box_size) + self.anchor_boundary_thresh = anchor_boundary_thresh + if isinstance(loss_weight, float): + loss_weight = {"loss_rpn_cls": loss_weight, "loss_rpn_loc": loss_weight} + self.loss_weight = loss_weight + self.box_reg_loss_type = box_reg_loss_type + self.smooth_l1_beta = smooth_l1_beta + + @classmethod + def from_config(cls, cfg, input_shape: Dict[str, ShapeSpec]): + in_features = cfg.MODEL.RPN.IN_FEATURES + ret = { + "in_features": in_features, + "min_box_size": cfg.MODEL.PROPOSAL_GENERATOR.MIN_SIZE, + "nms_thresh": cfg.MODEL.RPN.NMS_THRESH, + "batch_size_per_image": cfg.MODEL.RPN.BATCH_SIZE_PER_IMAGE, + "positive_fraction": cfg.MODEL.RPN.POSITIVE_FRACTION, + "loss_weight": { + "loss_rpn_cls": cfg.MODEL.RPN.LOSS_WEIGHT, + "loss_rpn_loc": cfg.MODEL.RPN.BBOX_REG_LOSS_WEIGHT * cfg.MODEL.RPN.LOSS_WEIGHT, + }, + "anchor_boundary_thresh": cfg.MODEL.RPN.BOUNDARY_THRESH, + "box2box_transform": Box2BoxTransform(weights=cfg.MODEL.RPN.BBOX_REG_WEIGHTS), + "box_reg_loss_type": cfg.MODEL.RPN.BBOX_REG_LOSS_TYPE, + "smooth_l1_beta": cfg.MODEL.RPN.SMOOTH_L1_BETA, + } + + ret["pre_nms_topk"] = (cfg.MODEL.RPN.PRE_NMS_TOPK_TRAIN, cfg.MODEL.RPN.PRE_NMS_TOPK_TEST) + ret["post_nms_topk"] = (cfg.MODEL.RPN.POST_NMS_TOPK_TRAIN, cfg.MODEL.RPN.POST_NMS_TOPK_TEST) + + ret["anchor_generator"] = build_anchor_generator(cfg, [input_shape[f] for f in in_features]) + ret["anchor_matcher"] = Matcher( + cfg.MODEL.RPN.IOU_THRESHOLDS, cfg.MODEL.RPN.IOU_LABELS, allow_low_quality_matches=True + ) + ret["head"] = build_rpn_head(cfg, [input_shape[f] for f in in_features]) + return ret + + def _subsample_labels(self, label): + """ + Randomly sample a subset of positive and negative examples, and overwrite + the label vector to the ignore value (-1) for all elements that are not + included in the sample. + + Args: + labels (Tensor): a vector of -1, 0, 1. Will be modified in-place and returned. + """ + pos_idx, neg_idx = subsample_labels( + label, self.batch_size_per_image, self.positive_fraction, 0 + ) + # Fill with the ignore label (-1), then set positive and negative labels + label.fill_(-1) + label.scatter_(0, pos_idx, 1) + label.scatter_(0, neg_idx, 0) + return label + + @torch.jit.unused + @torch.no_grad() + def label_and_sample_anchors( + self, anchors: List[Boxes], gt_instances: List[Instances] + ) -> Tuple[List[torch.Tensor], List[torch.Tensor]]: + """ + Args: + anchors (list[Boxes]): anchors for each feature map. + gt_instances: the ground-truth instances for each image. + + Returns: + list[Tensor]: + List of #img tensors. i-th element is a vector of labels whose length is + the total number of anchors across all feature maps R = sum(Hi * Wi * A). + Label values are in {-1, 0, 1}, with meanings: -1 = ignore; 0 = negative + class; 1 = positive class. + list[Tensor]: + i-th element is a Rx4 tensor. The values are the matched gt boxes for each + anchor. Values are undefined for those anchors not labeled as 1. + """ + anchors = Boxes.cat(anchors) + + gt_boxes = [x.gt_boxes for x in gt_instances] + image_sizes = [x.image_size for x in gt_instances] + del gt_instances + + gt_labels = [] + matched_gt_boxes = [] + for image_size_i, gt_boxes_i in zip(image_sizes, gt_boxes): + """ + image_size_i: (h, w) for the i-th image + gt_boxes_i: ground-truth boxes for i-th image + """ + + match_quality_matrix = retry_if_cuda_oom(pairwise_iou)(gt_boxes_i, anchors) + matched_idxs, gt_labels_i = retry_if_cuda_oom(self.anchor_matcher)(match_quality_matrix) + # Matching is memory-expensive and may result in CPU tensors. But the result is small + gt_labels_i = gt_labels_i.to(device=gt_boxes_i.device) + del match_quality_matrix + + if self.anchor_boundary_thresh >= 0: + # Discard anchors that go out of the boundaries of the image + # NOTE: This is legacy functionality that is turned off by default in Detectron2 + anchors_inside_image = anchors.inside_box(image_size_i, self.anchor_boundary_thresh) + gt_labels_i[~anchors_inside_image] = -1 + + # A vector of labels (-1, 0, 1) for each anchor + gt_labels_i = self._subsample_labels(gt_labels_i) + + if len(gt_boxes_i) == 0: + # These values won't be used anyway since the anchor is labeled as background + matched_gt_boxes_i = torch.zeros_like(anchors.tensor) + else: + # TODO wasted indexing computation for ignored boxes + matched_gt_boxes_i = gt_boxes_i[matched_idxs].tensor + + gt_labels.append(gt_labels_i) # N,AHW + matched_gt_boxes.append(matched_gt_boxes_i) + return gt_labels, matched_gt_boxes + + @torch.jit.unused + def losses( + self, + anchors: List[Boxes], + pred_objectness_logits: List[torch.Tensor], + gt_labels: List[torch.Tensor], + pred_anchor_deltas: List[torch.Tensor], + gt_boxes: List[torch.Tensor], + ) -> Dict[str, torch.Tensor]: + """ + Return the losses from a set of RPN predictions and their associated ground-truth. + + Args: + anchors (list[Boxes or RotatedBoxes]): anchors for each feature map, each + has shape (Hi*Wi*A, B), where B is box dimension (4 or 5). + pred_objectness_logits (list[Tensor]): A list of L elements. + Element i is a tensor of shape (N, Hi*Wi*A) representing + the predicted objectness logits for all anchors. + gt_labels (list[Tensor]): Output of :meth:`label_and_sample_anchors`. + pred_anchor_deltas (list[Tensor]): A list of L elements. Element i is a tensor of shape + (N, Hi*Wi*A, 4 or 5) representing the predicted "deltas" used to transform anchors + to proposals. + gt_boxes (list[Tensor]): Output of :meth:`label_and_sample_anchors`. + + Returns: + dict[loss name -> loss value]: A dict mapping from loss name to loss value. + Loss names are: `loss_rpn_cls` for objectness classification and + `loss_rpn_loc` for proposal localization. + """ + num_images = len(gt_labels) + gt_labels = torch.stack(gt_labels) # (N, sum(Hi*Wi*Ai)) + + # Log the number of positive/negative anchors per-image that's used in training + pos_mask = gt_labels == 1 + num_pos_anchors = pos_mask.sum().item() + num_neg_anchors = (gt_labels == 0).sum().item() + storage = get_event_storage() + storage.put_scalar("rpn/num_pos_anchors", num_pos_anchors / num_images) + storage.put_scalar("rpn/num_neg_anchors", num_neg_anchors / num_images) + + localization_loss = _dense_box_regression_loss( + anchors, + self.box2box_transform, + pred_anchor_deltas, + gt_boxes, + pos_mask, + box_reg_loss_type=self.box_reg_loss_type, + smooth_l1_beta=self.smooth_l1_beta, + ) + + valid_mask = gt_labels >= 0 + objectness_loss = F.binary_cross_entropy_with_logits( + cat(pred_objectness_logits, dim=1)[valid_mask], + gt_labels[valid_mask].to(torch.float32), + reduction="sum", + ) + normalizer = self.batch_size_per_image * num_images + losses = { + "loss_rpn_cls": objectness_loss / normalizer, + # The original Faster R-CNN paper uses a slightly different normalizer + # for loc loss. But it doesn't matter in practice + "loss_rpn_loc": localization_loss / normalizer, + } + losses = {k: v * self.loss_weight.get(k, 1.0) for k, v in losses.items()} + return losses + + def forward( + self, + images: ImageList, + features: Dict[str, torch.Tensor], + gt_instances: Optional[List[Instances]] = None, + ): + """ + Args: + images (ImageList): input images of length `N` + features (dict[str, Tensor]): input data as a mapping from feature + map name to tensor. Axis 0 represents the number of images `N` in + the input data; axes 1-3 are channels, height, and width, which may + vary between feature maps (e.g., if a feature pyramid is used). + gt_instances (list[Instances], optional): a length `N` list of `Instances`s. + Each `Instances` stores ground-truth instances for the corresponding image. + + Returns: + proposals: list[Instances]: contains fields "proposal_boxes", "objectness_logits" + loss: dict[Tensor] or None + """ + features = [features[f] for f in self.in_features] + anchors = self.anchor_generator(features) + + pred_objectness_logits, pred_anchor_deltas = self.rpn_head(features) + # Transpose the Hi*Wi*A dimension to the middle: + pred_objectness_logits = [ + # (N, A, Hi, Wi) -> (N, Hi, Wi, A) -> (N, Hi*Wi*A) + score.permute(0, 2, 3, 1).flatten(1) + for score in pred_objectness_logits + ] + pred_anchor_deltas = [ + # (N, A*B, Hi, Wi) -> (N, A, B, Hi, Wi) -> (N, Hi, Wi, A, B) -> (N, Hi*Wi*A, B) + x.view(x.shape[0], -1, self.anchor_generator.box_dim, x.shape[-2], x.shape[-1]) + .permute(0, 3, 4, 1, 2) + .flatten(1, -2) + for x in pred_anchor_deltas + ] + + if self.training: + assert gt_instances is not None, "RPN requires gt_instances in training!" + gt_labels, gt_boxes = self.label_and_sample_anchors(anchors, gt_instances) + losses = self.losses( + anchors, pred_objectness_logits, gt_labels, pred_anchor_deltas, gt_boxes + ) + else: + losses = {} + proposals = self.predict_proposals( + anchors, pred_objectness_logits, pred_anchor_deltas, images.image_sizes + ) + return proposals, losses + + def predict_proposals( + self, + anchors: List[Boxes], + pred_objectness_logits: List[torch.Tensor], + pred_anchor_deltas: List[torch.Tensor], + image_sizes: List[Tuple[int, int]], + ): + """ + Decode all the predicted box regression deltas to proposals. Find the top proposals + by applying NMS and removing boxes that are too small. + + Returns: + proposals (list[Instances]): list of N Instances. The i-th Instances + stores post_nms_topk object proposals for image i, sorted by their + objectness score in descending order. + """ + # The proposals are treated as fixed for joint training with roi heads. + # This approach ignores the derivative w.r.t. the proposal boxes’ coordinates that + # are also network responses. + with torch.no_grad(): + pred_proposals = self._decode_proposals(anchors, pred_anchor_deltas) + return find_top_rpn_proposals( + pred_proposals, + pred_objectness_logits, + image_sizes, + self.nms_thresh, + self.pre_nms_topk[self.training], + self.post_nms_topk[self.training], + self.min_box_size, + self.training, + ) + + def _decode_proposals(self, anchors: List[Boxes], pred_anchor_deltas: List[torch.Tensor]): + """ + Transform anchors into proposals by applying the predicted anchor deltas. + + Returns: + proposals (list[Tensor]): A list of L tensors. Tensor i has shape + (N, Hi*Wi*A, B) + """ + N = pred_anchor_deltas[0].shape[0] + proposals = [] + # For each feature map + for anchors_i, pred_anchor_deltas_i in zip(anchors, pred_anchor_deltas): + B = anchors_i.tensor.size(1) + pred_anchor_deltas_i = pred_anchor_deltas_i.reshape(-1, B) + # Expand anchors to shape (N*Hi*Wi*A, B) + anchors_i = anchors_i.tensor.unsqueeze(0).expand(N, -1, -1).reshape(-1, B) + proposals_i = self.box2box_transform.apply_deltas(pred_anchor_deltas_i, anchors_i) + # Append feature map proposals with shape (N, Hi*Wi*A, B) + proposals.append(proposals_i.view(N, -1, B)) + return proposals diff --git a/detectron2/modeling/proposal_generator/rrpn.py b/detectron2/modeling/proposal_generator/rrpn.py new file mode 100644 index 0000000000000000000000000000000000000000..6ee4d8fd70430c5242cc02a1df8400493ffd75b7 --- /dev/null +++ b/detectron2/modeling/proposal_generator/rrpn.py @@ -0,0 +1,203 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import itertools +import logging +from typing import Dict, List +import torch + +from detectron2.config import configurable +from detectron2.layers import ShapeSpec, batched_nms_rotated, cat +from detectron2.structures import Instances, RotatedBoxes, pairwise_iou_rotated +from detectron2.utils.memory import retry_if_cuda_oom + +from ..box_regression import Box2BoxTransformRotated +from .build import PROPOSAL_GENERATOR_REGISTRY +from .rpn import RPN + +logger = logging.getLogger(__name__) + + +def find_top_rrpn_proposals( + proposals, + pred_objectness_logits, + image_sizes, + nms_thresh, + pre_nms_topk, + post_nms_topk, + min_box_size, + training, +): + """ + For each feature map, select the `pre_nms_topk` highest scoring proposals, + apply NMS, clip proposals, and remove small boxes. Return the `post_nms_topk` + highest scoring proposals among all the feature maps if `training` is True, + otherwise, returns the highest `post_nms_topk` scoring proposals for each + feature map. + + Args: + proposals (list[Tensor]): A list of L tensors. Tensor i has shape (N, Hi*Wi*A, 5). + All proposal predictions on the feature maps. + pred_objectness_logits (list[Tensor]): A list of L tensors. Tensor i has shape (N, Hi*Wi*A). + image_sizes (list[tuple]): sizes (h, w) for each image + nms_thresh (float): IoU threshold to use for NMS + pre_nms_topk (int): number of top k scoring proposals to keep before applying NMS. + When RRPN is run on multiple feature maps (as in FPN) this number is per + feature map. + post_nms_topk (int): number of top k scoring proposals to keep after applying NMS. + When RRPN is run on multiple feature maps (as in FPN) this number is total, + over all feature maps. + min_box_size(float): minimum proposal box side length in pixels (absolute units wrt + input images). + training (bool): True if proposals are to be used in training, otherwise False. + This arg exists only to support a legacy bug; look for the "NB: Legacy bug ..." + comment. + + Returns: + proposals (list[Instances]): list of N Instances. The i-th Instances + stores post_nms_topk object proposals for image i. + """ + num_images = len(image_sizes) + device = proposals[0].device + + # 1. Select top-k anchor for every level and every image + topk_scores = [] # #lvl Tensor, each of shape N x topk + topk_proposals = [] + level_ids = [] # #lvl Tensor, each of shape (topk,) + batch_idx = torch.arange(num_images, device=device) + for level_id, proposals_i, logits_i in zip( + itertools.count(), proposals, pred_objectness_logits + ): + Hi_Wi_A = logits_i.shape[1] + num_proposals_i = min(pre_nms_topk, Hi_Wi_A) + + # sort is faster than topk (https://github.com/pytorch/pytorch/issues/22812) + # topk_scores_i, topk_idx = logits_i.topk(num_proposals_i, dim=1) + logits_i, idx = logits_i.sort(descending=True, dim=1) + topk_scores_i = logits_i[batch_idx, :num_proposals_i] + topk_idx = idx[batch_idx, :num_proposals_i] + + # each is N x topk + topk_proposals_i = proposals_i[batch_idx[:, None], topk_idx] # N x topk x 5 + + topk_proposals.append(topk_proposals_i) + topk_scores.append(topk_scores_i) + level_ids.append(torch.full((num_proposals_i,), level_id, dtype=torch.int64, device=device)) + + # 2. Concat all levels together + topk_scores = cat(topk_scores, dim=1) + topk_proposals = cat(topk_proposals, dim=1) + level_ids = cat(level_ids, dim=0) + + # 3. For each image, run a per-level NMS, and choose topk results. + results = [] + for n, image_size in enumerate(image_sizes): + boxes = RotatedBoxes(topk_proposals[n]) + scores_per_img = topk_scores[n] + valid_mask = torch.isfinite(boxes.tensor).all(dim=1) & torch.isfinite(scores_per_img) + if not valid_mask.all(): + boxes = boxes[valid_mask] + scores_per_img = scores_per_img[valid_mask] + boxes.clip(image_size) + + # filter empty boxes + keep = boxes.nonempty(threshold=min_box_size) + lvl = level_ids + if keep.sum().item() != len(boxes): + boxes, scores_per_img, lvl = (boxes[keep], scores_per_img[keep], level_ids[keep]) + + keep = batched_nms_rotated(boxes.tensor, scores_per_img, lvl, nms_thresh) + # In Detectron1, there was different behavior during training vs. testing. + # (https://github.com/facebookresearch/Detectron/issues/459) + # During training, topk is over the proposals from *all* images in the training batch. + # During testing, it is over the proposals for each image separately. + # As a result, the training behavior becomes batch-dependent, + # and the configuration "POST_NMS_TOPK_TRAIN" end up relying on the batch size. + # This bug is addressed in Detectron2 to make the behavior independent of batch size. + keep = keep[:post_nms_topk] + + res = Instances(image_size) + res.proposal_boxes = boxes[keep] + res.objectness_logits = scores_per_img[keep] + results.append(res) + return results + + +@PROPOSAL_GENERATOR_REGISTRY.register() +class RRPN(RPN): + """ + Rotated Region Proposal Network described in :paper:`RRPN`. + """ + + @configurable + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + if self.anchor_boundary_thresh >= 0: + raise NotImplementedError( + "anchor_boundary_thresh is a legacy option not implemented for RRPN." + ) + + @classmethod + def from_config(cls, cfg, input_shape: Dict[str, ShapeSpec]): + ret = super().from_config(cfg, input_shape) + ret["box2box_transform"] = Box2BoxTransformRotated(weights=cfg.MODEL.RPN.BBOX_REG_WEIGHTS) + return ret + + @torch.no_grad() + def label_and_sample_anchors(self, anchors: List[RotatedBoxes], gt_instances: List[Instances]): + """ + Args: + anchors (list[RotatedBoxes]): anchors for each feature map. + gt_instances: the ground-truth instances for each image. + + Returns: + list[Tensor]: + List of #img tensors. i-th element is a vector of labels whose length is + the total number of anchors across feature maps. Label values are in {-1, 0, 1}, + with meanings: -1 = ignore; 0 = negative class; 1 = positive class. + list[Tensor]: + i-th element is a Nx5 tensor, where N is the total number of anchors across + feature maps. The values are the matched gt boxes for each anchor. + Values are undefined for those anchors not labeled as 1. + """ + anchors = RotatedBoxes.cat(anchors) + + gt_boxes = [x.gt_boxes for x in gt_instances] + del gt_instances + + gt_labels = [] + matched_gt_boxes = [] + for gt_boxes_i in gt_boxes: + """ + gt_boxes_i: ground-truth boxes for i-th image + """ + match_quality_matrix = retry_if_cuda_oom(pairwise_iou_rotated)(gt_boxes_i, anchors) + matched_idxs, gt_labels_i = retry_if_cuda_oom(self.anchor_matcher)(match_quality_matrix) + # Matching is memory-expensive and may result in CPU tensors. But the result is small + gt_labels_i = gt_labels_i.to(device=gt_boxes_i.device) + + # A vector of labels (-1, 0, 1) for each anchor + gt_labels_i = self._subsample_labels(gt_labels_i) + + if len(gt_boxes_i) == 0: + # These values won't be used anyway since the anchor is labeled as background + matched_gt_boxes_i = torch.zeros_like(anchors.tensor) + else: + # TODO wasted indexing computation for ignored boxes + matched_gt_boxes_i = gt_boxes_i[matched_idxs].tensor + + gt_labels.append(gt_labels_i) # N,AHW + matched_gt_boxes.append(matched_gt_boxes_i) + return gt_labels, matched_gt_boxes + + @torch.no_grad() + def predict_proposals(self, anchors, pred_objectness_logits, pred_anchor_deltas, image_sizes): + pred_proposals = self._decode_proposals(anchors, pred_anchor_deltas) + return find_top_rrpn_proposals( + pred_proposals, + pred_objectness_logits, + image_sizes, + self.nms_thresh, + self.pre_nms_topk[self.training], + self.post_nms_topk[self.training], + self.min_box_size, + self.training, + ) diff --git a/detectron2/modeling/roi_heads/__init__.py b/detectron2/modeling/roi_heads/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3950e1bc22dfd9024b5371ae9fdb0fe4a45ab0e1 --- /dev/null +++ b/detectron2/modeling/roi_heads/__init__.py @@ -0,0 +1,35 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .box_head import ROI_BOX_HEAD_REGISTRY, build_box_head, FastRCNNConvFCHead +from .keypoint_head import ( + ROI_KEYPOINT_HEAD_REGISTRY, + build_keypoint_head, + BaseKeypointRCNNHead, + KRCNNConvDeconvUpsampleHead, +) +from .mask_head import ( + ROI_MASK_HEAD_REGISTRY, + build_mask_head, + BaseMaskRCNNHead, + MaskRCNNConvUpsampleHead, +) +from .roi_heads import ( + ROI_HEADS_REGISTRY, + ROIHeads, + Res5ROIHeads, + StandardROIHeads, + build_roi_heads, + select_foreground_proposals, +) +from .clip_roi_heads import ( + CLIPRes5ROIHeads, + CLIPSwinROIHeads, + PretrainRes5ROIHeads, + CLIPStandardROIHeads, +) +from .cascade_rcnn import CascadeROIHeads +from .rotated_fast_rcnn import RROIHeads +from .fast_rcnn import FastRCNNOutputLayers + +from . import cascade_rcnn # isort:skip + +__all__ = list(globals().keys()) diff --git a/detectron2/modeling/roi_heads/__pycache__/__init__.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..4dfa6694b47cebcb9995beea2e8ee8417ad0172e Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/__pycache__/box_head.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/box_head.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a40a44dd01253f1198e7f1feb3950f525250b982 Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/box_head.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/__pycache__/cascade_rcnn.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/cascade_rcnn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..af576dabe7e71bf6f861e55c862a31081257680a Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/cascade_rcnn.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/__pycache__/clip_roi_heads.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/clip_roi_heads.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..37e17bb00cbce274ae608ad8655e4f95c56415ee Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/clip_roi_heads.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/__pycache__/fast_rcnn.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/fast_rcnn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..996533a2a855d495fb2d13173583bfe41f528db7 Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/fast_rcnn.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/__pycache__/keypoint_head.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/keypoint_head.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c376674bc4d3fac4b23918b3515069c1762ef611 Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/keypoint_head.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/__pycache__/mask_head.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/mask_head.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a9fdef37dce0fc55a421946f072860e916511e25 Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/mask_head.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/__pycache__/roi_heads.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/roi_heads.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ba96ecb311ed4cb76f7337dd5484729b93eb87b1 Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/roi_heads.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/__pycache__/rotated_fast_rcnn.cpython-39.pyc b/detectron2/modeling/roi_heads/__pycache__/rotated_fast_rcnn.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..08833c8e09688e5380625efffaa70c850e684b61 Binary files /dev/null and b/detectron2/modeling/roi_heads/__pycache__/rotated_fast_rcnn.cpython-39.pyc differ diff --git a/detectron2/modeling/roi_heads/box_head.py b/detectron2/modeling/roi_heads/box_head.py new file mode 100644 index 0000000000000000000000000000000000000000..5d0370b0400d9268f13c905e4096a84ce42e9bfd --- /dev/null +++ b/detectron2/modeling/roi_heads/box_head.py @@ -0,0 +1,118 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import numpy as np +from typing import List +import fvcore.nn.weight_init as weight_init +import torch +from torch import nn + +from detectron2.config import configurable +from detectron2.layers import Conv2d, ShapeSpec, get_norm +from detectron2.utils.registry import Registry + +__all__ = ["FastRCNNConvFCHead", "build_box_head", "ROI_BOX_HEAD_REGISTRY"] + +ROI_BOX_HEAD_REGISTRY = Registry("ROI_BOX_HEAD") +ROI_BOX_HEAD_REGISTRY.__doc__ = """ +Registry for box heads, which make box predictions from per-region features. + +The registered object will be called with `obj(cfg, input_shape)`. +""" + + +# To get torchscript support, we make the head a subclass of `nn.Sequential`. +# Therefore, to add new layers in this head class, please make sure they are +# added in the order they will be used in forward(). +@ROI_BOX_HEAD_REGISTRY.register() +class FastRCNNConvFCHead(nn.Sequential): + """ + A head with several 3x3 conv layers (each followed by norm & relu) and then + several fc layers (each followed by relu). + """ + + @configurable + def __init__( + self, input_shape: ShapeSpec, *, conv_dims: List[int], fc_dims: List[int], conv_norm="" + ): + """ + NOTE: this interface is experimental. + + Args: + input_shape (ShapeSpec): shape of the input feature. + conv_dims (list[int]): the output dimensions of the conv layers + fc_dims (list[int]): the output dimensions of the fc layers + conv_norm (str or callable): normalization for the conv layers. + See :func:`detectron2.layers.get_norm` for supported types. + """ + super().__init__() + assert len(conv_dims) + len(fc_dims) > 0 + + self._output_size = (input_shape.channels, input_shape.height, input_shape.width) + + self.conv_norm_relus = [] + for k, conv_dim in enumerate(conv_dims): + conv = Conv2d( + self._output_size[0], + conv_dim, + kernel_size=3, + padding=1, + bias=not conv_norm, + norm=get_norm(conv_norm, conv_dim), + activation=nn.ReLU(), + ) + self.add_module("conv{}".format(k + 1), conv) + self.conv_norm_relus.append(conv) + self._output_size = (conv_dim, self._output_size[1], self._output_size[2]) + + self.fcs = [] + for k, fc_dim in enumerate(fc_dims): + if k == 0: + self.add_module("flatten", nn.Flatten()) + fc = nn.Linear(int(np.prod(self._output_size)), fc_dim) + self.add_module("fc{}".format(k + 1), fc) + self.add_module("fc_relu{}".format(k + 1), nn.ReLU()) + self.fcs.append(fc) + self._output_size = fc_dim + + for layer in self.conv_norm_relus: + weight_init.c2_msra_fill(layer) + for layer in self.fcs: + weight_init.c2_xavier_fill(layer) + + @classmethod + def from_config(cls, cfg, input_shape): + num_conv = cfg.MODEL.ROI_BOX_HEAD.NUM_CONV + conv_dim = cfg.MODEL.ROI_BOX_HEAD.CONV_DIM + num_fc = cfg.MODEL.ROI_BOX_HEAD.NUM_FC + fc_dim = cfg.MODEL.ROI_BOX_HEAD.FC_DIM + return { + "input_shape": input_shape, + "conv_dims": [conv_dim] * num_conv, + "fc_dims": [fc_dim] * num_fc, + "conv_norm": cfg.MODEL.ROI_BOX_HEAD.NORM, + } + + def forward(self, x): + for layer in self: + x = layer(x) + return x + + @property + @torch.jit.unused + def output_shape(self): + """ + Returns: + ShapeSpec: the output feature shape + """ + o = self._output_size + if isinstance(o, int): + return ShapeSpec(channels=o) + else: + return ShapeSpec(channels=o[0], height=o[1], width=o[2]) + + +def build_box_head(cfg, input_shape): + """ + Build a box head defined by `cfg.MODEL.ROI_BOX_HEAD.NAME`. + """ + name = cfg.MODEL.ROI_BOX_HEAD.NAME + return ROI_BOX_HEAD_REGISTRY.get(name)(cfg, input_shape) diff --git a/detectron2/modeling/roi_heads/cascade_rcnn.py b/detectron2/modeling/roi_heads/cascade_rcnn.py new file mode 100644 index 0000000000000000000000000000000000000000..bc110653052522bfcf4a01a835d938a19e5a8b2d --- /dev/null +++ b/detectron2/modeling/roi_heads/cascade_rcnn.py @@ -0,0 +1,298 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from typing import List +import torch +from torch import nn +from torch.autograd.function import Function + +from detectron2.config import configurable +from detectron2.layers import ShapeSpec +from detectron2.structures import Boxes, Instances, pairwise_iou +from detectron2.utils.events import get_event_storage + +from ..box_regression import Box2BoxTransform +from ..matcher import Matcher +from ..poolers import ROIPooler +from .box_head import build_box_head +from .fast_rcnn import FastRCNNOutputLayers, fast_rcnn_inference +from .roi_heads import ROI_HEADS_REGISTRY, StandardROIHeads + + +class _ScaleGradient(Function): + @staticmethod + def forward(ctx, input, scale): + ctx.scale = scale + return input + + @staticmethod + def backward(ctx, grad_output): + return grad_output * ctx.scale, None + + +@ROI_HEADS_REGISTRY.register() +class CascadeROIHeads(StandardROIHeads): + """ + The ROI heads that implement :paper:`Cascade R-CNN`. + """ + + @configurable + def __init__( + self, + *, + box_in_features: List[str], + box_pooler: ROIPooler, + box_heads: List[nn.Module], + box_predictors: List[nn.Module], + proposal_matchers: List[Matcher], + **kwargs, + ): + """ + NOTE: this interface is experimental. + + Args: + box_pooler (ROIPooler): pooler that extracts region features from given boxes + box_heads (list[nn.Module]): box head for each cascade stage + box_predictors (list[nn.Module]): box predictor for each cascade stage + proposal_matchers (list[Matcher]): matcher with different IoU thresholds to + match boxes with ground truth for each stage. The first matcher matches + RPN proposals with ground truth, the other matchers use boxes predicted + by the previous stage as proposals and match them with ground truth. + """ + assert "proposal_matcher" not in kwargs, ( + "CascadeROIHeads takes 'proposal_matchers=' for each stage instead " + "of one 'proposal_matcher='." + ) + # The first matcher matches RPN proposals with ground truth, done in the base class + kwargs["proposal_matcher"] = proposal_matchers[0] + num_stages = self.num_cascade_stages = len(box_heads) + box_heads = nn.ModuleList(box_heads) + box_predictors = nn.ModuleList(box_predictors) + assert len(box_predictors) == num_stages, f"{len(box_predictors)} != {num_stages}!" + assert len(proposal_matchers) == num_stages, f"{len(proposal_matchers)} != {num_stages}!" + super().__init__( + box_in_features=box_in_features, + box_pooler=box_pooler, + box_head=box_heads, + box_predictor=box_predictors, + **kwargs, + ) + self.proposal_matchers = proposal_matchers + + @classmethod + def from_config(cls, cfg, input_shape): + ret = super().from_config(cfg, input_shape) + ret.pop("proposal_matcher") + return ret + + @classmethod + def _init_box_head(cls, cfg, input_shape): + # fmt: off + in_features = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_scales = tuple(1.0 / input_shape[k].stride for k in in_features) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + cascade_bbox_reg_weights = cfg.MODEL.ROI_BOX_CASCADE_HEAD.BBOX_REG_WEIGHTS + cascade_ious = cfg.MODEL.ROI_BOX_CASCADE_HEAD.IOUS + assert len(cascade_bbox_reg_weights) == len(cascade_ious) + assert cfg.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG, \ + "CascadeROIHeads only support class-agnostic regression now!" + assert cascade_ious[0] == cfg.MODEL.ROI_HEADS.IOU_THRESHOLDS[0] + # fmt: on + + in_channels = [input_shape[f].channels for f in in_features] + # Check all channel counts are equal + assert len(set(in_channels)) == 1, in_channels + in_channels = in_channels[0] + + box_pooler = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + pooled_shape = ShapeSpec( + channels=in_channels, width=pooler_resolution, height=pooler_resolution + ) + + box_heads, box_predictors, proposal_matchers = [], [], [] + for match_iou, bbox_reg_weights in zip(cascade_ious, cascade_bbox_reg_weights): + box_head = build_box_head(cfg, pooled_shape) + box_heads.append(box_head) + box_predictors.append( + FastRCNNOutputLayers( + cfg, + box_head.output_shape, + box2box_transform=Box2BoxTransform(weights=bbox_reg_weights), + ) + ) + proposal_matchers.append(Matcher([match_iou], [0, 1], allow_low_quality_matches=False)) + return { + "box_in_features": in_features, + "box_pooler": box_pooler, + "box_heads": box_heads, + "box_predictors": box_predictors, + "proposal_matchers": proposal_matchers, + } + + def forward(self, images, features, proposals, targets=None): + del images + if self.training: + proposals = self.label_and_sample_proposals(proposals, targets) + + if self.training: + # Need targets to box head + losses = self._forward_box(features, proposals, targets) + losses.update(self._forward_mask(features, proposals)) + losses.update(self._forward_keypoint(features, proposals)) + return proposals, losses + else: + pred_instances = self._forward_box(features, proposals) + pred_instances = self.forward_with_given_boxes(features, pred_instances) + return pred_instances, {} + + def _forward_box(self, features, proposals, targets=None): + """ + Args: + features, targets: the same as in + Same as in :meth:`ROIHeads.forward`. + proposals (list[Instances]): the per-image object proposals with + their matching ground truth. + Each has fields "proposal_boxes", and "objectness_logits", + "gt_classes", "gt_boxes". + """ + features = [features[f] for f in self.box_in_features] + head_outputs = [] # (predictor, predictions, proposals) + prev_pred_boxes = None + image_sizes = [x.image_size for x in proposals] + for k in range(self.num_cascade_stages): + if k > 0: + # The output boxes of the previous stage are used to create the input + # proposals of the next stage. + proposals = self._create_proposals_from_boxes(prev_pred_boxes, image_sizes) + if self.training: + proposals = self._match_and_label_boxes(proposals, k, targets) + predictions = self._run_stage(features, proposals, k) + prev_pred_boxes = self.box_predictor[k].predict_boxes(predictions, proposals) + head_outputs.append((self.box_predictor[k], predictions, proposals)) + + if self.training: + losses = {} + storage = get_event_storage() + for stage, (predictor, predictions, proposals) in enumerate(head_outputs): + with storage.name_scope("stage{}".format(stage)): + stage_losses = predictor.losses(predictions, proposals) + losses.update({k + "_stage{}".format(stage): v for k, v in stage_losses.items()}) + return losses + else: + # Each is a list[Tensor] of length #image. Each tensor is Ri x (K+1) + scores_per_stage = [h[0].predict_probs(h[1], h[2]) for h in head_outputs] + + # Average the scores across heads + scores = [ + sum(list(scores_per_image)) * (1.0 / self.num_cascade_stages) + for scores_per_image in zip(*scores_per_stage) + ] + # Use the boxes of the last head + predictor, predictions, proposals = head_outputs[-1] + boxes = predictor.predict_boxes(predictions, proposals) + pred_instances, _ = fast_rcnn_inference( + boxes, + scores, + image_sizes, + predictor.test_score_thresh, + predictor.test_nms_thresh, + predictor.test_topk_per_image, + ) + return pred_instances + + @torch.no_grad() + def _match_and_label_boxes(self, proposals, stage, targets): + """ + Match proposals with groundtruth using the matcher at the given stage. + Label the proposals as foreground or background based on the match. + + Args: + proposals (list[Instances]): One Instances for each image, with + the field "proposal_boxes". + stage (int): the current stage + targets (list[Instances]): the ground truth instances + + Returns: + list[Instances]: the same proposals, but with fields "gt_classes" and "gt_boxes" + """ + num_fg_samples, num_bg_samples = [], [] + for proposals_per_image, targets_per_image in zip(proposals, targets): + match_quality_matrix = pairwise_iou( + targets_per_image.gt_boxes, proposals_per_image.proposal_boxes + ) + # proposal_labels are 0 or 1 + matched_idxs, proposal_labels = self.proposal_matchers[stage](match_quality_matrix) + if len(targets_per_image) > 0: + gt_classes = targets_per_image.gt_classes[matched_idxs] + # Label unmatched proposals (0 label from matcher) as background (label=num_classes) + gt_classes[proposal_labels == 0] = self.num_classes + gt_boxes = targets_per_image.gt_boxes[matched_idxs] + else: + gt_classes = torch.zeros_like(matched_idxs) + self.num_classes + gt_boxes = Boxes( + targets_per_image.gt_boxes.tensor.new_zeros((len(proposals_per_image), 4)) + ) + proposals_per_image.gt_classes = gt_classes + proposals_per_image.gt_boxes = gt_boxes + + num_fg_samples.append((proposal_labels == 1).sum().item()) + num_bg_samples.append(proposal_labels.numel() - num_fg_samples[-1]) + + # Log the number of fg/bg samples in each stage + storage = get_event_storage() + storage.put_scalar( + "stage{}/roi_head/num_fg_samples".format(stage), + sum(num_fg_samples) / len(num_fg_samples), + ) + storage.put_scalar( + "stage{}/roi_head/num_bg_samples".format(stage), + sum(num_bg_samples) / len(num_bg_samples), + ) + return proposals + + def _run_stage(self, features, proposals, stage): + """ + Args: + features (list[Tensor]): #lvl input features to ROIHeads + proposals (list[Instances]): #image Instances, with the field "proposal_boxes" + stage (int): the current stage + + Returns: + Same output as `FastRCNNOutputLayers.forward()`. + """ + box_features = self.box_pooler(features, [x.proposal_boxes for x in proposals]) + # The original implementation averages the losses among heads, + # but scale up the parameter gradients of the heads. + # This is equivalent to adding the losses among heads, + # but scale down the gradients on features. + box_features = _ScaleGradient.apply(box_features, 1.0 / self.num_cascade_stages) + box_features = self.box_head[stage](box_features) + return self.box_predictor[stage](box_features) + + def _create_proposals_from_boxes(self, boxes, image_sizes): + """ + Args: + boxes (list[Tensor]): per-image predicted boxes, each of shape Ri x 4 + image_sizes (list[tuple]): list of image shapes in (h, w) + + Returns: + list[Instances]: per-image proposals with the given boxes. + """ + # Just like RPN, the proposals should not have gradients + boxes = [Boxes(b.detach()) for b in boxes] + proposals = [] + for boxes_per_image, image_size in zip(boxes, image_sizes): + boxes_per_image.clip(image_size) + if self.training: + # do not filter empty boxes at inference time, + # because the scores from each stage need to be aligned and added later + boxes_per_image = boxes_per_image[boxes_per_image.nonempty()] + prop = Instances(image_size) + prop.proposal_boxes = boxes_per_image + proposals.append(prop) + return proposals diff --git a/detectron2/modeling/roi_heads/clip_roi_heads.py b/detectron2/modeling/roi_heads/clip_roi_heads.py new file mode 100644 index 0000000000000000000000000000000000000000..7144e8f3b1f25b395c290c218ddee129148ea3b5 --- /dev/null +++ b/detectron2/modeling/roi_heads/clip_roi_heads.py @@ -0,0 +1,747 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import inspect +import logging +import numpy as np +from typing import Dict, List, Optional, Tuple +import torch +from torch import nn + +from detectron2.config import configurable +from detectron2.layers import ShapeSpec, nonzero_tuple +from detectron2.structures import Boxes, ImageList, Instances, pairwise_iou +from detectron2.utils.events import get_event_storage +from detectron2.utils.registry import Registry + +from ..backbone.resnet import BottleneckBlock, ResNet +from ..matcher import Matcher +from ..poolers import ROIPooler +from ..proposal_generator.proposal_utils import add_ground_truth_to_proposals +from ..sampling import subsample_labels +from .box_head import build_box_head +from .fast_rcnn import FastRCNNOutputLayers +from .keypoint_head import build_keypoint_head +from .mask_head import build_mask_head + +from .roi_heads import ROI_HEADS_REGISTRY, select_foreground_proposals, ROIHeads + +@ROI_HEADS_REGISTRY.register() +class CLIPRes5ROIHeads(ROIHeads): + """ + Created for CLIP ResNet. This head uses the last resnet layer from backbone. + The ROIHeads in a typical "C4" R-CNN model, where + the box and mask head share the cropping and + the per-region feature computation by a Res5 block. + See :paper:`ResNet` Appendix A. + """ + + @configurable + def __init__( + self, + *, + in_features: List[str], + pooler: ROIPooler, + res5: None, + box_predictor: nn.Module, + mask_head: Optional[nn.Module] = None, + **kwargs, + ): + """ + NOTE: this interface is experimental. + + Args: + in_features (list[str]): list of backbone feature map names to use for + feature extraction + pooler (ROIPooler): pooler to extra region features from backbone + res5 (nn.Sequential): a CNN to compute per-region features, to be used by + ``box_predictor`` and ``mask_head``. Typically this is a "res5" + block from a ResNet. + box_predictor (nn.Module): make box predictions from the feature. + Should have the same interface as :class:`FastRCNNOutputLayers`. + mask_head (nn.Module): transform features to make mask predictions + """ + super().__init__(**kwargs) + self.in_features = in_features + self.pooler = pooler + # if isinstance(res5, (list, tuple)): + # res5 = nn.Sequential(*res5) + self.res5 = res5 # None, this head uses the res5 from backbone + self.box_predictor = box_predictor + self.mask_on = mask_head is not None + if self.mask_on: + self.mask_head = mask_head + + @classmethod + def from_config(cls, cfg, input_shape): + # fmt: off + ret = super().from_config(cfg) + in_features = ret["in_features"] = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + pooler_scales = (1.0 / input_shape[in_features[0]].stride, ) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + mask_on = cfg.MODEL.MASK_ON + # fmt: on + assert not cfg.MODEL.KEYPOINT_ON + assert len(in_features) == 1 + + ret["pooler"] = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + + # Compatbility with old moco code. Might be useful. + # See notes in StandardROIHeads.from_config + # if not inspect.ismethod(cls._build_res5_block): + # logger.warning( + # "The behavior of _build_res5_block may change. " + # "Please do not depend on private methods." + # ) + # cls._build_res5_block = classmethod(cls._build_res5_block) + + ret["res5"], out_channels = None, cfg.MODEL.RESNETS.RES2_OUT_CHANNELS * 8 # cls._build_res5_block(cfg) + ret["box_predictor"] = FastRCNNOutputLayers( + cfg, ShapeSpec(channels=out_channels, height=1, width=1) + ) + + if mask_on: + ret["mask_head"] = build_mask_head( + cfg, + ShapeSpec(channels=out_channels, width=pooler_resolution, height=pooler_resolution), + ) + return ret + + def _shared_roi_transform(self, features, boxes, backbone_res5): + x = self.pooler(features, boxes) + return backbone_res5(x) + + def forward(self, images, features, proposals, queries, targets=None, + res5=None, ds=None, norm=None, vision_projection=None, attnpool=None): + """ + See :meth:`ROIHeads.forward`. + """ + del images + + if self.training: + assert targets + proposals = self.label_and_sample_proposals(proposals, targets) + del targets + + proposal_boxes = [x.proposal_boxes for x in proposals] + box_features = self._shared_roi_transform( + [features[f] for f in self.in_features], proposal_boxes, res5 + ) + if attnpool: # att pooling + att_feats = attnpool(box_features) + predictions = self.box_predictor(att_feats, queries) + else: # mean pooling + predictions = self.box_predictor(box_features.mean(dim=[2, 3])) + if self.training: + del features + losses = self.box_predictor.losses(predictions, proposals) + if self.mask_on: + proposals, fg_selection_masks = select_foreground_proposals( + proposals, self.num_classes + ) + # Since the ROI feature transform is shared between boxes and masks, + # we don't need to recompute features. The mask loss is only defined + # on foreground proposals, so we need to select out the foreground + # features. + mask_features = box_features[torch.cat(fg_selection_masks, dim=0)] + del box_features + losses.update(self.mask_head(mask_features, proposals)) + return [], losses + else: + pred_instances, _ = self.box_predictor.inference(predictions, proposals) + pred_instances = self.forward_with_given_boxes(features, pred_instances, res5) + return pred_instances, {} + + def forward_with_given_boxes(self, features, instances, res5=None): + """ + Use the given boxes in `instances` to produce other (non-box) per-ROI outputs. + + Args: + features: same as in `forward()` + instances (list[Instances]): instances to predict other outputs. Expect the keys + "pred_boxes" and "pred_classes" to exist. + + Returns: + instances (Instances): + the same `Instances` object, with extra + fields such as `pred_masks` or `pred_keypoints`. + """ + assert not self.training + assert instances[0].has("pred_boxes") and instances[0].has("pred_classes") + + if self.mask_on: + features = [features[f] for f in self.in_features] + x = self._shared_roi_transform(features, [x.pred_boxes for x in instances], res5) + return self.mask_head(x, instances) + else: + return instances + +@ROI_HEADS_REGISTRY.register() +class CLIPSwinROIHeads(ROIHeads): + """ + Created for CLIP ResNet. This head uses the last resnet layer from backbone. + The ROIHeads in a typical "C4" R-CNN model, where + the box and mask head share the cropping and + the per-region feature computation by a Res5 block. + See :paper:`ResNet` Appendix A. + """ + + @configurable + def __init__( + self, + *, + in_features: List[str], + pooler: ROIPooler, + res5: None, + box_predictor: nn.Module, + mask_head: Optional[nn.Module] = None, + **kwargs, + ): + """ + NOTE: this interface is experimental. + + Args: + in_features (list[str]): list of backbone feature map names to use for + feature extraction + pooler (ROIPooler): pooler to extra region features from backbone + res5 (nn.Sequential): a CNN to compute per-region features, to be used by + ``box_predictor`` and ``mask_head``. Typically this is a "res5" + block from a ResNet. + box_predictor (nn.Module): make box predictions from the feature. + Should have the same interface as :class:`FastRCNNOutputLayers`. + mask_head (nn.Module): transform features to make mask predictions + """ + super().__init__(**kwargs) + self.in_features = in_features + self.pooler = pooler + # if isinstance(res5, (list, tuple)): + # res5 = nn.Sequential(*res5) + self.res5 = res5 # None, this head uses the res5 from backbone + self.box_predictor = box_predictor + self.mask_on = mask_head is not None + if self.mask_on: + self.mask_head = mask_head + + @classmethod + def from_config(cls, cfg, input_shape): + # fmt: off + ret = super().from_config(cfg) + in_features = ret["in_features"] = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + pooler_scales = (1.0 / input_shape[in_features[0]].stride, ) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + mask_on = cfg.MODEL.MASK_ON + # fmt: on + assert not cfg.MODEL.KEYPOINT_ON + assert len(in_features) == 1 + + ret["pooler"] = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + + # Compatbility with old moco code. Might be useful. + # See notes in StandardROIHeads.from_config + # if not inspect.ismethod(cls._build_res5_block): + # logger.warning( + # "The behavior of _build_res5_block may change. " + # "Please do not depend on private methods." + # ) + # cls._build_res5_block = classmethod(cls._build_res5_block) + + ret["res5"], out_channels = None, cfg.MODEL.RESNETS.RES2_OUT_CHANNELS * 8 # cls._build_res5_block(cfg) + ret["box_predictor"] = FastRCNNOutputLayers( + cfg, ShapeSpec(channels=out_channels, height=1, width=1) + ) + + if mask_on: + ret["mask_head"] = build_mask_head( + cfg, + ShapeSpec(channels=out_channels, width=pooler_resolution, height=pooler_resolution), + ) + return ret + + def _shared_roi_transform(self, features, boxes, backbone_res5, backbone_ds): + x = self.pooler(features, boxes) + if backbone_ds: + x_flattened = x.flatten(2).transpose(1, 2) + x_ds = backbone_ds(x_flattened, x.shape[2], x.shape[3]) + return backbone_res5(x_ds, x.shape[2] // 2, x.shape[3] // 2) + else: + return backbone_res5(x) + + def forward(self, images, features, proposals, queries, targets=None, + res5=None, ds=None, norm=None, vision_projection=None, attnpool=None): + """ + See :meth:`ROIHeads.forward`. + """ + del images + + if self.training: + assert targets + proposals = self.label_and_sample_proposals(proposals, targets) + del targets + + proposal_boxes = [x.proposal_boxes for x in proposals] + box_features = self._shared_roi_transform( + [features[f] for f in self.in_features], proposal_boxes, res5, ds, + ) + if isinstance(box_features, tuple): + box_features = norm(box_features[0]).mean(1) + box_features = box_features @ vision_projection + box_features = box_features / box_features.norm(dim=-1, keepdim=True) + + if attnpool: # att pooling + att_feats = attnpool(box_features) + predictions = self.box_predictor(att_feats) + else: # mean pooling + predictions = self.box_predictor(box_features, queries) + + if self.training: + del features + losses = self.box_predictor.losses(predictions, proposals) + if self.mask_on: + proposals, fg_selection_masks = select_foreground_proposals( + proposals, self.num_classes + ) + # Since the ROI feature transform is shared between boxes and masks, + # we don't need to recompute features. The mask loss is only defined + # on foreground proposals, so we need to select out the foreground + # features. + mask_features = box_features[torch.cat(fg_selection_masks, dim=0)] + del box_features + losses.update(self.mask_head(mask_features, proposals)) + return [], losses + else: + pred_instances, _ = self.box_predictor.inference(predictions, proposals) + # pred_instances = self.forward_with_given_boxes(features, pred_instances, res5) + return pred_instances, {} + + def forward_with_given_boxes(self, features, instances, res5=None): + """ + Use the given boxes in `instances` to produce other (non-box) per-ROI outputs. + + Args: + features: same as in `forward()` + instances (list[Instances]): instances to predict other outputs. Expect the keys + "pred_boxes" and "pred_classes" to exist. + + Returns: + instances (Instances): + the same `Instances` object, with extra + fields such as `pred_masks` or `pred_keypoints`. + """ + assert not self.training + assert instances[0].has("pred_boxes") and instances[0].has("pred_classes") + + if self.mask_on: + features = [features[f] for f in self.in_features] + x = self._shared_roi_transform(features, [x.pred_boxes for x in instances], res5) + return self.mask_head(x, instances) + else: + return instances + +@ROI_HEADS_REGISTRY.register() +class PretrainRes5ROIHeads(ROIHeads): + """ + Created for pretraining CLIP ResNet without box_predictor. This head uses the last resnet layer from backbone. + The ROIHeads in a typical "C4" R-CNN model, where + the box and mask head share the cropping and + the per-region feature computation by a Res5 block. + See :paper:`ResNet` Appendix A. + """ + + @configurable + def __init__( + self, + *, + in_features: List[str], + pooler: ROIPooler, + res5: None, + box_predictor: Optional[nn.Module] = None, + mask_head: Optional[nn.Module] = None, + **kwargs, + ): + """ + NOTE: this interface is experimental. + + Args: + in_features (list[str]): list of backbone feature map names to use for + feature extraction + pooler (ROIPooler): pooler to extra region features from backbone + res5 (nn.Sequential): a CNN to compute per-region features, to be used by + ``box_predictor`` and ``mask_head``. Typically this is a "res5" + block from a ResNet. + box_predictor (nn.Module): make box predictions from the feature. + Should have the same interface as :class:`FastRCNNOutputLayers`. + mask_head (nn.Module): transform features to make mask predictions + """ + super().__init__(**kwargs) + self.in_features = in_features + self.pooler = pooler + # if isinstance(res5, (list, tuple)): + # res5 = nn.Sequential(*res5) + self.res5 = res5 # None, this head uses the res5 from backbone + self.box_predictor = None + self.mask_on = None + + @classmethod + def from_config(cls, cfg, input_shape): + # fmt: off + ret = super().from_config(cfg) + in_features = ret["in_features"] = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + pooler_scales = (1.0 / input_shape[in_features[0]].stride, ) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + mask_on = cfg.MODEL.MASK_ON + # fmt: on + assert not cfg.MODEL.KEYPOINT_ON + assert len(in_features) == 1 + + ret["pooler"] = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + + ret["res5"], out_channels = None, cfg.MODEL.RESNETS.RES2_OUT_CHANNELS * 8 # cls._build_res5_block(cfg) + ret["box_predictor"] = None + ret["mask_head"] = None + return ret + + def _shared_roi_transform(self, features, boxes, backbone_res5, backbone_ds): + x = self.pooler(features, boxes) + if backbone_ds: + return backbone_res5(backbone_ds(x)) + else: + return backbone_res5(x) + + def forward(self, images, features, proposals, targets=None, res5=None, ds=None, attnpool=None): + """ + See :meth:`ROIHeads.forward`. + """ + # if self.training: + # assert targets + # proposals = self.label_and_sample_proposals(proposals, targets) + # del targets + if isinstance(proposals[0], Boxes): # grid boxes + proposal_boxes = proposals + else: # object proposals + proposal_boxes = [x.proposal_boxes for x in proposals] + box_features = self._shared_roi_transform( + [features[f] for f in self.in_features], proposal_boxes, res5 + ) + if attnpool: # att pooling + att_feats = attnpool(box_features) + region_feats = att_feats # self.box_predictor(att_feats) + else: # mean pooling + region_feats = box_features.mean(dim=[2, 3]) # self.box_predictor(box_features.mean(dim=[2, 3])) + + return region_feats + + def forward_with_given_boxes(self, features, instances, res5=None): + """ + Use the given boxes in `instances` to produce other (non-box) per-ROI outputs. + + Args: + features: same as in `forward()` + instances (list[Instances]): instances to predict other outputs. Expect the keys + "pred_boxes" and "pred_classes" to exist. + + Returns: + instances (Instances): + the same `Instances` object, with extra + fields such as `pred_masks` or `pred_keypoints`. + """ + assert not self.training + assert instances[0].has("pred_boxes") and instances[0].has("pred_classes") + + return instances + +@ROI_HEADS_REGISTRY.register() +class CLIPStandardROIHeads(ROIHeads): + """ + Created for CLIP ResNet. This head uses the attention pool layers from backbone. + It's "standard" in a sense that there is no ROI transform sharing + or feature sharing between tasks. + Each head independently processes the input features by each head's + own pooler and head. + + This class is used by most models, such as FPN and C5. + To implement more models, you can subclass it and implement a different + :meth:`forward()` or a head. + """ + + @configurable + def __init__( + self, + *, + box_in_features: List[str], + box_pooler: ROIPooler, + box_head: nn.Module, + box_predictor: nn.Module, + mask_in_features: Optional[List[str]] = None, + mask_pooler: Optional[ROIPooler] = None, + mask_head: Optional[nn.Module] = None, + train_on_pred_boxes: bool = False, + **kwargs, + ): + """ + NOTE: this interface is experimental. + + Args: + box_in_features (list[str]): list of feature names to use for the box head. + box_pooler (ROIPooler): pooler to extra region features for box head + box_head (nn.Module): transform features to make box predictions + box_predictor (nn.Module): make box predictions from the feature. + Should have the same interface as :class:`FastRCNNOutputLayers`. + mask_in_features (list[str]): list of feature names to use for the mask + pooler or mask head. None if not using mask head. + mask_pooler (ROIPooler): pooler to extract region features from image features. + The mask head will then take region features to make predictions. + If None, the mask head will directly take the dict of image features + defined by `mask_in_features` + mask_head (nn.Module): transform features to make mask predictions + keypoint_in_features, keypoint_pooler, keypoint_head: similar to ``mask_*``. + train_on_pred_boxes (bool): whether to use proposal boxes or + predicted boxes from the box head to train other heads. + """ + super().__init__(**kwargs) + # keep self.in_features for backward compatibility + self.in_features = self.box_in_features = box_in_features + self.box_pooler = box_pooler + self.box_head = box_head + self.box_predictor = box_predictor + + self.mask_on = mask_in_features is not None + if self.mask_on: + self.mask_in_features = mask_in_features + self.mask_pooler = mask_pooler + self.mask_head = mask_head + + self.train_on_pred_boxes = train_on_pred_boxes + + @classmethod + def from_config(cls, cfg, input_shape): + ret = super().from_config(cfg) + ret["train_on_pred_boxes"] = cfg.MODEL.ROI_BOX_HEAD.TRAIN_ON_PRED_BOXES + # Subclasses that have not been updated to use from_config style construction + # may have overridden _init_*_head methods. In this case, those overridden methods + # will not be classmethods and we need to avoid trying to call them here. + # We test for this with ismethod which only returns True for bound methods of cls. + # Such subclasses will need to handle calling their overridden _init_*_head methods. + if inspect.ismethod(cls._init_box_head): + ret.update(cls._init_box_head(cfg, input_shape)) + if inspect.ismethod(cls._init_mask_head): + ret.update(cls._init_mask_head(cfg, input_shape)) + return ret + + @classmethod + def _init_box_head(cls, cfg, input_shape): + # fmt: off + in_features = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_scales = tuple(1.0 / input_shape[k].stride for k in in_features) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + # fmt: on + + # If StandardROIHeads is applied on multiple feature maps (as in FPN), + # then we share the same predictors and therefore the channel counts must be the same + in_channels = [input_shape[f].channels for f in in_features] + # Check all channel counts are equal + assert len(set(in_channels)) == 1, in_channels + in_channels = in_channels[0] + + box_pooler = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + # Here we split "box head" and "box predictor", which is mainly due to historical reasons. + # They are used together so the "box predictor" layers should be part of the "box head". + # New subclasses of ROIHeads do not need "box predictor"s. + box_head = None if cfg.MODEL.CLIP.USE_TEXT_EMB_CLASSIFIER else build_box_head( + cfg, ShapeSpec(channels=in_channels, height=pooler_resolution, width=pooler_resolution) + ) + box_head_output_shape = 1024 + box_predictor = FastRCNNOutputLayers(cfg, box_head_output_shape) + return { + "box_in_features": in_features, + "box_pooler": box_pooler, + "box_head": box_head, + "box_predictor": box_predictor, + } + + @classmethod + def _init_mask_head(cls, cfg, input_shape): + if not cfg.MODEL.MASK_ON: + return {} + # fmt: off + in_features = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_MASK_HEAD.POOLER_RESOLUTION + pooler_scales = tuple(1.0 / input_shape[k].stride for k in in_features) + sampling_ratio = cfg.MODEL.ROI_MASK_HEAD.POOLER_SAMPLING_RATIO + pooler_type = cfg.MODEL.ROI_MASK_HEAD.POOLER_TYPE + # fmt: on + + in_channels = [input_shape[f].channels for f in in_features][0] + + ret = {"mask_in_features": in_features} + ret["mask_pooler"] = ( + ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + if pooler_type + else None + ) + if pooler_type: + shape = ShapeSpec( + channels=in_channels, width=pooler_resolution, height=pooler_resolution + ) + else: + shape = {f: input_shape[f] for f in in_features} + ret["mask_head"] = build_mask_head(cfg, shape) + return ret + + def forward( + self, + images: ImageList, + features: Dict[str, torch.Tensor], + proposals: List[Instances], + targets: Optional[List[Instances]] = None, + attnpool=None, + ) -> Tuple[List[Instances], Dict[str, torch.Tensor]]: + """ + See :class:`ROIHeads.forward`. + """ + del images + if self.training: + assert targets, "'targets' argument is required during training" + proposals = self.label_and_sample_proposals(proposals, targets) + del targets + + if self.training: + losses = self._forward_box(features, proposals, attnpool=attnpool) + # Usually the original proposals used by the box head are used by the mask, keypoint + # heads. But when `self.train_on_pred_boxes is True`, proposals will contain boxes + # predicted by the box head. + losses.update(self._forward_mask(features, proposals)) + return proposals, losses + else: + pred_instances = self._forward_box(features, proposals, attnpool=attnpool) + # During inference cascaded prediction is used: the mask and keypoints heads are only + # applied to the top scoring box detections. + pred_instances = self.forward_with_given_boxes(features, pred_instances) + return pred_instances, {} + + def forward_with_given_boxes( + self, features: Dict[str, torch.Tensor], instances: List[Instances] + ) -> List[Instances]: + """ + Use the given boxes in `instances` to produce other (non-box) per-ROI outputs. + + This is useful for downstream tasks where a box is known, but need to obtain + other attributes (outputs of other heads). + Test-time augmentation also uses this. + + Args: + features: same as in `forward()` + instances (list[Instances]): instances to predict other outputs. Expect the keys + "pred_boxes" and "pred_classes" to exist. + + Returns: + list[Instances]: + the same `Instances` objects, with extra + fields such as `pred_masks` or `pred_keypoints`. + """ + assert not self.training + assert instances[0].has("pred_boxes") and instances[0].has("pred_classes") + + instances = self._forward_mask(features, instances) + return instances + + def _forward_box(self, features: Dict[str, torch.Tensor], proposals: List[Instances], attnpool=None): + """ + Forward logic of the box prediction branch. If `self.train_on_pred_boxes is True`, + the function puts predicted boxes in the `proposal_boxes` field of `proposals` argument. + + Args: + features (dict[str, Tensor]): mapping from feature map names to tensor. + Same as in :meth:`ROIHeads.forward`. + proposals (list[Instances]): the per-image object proposals with + their matching ground truth. + Each has fields "proposal_boxes", and "objectness_logits", + "gt_classes", "gt_boxes". + + Returns: + In training, a dict of losses. + In inference, a list of `Instances`, the predicted instances. + """ + features = [features[f] for f in self.box_in_features] + box_features = self.box_pooler(features, [x.proposal_boxes for x in proposals]) + if attnpool: # att pooling + box_features = attnpool(box_features) + else: # default FPN pooling (FastRCNNConvFCHead) + box_features = self.box_head(box_features) + predictions = self.box_predictor(box_features) + del box_features + + if self.training: + losses = self.box_predictor.losses(predictions, proposals) + # proposals is modified in-place below, so losses must be computed first. + if self.train_on_pred_boxes: + with torch.no_grad(): + pred_boxes = self.box_predictor.predict_boxes_for_gt_classes( + predictions, proposals + ) + for proposals_per_image, pred_boxes_per_image in zip(proposals, pred_boxes): + proposals_per_image.proposal_boxes = Boxes(pred_boxes_per_image) + return losses + else: + pred_instances, _ = self.box_predictor.inference(predictions, proposals) + return pred_instances + + def _forward_mask(self, features: Dict[str, torch.Tensor], instances: List[Instances]): + """ + Forward logic of the mask prediction branch. + + Args: + features (dict[str, Tensor]): mapping from feature map names to tensor. + Same as in :meth:`ROIHeads.forward`. + instances (list[Instances]): the per-image instances to train/predict masks. + In training, they can be the proposals. + In inference, they can be the boxes predicted by R-CNN box head. + + Returns: + In training, a dict of losses. + In inference, update `instances` with new fields "pred_masks" and return it. + """ + if not self.mask_on: + return {} if self.training else instances + + if self.training: + # head is only trained on positive proposals. + instances, _ = select_foreground_proposals(instances, self.num_classes) + + if self.mask_pooler is not None: + features = [features[f] for f in self.mask_in_features] + boxes = [x.proposal_boxes if self.training else x.pred_boxes for x in instances] + features = self.mask_pooler(features, boxes) + else: + features = {f: features[f] for f in self.mask_in_features} + return self.mask_head(features, instances) \ No newline at end of file diff --git a/detectron2/modeling/roi_heads/fast_rcnn.py b/detectron2/modeling/roi_heads/fast_rcnn.py new file mode 100644 index 0000000000000000000000000000000000000000..5fe09c42b06a766e20d96cfcd0ce6356ff90bdfc --- /dev/null +++ b/detectron2/modeling/roi_heads/fast_rcnn.py @@ -0,0 +1,1086 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +from typing import Dict, List, Tuple, Union +import torch +from fvcore.nn import giou_loss, smooth_l1_loss +from torch import nn +from torch.nn import functional as F + +from detectron2.config import configurable +from detectron2.layers import ShapeSpec, batched_nms, cat, cross_entropy, nonzero_tuple +from detectron2.layers.soft_nms import batched_soft_nms +from detectron2.modeling.box_regression import Box2BoxTransform +from detectron2.structures import Boxes, Instances +from detectron2.utils.events import get_event_storage + +__all__ = ["fast_rcnn_inference", "FastRCNNOutputLayers", "CLIPFastRCNNOutputLayers"] + + +logger = logging.getLogger(__name__) + +""" +Shape shorthand in this module: + + N: number of images in the minibatch + R: number of ROIs, combined over all images, in the minibatch + Ri: number of ROIs in image i + K: number of foreground classes. E.g.,there are 80 foreground classes in COCO. + +Naming convention: + + deltas: refers to the 4-d (dx, dy, dw, dh) deltas that parameterize the box2box + transform (see :class:`box_regression.Box2BoxTransform`). + + pred_class_logits: predicted class scores in [-inf, +inf]; use + softmax(pred_class_logits) to estimate P(class). + + gt_classes: ground-truth classification labels in [0, K], where [0, K) represent + foreground object classes and K represents the background class. + + pred_proposal_deltas: predicted box2box transform deltas for transforming proposals + to detection box predictions. + + gt_proposal_deltas: ground-truth box2box transform deltas +""" + + +def fast_rcnn_inference( + boxes: List[torch.Tensor], + scores: List[torch.Tensor], + image_shapes: List[Tuple[int, int]], + score_thresh: float, + nms_thresh: float, + soft_nms_enabled, + soft_nms_method, + soft_nms_sigma, + soft_nms_prune, + topk_per_image: int, + scores_bf_multiply, +): + """ + Call `fast_rcnn_inference_single_image` for all images. + + Args: + boxes (list[Tensor]): A list of Tensors of predicted class-specific or class-agnostic + boxes for each image. Element i has shape (Ri, K * 4) if doing + class-specific regression, or (Ri, 4) if doing class-agnostic + regression, where Ri is the number of predicted objects for image i. + This is compatible with the output of :meth:`FastRCNNOutputLayers.predict_boxes`. + scores (list[Tensor]): A list of Tensors of predicted class scores for each image. + Element i has shape (Ri, K + 1), where Ri is the number of predicted objects + for image i. Compatible with the output of :meth:`FastRCNNOutputLayers.predict_probs`. + image_shapes (list[tuple]): A list of (width, height) tuples for each image in the batch. + score_thresh (float): Only return detections with a confidence score exceeding this + threshold. + nms_thresh (float): The threshold to use for box non-maximum suppression. Value in [0, 1]. + soft_nms_enabled (bool): Indicate to use soft non-maximum suppression. + soft_nms_method: (str): One of ['gaussian', 'linear', 'hard'] + soft_nms_sigma: (float): Sigma for gaussian soft nms. Value in (0, inf) + soft_nms_prune: (float): Threshold for pruning during soft nms. Value in [0, 1] + topk_per_image (int): The number of top scoring detections to return. Set < 0 to return + all detections. + + Returns: + instances: (list[Instances]): A list of N instances, one for each image in the batch, + that stores the topk most confidence detections. + kept_indices: (list[Tensor]): A list of 1D tensor of length of N, each element indicates + the corresponding boxes/scores index in [0, Ri) from the input, for image i. + """ + result_per_image = [ + fast_rcnn_inference_single_image( + boxes_per_image, scores_per_image, image_shape, score_thresh, nms_thresh, + soft_nms_enabled, soft_nms_method, soft_nms_sigma, soft_nms_prune, topk_per_image, s_bf_per_img + ) + for scores_per_image, boxes_per_image, image_shape, s_bf_per_img in zip(scores, boxes, image_shapes, scores_bf_multiply) + ] + return [x[0] for x in result_per_image], [x[1] for x in result_per_image] + + +def _log_classification_stats(pred_logits, gt_classes, prefix="fast_rcnn"): + """ + Log the classification metrics to EventStorage. + + Args: + pred_logits: Rx(K+1) logits. The last column is for background class. + gt_classes: R labels + """ + num_instances = gt_classes.numel() + if num_instances == 0: + return + pred_classes = pred_logits.argmax(dim=1) + bg_class_ind = pred_logits.shape[1] - 1 + + fg_inds = (gt_classes >= 0) & (gt_classes < bg_class_ind) + num_fg = fg_inds.nonzero().numel() + fg_gt_classes = gt_classes[fg_inds] + fg_pred_classes = pred_classes[fg_inds] + + num_false_negative = (fg_pred_classes == bg_class_ind).nonzero().numel() + num_accurate = (pred_classes == gt_classes).nonzero().numel() + fg_num_accurate = (fg_pred_classes == fg_gt_classes).nonzero().numel() + + storage = get_event_storage() + storage.put_scalar(f"{prefix}/cls_accuracy", num_accurate / num_instances) + if num_fg > 0: + storage.put_scalar(f"{prefix}/fg_cls_accuracy", fg_num_accurate / num_fg) + storage.put_scalar(f"{prefix}/false_negative", num_false_negative / num_fg) + #print("cls_accuracy {:.2f}; fg_cls_accuracy {:.2f}; false_negative {:.2f}".format(num_accurate / num_instances, fg_num_accurate / num_fg, num_false_negative / num_fg)) + + +def fast_rcnn_inference_single_image( + boxes, + scores, + image_shape: Tuple[int, int], + score_thresh: float, + nms_thresh: float, + soft_nms_enabled, + soft_nms_method, + soft_nms_sigma, + soft_nms_prune, + topk_per_image: int, + scores_bf_multiply: None, +): + """ + Single-image inference. Return bounding-box detection results by thresholding + on scores and applying non-maximum suppression (NMS). + + Args: + Same as `fast_rcnn_inference`, but with boxes, scores, and image shapes + per image. + + Returns: + Same as `fast_rcnn_inference`, but for only one image. + """ + valid_mask = torch.isfinite(boxes).all(dim=1) & torch.isfinite(scores).all(dim=1) + if not valid_mask.all(): + boxes = boxes[valid_mask] + scores = scores[valid_mask] + scores_bf_multiply = scores_bf_multiply[valid_mask] + + # scores = scores[:, :-1] + # scores_bf_multiply = scores_bf_multiply[:, :-1] + num_bbox_reg_classes = boxes.shape[1] // 4 + # Convert to Boxes to use the `clip` function ... + boxes = Boxes(boxes.reshape(-1, 4)) + boxes.clip(image_shape) + boxes = boxes.tensor.view(-1, num_bbox_reg_classes, 4) # R x C x 4 + + # 1. Filter results based on detection scores. It can make NMS more efficient + # by filtering out low-confidence detections. + filter_mask = scores > score_thresh # R x K + # R' x 2. First column contains indices of the R predictions; + # Second column contains indices of classes. + filter_inds = filter_mask.nonzero() + if num_bbox_reg_classes == 1: + boxes = boxes[filter_inds[:, 0], 0] + else: + boxes = boxes[filter_mask] + scores = scores[filter_mask] + scores_bf_multiply = scores_bf_multiply[filter_mask] + + # 2. Apply NMS for each class independently. + if not soft_nms_enabled: + keep = batched_nms(boxes, scores, filter_inds[:, 1], nms_thresh) + else: + keep, soft_nms_scores = batched_soft_nms( + boxes, + scores, + filter_inds[:, 1], + soft_nms_method, + soft_nms_sigma, + nms_thresh, + soft_nms_prune, + ) + scores[keep] = soft_nms_scores + # scores_bf_multiply? (TBD) + scores_bf_multiply = scores + if topk_per_image >= 0: + keep = keep[:topk_per_image] + boxes, scores, filter_inds = boxes[keep], scores[keep], filter_inds[keep] + scores_bf_multiply = scores_bf_multiply[keep] + + result = Instances(image_shape) + result.pred_boxes = Boxes(boxes) + result.scores = scores + result.scores = scores_bf_multiply # convert to the original scores before multiplying RPN scores + result.pred_classes = filter_inds[:, 1] + return result, filter_inds[:, 0] + + +class FastRCNNOutputs: + """ + An internal implementation that stores information about outputs of a Fast R-CNN head, + and provides methods that are used to decode the outputs of a Fast R-CNN head. + """ + + def __init__( + self, + box2box_transform, + pred_class_logits, + pred_proposal_deltas, + proposals, + smooth_l1_beta=0.0, + box_reg_loss_type="smooth_l1", + ): + """ + Args: + box2box_transform (Box2BoxTransform/Box2BoxTransformRotated): + box2box transform instance for proposal-to-detection transformations. + pred_class_logits (Tensor): A tensor of shape (R, K + 1) storing the predicted class + logits for all R predicted object instances. + Each row corresponds to a predicted object instance. + pred_proposal_deltas (Tensor): A tensor of shape (R, K * B) or (R, B) for + class-specific or class-agnostic regression. It stores the predicted deltas that + transform proposals into final box detections. + B is the box dimension (4 or 5). + When B is 4, each row is [dx, dy, dw, dh (, ....)]. + When B is 5, each row is [dx, dy, dw, dh, da (, ....)]. + proposals (list[Instances]): A list of N Instances, where Instances i stores the + proposals for image i, in the field "proposal_boxes". + When training, each Instances must have ground-truth labels + stored in the field "gt_classes" and "gt_boxes". + The total number of all instances must be equal to R. + smooth_l1_beta (float): The transition point between L1 and L2 loss in + the smooth L1 loss function. When set to 0, the loss becomes L1. When + set to +inf, the loss becomes constant 0. + box_reg_loss_type (str): Box regression loss type. One of: "smooth_l1", "giou" + """ + self.box2box_transform = box2box_transform + self.num_preds_per_image = [len(p) for p in proposals] + self.pred_class_logits = pred_class_logits + self.pred_proposal_deltas = pred_proposal_deltas + self.smooth_l1_beta = smooth_l1_beta + self.box_reg_loss_type = box_reg_loss_type + + self.image_shapes = [x.image_size for x in proposals] + + if len(proposals): + box_type = type(proposals[0].proposal_boxes) + # cat(..., dim=0) concatenates over all images in the batch + self.proposals = box_type.cat([p.proposal_boxes for p in proposals]) + assert ( + not self.proposals.tensor.requires_grad + ), "Proposals should not require gradients!" + + # "gt_classes" exists if and only if training. But other gt fields may + # not necessarily exist in training for images that have no groundtruth. + if proposals[0].has("gt_classes"): + self.gt_classes = cat([p.gt_classes for p in proposals], dim=0) + + # If "gt_boxes" does not exist, the proposals must be all negative and + # should not be included in regression loss computation. + # Here we just use proposal_boxes as an arbitrary placeholder because its + # value won't be used in self.box_reg_loss(). + gt_boxes = [ + p.gt_boxes if p.has("gt_boxes") else p.proposal_boxes for p in proposals + ] + self.gt_boxes = box_type.cat(gt_boxes) + else: + self.proposals = Boxes(torch.zeros(0, 4, device=self.pred_proposal_deltas.device)) + self._no_instances = len(self.proposals) == 0 # no instances found + + def softmax_cross_entropy_loss(self): + """ + Deprecated + """ + _log_classification_stats(self.pred_class_logits, self.gt_classes) + return cross_entropy(self.pred_class_logits, self.gt_classes, reduction="mean") + + def box_reg_loss(self): + """ + Deprecated + """ + if self._no_instances: + return 0.0 * self.pred_proposal_deltas.sum() + + box_dim = self.proposals.tensor.size(1) # 4 or 5 + cls_agnostic_bbox_reg = self.pred_proposal_deltas.size(1) == box_dim + device = self.pred_proposal_deltas.device + + bg_class_ind = self.pred_class_logits.shape[1] - 1 + # Box delta loss is only computed between the prediction for the gt class k + # (if 0 <= k < bg_class_ind) and the target; there is no loss defined on predictions + # for non-gt classes and background. + # Empty fg_inds should produce a valid loss of zero because reduction=sum. + fg_inds = nonzero_tuple((self.gt_classes >= 0) & (self.gt_classes < bg_class_ind))[0] + + if cls_agnostic_bbox_reg: + # pred_proposal_deltas only corresponds to foreground class for agnostic + gt_class_cols = torch.arange(box_dim, device=device) + else: + # pred_proposal_deltas for class k are located in columns [b * k : b * k + b], + # where b is the dimension of box representation (4 or 5) + # Note that compared to Detectron1, + # we do not perform bounding box regression for background classes. + gt_class_cols = box_dim * self.gt_classes[fg_inds, None] + torch.arange( + box_dim, device=device + ) + + if self.box_reg_loss_type == "smooth_l1": + gt_proposal_deltas = self.box2box_transform.get_deltas( + self.proposals.tensor, self.gt_boxes.tensor + ) + loss_box_reg = smooth_l1_loss( + self.pred_proposal_deltas[fg_inds[:, None], gt_class_cols], + gt_proposal_deltas[fg_inds], + self.smooth_l1_beta, + reduction="sum", + ) + elif self.box_reg_loss_type == "giou": + fg_pred_boxes = self.box2box_transform.apply_deltas( + self.pred_proposal_deltas[fg_inds[:, None], gt_class_cols], + self.proposals.tensor[fg_inds], + ) + loss_box_reg = giou_loss( + fg_pred_boxes, + self.gt_boxes.tensor[fg_inds], + reduction="sum", + ) + else: + raise ValueError(f"Invalid bbox reg loss type '{self.box_reg_loss_type}'") + + loss_box_reg = loss_box_reg / self.gt_classes.numel() + return loss_box_reg + + def losses(self): + """ + Deprecated + """ + return {"loss_cls": self.softmax_cross_entropy_loss(), "loss_box_reg": self.box_reg_loss()} + + def predict_boxes(self): + """ + Deprecated + """ + pred = self.box2box_transform.apply_deltas(self.pred_proposal_deltas, self.proposals.tensor) + return pred.split(self.num_preds_per_image, dim=0) + + def predict_probs(self): + """ + Deprecated + """ + probs = F.softmax(self.pred_class_logits, dim=-1) + return probs.split(self.num_preds_per_image, dim=0) + + +class FastRCNNOutputLayers(nn.Module): + """ + Two linear layers for predicting Fast R-CNN outputs: + + 1. proposal-to-detection box regression deltas + 2. classification scores + """ + + @configurable + def __init__( + self, + input_shape: ShapeSpec, + *, + box2box_transform, + num_classes: int, + test_score_thresh: float = 0.0, + test_nms_thresh: float = 0.5, + soft_nms_enabled=False, + soft_nms_method="gaussian", + soft_nms_sigma=0.5, + soft_nms_prune=0.001, + test_topk_per_image: int = 100, + cls_agnostic_bbox_reg: bool = False, + smooth_l1_beta: float = 0.0, + box_reg_loss_type: str = "smooth_l1", + loss_weight: Union[float, Dict[str, float]] = 1.0, + clip_cls_emb: tuple = (False, None), + no_box_delta: bool = False, + bg_cls_loss_weight: None, + multiply_rpn_score: False, + openset_test: None, + ): + """ + NOTE: this interface is experimental. + + Args: + input_shape (ShapeSpec): shape of the input feature to this module + box2box_transform (Box2BoxTransform or Box2BoxTransformRotated): + num_classes (int): number of foreground classes + test_score_thresh (float): threshold to filter predictions results. + test_nms_thresh (float): NMS threshold for prediction results. + test_topk_per_image (int): number of top predictions to produce per image. + cls_agnostic_bbox_reg (bool): whether to use class agnostic for bbox regression + smooth_l1_beta (float): transition point from L1 to L2 loss. Only used if + `box_reg_loss_type` is "smooth_l1" + box_reg_loss_type (str): Box regression loss type. One of: "smooth_l1", "giou" + loss_weight (float|dict): weights to use for losses. Can be single float for weighting + all losses, or a dict of individual weightings. Valid dict keys are: + * "loss_cls": applied to classification loss + * "loss_box_reg": applied to box regression loss + """ + super().__init__() + if isinstance(input_shape, int): # some backward compatibility + input_shape = ShapeSpec(channels=input_shape) + self.num_classes = num_classes + input_size = input_shape.channels * (input_shape.width or 1) * (input_shape.height or 1) + if clip_cls_emb[0]: # if combine {C4, text emb as classifier}, then has to use att_pool to match dimension + input_size = clip_cls_emb[3] if clip_cls_emb[2] in ['CLIPRes5ROIHeads', 'CLIPStandardROIHeads'] else input_size + # prediction layer for num_classes foreground classes and one background class (hence + 1) + self.cls_score = nn.Linear(input_size, num_classes + 1) + num_bbox_reg_classes = 1 if cls_agnostic_bbox_reg else num_classes + box_dim = len(box2box_transform.weights) + self.bbox_pred = nn.Linear(input_size, num_bbox_reg_classes * box_dim) + + nn.init.normal_(self.cls_score.weight, std=0.01) + nn.init.normal_(self.bbox_pred.weight, std=0.001) + for l in [self.cls_score, self.bbox_pred]: + nn.init.constant_(l.bias, 0) + + self.box2box_transform = box2box_transform + self.smooth_l1_beta = smooth_l1_beta + self.test_score_thresh = test_score_thresh + self.test_nms_thresh = test_nms_thresh + self.soft_nms_enabled = soft_nms_enabled + self.soft_nms_method = soft_nms_method + self.soft_nms_sigma = soft_nms_sigma + self.soft_nms_prune = soft_nms_prune + self.test_topk_per_image = test_topk_per_image + self.box_reg_loss_type = box_reg_loss_type + if isinstance(loss_weight, float): + loss_weight = {"loss_cls": loss_weight, "loss_box_reg": loss_weight} + self.loss_weight = loss_weight + + # use clip text embeddings as classifier's weights + self.use_clip_cls_emb = clip_cls_emb[0] + if self.use_clip_cls_emb: + ######### V2L projection layer in CVPR OVR model ######### + if openset_test[3]: # run CVPR model + self.emb_pred = nn.Linear(input_size, 768) + self.emb_pred.weight.requires_grad = False + self.emb_pred.bias.requires_grad = False + input_size = 768 + else: + self.emb_pred = None + ######### V2L projection layer in CVPR OVR model ######### + text_emb_require_grad = False + self.use_bias = False + self.tempurature = openset_test[2] # 0.01 # the smaller, the bigger difference among probs after softmax + self.no_box_delta = no_box_delta + if bg_cls_loss_weight is not None: # loss weigh for bg regions + self.cls_loss_weight = torch.ones(num_classes + 1) + self.cls_loss_weight[-1] = bg_cls_loss_weight + else: + self.cls_loss_weight = None + self.multiply_rpn_score = multiply_rpn_score + self.focal_scaled_loss = openset_test[4] + + @classmethod + def from_config(cls, cfg, input_shape): + # if cfg.MODEL.CLIP.CROP_REGION_TYPE == "RPN": + # assert cfg.MODEL.CLIP.NO_BOX_DELTA is False + return { + "input_shape": input_shape, + "box2box_transform": Box2BoxTransform(weights=cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS), + # fmt: off + "num_classes" : cfg.MODEL.ROI_HEADS.NUM_CLASSES, + "cls_agnostic_bbox_reg" : cfg.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG, + "smooth_l1_beta" : cfg.MODEL.ROI_BOX_HEAD.SMOOTH_L1_BETA, + "test_score_thresh" : cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST, + "test_nms_thresh" : cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST, + "soft_nms_enabled" : cfg.MODEL.ROI_HEADS.SOFT_NMS_ENABLED, + "soft_nms_method" : cfg.MODEL.ROI_HEADS.SOFT_NMS_METHOD, + "soft_nms_sigma" : cfg.MODEL.ROI_HEADS.SOFT_NMS_SIGMA, + "soft_nms_prune" : cfg.MODEL.ROI_HEADS.SOFT_NMS_PRUNE, + "test_topk_per_image" : cfg.TEST.DETECTIONS_PER_IMAGE, + "box_reg_loss_type" : cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_TYPE, + "loss_weight" : {"loss_box_reg": cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_WEIGHT}, + "clip_cls_emb" : (cfg.MODEL.CLIP.USE_TEXT_EMB_CLASSIFIER, cfg.MODEL.CLIP.TEXT_EMB_PATH, cfg.MODEL.ROI_HEADS.NAME, cfg.MODEL.CLIP.TEXT_EMB_DIM), + "no_box_delta" : cfg.MODEL.CLIP.NO_BOX_DELTA or cfg.MODEL.CLIP.CROP_REGION_TYPE == 'GT', + "bg_cls_loss_weight" : cfg.MODEL.CLIP.BG_CLS_LOSS_WEIGHT, + "multiply_rpn_score" : cfg.MODEL.CLIP.MULTIPLY_RPN_SCORE, + "openset_test" : (cfg.MODEL.CLIP.OPENSET_TEST_NUM_CLASSES, cfg.MODEL.CLIP.OPENSET_TEST_TEXT_EMB_PATH, \ + cfg.MODEL.CLIP.CLSS_TEMP, cfg.MODEL.CLIP.RUN_CVPR_OVR, cfg.MODEL.CLIP.FOCAL_SCALED_LOSS) + # fmt: on + } + + def forward(self, x, queries): + """ + Args: + x: per-region features of shape (N, ...) for N bounding boxes to predict. + + Returns: + (Tensor, Tensor): + First tensor: shape (N,K+1), scores for each of the N box. Each row contains the + scores for K object categories and 1 background class. + + Second tensor: bounding box regression deltas for each box. Shape is shape (N,Kx4), + or (N,4) for class-agnostic regression. + """ + if x.dim() > 2: + x = torch.flatten(x, start_dim=1) + if self.use_clip_cls_emb: # use clip text embeddings as classifier's weights + normalized_x = F.normalize(x, p=2.0, dim=1) + cls_scores = normalized_x @ queries.t() + bg_cls_scores = cls_scores.new(cls_scores.shape[0], 1).fill_(0.3) + scores = cls_scores # torch.cat((cls_scores, bg_cls_scores), 1) + else: # default setting + scores = self.cls_score(x) + proposal_deltas = scores.new(scores.shape[0], 4).fill_(0) # self.bbox_pred(x) + return scores, proposal_deltas + + def losses(self, predictions, proposals): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were used + to compute predictions. The fields ``proposal_boxes``, ``gt_boxes``, + ``gt_classes`` are expected. + + Returns: + Dict[str, Tensor]: dict of losses + """ + scores, proposal_deltas = predictions + + # parse classification outputs + gt_classes = ( + cat([p.gt_classes for p in proposals], dim=0) if len(proposals) else torch.empty(0) + ) + _log_classification_stats(scores, gt_classes) + + # parse box regression outputs + if len(proposals): + proposal_boxes = cat([p.proposal_boxes.tensor for p in proposals], dim=0) # Nx4 + assert not proposal_boxes.requires_grad, "Proposals should not require gradients!" + # If "gt_boxes" does not exist, the proposals must be all negative and + # should not be included in regression loss computation. + # Here we just use proposal_boxes as an arbitrary placeholder because its + # value won't be used in self.box_reg_loss(). + gt_boxes = cat( + [(p.gt_boxes if p.has("gt_boxes") else p.proposal_boxes).tensor for p in proposals], + dim=0, + ) + else: + proposal_boxes = gt_boxes = torch.empty((0, 4), device=proposal_deltas.device) + + # loss weights + if self.cls_loss_weight is not None and self.cls_loss_weight.device != scores.device: + self.cls_loss_weight = self.cls_loss_weight.to(scores.device) + if self.focal_scaled_loss is not None: + loss_cls = self.focal_loss(scores, gt_classes, gamma=self.focal_scaled_loss) + else: + loss_cls = cross_entropy(scores, gt_classes, reduction="mean") if self.cls_loss_weight is None else \ + cross_entropy(scores, gt_classes, reduction="mean", weight=self.cls_loss_weight) + losses = { + "loss_cls": loss_cls, + "loss_box_reg": self.box_reg_loss( + proposal_boxes, gt_boxes, proposal_deltas, gt_classes + ), + } + return {k: v * self.loss_weight.get(k, 1.0) for k, v in losses.items()} + + def focal_loss(self, inputs, targets, alpha=0.25, gamma=0.5, reduction="mean", mode='softmax'): + """Inspired by RetinaNet implementation""" + if mode == 'sigmoid': # original focal loss implementation, except we include bg loss + targets = F.one_hot(targets, num_classes=self.num_classes + 1).to(inputs.dtype) # create binary label for each logit entry, including bg loss + p = torch.sigmoid(inputs) + ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none") + p_t = p * targets + (1 - p) * (1 - targets) + loss = ce_loss * ((1 - p_t) ** gamma) + + if alpha >= 0: + alpha_t = alpha * targets + (1 - alpha) * (1 - targets) + loss = alpha_t * loss + elif mode == 'softmax': + only_fg = False # if True, only fg rois are attached the focal loss scaling + #gamma = 0.3 # 0.5 # 0.8 # 1.5 # 1.0 + alpha = -1 # no binary target in this case; instead, we can use bg loss weight + if targets.numel() == 0 and reduction == "mean": + return input.sum() * 0.0 # connect the gradient + ce_loss = F.cross_entropy(inputs, targets, reduction="none") + p = F.softmax(inputs, dim=-1) + p_t = p[torch.arange(p.size(0)).to(p.device), targets] # get prob of target class + if only_fg: # apply scaling to only fg rois + roi_wise_gamma = torch.zeros(p.size(0)).to(p.device) + roi_wise_gamma[targets != self.num_classes] = gamma + gamma = roi_wise_gamma + loss = ce_loss * ((1 - p_t) ** gamma) + + # if alpha >= 0: + # alpha_t = alpha * targets + (1 - alpha) * (1 - targets) + # loss = alpha_t * loss + # bg loss weight + if self.cls_loss_weight is not None: + loss_weight = torch.ones(loss.size(0)).to(p.device) + loss_weight[targets == self.num_classes] = self.cls_loss_weight[-1].item() + loss = loss * loss_weight + + if reduction == "mean": + loss = loss.mean() + elif reduction == "sum": + loss = loss.sum() + + return loss + + def box_reg_loss(self, proposal_boxes, gt_boxes, pred_deltas, gt_classes): + """ + Args: + All boxes are tensors with the same shape Rx(4 or 5). + gt_classes is a long tensor of shape R, the gt class label of each proposal. + R shall be the number of proposals. + """ + box_dim = proposal_boxes.shape[1] # 4 or 5 + # Regression loss is only computed for foreground proposals (those matched to a GT) + fg_inds = nonzero_tuple((gt_classes >= 0) & (gt_classes < self.num_classes))[0] + if pred_deltas.shape[1] == box_dim: # cls-agnostic regression + fg_pred_deltas = pred_deltas[fg_inds] + else: + fg_pred_deltas = pred_deltas.view(-1, self.num_classes, box_dim)[ + fg_inds, gt_classes[fg_inds] + ] + + if self.box_reg_loss_type == "smooth_l1": + gt_pred_deltas = self.box2box_transform.get_deltas( + proposal_boxes[fg_inds], + gt_boxes[fg_inds], + ) + loss_box_reg = smooth_l1_loss( + fg_pred_deltas, gt_pred_deltas, self.smooth_l1_beta, reduction="sum" + ) + elif self.box_reg_loss_type == "giou": + fg_pred_boxes = self.box2box_transform.apply_deltas( + fg_pred_deltas, proposal_boxes[fg_inds] + ) + loss_box_reg = giou_loss(fg_pred_boxes, gt_boxes[fg_inds], reduction="sum") + else: + raise ValueError(f"Invalid bbox reg loss type '{self.box_reg_loss_type}'") + # The reg loss is normalized using the total number of regions (R), not the number + # of foreground regions even though the box regression loss is only defined on + # foreground regions. Why? Because doing so gives equal training influence to + # each foreground example. To see how, consider two different minibatches: + # (1) Contains a single foreground region + # (2) Contains 100 foreground regions + # If we normalize by the number of foreground regions, the single example in + # minibatch (1) will be given 100 times as much influence as each foreground + # example in minibatch (2). Normalizing by the total number of regions, R, + # means that the single example in minibatch (1) and each of the 100 examples + # in minibatch (2) are given equal influence. + return loss_box_reg / max(gt_classes.numel(), 1.0) # return 0 if empty + + def inference(self, predictions: Tuple[torch.Tensor, torch.Tensor], proposals: List[Instances]): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were + used to compute predictions. The ``proposal_boxes`` field is expected. + + Returns: + list[Instances]: same as `fast_rcnn_inference`. + list[Tensor]: same as `fast_rcnn_inference`. + """ + boxes = self.predict_boxes(predictions, proposals) + scores = self.predict_probs(predictions, proposals) + image_shapes = [x.image_size for x in proposals] + scores_bf_multiply = scores # as a backup + if self.multiply_rpn_score: + rpn_scores = [p.get('objectness_logits') for p in proposals] + # filter based on rpn_scores + # boxes = (boxes[0][rpn_scores[0] > 0.9],) + # scores = (scores[0][rpn_scores[0] > 0.9],) + # rpn_scores = [rpn_scores[0][rpn_scores[0] > 0.9]] + # scores_bf_multiply = scores # as a backup + #rpn_scores = [p.get('objectness_logits').sigmoid() for p in proposals] + scores = [(torch.sigmoid(s) * torch.sigmoid(rpn_s[:, None])) ** 0.5 for s, rpn_s in zip(scores, rpn_scores)] + return fast_rcnn_inference( + boxes, + scores, + image_shapes, + self.test_score_thresh, + self.test_nms_thresh, + self.soft_nms_enabled, + self.soft_nms_method, + self.soft_nms_sigma, + self.soft_nms_prune, + self.test_topk_per_image, + scores_bf_multiply = scores_bf_multiply if self.multiply_rpn_score else None, + ) + + def predict_boxes_for_gt_classes(self, predictions, proposals): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were used + to compute predictions. The fields ``proposal_boxes``, ``gt_classes`` are expected. + + Returns: + list[Tensor]: + A list of Tensors of predicted boxes for GT classes in case of + class-specific box head. Element i of the list has shape (Ri, B), where Ri is + the number of proposals for image i and B is the box dimension (4 or 5) + """ + if not len(proposals): + return [] + scores, proposal_deltas = predictions + proposal_boxes = cat([p.proposal_boxes.tensor for p in proposals], dim=0) + N, B = proposal_boxes.shape + predict_boxes = self.box2box_transform.apply_deltas( + proposal_deltas, proposal_boxes + ) # Nx(KxB) + + K = predict_boxes.shape[1] // B + if K > 1: + gt_classes = torch.cat([p.gt_classes for p in proposals], dim=0) + # Some proposals are ignored or have a background class. Their gt_classes + # cannot be used as index. + gt_classes = gt_classes.clamp_(0, K - 1) + + predict_boxes = predict_boxes.view(N, K, B)[ + torch.arange(N, dtype=torch.long, device=predict_boxes.device), gt_classes + ] + num_prop_per_image = [len(p) for p in proposals] + return predict_boxes.split(num_prop_per_image) + + def predict_boxes( + self, predictions: Tuple[torch.Tensor, torch.Tensor], proposals: List[Instances] + ): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were + used to compute predictions. The ``proposal_boxes`` field is expected. + + Returns: + list[Tensor]: + A list of Tensors of predicted class-specific or class-agnostic boxes + for each image. Element i has shape (Ri, K * B) or (Ri, B), where Ri is + the number of proposals for image i and B is the box dimension (4 or 5) + """ + if not len(proposals): + return [] + _, proposal_deltas = predictions + num_prop_per_image = [len(p) for p in proposals] + proposal_boxes = cat([p.proposal_boxes.tensor for p in proposals], dim=0) + if self.no_box_delta: + predict_boxes = proposal_boxes + else: + predict_boxes = self.box2box_transform.apply_deltas( + proposal_deltas, + proposal_boxes, + ) # Nx(KxB) + return predict_boxes.split(num_prop_per_image) + + def predict_probs( + self, predictions: Tuple[torch.Tensor, torch.Tensor], proposals: List[Instances] + ): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were + used to compute predictions. + + Returns: + list[Tensor]: + A list of Tensors of predicted class probabilities for each image. + Element i has shape (Ri, K + 1), where Ri is the number of proposals for image i. + """ + scores, _ = predictions + num_inst_per_image = [len(p) for p in proposals] + # probs = F.softmax(scores, dim=-1) + probs = scores + return probs.split(num_inst_per_image, dim=0) + + +class OLDFastRCNNOutputLayers(nn.Module): + """ + Two linear layers for predicting Fast R-CNN outputs: + + 1. proposal-to-detection box regression deltas + 2. classification scores + """ + + @configurable + def __init__( + self, + input_shape: ShapeSpec, + *, + box2box_transform, + num_classes: int, + test_score_thresh: float = 0.0, + test_nms_thresh: float = 0.5, + test_topk_per_image: int = 100, + cls_agnostic_bbox_reg: bool = False, + smooth_l1_beta: float = 0.0, + box_reg_loss_type: str = "smooth_l1", + loss_weight: Union[float, Dict[str, float]] = 1.0, + no_box_delta: bool = False, + ): + """ + NOTE: this interface is experimental. + + Args: + input_shape (ShapeSpec): shape of the input feature to this module + box2box_transform (Box2BoxTransform or Box2BoxTransformRotated): + num_classes (int): number of foreground classes + test_score_thresh (float): threshold to filter predictions results. + test_nms_thresh (float): NMS threshold for prediction results. + test_topk_per_image (int): number of top predictions to produce per image. + cls_agnostic_bbox_reg (bool): whether to use class agnostic for bbox regression + smooth_l1_beta (float): transition point from L1 to L2 loss. Only used if + `box_reg_loss_type` is "smooth_l1" + box_reg_loss_type (str): Box regression loss type. One of: "smooth_l1", "giou" + loss_weight (float|dict): weights to use for losses. Can be single float for weighting + all losses, or a dict of individual weightings. Valid dict keys are: + * "loss_cls": applied to classification loss + * "loss_box_reg": applied to box regression loss + """ + super().__init__() + if isinstance(input_shape, int): # some backward compatibility + input_shape = ShapeSpec(channels=input_shape) + self.num_classes = num_classes + input_size = input_shape.channels * (input_shape.width or 1) * (input_shape.height or 1) + # prediction layer for num_classes foreground classes and one background class (hence + 1) + self.cls_score = nn.Linear(input_size, num_classes + 1) + num_bbox_reg_classes = 1 if cls_agnostic_bbox_reg else num_classes + box_dim = len(box2box_transform.weights) + self.bbox_pred = nn.Linear(input_size, num_bbox_reg_classes * box_dim) + + nn.init.normal_(self.cls_score.weight, std=0.01) + nn.init.normal_(self.bbox_pred.weight, std=0.001) + for l in [self.cls_score, self.bbox_pred]: + nn.init.constant_(l.bias, 0) + + self.box2box_transform = box2box_transform + self.smooth_l1_beta = smooth_l1_beta + self.test_score_thresh = test_score_thresh + self.test_nms_thresh = test_nms_thresh + self.test_topk_per_image = test_topk_per_image + self.box_reg_loss_type = box_reg_loss_type + if isinstance(loss_weight, float): + loss_weight = {"loss_cls": loss_weight, "loss_box_reg": loss_weight} + self.loss_weight = loss_weight + self.no_box_delta = no_box_delta + + @classmethod + def from_config(cls, cfg, input_shape): + return { + "input_shape": input_shape, + "box2box_transform": Box2BoxTransform(weights=cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS), + # fmt: off + "num_classes" : cfg.MODEL.ROI_HEADS.NUM_CLASSES, + "cls_agnostic_bbox_reg" : cfg.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG, + "smooth_l1_beta" : cfg.MODEL.ROI_BOX_HEAD.SMOOTH_L1_BETA, + "test_score_thresh" : cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST, + "test_nms_thresh" : cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST, + "test_topk_per_image" : cfg.TEST.DETECTIONS_PER_IMAGE, + "box_reg_loss_type" : cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_TYPE, + "loss_weight" : {"loss_box_reg": cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_WEIGHT}, + "no_box_delta" : cfg.MODEL.CLIP.NO_BOX_DELTA or cfg.MODEL.CLIP.CROP_REGION_TYPE == 'GT', + # fmt: on + } + + def forward(self, x): + """ + Args: + x: per-region features of shape (N, ...) for N bounding boxes to predict. + + Returns: + (Tensor, Tensor): + First tensor: shape (N,K+1), scores for each of the N box. Each row contains the + scores for K object categories and 1 background class. + + Second tensor: bounding box regression deltas for each box. Shape is shape (N,Kx4), + or (N,4) for class-agnostic regression. + """ + if x.dim() > 2: + x = torch.flatten(x, start_dim=1) + scores = self.cls_score(x) + proposal_deltas = self.bbox_pred(x) + return scores, proposal_deltas + + def losses(self, predictions, proposals): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were used + to compute predictions. The fields ``proposal_boxes``, ``gt_boxes``, + ``gt_classes`` are expected. + + Returns: + Dict[str, Tensor]: dict of losses + """ + scores, proposal_deltas = predictions + + # parse classification outputs + gt_classes = ( + cat([p.gt_classes for p in proposals], dim=0) if len(proposals) else torch.empty(0) + ) + _log_classification_stats(scores, gt_classes) + + # parse box regression outputs + if len(proposals): + proposal_boxes = cat([p.proposal_boxes.tensor for p in proposals], dim=0) # Nx4 + assert not proposal_boxes.requires_grad, "Proposals should not require gradients!" + # If "gt_boxes" does not exist, the proposals must be all negative and + # should not be included in regression loss computation. + # Here we just use proposal_boxes as an arbitrary placeholder because its + # value won't be used in self.box_reg_loss(). + gt_boxes = cat( + [(p.gt_boxes if p.has("gt_boxes") else p.proposal_boxes).tensor for p in proposals], + dim=0, + ) + else: + proposal_boxes = gt_boxes = torch.empty((0, 4), device=proposal_deltas.device) + + losses = { + "loss_cls": cross_entropy(scores, gt_classes, reduction="mean"), + "loss_box_reg": self.box_reg_loss( + proposal_boxes, gt_boxes, proposal_deltas, gt_classes + ), + } + return {k: v * self.loss_weight.get(k, 1.0) for k, v in losses.items()} + + def box_reg_loss(self, proposal_boxes, gt_boxes, pred_deltas, gt_classes): + """ + Args: + All boxes are tensors with the same shape Rx(4 or 5). + gt_classes is a long tensor of shape R, the gt class label of each proposal. + R shall be the number of proposals. + """ + box_dim = proposal_boxes.shape[1] # 4 or 5 + # Regression loss is only computed for foreground proposals (those matched to a GT) + fg_inds = nonzero_tuple((gt_classes >= 0) & (gt_classes < self.num_classes))[0] + if pred_deltas.shape[1] == box_dim: # cls-agnostic regression + fg_pred_deltas = pred_deltas[fg_inds] + else: + fg_pred_deltas = pred_deltas.view(-1, self.num_classes, box_dim)[ + fg_inds, gt_classes[fg_inds] + ] + + if self.box_reg_loss_type == "smooth_l1": + gt_pred_deltas = self.box2box_transform.get_deltas( + proposal_boxes[fg_inds], + gt_boxes[fg_inds], + ) + loss_box_reg = smooth_l1_loss( + fg_pred_deltas, gt_pred_deltas, self.smooth_l1_beta, reduction="sum" + ) + elif self.box_reg_loss_type == "giou": + fg_pred_boxes = self.box2box_transform.apply_deltas( + fg_pred_deltas, proposal_boxes[fg_inds] + ) + loss_box_reg = giou_loss(fg_pred_boxes, gt_boxes[fg_inds], reduction="sum") + else: + raise ValueError(f"Invalid bbox reg loss type '{self.box_reg_loss_type}'") + # The reg loss is normalized using the total number of regions (R), not the number + # of foreground regions even though the box regression loss is only defined on + # foreground regions. Why? Because doing so gives equal training influence to + # each foreground example. To see how, consider two different minibatches: + # (1) Contains a single foreground region + # (2) Contains 100 foreground regions + # If we normalize by the number of foreground regions, the single example in + # minibatch (1) will be given 100 times as much influence as each foreground + # example in minibatch (2). Normalizing by the total number of regions, R, + # means that the single example in minibatch (1) and each of the 100 examples + # in minibatch (2) are given equal influence. + return loss_box_reg / max(gt_classes.numel(), 1.0) # return 0 if empty + + def inference(self, predictions: Tuple[torch.Tensor, torch.Tensor], proposals: List[Instances]): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were + used to compute predictions. The ``proposal_boxes`` field is expected. + + Returns: + list[Instances]: same as `fast_rcnn_inference`. + list[Tensor]: same as `fast_rcnn_inference`. + """ + boxes = self.predict_boxes(predictions, proposals) + scores = self.predict_probs(predictions, proposals) + image_shapes = [x.image_size for x in proposals] + return fast_rcnn_inference( + boxes, + scores, + image_shapes, + self.test_score_thresh, + self.test_nms_thresh, + self.test_topk_per_image, + ) + + def predict_boxes_for_gt_classes(self, predictions, proposals): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were used + to compute predictions. The fields ``proposal_boxes``, ``gt_classes`` are expected. + + Returns: + list[Tensor]: + A list of Tensors of predicted boxes for GT classes in case of + class-specific box head. Element i of the list has shape (Ri, B), where Ri is + the number of proposals for image i and B is the box dimension (4 or 5) + """ + if not len(proposals): + return [] + scores, proposal_deltas = predictions + proposal_boxes = cat([p.proposal_boxes.tensor for p in proposals], dim=0) + N, B = proposal_boxes.shape + predict_boxes = self.box2box_transform.apply_deltas( + proposal_deltas, proposal_boxes + ) # Nx(KxB) + + K = predict_boxes.shape[1] // B + if K > 1: + gt_classes = torch.cat([p.gt_classes for p in proposals], dim=0) + # Some proposals are ignored or have a background class. Their gt_classes + # cannot be used as index. + gt_classes = gt_classes.clamp_(0, K - 1) + + predict_boxes = predict_boxes.view(N, K, B)[ + torch.arange(N, dtype=torch.long, device=predict_boxes.device), gt_classes + ] + num_prop_per_image = [len(p) for p in proposals] + return predict_boxes.split(num_prop_per_image) + + def predict_boxes( + self, predictions: Tuple[torch.Tensor, torch.Tensor], proposals: List[Instances] + ): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were + used to compute predictions. The ``proposal_boxes`` field is expected. + + Returns: + list[Tensor]: + A list of Tensors of predicted class-specific or class-agnostic boxes + for each image. Element i has shape (Ri, K * B) or (Ri, B), where Ri is + the number of proposals for image i and B is the box dimension (4 or 5) + """ + if not len(proposals): + return [] + _, proposal_deltas = predictions + num_prop_per_image = [len(p) for p in proposals] + proposal_boxes = cat([p.proposal_boxes.tensor for p in proposals], dim=0) + if self.no_box_delta: + predict_boxes = proposal_boxes + else: + predict_boxes = self.box2box_transform.apply_deltas( + proposal_deltas, + proposal_boxes, + ) # Nx(KxB) + return predict_boxes.split(num_prop_per_image) + + def predict_probs( + self, predictions: Tuple[torch.Tensor, torch.Tensor], proposals: List[Instances] + ): + """ + Args: + predictions: return values of :meth:`forward()`. + proposals (list[Instances]): proposals that match the features that were + used to compute predictions. + + Returns: + list[Tensor]: + A list of Tensors of predicted class probabilities for each image. + Element i has shape (Ri, K + 1), where Ri is the number of proposals for image i. + """ + scores, _ = predictions + num_inst_per_image = [len(p) for p in proposals] + probs = F.softmax(scores, dim=-1) + return probs.split(num_inst_per_image, dim=0) \ No newline at end of file diff --git a/detectron2/modeling/roi_heads/keypoint_head.py b/detectron2/modeling/roi_heads/keypoint_head.py new file mode 100644 index 0000000000000000000000000000000000000000..e0acc138e72fcb188e4ffb3d156358b8ca59babf --- /dev/null +++ b/detectron2/modeling/roi_heads/keypoint_head.py @@ -0,0 +1,272 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from typing import List +import torch +from torch import nn +from torch.nn import functional as F + +from detectron2.config import configurable +from detectron2.layers import Conv2d, ConvTranspose2d, cat, interpolate +from detectron2.structures import Instances, heatmaps_to_keypoints +from detectron2.utils.events import get_event_storage +from detectron2.utils.registry import Registry + +_TOTAL_SKIPPED = 0 + + +__all__ = [ + "ROI_KEYPOINT_HEAD_REGISTRY", + "build_keypoint_head", + "BaseKeypointRCNNHead", + "KRCNNConvDeconvUpsampleHead", +] + + +ROI_KEYPOINT_HEAD_REGISTRY = Registry("ROI_KEYPOINT_HEAD") +ROI_KEYPOINT_HEAD_REGISTRY.__doc__ = """ +Registry for keypoint heads, which make keypoint predictions from per-region features. + +The registered object will be called with `obj(cfg, input_shape)`. +""" + + +def build_keypoint_head(cfg, input_shape): + """ + Build a keypoint head from `cfg.MODEL.ROI_KEYPOINT_HEAD.NAME`. + """ + name = cfg.MODEL.ROI_KEYPOINT_HEAD.NAME + return ROI_KEYPOINT_HEAD_REGISTRY.get(name)(cfg, input_shape) + + +def keypoint_rcnn_loss(pred_keypoint_logits, instances, normalizer): + """ + Arguments: + pred_keypoint_logits (Tensor): A tensor of shape (N, K, S, S) where N is the total number + of instances in the batch, K is the number of keypoints, and S is the side length + of the keypoint heatmap. The values are spatial logits. + instances (list[Instances]): A list of M Instances, where M is the batch size. + These instances are predictions from the model + that are in 1:1 correspondence with pred_keypoint_logits. + Each Instances should contain a `gt_keypoints` field containing a `structures.Keypoint` + instance. + normalizer (float): Normalize the loss by this amount. + If not specified, we normalize by the number of visible keypoints in the minibatch. + + Returns a scalar tensor containing the loss. + """ + heatmaps = [] + valid = [] + + keypoint_side_len = pred_keypoint_logits.shape[2] + for instances_per_image in instances: + if len(instances_per_image) == 0: + continue + keypoints = instances_per_image.gt_keypoints + heatmaps_per_image, valid_per_image = keypoints.to_heatmap( + instances_per_image.proposal_boxes.tensor, keypoint_side_len + ) + heatmaps.append(heatmaps_per_image.view(-1)) + valid.append(valid_per_image.view(-1)) + + if len(heatmaps): + keypoint_targets = cat(heatmaps, dim=0) + valid = cat(valid, dim=0).to(dtype=torch.uint8) + valid = torch.nonzero(valid).squeeze(1) + + # torch.mean (in binary_cross_entropy_with_logits) doesn't + # accept empty tensors, so handle it separately + if len(heatmaps) == 0 or valid.numel() == 0: + global _TOTAL_SKIPPED + _TOTAL_SKIPPED += 1 + storage = get_event_storage() + storage.put_scalar("kpts_num_skipped_batches", _TOTAL_SKIPPED, smoothing_hint=False) + return pred_keypoint_logits.sum() * 0 + + N, K, H, W = pred_keypoint_logits.shape + pred_keypoint_logits = pred_keypoint_logits.view(N * K, H * W) + + keypoint_loss = F.cross_entropy( + pred_keypoint_logits[valid], keypoint_targets[valid], reduction="sum" + ) + + # If a normalizer isn't specified, normalize by the number of visible keypoints in the minibatch + if normalizer is None: + normalizer = valid.numel() + keypoint_loss /= normalizer + + return keypoint_loss + + +def keypoint_rcnn_inference(pred_keypoint_logits: torch.Tensor, pred_instances: List[Instances]): + """ + Post process each predicted keypoint heatmap in `pred_keypoint_logits` into (x, y, score) + and add it to the `pred_instances` as a `pred_keypoints` field. + + Args: + pred_keypoint_logits (Tensor): A tensor of shape (R, K, S, S) where R is the total number + of instances in the batch, K is the number of keypoints, and S is the side length of + the keypoint heatmap. The values are spatial logits. + pred_instances (list[Instances]): A list of N Instances, where N is the number of images. + + Returns: + None. Each element in pred_instances will contain extra "pred_keypoints" and + "pred_keypoint_heatmaps" fields. "pred_keypoints" is a tensor of shape + (#instance, K, 3) where the last dimension corresponds to (x, y, score). + The scores are larger than 0. "pred_keypoint_heatmaps" contains the raw + keypoint logits as passed to this function. + """ + # flatten all bboxes from all images together (list[Boxes] -> Rx4 tensor) + bboxes_flat = cat([b.pred_boxes.tensor for b in pred_instances], dim=0) + + pred_keypoint_logits = pred_keypoint_logits.detach() + keypoint_results = heatmaps_to_keypoints(pred_keypoint_logits, bboxes_flat.detach()) + num_instances_per_image = [len(i) for i in pred_instances] + keypoint_results = keypoint_results[:, :, [0, 1, 3]].split(num_instances_per_image, dim=0) + heatmap_results = pred_keypoint_logits.split(num_instances_per_image, dim=0) + + for keypoint_results_per_image, heatmap_results_per_image, instances_per_image in zip( + keypoint_results, heatmap_results, pred_instances + ): + # keypoint_results_per_image is (num instances)x(num keypoints)x(x, y, score) + # heatmap_results_per_image is (num instances)x(num keypoints)x(side)x(side) + instances_per_image.pred_keypoints = keypoint_results_per_image + instances_per_image.pred_keypoint_heatmaps = heatmap_results_per_image + + +class BaseKeypointRCNNHead(nn.Module): + """ + Implement the basic Keypoint R-CNN losses and inference logic described in + Sec. 5 of :paper:`Mask R-CNN`. + """ + + @configurable + def __init__(self, *, num_keypoints, loss_weight=1.0, loss_normalizer=1.0): + """ + NOTE: this interface is experimental. + + Args: + num_keypoints (int): number of keypoints to predict + loss_weight (float): weight to multiple on the keypoint loss + loss_normalizer (float or str): + If float, divide the loss by `loss_normalizer * #images`. + If 'visible', the loss is normalized by the total number of + visible keypoints across images. + """ + super().__init__() + self.num_keypoints = num_keypoints + self.loss_weight = loss_weight + assert loss_normalizer == "visible" or isinstance(loss_normalizer, float), loss_normalizer + self.loss_normalizer = loss_normalizer + + @classmethod + def from_config(cls, cfg, input_shape): + ret = { + "loss_weight": cfg.MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT, + "num_keypoints": cfg.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS, + } + normalize_by_visible = ( + cfg.MODEL.ROI_KEYPOINT_HEAD.NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS + ) # noqa + if not normalize_by_visible: + batch_size_per_image = cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE + positive_sample_fraction = cfg.MODEL.ROI_HEADS.POSITIVE_FRACTION + ret["loss_normalizer"] = ( + ret["num_keypoints"] * batch_size_per_image * positive_sample_fraction + ) + else: + ret["loss_normalizer"] = "visible" + return ret + + def forward(self, x, instances: List[Instances]): + """ + Args: + x: input 4D region feature(s) provided by :class:`ROIHeads`. + instances (list[Instances]): contains the boxes & labels corresponding + to the input features. + Exact format is up to its caller to decide. + Typically, this is the foreground instances in training, with + "proposal_boxes" field and other gt annotations. + In inference, it contains boxes that are already predicted. + + Returns: + A dict of losses if in training. The predicted "instances" if in inference. + """ + x = self.layers(x) + if self.training: + num_images = len(instances) + normalizer = ( + None if self.loss_normalizer == "visible" else num_images * self.loss_normalizer + ) + return { + "loss_keypoint": keypoint_rcnn_loss(x, instances, normalizer=normalizer) + * self.loss_weight + } + else: + keypoint_rcnn_inference(x, instances) + return instances + + def layers(self, x): + """ + Neural network layers that makes predictions from regional input features. + """ + raise NotImplementedError + + +# To get torchscript support, we make the head a subclass of `nn.Sequential`. +# Therefore, to add new layers in this head class, please make sure they are +# added in the order they will be used in forward(). +@ROI_KEYPOINT_HEAD_REGISTRY.register() +class KRCNNConvDeconvUpsampleHead(BaseKeypointRCNNHead, nn.Sequential): + """ + A standard keypoint head containing a series of 3x3 convs, followed by + a transpose convolution and bilinear interpolation for upsampling. + It is described in Sec. 5 of :paper:`Mask R-CNN`. + """ + + @configurable + def __init__(self, input_shape, *, num_keypoints, conv_dims, **kwargs): + """ + NOTE: this interface is experimental. + + Args: + input_shape (ShapeSpec): shape of the input feature + conv_dims: an iterable of output channel counts for each conv in the head + e.g. (512, 512, 512) for three convs outputting 512 channels. + """ + super().__init__(num_keypoints=num_keypoints, **kwargs) + + # default up_scale to 2.0 (this can be made an option) + up_scale = 2.0 + in_channels = input_shape.channels + + for idx, layer_channels in enumerate(conv_dims, 1): + module = Conv2d(in_channels, layer_channels, 3, stride=1, padding=1) + self.add_module("conv_fcn{}".format(idx), module) + self.add_module("conv_fcn_relu{}".format(idx), nn.ReLU()) + in_channels = layer_channels + + deconv_kernel = 4 + self.score_lowres = ConvTranspose2d( + in_channels, num_keypoints, deconv_kernel, stride=2, padding=deconv_kernel // 2 - 1 + ) + self.up_scale = up_scale + + for name, param in self.named_parameters(): + if "bias" in name: + nn.init.constant_(param, 0) + elif "weight" in name: + # Caffe2 implementation uses MSRAFill, which in fact + # corresponds to kaiming_normal_ in PyTorch + nn.init.kaiming_normal_(param, mode="fan_out", nonlinearity="relu") + + @classmethod + def from_config(cls, cfg, input_shape): + ret = super().from_config(cfg, input_shape) + ret["input_shape"] = input_shape + ret["conv_dims"] = cfg.MODEL.ROI_KEYPOINT_HEAD.CONV_DIMS + return ret + + def layers(self, x): + for layer in self: + x = layer(x) + x = interpolate(x, scale_factor=self.up_scale, mode="bilinear", align_corners=False) + return x diff --git a/detectron2/modeling/roi_heads/mask_head.py b/detectron2/modeling/roi_heads/mask_head.py new file mode 100644 index 0000000000000000000000000000000000000000..5ac5c4b9aaa34653d6c50e512a5a4300da450c7f --- /dev/null +++ b/detectron2/modeling/roi_heads/mask_head.py @@ -0,0 +1,292 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from typing import List +import fvcore.nn.weight_init as weight_init +import torch +from torch import nn +from torch.nn import functional as F + +from detectron2.config import configurable +from detectron2.layers import Conv2d, ConvTranspose2d, ShapeSpec, cat, get_norm +from detectron2.structures import Instances +from detectron2.utils.events import get_event_storage +from detectron2.utils.registry import Registry + +__all__ = [ + "BaseMaskRCNNHead", + "MaskRCNNConvUpsampleHead", + "build_mask_head", + "ROI_MASK_HEAD_REGISTRY", +] + + +ROI_MASK_HEAD_REGISTRY = Registry("ROI_MASK_HEAD") +ROI_MASK_HEAD_REGISTRY.__doc__ = """ +Registry for mask heads, which predicts instance masks given +per-region features. + +The registered object will be called with `obj(cfg, input_shape)`. +""" + + +@torch.jit.unused +def mask_rcnn_loss(pred_mask_logits: torch.Tensor, instances: List[Instances], vis_period: int = 0): + """ + Compute the mask prediction loss defined in the Mask R-CNN paper. + + Args: + pred_mask_logits (Tensor): A tensor of shape (B, C, Hmask, Wmask) or (B, 1, Hmask, Wmask) + for class-specific or class-agnostic, where B is the total number of predicted masks + in all images, C is the number of foreground classes, and Hmask, Wmask are the height + and width of the mask predictions. The values are logits. + instances (list[Instances]): A list of N Instances, where N is the number of images + in the batch. These instances are in 1:1 + correspondence with the pred_mask_logits. The ground-truth labels (class, box, mask, + ...) associated with each instance are stored in fields. + vis_period (int): the period (in steps) to dump visualization. + + Returns: + mask_loss (Tensor): A scalar tensor containing the loss. + """ + cls_agnostic_mask = pred_mask_logits.size(1) == 1 + total_num_masks = pred_mask_logits.size(0) + mask_side_len = pred_mask_logits.size(2) + assert pred_mask_logits.size(2) == pred_mask_logits.size(3), "Mask prediction must be square!" + + gt_classes = [] + gt_masks = [] + for instances_per_image in instances: + if len(instances_per_image) == 0: + continue + if not cls_agnostic_mask: + gt_classes_per_image = instances_per_image.gt_classes.to(dtype=torch.int64) + gt_classes.append(gt_classes_per_image) + + gt_masks_per_image = instances_per_image.gt_masks.crop_and_resize( + instances_per_image.proposal_boxes.tensor, mask_side_len + ).to(device=pred_mask_logits.device) + # A tensor of shape (N, M, M), N=#instances in the image; M=mask_side_len + gt_masks.append(gt_masks_per_image) + + if len(gt_masks) == 0: + return pred_mask_logits.sum() * 0 + + gt_masks = cat(gt_masks, dim=0) + + if cls_agnostic_mask: + pred_mask_logits = pred_mask_logits[:, 0] + else: + indices = torch.arange(total_num_masks) + gt_classes = cat(gt_classes, dim=0) + pred_mask_logits = pred_mask_logits[indices, gt_classes] + + if gt_masks.dtype == torch.bool: + gt_masks_bool = gt_masks + else: + # Here we allow gt_masks to be float as well (depend on the implementation of rasterize()) + gt_masks_bool = gt_masks > 0.5 + gt_masks = gt_masks.to(dtype=torch.float32) + + # Log the training accuracy (using gt classes and 0.5 threshold) + mask_incorrect = (pred_mask_logits > 0.0) != gt_masks_bool + mask_accuracy = 1 - (mask_incorrect.sum().item() / max(mask_incorrect.numel(), 1.0)) + num_positive = gt_masks_bool.sum().item() + false_positive = (mask_incorrect & ~gt_masks_bool).sum().item() / max( + gt_masks_bool.numel() - num_positive, 1.0 + ) + false_negative = (mask_incorrect & gt_masks_bool).sum().item() / max(num_positive, 1.0) + + storage = get_event_storage() + storage.put_scalar("mask_rcnn/accuracy", mask_accuracy) + storage.put_scalar("mask_rcnn/false_positive", false_positive) + storage.put_scalar("mask_rcnn/false_negative", false_negative) + if vis_period > 0 and storage.iter % vis_period == 0: + pred_masks = pred_mask_logits.sigmoid() + vis_masks = torch.cat([pred_masks, gt_masks], axis=2) + name = "Left: mask prediction; Right: mask GT" + for idx, vis_mask in enumerate(vis_masks): + vis_mask = torch.stack([vis_mask] * 3, axis=0) + storage.put_image(name + f" ({idx})", vis_mask) + + mask_loss = F.binary_cross_entropy_with_logits(pred_mask_logits, gt_masks, reduction="mean") + return mask_loss + + +def mask_rcnn_inference(pred_mask_logits: torch.Tensor, pred_instances: List[Instances]): + """ + Convert pred_mask_logits to estimated foreground probability masks while also + extracting only the masks for the predicted classes in pred_instances. For each + predicted box, the mask of the same class is attached to the instance by adding a + new "pred_masks" field to pred_instances. + + Args: + pred_mask_logits (Tensor): A tensor of shape (B, C, Hmask, Wmask) or (B, 1, Hmask, Wmask) + for class-specific or class-agnostic, where B is the total number of predicted masks + in all images, C is the number of foreground classes, and Hmask, Wmask are the height + and width of the mask predictions. The values are logits. + pred_instances (list[Instances]): A list of N Instances, where N is the number of images + in the batch. Each Instances must have field "pred_classes". + + Returns: + None. pred_instances will contain an extra "pred_masks" field storing a mask of size (Hmask, + Wmask) for predicted class. Note that the masks are returned as a soft (non-quantized) + masks the resolution predicted by the network; post-processing steps, such as resizing + the predicted masks to the original image resolution and/or binarizing them, is left + to the caller. + """ + cls_agnostic_mask = pred_mask_logits.size(1) == 1 + + if cls_agnostic_mask: + mask_probs_pred = pred_mask_logits.sigmoid() + else: + # Select masks corresponding to the predicted classes + num_masks = pred_mask_logits.shape[0] + class_pred = cat([i.pred_classes for i in pred_instances]) + indices = torch.arange(num_masks, device=class_pred.device) + mask_probs_pred = pred_mask_logits[indices, class_pred][:, None].sigmoid() + # mask_probs_pred.shape: (B, 1, Hmask, Wmask) + + num_boxes_per_image = [len(i) for i in pred_instances] + mask_probs_pred = mask_probs_pred.split(num_boxes_per_image, dim=0) + + for prob, instances in zip(mask_probs_pred, pred_instances): + instances.pred_masks = prob # (1, Hmask, Wmask) + + +class BaseMaskRCNNHead(nn.Module): + """ + Implement the basic Mask R-CNN losses and inference logic described in :paper:`Mask R-CNN` + """ + + @configurable + def __init__(self, *, loss_weight: float = 1.0, vis_period: int = 0): + """ + NOTE: this interface is experimental. + + Args: + loss_weight (float): multiplier of the loss + vis_period (int): visualization period + """ + super().__init__() + self.vis_period = vis_period + self.loss_weight = loss_weight + + @classmethod + def from_config(cls, cfg, input_shape): + return {"vis_period": cfg.VIS_PERIOD} + + def forward(self, x, instances: List[Instances]): + """ + Args: + x: input region feature(s) provided by :class:`ROIHeads`. + instances (list[Instances]): contains the boxes & labels corresponding + to the input features. + Exact format is up to its caller to decide. + Typically, this is the foreground instances in training, with + "proposal_boxes" field and other gt annotations. + In inference, it contains boxes that are already predicted. + + Returns: + A dict of losses in training. The predicted "instances" in inference. + """ + x = self.layers(x) + if self.training: + return {"loss_mask": mask_rcnn_loss(x, instances, self.vis_period) * self.loss_weight} + else: + mask_rcnn_inference(x, instances) + return instances + + def layers(self, x): + """ + Neural network layers that makes predictions from input features. + """ + raise NotImplementedError + + +# To get torchscript support, we make the head a subclass of `nn.Sequential`. +# Therefore, to add new layers in this head class, please make sure they are +# added in the order they will be used in forward(). +@ROI_MASK_HEAD_REGISTRY.register() +class MaskRCNNConvUpsampleHead(BaseMaskRCNNHead, nn.Sequential): + """ + A mask head with several conv layers, plus an upsample layer (with `ConvTranspose2d`). + Predictions are made with a final 1x1 conv layer. + """ + + @configurable + def __init__(self, input_shape: ShapeSpec, *, num_classes, conv_dims, conv_norm="", **kwargs): + """ + NOTE: this interface is experimental. + + Args: + input_shape (ShapeSpec): shape of the input feature + num_classes (int): the number of foreground classes (i.e. background is not + included). 1 if using class agnostic prediction. + conv_dims (list[int]): a list of N>0 integers representing the output dimensions + of N-1 conv layers and the last upsample layer. + conv_norm (str or callable): normalization for the conv layers. + See :func:`detectron2.layers.get_norm` for supported types. + """ + super().__init__(**kwargs) + assert len(conv_dims) >= 1, "conv_dims have to be non-empty!" + + self.conv_norm_relus = [] + + cur_channels = input_shape.channels + for k, conv_dim in enumerate(conv_dims[:-1]): + conv = Conv2d( + cur_channels, + conv_dim, + kernel_size=3, + stride=1, + padding=1, + bias=not conv_norm, + norm=get_norm(conv_norm, conv_dim), + activation=nn.ReLU(), + ) + self.add_module("mask_fcn{}".format(k + 1), conv) + self.conv_norm_relus.append(conv) + cur_channels = conv_dim + + self.deconv = ConvTranspose2d( + cur_channels, conv_dims[-1], kernel_size=2, stride=2, padding=0 + ) + self.add_module("deconv_relu", nn.ReLU()) + cur_channels = conv_dims[-1] + + self.predictor = Conv2d(cur_channels, num_classes, kernel_size=1, stride=1, padding=0) + + for layer in self.conv_norm_relus + [self.deconv]: + weight_init.c2_msra_fill(layer) + # use normal distribution initialization for mask prediction layer + nn.init.normal_(self.predictor.weight, std=0.001) + if self.predictor.bias is not None: + nn.init.constant_(self.predictor.bias, 0) + + @classmethod + def from_config(cls, cfg, input_shape): + ret = super().from_config(cfg, input_shape) + conv_dim = cfg.MODEL.ROI_MASK_HEAD.CONV_DIM + num_conv = cfg.MODEL.ROI_MASK_HEAD.NUM_CONV + ret.update( + conv_dims=[conv_dim] * (num_conv + 1), # +1 for ConvTranspose + conv_norm=cfg.MODEL.ROI_MASK_HEAD.NORM, + input_shape=input_shape, + ) + if cfg.MODEL.ROI_MASK_HEAD.CLS_AGNOSTIC_MASK: + ret["num_classes"] = 1 + else: + ret["num_classes"] = cfg.MODEL.ROI_HEADS.NUM_CLASSES + return ret + + def layers(self, x): + for layer in self: + x = layer(x) + return x + + +def build_mask_head(cfg, input_shape): + """ + Build a mask head defined by `cfg.MODEL.ROI_MASK_HEAD.NAME`. + """ + name = cfg.MODEL.ROI_MASK_HEAD.NAME + return ROI_MASK_HEAD_REGISTRY.get(name)(cfg, input_shape) diff --git a/detectron2/modeling/roi_heads/roi_heads.py b/detectron2/modeling/roi_heads/roi_heads.py new file mode 100644 index 0000000000000000000000000000000000000000..6ef2cab98374ab279d49640b5644c9f7ecb0ee45 --- /dev/null +++ b/detectron2/modeling/roi_heads/roi_heads.py @@ -0,0 +1,887 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import inspect +import logging +import numpy as np +from typing import Dict, List, Optional, Tuple +import torch +from torch import nn + +from detectron2.config import configurable +from detectron2.layers import ShapeSpec, nonzero_tuple +from detectron2.structures import Boxes, ImageList, Instances, pairwise_iou +from detectron2.utils.events import get_event_storage +from detectron2.utils.registry import Registry + +from ..backbone.resnet import BottleneckBlock, ResNet +from ..matcher import Matcher +from ..poolers import ROIPooler +from ..proposal_generator.proposal_utils import add_ground_truth_to_proposals +from ..sampling import subsample_labels +from .box_head import build_box_head +from .fast_rcnn import FastRCNNOutputLayers +from .keypoint_head import build_keypoint_head +from .mask_head import build_mask_head + +ROI_HEADS_REGISTRY = Registry("ROI_HEADS") +ROI_HEADS_REGISTRY.__doc__ = """ +Registry for ROI heads in a generalized R-CNN model. +ROIHeads take feature maps and region proposals, and +perform per-region computation. + +The registered object will be called with `obj(cfg, input_shape)`. +The call is expected to return an :class:`ROIHeads`. +""" + +logger = logging.getLogger(__name__) + + +def build_roi_heads(cfg, input_shape): + """ + Build ROIHeads defined by `cfg.MODEL.ROI_HEADS.NAME`. + """ + name = cfg.MODEL.ROI_HEADS.NAME + return ROI_HEADS_REGISTRY.get(name)(cfg, input_shape) + + +def select_foreground_proposals( + proposals: List[Instances], bg_label: int +) -> Tuple[List[Instances], List[torch.Tensor]]: + """ + Given a list of N Instances (for N images), each containing a `gt_classes` field, + return a list of Instances that contain only instances with `gt_classes != -1 && + gt_classes != bg_label`. + + Args: + proposals (list[Instances]): A list of N Instances, where N is the number of + images in the batch. + bg_label: label index of background class. + + Returns: + list[Instances]: N Instances, each contains only the selected foreground instances. + list[Tensor]: N boolean vector, correspond to the selection mask of + each Instances object. True for selected instances. + """ + assert isinstance(proposals, (list, tuple)) + assert isinstance(proposals[0], Instances) + assert proposals[0].has("gt_classes") + fg_proposals = [] + fg_selection_masks = [] + for proposals_per_image in proposals: + gt_classes = proposals_per_image.gt_classes + fg_selection_mask = (gt_classes != -1) & (gt_classes != bg_label) + fg_idxs = fg_selection_mask.nonzero().squeeze(1) + fg_proposals.append(proposals_per_image[fg_idxs]) + fg_selection_masks.append(fg_selection_mask) + return fg_proposals, fg_selection_masks + + +def select_proposals_with_visible_keypoints(proposals: List[Instances]) -> List[Instances]: + """ + Args: + proposals (list[Instances]): a list of N Instances, where N is the + number of images. + + Returns: + proposals: only contains proposals with at least one visible keypoint. + + Note that this is still slightly different from Detectron. + In Detectron, proposals for training keypoint head are re-sampled from + all the proposals with IOU>threshold & >=1 visible keypoint. + + Here, the proposals are first sampled from all proposals with + IOU>threshold, then proposals with no visible keypoint are filtered out. + This strategy seems to make no difference on Detectron and is easier to implement. + """ + ret = [] + all_num_fg = [] + for proposals_per_image in proposals: + # If empty/unannotated image (hard negatives), skip filtering for train + if len(proposals_per_image) == 0: + ret.append(proposals_per_image) + continue + gt_keypoints = proposals_per_image.gt_keypoints.tensor + # #fg x K x 3 + vis_mask = gt_keypoints[:, :, 2] >= 1 + xs, ys = gt_keypoints[:, :, 0], gt_keypoints[:, :, 1] + proposal_boxes = proposals_per_image.proposal_boxes.tensor.unsqueeze(dim=1) # #fg x 1 x 4 + kp_in_box = ( + (xs >= proposal_boxes[:, :, 0]) + & (xs <= proposal_boxes[:, :, 2]) + & (ys >= proposal_boxes[:, :, 1]) + & (ys <= proposal_boxes[:, :, 3]) + ) + selection = (kp_in_box & vis_mask).any(dim=1) + selection_idxs = nonzero_tuple(selection)[0] + all_num_fg.append(selection_idxs.numel()) + ret.append(proposals_per_image[selection_idxs]) + + storage = get_event_storage() + storage.put_scalar("keypoint_head/num_fg_samples", np.mean(all_num_fg)) + return ret + + +class ROIHeads(torch.nn.Module): + """ + ROIHeads perform all per-region computation in an R-CNN. + + It typically contains logic to + + 1. (in training only) match proposals with ground truth and sample them + 2. crop the regions and extract per-region features using proposals + 3. make per-region predictions with different heads + + It can have many variants, implemented as subclasses of this class. + This base class contains the logic to match/sample proposals. + But it is not necessary to inherit this class if the sampling logic is not needed. + """ + + @configurable + def __init__( + self, + *, + num_classes, + batch_size_per_image, + positive_fraction, + proposal_matcher, + proposal_append_gt=True, + only_sample_fg_proposals=False, + ): + """ + NOTE: this interface is experimental. + + Args: + num_classes (int): number of foreground classes (i.e. background is not included) + batch_size_per_image (int): number of proposals to sample for training + positive_fraction (float): fraction of positive (foreground) proposals + to sample for training. + proposal_matcher (Matcher): matcher that matches proposals and ground truth + proposal_append_gt (bool): whether to include ground truth as proposals as well + """ + super().__init__() + self.batch_size_per_image = batch_size_per_image + self.positive_fraction = positive_fraction + self.num_classes = num_classes + self.proposal_matcher = proposal_matcher + self.proposal_append_gt = proposal_append_gt + self.only_sample_fg_proposals = only_sample_fg_proposals + + @classmethod + def from_config(cls, cfg): + return { + "batch_size_per_image": cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE, + "positive_fraction": cfg.MODEL.ROI_HEADS.POSITIVE_FRACTION, + "num_classes": cfg.MODEL.ROI_HEADS.NUM_CLASSES, + "proposal_append_gt": cfg.MODEL.ROI_HEADS.PROPOSAL_APPEND_GT, + # Matcher to assign box proposals to gt boxes + "proposal_matcher": Matcher( + cfg.MODEL.ROI_HEADS.IOU_THRESHOLDS, + cfg.MODEL.ROI_HEADS.IOU_LABELS, + allow_low_quality_matches=False, + ), + "only_sample_fg_proposals": cfg.MODEL.CLIP.ONLY_SAMPLE_FG_PROPOSALS, + } + + def _sample_proposals( + self, matched_idxs: torch.Tensor, matched_labels: torch.Tensor, gt_classes: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Based on the matching between N proposals and M groundtruth, + sample the proposals and set their classification labels. + + Args: + matched_idxs (Tensor): a vector of length N, each is the best-matched + gt index in [0, M) for each proposal. + matched_labels (Tensor): a vector of length N, the matcher's label + (one of cfg.MODEL.ROI_HEADS.IOU_LABELS) for each proposal. + gt_classes (Tensor): a vector of length M. + + Returns: + Tensor: a vector of indices of sampled proposals. Each is in [0, N). + Tensor: a vector of the same length, the classification label for + each sampled proposal. Each sample is labeled as either a category in + [0, num_classes) or the background (num_classes). + """ + has_gt = gt_classes.numel() > 0 + # Get the corresponding GT for each proposal + if has_gt: + gt_classes = gt_classes[matched_idxs] + # Label unmatched proposals (0 label from matcher) as background (label=num_classes) + gt_classes[matched_labels == 0] = self.num_classes + # Label ignore proposals (-1 label) + gt_classes[matched_labels == -1] = -1 + else: + gt_classes = torch.zeros_like(matched_idxs) + self.num_classes + + # only sample fg proposals to train recognition branch (ref to subsample_labels) + if self.only_sample_fg_proposals: + if has_gt: + positive = nonzero_tuple((gt_classes != -1) & (gt_classes != self.num_classes))[0] + num_pos = int(self.batch_size_per_image * self.positive_fraction) + # protect against not enough positive examples + num_pos = min(positive.numel(), num_pos) + # randomly select positive and negative examples + perm1 = torch.randperm(positive.numel(), device=positive.device)[:num_pos] + sampled_idxs = positive[perm1] + else: # no gt, only keep 1 bg proposal to fill the slot + sampled_idxs = torch.zeros_like(matched_idxs[0:1]) + return sampled_idxs, gt_classes[sampled_idxs] + + sampled_fg_idxs, sampled_bg_idxs = subsample_labels( + gt_classes, self.batch_size_per_image, self.positive_fraction, self.num_classes + ) + + sampled_idxs = torch.cat([sampled_fg_idxs, sampled_bg_idxs], dim=0) + return sampled_idxs, gt_classes[sampled_idxs] + + @torch.no_grad() + def label_and_sample_proposals( + self, proposals: List[Instances], targets: List[Instances] + ) -> List[Instances]: + """ + Prepare some proposals to be used to train the ROI heads. + It performs box matching between `proposals` and `targets`, and assigns + training labels to the proposals. + It returns ``self.batch_size_per_image`` random samples from proposals and groundtruth + boxes, with a fraction of positives that is no larger than + ``self.positive_fraction``. + + Args: + See :meth:`ROIHeads.forward` + + Returns: + list[Instances]: + length `N` list of `Instances`s containing the proposals + sampled for training. Each `Instances` has the following fields: + + - proposal_boxes: the proposal boxes + - gt_boxes: the ground-truth box that the proposal is assigned to + (this is only meaningful if the proposal has a label > 0; if label = 0 + then the ground-truth box is random) + + Other fields such as "gt_classes", "gt_masks", that's included in `targets`. + """ + # Augment proposals with ground-truth boxes. + # In the case of learned proposals (e.g., RPN), when training starts + # the proposals will be low quality due to random initialization. + # It's possible that none of these initial + # proposals have high enough overlap with the gt objects to be used + # as positive examples for the second stage components (box head, + # cls head, mask head). Adding the gt boxes to the set of proposals + # ensures that the second stage components will have some positive + # examples from the start of training. For RPN, this augmentation improves + # convergence and empirically improves box AP on COCO by about 0.5 + # points (under one tested configuration). + if self.proposal_append_gt: + proposals = add_ground_truth_to_proposals(targets, proposals) + + proposals_with_gt = [] + + num_fg_samples = [] + num_bg_samples = [] + for proposals_per_image, targets_per_image in zip(proposals, targets): + has_gt = len(targets_per_image) > 0 + match_quality_matrix = pairwise_iou( + targets_per_image.gt_boxes, proposals_per_image.proposal_boxes + ) + matched_idxs, matched_labels = self.proposal_matcher(match_quality_matrix) + sampled_idxs, gt_classes = self._sample_proposals( + matched_idxs, matched_labels, targets_per_image.gt_classes + ) + + # Set target attributes of the sampled proposals: + proposals_per_image = proposals_per_image[sampled_idxs] + proposals_per_image.gt_classes = gt_classes + + if has_gt: + sampled_targets = matched_idxs[sampled_idxs] + # We index all the attributes of targets that start with "gt_" + # and have not been added to proposals yet (="gt_classes"). + # NOTE: here the indexing waste some compute, because heads + # like masks, keypoints, etc, will filter the proposals again, + # (by foreground/background, or number of keypoints in the image, etc) + # so we essentially index the data twice. + for (trg_name, trg_value) in targets_per_image.get_fields().items(): + if trg_name.startswith("gt_") and not proposals_per_image.has(trg_name): + proposals_per_image.set(trg_name, trg_value[sampled_targets]) + # If no GT is given in the image, we don't know what a dummy gt value can be. + # Therefore the returned proposals won't have any gt_* fields, except for a + # gt_classes full of background label. + + num_bg_samples.append((gt_classes == self.num_classes).sum().item()) + num_fg_samples.append(gt_classes.numel() - num_bg_samples[-1]) + proposals_with_gt.append(proposals_per_image) + + # Log the number of fg/bg samples that are selected for training ROI heads + storage = get_event_storage() + storage.put_scalar("roi_head/num_fg_samples", np.mean(num_fg_samples)) + storage.put_scalar("roi_head/num_bg_samples", np.mean(num_bg_samples)) + #print("num_fg: {}; num_bg: {}".format(num_fg_samples, num_bg_samples)) + + return proposals_with_gt + + def forward( + self, + images: ImageList, + features: Dict[str, torch.Tensor], + proposals: List[Instances], + targets: Optional[List[Instances]] = None, + ) -> Tuple[List[Instances], Dict[str, torch.Tensor]]: + """ + Args: + images (ImageList): + features (dict[str,Tensor]): input data as a mapping from feature + map name to tensor. Axis 0 represents the number of images `N` in + the input data; axes 1-3 are channels, height, and width, which may + vary between feature maps (e.g., if a feature pyramid is used). + proposals (list[Instances]): length `N` list of `Instances`. The i-th + `Instances` contains object proposals for the i-th input image, + with fields "proposal_boxes" and "objectness_logits". + targets (list[Instances], optional): length `N` list of `Instances`. The i-th + `Instances` contains the ground-truth per-instance annotations + for the i-th input image. Specify `targets` during training only. + It may have the following fields: + + - gt_boxes: the bounding box of each instance. + - gt_classes: the label for each instance with a category ranging in [0, #class]. + - gt_masks: PolygonMasks or BitMasks, the ground-truth masks of each instance. + - gt_keypoints: NxKx3, the groud-truth keypoints for each instance. + + Returns: + list[Instances]: length `N` list of `Instances` containing the + detected instances. Returned during inference only; may be [] during training. + + dict[str->Tensor]: + mapping from a named loss to a tensor storing the loss. Used during training only. + """ + raise NotImplementedError() + + +@ROI_HEADS_REGISTRY.register() +class Res5ROIHeads(ROIHeads): + """ + The ROIHeads in a typical "C4" R-CNN model, where + the box and mask head share the cropping and + the per-region feature computation by a Res5 block. + See :paper:`ResNet` Appendix A. + """ + + @configurable + def __init__( + self, + *, + in_features: List[str], + pooler: ROIPooler, + res5: nn.Module, + box_predictor: nn.Module, + mask_head: Optional[nn.Module] = None, + **kwargs, + ): + """ + NOTE: this interface is experimental. + + Args: + in_features (list[str]): list of backbone feature map names to use for + feature extraction + pooler (ROIPooler): pooler to extra region features from backbone + res5 (nn.Sequential): a CNN to compute per-region features, to be used by + ``box_predictor`` and ``mask_head``. Typically this is a "res5" + block from a ResNet. + box_predictor (nn.Module): make box predictions from the feature. + Should have the same interface as :class:`FastRCNNOutputLayers`. + mask_head (nn.Module): transform features to make mask predictions + """ + super().__init__(**kwargs) + self.in_features = in_features + self.pooler = pooler + if isinstance(res5, (list, tuple)): + res5 = nn.Sequential(*res5) + self.res5 = res5 + self.box_predictor = box_predictor + self.mask_on = mask_head is not None + if self.mask_on: + self.mask_head = mask_head + + @classmethod + def from_config(cls, cfg, input_shape): + # fmt: off + ret = super().from_config(cfg) + in_features = ret["in_features"] = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + pooler_scales = (1.0 / input_shape[in_features[0]].stride, ) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + mask_on = cfg.MODEL.MASK_ON + # fmt: on + assert not cfg.MODEL.KEYPOINT_ON + assert len(in_features) == 1 + + ret["pooler"] = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + + # Compatbility with old moco code. Might be useful. + # See notes in StandardROIHeads.from_config + if not inspect.ismethod(cls._build_res5_block): + logger.warning( + "The behavior of _build_res5_block may change. " + "Please do not depend on private methods." + ) + cls._build_res5_block = classmethod(cls._build_res5_block) + + ret["res5"], out_channels = cls._build_res5_block(cfg) + ret["box_predictor"] = FastRCNNOutputLayers( + cfg, ShapeSpec(channels=out_channels, height=1, width=1) + ) + + if mask_on: + ret["mask_head"] = build_mask_head( + cfg, + ShapeSpec(channels=out_channels, width=pooler_resolution, height=pooler_resolution), + ) + return ret + + @classmethod + def _build_res5_block(cls, cfg): + # fmt: off + stage_channel_factor = 2 ** 3 # res5 is 8x res2 + num_groups = cfg.MODEL.RESNETS.NUM_GROUPS + width_per_group = cfg.MODEL.RESNETS.WIDTH_PER_GROUP + bottleneck_channels = num_groups * width_per_group * stage_channel_factor + out_channels = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS * stage_channel_factor + stride_in_1x1 = cfg.MODEL.RESNETS.STRIDE_IN_1X1 + norm = cfg.MODEL.RESNETS.NORM + assert not cfg.MODEL.RESNETS.DEFORM_ON_PER_STAGE[-1], \ + "Deformable conv is not yet supported in res5 head." + # fmt: on + + blocks = ResNet.make_stage( + BottleneckBlock, + 3, + stride_per_block=[2, 1, 1], + in_channels=out_channels // 2, + bottleneck_channels=bottleneck_channels, + out_channels=out_channels, + num_groups=num_groups, + norm=norm, + stride_in_1x1=stride_in_1x1, + ) + return nn.Sequential(*blocks), out_channels + + def _shared_roi_transform(self, features, boxes): + x = self.pooler(features, boxes) + return self.res5(x) + + def forward(self, images, features, proposals, targets=None): + """ + See :meth:`ROIHeads.forward`. + """ + del images + + if self.training: + assert targets + proposals = self.label_and_sample_proposals(proposals, targets) + del targets + + proposal_boxes = [x.proposal_boxes for x in proposals] + box_features = self._shared_roi_transform( + [features[f] for f in self.in_features], proposal_boxes + ) + predictions = self.box_predictor(box_features.mean(dim=[2, 3])) + + if self.training: + del features + losses = self.box_predictor.losses(predictions, proposals) + if self.mask_on: + proposals, fg_selection_masks = select_foreground_proposals( + proposals, self.num_classes + ) + # Since the ROI feature transform is shared between boxes and masks, + # we don't need to recompute features. The mask loss is only defined + # on foreground proposals, so we need to select out the foreground + # features. + mask_features = box_features[torch.cat(fg_selection_masks, dim=0)] + del box_features + losses.update(self.mask_head(mask_features, proposals)) + return [], losses + else: + pred_instances, _ = self.box_predictor.inference(predictions, proposals) + pred_instances = self.forward_with_given_boxes(features, pred_instances) + return pred_instances, {} + + def forward_with_given_boxes(self, features, instances): + """ + Use the given boxes in `instances` to produce other (non-box) per-ROI outputs. + + Args: + features: same as in `forward()` + instances (list[Instances]): instances to predict other outputs. Expect the keys + "pred_boxes" and "pred_classes" to exist. + + Returns: + instances (Instances): + the same `Instances` object, with extra + fields such as `pred_masks` or `pred_keypoints`. + """ + assert not self.training + assert instances[0].has("pred_boxes") and instances[0].has("pred_classes") + + if self.mask_on: + features = [features[f] for f in self.in_features] + x = self._shared_roi_transform(features, [x.pred_boxes for x in instances]) + return self.mask_head(x, instances) + else: + return instances + + +@ROI_HEADS_REGISTRY.register() +class StandardROIHeads(ROIHeads): + """ + It's "standard" in a sense that there is no ROI transform sharing + or feature sharing between tasks. + Each head independently processes the input features by each head's + own pooler and head. + + This class is used by most models, such as FPN and C5. + To implement more models, you can subclass it and implement a different + :meth:`forward()` or a head. + """ + + @configurable + def __init__( + self, + *, + box_in_features: List[str], + box_pooler: ROIPooler, + box_head: nn.Module, + box_predictor: nn.Module, + mask_in_features: Optional[List[str]] = None, + mask_pooler: Optional[ROIPooler] = None, + mask_head: Optional[nn.Module] = None, + keypoint_in_features: Optional[List[str]] = None, + keypoint_pooler: Optional[ROIPooler] = None, + keypoint_head: Optional[nn.Module] = None, + train_on_pred_boxes: bool = False, + **kwargs, + ): + """ + NOTE: this interface is experimental. + + Args: + box_in_features (list[str]): list of feature names to use for the box head. + box_pooler (ROIPooler): pooler to extra region features for box head + box_head (nn.Module): transform features to make box predictions + box_predictor (nn.Module): make box predictions from the feature. + Should have the same interface as :class:`FastRCNNOutputLayers`. + mask_in_features (list[str]): list of feature names to use for the mask + pooler or mask head. None if not using mask head. + mask_pooler (ROIPooler): pooler to extract region features from image features. + The mask head will then take region features to make predictions. + If None, the mask head will directly take the dict of image features + defined by `mask_in_features` + mask_head (nn.Module): transform features to make mask predictions + keypoint_in_features, keypoint_pooler, keypoint_head: similar to ``mask_*``. + train_on_pred_boxes (bool): whether to use proposal boxes or + predicted boxes from the box head to train other heads. + """ + super().__init__(**kwargs) + # keep self.in_features for backward compatibility + self.in_features = self.box_in_features = box_in_features + self.box_pooler = box_pooler + self.box_head = box_head + self.box_predictor = box_predictor + + self.mask_on = mask_in_features is not None + if self.mask_on: + self.mask_in_features = mask_in_features + self.mask_pooler = mask_pooler + self.mask_head = mask_head + + self.keypoint_on = keypoint_in_features is not None + if self.keypoint_on: + self.keypoint_in_features = keypoint_in_features + self.keypoint_pooler = keypoint_pooler + self.keypoint_head = keypoint_head + + self.train_on_pred_boxes = train_on_pred_boxes + + @classmethod + def from_config(cls, cfg, input_shape): + ret = super().from_config(cfg) + ret["train_on_pred_boxes"] = cfg.MODEL.ROI_BOX_HEAD.TRAIN_ON_PRED_BOXES + # Subclasses that have not been updated to use from_config style construction + # may have overridden _init_*_head methods. In this case, those overridden methods + # will not be classmethods and we need to avoid trying to call them here. + # We test for this with ismethod which only returns True for bound methods of cls. + # Such subclasses will need to handle calling their overridden _init_*_head methods. + if inspect.ismethod(cls._init_box_head): + ret.update(cls._init_box_head(cfg, input_shape)) + if inspect.ismethod(cls._init_mask_head): + ret.update(cls._init_mask_head(cfg, input_shape)) + if inspect.ismethod(cls._init_keypoint_head): + ret.update(cls._init_keypoint_head(cfg, input_shape)) + return ret + + @classmethod + def _init_box_head(cls, cfg, input_shape): + # fmt: off + in_features = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_scales = tuple(1.0 / input_shape[k].stride for k in in_features) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + # fmt: on + + # If StandardROIHeads is applied on multiple feature maps (as in FPN), + # then we share the same predictors and therefore the channel counts must be the same + in_channels = [input_shape[f].channels for f in in_features] + # Check all channel counts are equal + assert len(set(in_channels)) == 1, in_channels + in_channels = in_channels[0] + + box_pooler = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + # Here we split "box head" and "box predictor", which is mainly due to historical reasons. + # They are used together so the "box predictor" layers should be part of the "box head". + # New subclasses of ROIHeads do not need "box predictor"s. + box_head = build_box_head( + cfg, ShapeSpec(channels=in_channels, height=pooler_resolution, width=pooler_resolution) + ) + box_predictor = FastRCNNOutputLayers(cfg, box_head.output_shape) + return { + "box_in_features": in_features, + "box_pooler": box_pooler, + "box_head": box_head, + "box_predictor": box_predictor, + } + + @classmethod + def _init_mask_head(cls, cfg, input_shape): + if not cfg.MODEL.MASK_ON: + return {} + # fmt: off + in_features = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_MASK_HEAD.POOLER_RESOLUTION + pooler_scales = tuple(1.0 / input_shape[k].stride for k in in_features) + sampling_ratio = cfg.MODEL.ROI_MASK_HEAD.POOLER_SAMPLING_RATIO + pooler_type = cfg.MODEL.ROI_MASK_HEAD.POOLER_TYPE + # fmt: on + + in_channels = [input_shape[f].channels for f in in_features][0] + + ret = {"mask_in_features": in_features} + ret["mask_pooler"] = ( + ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + if pooler_type + else None + ) + if pooler_type: + shape = ShapeSpec( + channels=in_channels, width=pooler_resolution, height=pooler_resolution + ) + else: + shape = {f: input_shape[f] for f in in_features} + ret["mask_head"] = build_mask_head(cfg, shape) + return ret + + @classmethod + def _init_keypoint_head(cls, cfg, input_shape): + if not cfg.MODEL.KEYPOINT_ON: + return {} + # fmt: off + in_features = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_RESOLUTION + pooler_scales = tuple(1.0 / input_shape[k].stride for k in in_features) # noqa + sampling_ratio = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_SAMPLING_RATIO + pooler_type = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_TYPE + # fmt: on + + in_channels = [input_shape[f].channels for f in in_features][0] + + ret = {"keypoint_in_features": in_features} + ret["keypoint_pooler"] = ( + ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + if pooler_type + else None + ) + if pooler_type: + shape = ShapeSpec( + channels=in_channels, width=pooler_resolution, height=pooler_resolution + ) + else: + shape = {f: input_shape[f] for f in in_features} + ret["keypoint_head"] = build_keypoint_head(cfg, shape) + return ret + + def forward( + self, + images: ImageList, + features: Dict[str, torch.Tensor], + proposals: List[Instances], + targets: Optional[List[Instances]] = None, + ) -> Tuple[List[Instances], Dict[str, torch.Tensor]]: + """ + See :class:`ROIHeads.forward`. + """ + del images + if self.training: + assert targets, "'targets' argument is required during training" + proposals = self.label_and_sample_proposals(proposals, targets) + del targets + + if self.training: + losses = self._forward_box(features, proposals) + # Usually the original proposals used by the box head are used by the mask, keypoint + # heads. But when `self.train_on_pred_boxes is True`, proposals will contain boxes + # predicted by the box head. + losses.update(self._forward_mask(features, proposals)) + losses.update(self._forward_keypoint(features, proposals)) + return proposals, losses + else: + pred_instances = self._forward_box(features, proposals) + # During inference cascaded prediction is used: the mask and keypoints heads are only + # applied to the top scoring box detections. + pred_instances = self.forward_with_given_boxes(features, pred_instances) + return pred_instances, {} + + def forward_with_given_boxes( + self, features: Dict[str, torch.Tensor], instances: List[Instances] + ) -> List[Instances]: + """ + Use the given boxes in `instances` to produce other (non-box) per-ROI outputs. + + This is useful for downstream tasks where a box is known, but need to obtain + other attributes (outputs of other heads). + Test-time augmentation also uses this. + + Args: + features: same as in `forward()` + instances (list[Instances]): instances to predict other outputs. Expect the keys + "pred_boxes" and "pred_classes" to exist. + + Returns: + list[Instances]: + the same `Instances` objects, with extra + fields such as `pred_masks` or `pred_keypoints`. + """ + assert not self.training + assert instances[0].has("pred_boxes") and instances[0].has("pred_classes") + + instances = self._forward_mask(features, instances) + instances = self._forward_keypoint(features, instances) + return instances + + def _forward_box(self, features: Dict[str, torch.Tensor], proposals: List[Instances]): + """ + Forward logic of the box prediction branch. If `self.train_on_pred_boxes is True`, + the function puts predicted boxes in the `proposal_boxes` field of `proposals` argument. + + Args: + features (dict[str, Tensor]): mapping from feature map names to tensor. + Same as in :meth:`ROIHeads.forward`. + proposals (list[Instances]): the per-image object proposals with + their matching ground truth. + Each has fields "proposal_boxes", and "objectness_logits", + "gt_classes", "gt_boxes". + + Returns: + In training, a dict of losses. + In inference, a list of `Instances`, the predicted instances. + """ + features = [features[f] for f in self.box_in_features] + box_features = self.box_pooler(features, [x.proposal_boxes for x in proposals]) + box_features = self.box_head(box_features) + predictions = self.box_predictor(box_features) + del box_features + + if self.training: + losses = self.box_predictor.losses(predictions, proposals) + # proposals is modified in-place below, so losses must be computed first. + if self.train_on_pred_boxes: + with torch.no_grad(): + pred_boxes = self.box_predictor.predict_boxes_for_gt_classes( + predictions, proposals + ) + for proposals_per_image, pred_boxes_per_image in zip(proposals, pred_boxes): + proposals_per_image.proposal_boxes = Boxes(pred_boxes_per_image) + return losses + else: + pred_instances, _ = self.box_predictor.inference(predictions, proposals) + return pred_instances + + def _forward_mask(self, features: Dict[str, torch.Tensor], instances: List[Instances]): + """ + Forward logic of the mask prediction branch. + + Args: + features (dict[str, Tensor]): mapping from feature map names to tensor. + Same as in :meth:`ROIHeads.forward`. + instances (list[Instances]): the per-image instances to train/predict masks. + In training, they can be the proposals. + In inference, they can be the boxes predicted by R-CNN box head. + + Returns: + In training, a dict of losses. + In inference, update `instances` with new fields "pred_masks" and return it. + """ + if not self.mask_on: + return {} if self.training else instances + + if self.training: + # head is only trained on positive proposals. + instances, _ = select_foreground_proposals(instances, self.num_classes) + + if self.mask_pooler is not None: + features = [features[f] for f in self.mask_in_features] + boxes = [x.proposal_boxes if self.training else x.pred_boxes for x in instances] + features = self.mask_pooler(features, boxes) + else: + features = {f: features[f] for f in self.mask_in_features} + return self.mask_head(features, instances) + + def _forward_keypoint(self, features: Dict[str, torch.Tensor], instances: List[Instances]): + """ + Forward logic of the keypoint prediction branch. + + Args: + features (dict[str, Tensor]): mapping from feature map names to tensor. + Same as in :meth:`ROIHeads.forward`. + instances (list[Instances]): the per-image instances to train/predict keypoints. + In training, they can be the proposals. + In inference, they can be the boxes predicted by R-CNN box head. + + Returns: + In training, a dict of losses. + In inference, update `instances` with new fields "pred_keypoints" and return it. + """ + if not self.keypoint_on: + return {} if self.training else instances + + if self.training: + # head is only trained on positive proposals with >=1 visible keypoints. + instances, _ = select_foreground_proposals(instances, self.num_classes) + instances = select_proposals_with_visible_keypoints(instances) + + if self.keypoint_pooler is not None: + features = [features[f] for f in self.keypoint_in_features] + boxes = [x.proposal_boxes if self.training else x.pred_boxes for x in instances] + features = self.keypoint_pooler(features, boxes) + else: + features = {f: features[f] for f in self.keypoint_in_features} + return self.keypoint_head(features, instances) diff --git a/detectron2/modeling/roi_heads/rotated_fast_rcnn.py b/detectron2/modeling/roi_heads/rotated_fast_rcnn.py new file mode 100644 index 0000000000000000000000000000000000000000..7d0a0a638e95f4b6cd83f9658efefc5fd50206d7 --- /dev/null +++ b/detectron2/modeling/roi_heads/rotated_fast_rcnn.py @@ -0,0 +1,270 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import numpy as np +import torch + +from detectron2.config import configurable +from detectron2.layers import ShapeSpec, batched_nms_rotated +from detectron2.structures import Instances, RotatedBoxes, pairwise_iou_rotated +from detectron2.utils.events import get_event_storage + +from ..box_regression import Box2BoxTransformRotated +from ..poolers import ROIPooler +from ..proposal_generator.proposal_utils import add_ground_truth_to_proposals +from .box_head import build_box_head +from .fast_rcnn import FastRCNNOutputLayers +from .roi_heads import ROI_HEADS_REGISTRY, StandardROIHeads + +logger = logging.getLogger(__name__) + +""" +Shape shorthand in this module: + + N: number of images in the minibatch + R: number of ROIs, combined over all images, in the minibatch + Ri: number of ROIs in image i + K: number of foreground classes. E.g.,there are 80 foreground classes in COCO. + +Naming convention: + + deltas: refers to the 5-d (dx, dy, dw, dh, da) deltas that parameterize the box2box + transform (see :class:`box_regression.Box2BoxTransformRotated`). + + pred_class_logits: predicted class scores in [-inf, +inf]; use + softmax(pred_class_logits) to estimate P(class). + + gt_classes: ground-truth classification labels in [0, K], where [0, K) represent + foreground object classes and K represents the background class. + + pred_proposal_deltas: predicted rotated box2box transform deltas for transforming proposals + to detection box predictions. + + gt_proposal_deltas: ground-truth rotated box2box transform deltas +""" + + +def fast_rcnn_inference_rotated( + boxes, scores, image_shapes, score_thresh, nms_thresh, topk_per_image +): + """ + Call `fast_rcnn_inference_single_image_rotated` for all images. + + Args: + boxes (list[Tensor]): A list of Tensors of predicted class-specific or class-agnostic + boxes for each image. Element i has shape (Ri, K * 5) if doing + class-specific regression, or (Ri, 5) if doing class-agnostic + regression, where Ri is the number of predicted objects for image i. + This is compatible with the output of :meth:`FastRCNNOutputs.predict_boxes`. + scores (list[Tensor]): A list of Tensors of predicted class scores for each image. + Element i has shape (Ri, K + 1), where Ri is the number of predicted objects + for image i. Compatible with the output of :meth:`FastRCNNOutputs.predict_probs`. + image_shapes (list[tuple]): A list of (width, height) tuples for each image in the batch. + score_thresh (float): Only return detections with a confidence score exceeding this + threshold. + nms_thresh (float): The threshold to use for box non-maximum suppression. Value in [0, 1]. + topk_per_image (int): The number of top scoring detections to return. Set < 0 to return + all detections. + + Returns: + instances: (list[Instances]): A list of N instances, one for each image in the batch, + that stores the topk most confidence detections. + kept_indices: (list[Tensor]): A list of 1D tensor of length of N, each element indicates + the corresponding boxes/scores index in [0, Ri) from the input, for image i. + """ + result_per_image = [ + fast_rcnn_inference_single_image_rotated( + boxes_per_image, scores_per_image, image_shape, score_thresh, nms_thresh, topk_per_image + ) + for scores_per_image, boxes_per_image, image_shape in zip(scores, boxes, image_shapes) + ] + return [x[0] for x in result_per_image], [x[1] for x in result_per_image] + + +def fast_rcnn_inference_single_image_rotated( + boxes, scores, image_shape, score_thresh, nms_thresh, topk_per_image +): + """ + Single-image inference. Return rotated bounding-box detection results by thresholding + on scores and applying rotated non-maximum suppression (Rotated NMS). + + Args: + Same as `fast_rcnn_inference_rotated`, but with rotated boxes, scores, and image shapes + per image. + + Returns: + Same as `fast_rcnn_inference_rotated`, but for only one image. + """ + valid_mask = torch.isfinite(boxes).all(dim=1) & torch.isfinite(scores).all(dim=1) + if not valid_mask.all(): + boxes = boxes[valid_mask] + scores = scores[valid_mask] + + B = 5 # box dimension + scores = scores[:, :-1] + num_bbox_reg_classes = boxes.shape[1] // B + # Convert to Boxes to use the `clip` function ... + boxes = RotatedBoxes(boxes.reshape(-1, B)) + boxes.clip(image_shape) + boxes = boxes.tensor.view(-1, num_bbox_reg_classes, B) # R x C x B + # Filter results based on detection scores + filter_mask = scores > score_thresh # R x K + # R' x 2. First column contains indices of the R predictions; + # Second column contains indices of classes. + filter_inds = filter_mask.nonzero() + if num_bbox_reg_classes == 1: + boxes = boxes[filter_inds[:, 0], 0] + else: + boxes = boxes[filter_mask] + scores = scores[filter_mask] + + # Apply per-class Rotated NMS + keep = batched_nms_rotated(boxes, scores, filter_inds[:, 1], nms_thresh) + if topk_per_image >= 0: + keep = keep[:topk_per_image] + boxes, scores, filter_inds = boxes[keep], scores[keep], filter_inds[keep] + + result = Instances(image_shape) + result.pred_boxes = RotatedBoxes(boxes) + result.scores = scores + result.pred_classes = filter_inds[:, 1] + + return result, filter_inds[:, 0] + + +class RotatedFastRCNNOutputLayers(FastRCNNOutputLayers): + """ + Two linear layers for predicting Rotated Fast R-CNN outputs. + """ + + @classmethod + def from_config(cls, cfg, input_shape): + args = super().from_config(cfg, input_shape) + args["box2box_transform"] = Box2BoxTransformRotated( + weights=cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS + ) + return args + + def inference(self, predictions, proposals): + """ + Returns: + list[Instances]: same as `fast_rcnn_inference_rotated`. + list[Tensor]: same as `fast_rcnn_inference_rotated`. + """ + boxes = self.predict_boxes(predictions, proposals) + scores = self.predict_probs(predictions, proposals) + image_shapes = [x.image_size for x in proposals] + + return fast_rcnn_inference_rotated( + boxes, + scores, + image_shapes, + self.test_score_thresh, + self.test_nms_thresh, + self.test_topk_per_image, + ) + + +@ROI_HEADS_REGISTRY.register() +class RROIHeads(StandardROIHeads): + """ + This class is used by Rotated Fast R-CNN to detect rotated boxes. + For now, it only supports box predictions but not mask or keypoints. + """ + + @configurable + def __init__(self, **kwargs): + """ + NOTE: this interface is experimental. + """ + super().__init__(**kwargs) + assert ( + not self.mask_on and not self.keypoint_on + ), "Mask/Keypoints not supported in Rotated ROIHeads." + assert not self.train_on_pred_boxes, "train_on_pred_boxes not implemented for RROIHeads!" + + @classmethod + def _init_box_head(cls, cfg, input_shape): + # fmt: off + in_features = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_scales = tuple(1.0 / input_shape[k].stride for k in in_features) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + # fmt: on + assert pooler_type in ["ROIAlignRotated"], pooler_type + # assume all channel counts are equal + in_channels = [input_shape[f].channels for f in in_features][0] + + box_pooler = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + box_head = build_box_head( + cfg, ShapeSpec(channels=in_channels, height=pooler_resolution, width=pooler_resolution) + ) + # This line is the only difference v.s. StandardROIHeads + box_predictor = RotatedFastRCNNOutputLayers(cfg, box_head.output_shape) + return { + "box_in_features": in_features, + "box_pooler": box_pooler, + "box_head": box_head, + "box_predictor": box_predictor, + } + + @torch.no_grad() + def label_and_sample_proposals(self, proposals, targets): + """ + Prepare some proposals to be used to train the RROI heads. + It performs box matching between `proposals` and `targets`, and assigns + training labels to the proposals. + It returns `self.batch_size_per_image` random samples from proposals and groundtruth boxes, + with a fraction of positives that is no larger than `self.positive_sample_fraction. + + Args: + See :meth:`StandardROIHeads.forward` + + Returns: + list[Instances]: length `N` list of `Instances`s containing the proposals + sampled for training. Each `Instances` has the following fields: + - proposal_boxes: the rotated proposal boxes + - gt_boxes: the ground-truth rotated boxes that the proposal is assigned to + (this is only meaningful if the proposal has a label > 0; if label = 0 + then the ground-truth box is random) + - gt_classes: the ground-truth classification lable for each proposal + """ + if self.proposal_append_gt: + proposals = add_ground_truth_to_proposals(targets, proposals) + + proposals_with_gt = [] + + num_fg_samples = [] + num_bg_samples = [] + for proposals_per_image, targets_per_image in zip(proposals, targets): + has_gt = len(targets_per_image) > 0 + match_quality_matrix = pairwise_iou_rotated( + targets_per_image.gt_boxes, proposals_per_image.proposal_boxes + ) + matched_idxs, matched_labels = self.proposal_matcher(match_quality_matrix) + sampled_idxs, gt_classes = self._sample_proposals( + matched_idxs, matched_labels, targets_per_image.gt_classes + ) + + proposals_per_image = proposals_per_image[sampled_idxs] + proposals_per_image.gt_classes = gt_classes + + if has_gt: + sampled_targets = matched_idxs[sampled_idxs] + proposals_per_image.gt_boxes = targets_per_image.gt_boxes[sampled_targets] + + num_bg_samples.append((gt_classes == self.num_classes).sum().item()) + num_fg_samples.append(gt_classes.numel() - num_bg_samples[-1]) + proposals_with_gt.append(proposals_per_image) + + # Log the number of fg/bg samples that are selected for training ROI heads + storage = get_event_storage() + storage.put_scalar("roi_head/num_fg_samples", np.mean(num_fg_samples)) + storage.put_scalar("roi_head/num_bg_samples", np.mean(num_bg_samples)) + + return proposals_with_gt diff --git a/detectron2/modeling/sampling.py b/detectron2/modeling/sampling.py new file mode 100644 index 0000000000000000000000000000000000000000..feb63e9804b98b5b048020596dccc0f1999e355c --- /dev/null +++ b/detectron2/modeling/sampling.py @@ -0,0 +1,54 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import torch + +from detectron2.layers import nonzero_tuple + +__all__ = ["subsample_labels"] + + +def subsample_labels( + labels: torch.Tensor, num_samples: int, positive_fraction: float, bg_label: int +): + """ + Return `num_samples` (or fewer, if not enough found) + random samples from `labels` which is a mixture of positives & negatives. + It will try to return as many positives as possible without + exceeding `positive_fraction * num_samples`, and then try to + fill the remaining slots with negatives. + + Args: + labels (Tensor): (N, ) label vector with values: + * -1: ignore + * bg_label: background ("negative") class + * otherwise: one or more foreground ("positive") classes + num_samples (int): The total number of labels with value >= 0 to return. + Values that are not sampled will be filled with -1 (ignore). + positive_fraction (float): The number of subsampled labels with values > 0 + is `min(num_positives, int(positive_fraction * num_samples))`. The number + of negatives sampled is `min(num_negatives, num_samples - num_positives_sampled)`. + In order words, if there are not enough positives, the sample is filled with + negatives. If there are also not enough negatives, then as many elements are + sampled as is possible. + bg_label (int): label index of background ("negative") class. + + Returns: + pos_idx, neg_idx (Tensor): + 1D vector of indices. The total length of both is `num_samples` or fewer. + """ + positive = nonzero_tuple((labels != -1) & (labels != bg_label))[0] + negative = nonzero_tuple(labels == bg_label)[0] + + num_pos = int(num_samples * positive_fraction) + # protect against not enough positive examples + num_pos = min(positive.numel(), num_pos) + num_neg = num_samples - num_pos + # protect against not enough negative examples + num_neg = min(negative.numel(), num_neg) + + # randomly select positive and negative examples + perm1 = torch.randperm(positive.numel(), device=positive.device)[:num_pos] + perm2 = torch.randperm(negative.numel())[:num_neg].to(negative.device) # torch.randperm(negative.numel(), device=negative.device)[:num_neg] + + pos_idx = positive[perm1] + neg_idx = negative[perm2] + return pos_idx, neg_idx diff --git a/detectron2/modeling/test_time_augmentation.py b/detectron2/modeling/test_time_augmentation.py new file mode 100644 index 0000000000000000000000000000000000000000..373e6bf00a39c040ff1da49d6dcd39a54a0b69a7 --- /dev/null +++ b/detectron2/modeling/test_time_augmentation.py @@ -0,0 +1,307 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import numpy as np +from contextlib import contextmanager +from itertools import count +from typing import List +import torch +from fvcore.transforms import HFlipTransform, NoOpTransform +from torch import nn +from torch.nn.parallel import DistributedDataParallel + +from detectron2.config import configurable +from detectron2.data.detection_utils import read_image +from detectron2.data.transforms import ( + RandomFlip, + ResizeShortestEdge, + ResizeTransform, + apply_augmentations, +) +from detectron2.structures import Boxes, Instances + +from .meta_arch import GeneralizedRCNN +from .postprocessing import detector_postprocess +from .roi_heads.fast_rcnn import fast_rcnn_inference_single_image + +__all__ = ["DatasetMapperTTA", "GeneralizedRCNNWithTTA"] + + +class DatasetMapperTTA: + """ + Implement test-time augmentation for detection data. + It is a callable which takes a dataset dict from a detection dataset, + and returns a list of dataset dicts where the images + are augmented from the input image by the transformations defined in the config. + This is used for test-time augmentation. + """ + + @configurable + def __init__(self, min_sizes: List[int], max_size: int, flip: bool): + """ + Args: + min_sizes: list of short-edge size to resize the image to + max_size: maximum height or width of resized images + flip: whether to apply flipping augmentation + """ + self.min_sizes = min_sizes + self.max_size = max_size + self.flip = flip + + @classmethod + def from_config(cls, cfg): + return { + "min_sizes": cfg.TEST.AUG.MIN_SIZES, + "max_size": cfg.TEST.AUG.MAX_SIZE, + "flip": cfg.TEST.AUG.FLIP, + } + + def __call__(self, dataset_dict): + """ + Args: + dict: a dict in standard model input format. See tutorials for details. + + Returns: + list[dict]: + a list of dicts, which contain augmented version of the input image. + The total number of dicts is ``len(min_sizes) * (2 if flip else 1)``. + Each dict has field "transforms" which is a TransformList, + containing the transforms that are used to generate this image. + """ + numpy_image = dataset_dict["image"].permute(1, 2, 0).numpy() + shape = numpy_image.shape + orig_shape = (dataset_dict["height"], dataset_dict["width"]) + if shape[:2] != orig_shape: + # It transforms the "original" image in the dataset to the input image + pre_tfm = ResizeTransform(orig_shape[0], orig_shape[1], shape[0], shape[1]) + else: + pre_tfm = NoOpTransform() + + # Create all combinations of augmentations to use + aug_candidates = [] # each element is a list[Augmentation] + for min_size in self.min_sizes: + resize = ResizeShortestEdge(min_size, self.max_size) + aug_candidates.append([resize]) # resize only + if self.flip: + flip = RandomFlip(prob=1.0) + aug_candidates.append([resize, flip]) # resize + flip + + # Apply all the augmentations + ret = [] + for aug in aug_candidates: + new_image, tfms = apply_augmentations(aug, np.copy(numpy_image)) + torch_image = torch.from_numpy(np.ascontiguousarray(new_image.transpose(2, 0, 1))) + + dic = copy.deepcopy(dataset_dict) + dic["transforms"] = pre_tfm + tfms + dic["image"] = torch_image + ret.append(dic) + return ret + + +class GeneralizedRCNNWithTTA(nn.Module): + """ + A GeneralizedRCNN with test-time augmentation enabled. + Its :meth:`__call__` method has the same interface as :meth:`GeneralizedRCNN.forward`. + """ + + def __init__(self, cfg, model, tta_mapper=None, batch_size=3): + """ + Args: + cfg (CfgNode): + model (GeneralizedRCNN): a GeneralizedRCNN to apply TTA on. + tta_mapper (callable): takes a dataset dict and returns a list of + augmented versions of the dataset dict. Defaults to + `DatasetMapperTTA(cfg)`. + batch_size (int): batch the augmented images into this batch size for inference. + """ + super().__init__() + if isinstance(model, DistributedDataParallel): + model = model.module + assert isinstance( + model, GeneralizedRCNN + ), "TTA is only supported on GeneralizedRCNN. Got a model of type {}".format(type(model)) + self.cfg = cfg.clone() + assert not self.cfg.MODEL.KEYPOINT_ON, "TTA for keypoint is not supported yet" + assert ( + not self.cfg.MODEL.LOAD_PROPOSALS + ), "TTA for pre-computed proposals is not supported yet" + + self.model = model + + if tta_mapper is None: + tta_mapper = DatasetMapperTTA(cfg) + self.tta_mapper = tta_mapper + self.batch_size = batch_size + + @contextmanager + def _turn_off_roi_heads(self, attrs): + """ + Open a context where some heads in `model.roi_heads` are temporarily turned off. + Args: + attr (list[str]): the attribute in `model.roi_heads` which can be used + to turn off a specific head, e.g., "mask_on", "keypoint_on". + """ + roi_heads = self.model.roi_heads + old = {} + for attr in attrs: + try: + old[attr] = getattr(roi_heads, attr) + except AttributeError: + # The head may not be implemented in certain ROIHeads + pass + + if len(old.keys()) == 0: + yield + else: + for attr in old.keys(): + setattr(roi_heads, attr, False) + yield + for attr in old.keys(): + setattr(roi_heads, attr, old[attr]) + + def _batch_inference(self, batched_inputs, detected_instances=None): + """ + Execute inference on a list of inputs, + using batch size = self.batch_size, instead of the length of the list. + + Inputs & outputs have the same format as :meth:`GeneralizedRCNN.inference` + """ + if detected_instances is None: + detected_instances = [None] * len(batched_inputs) + + outputs = [] + inputs, instances = [], [] + for idx, input, instance in zip(count(), batched_inputs, detected_instances): + inputs.append(input) + instances.append(instance) + if len(inputs) == self.batch_size or idx == len(batched_inputs) - 1: + outputs.extend( + self.model.inference( + inputs, + instances if instances[0] is not None else None, + do_postprocess=False, + ) + ) + inputs, instances = [], [] + return outputs + + def __call__(self, batched_inputs): + """ + Same input/output format as :meth:`GeneralizedRCNN.forward` + """ + + def _maybe_read_image(dataset_dict): + ret = copy.copy(dataset_dict) + if "image" not in ret: + image = read_image(ret.pop("file_name"), self.model.input_format) + image = torch.from_numpy(np.ascontiguousarray(image.transpose(2, 0, 1))) # CHW + ret["image"] = image + if "height" not in ret and "width" not in ret: + ret["height"] = image.shape[1] + ret["width"] = image.shape[2] + return ret + + return [self._inference_one_image(_maybe_read_image(x)) for x in batched_inputs] + + def _inference_one_image(self, input): + """ + Args: + input (dict): one dataset dict with "image" field being a CHW tensor + + Returns: + dict: one output dict + """ + orig_shape = (input["height"], input["width"]) + augmented_inputs, tfms = self._get_augmented_inputs(input) + # Detect boxes from all augmented versions + with self._turn_off_roi_heads(["mask_on", "keypoint_on"]): + # temporarily disable roi heads + all_boxes, all_scores, all_classes = self._get_augmented_boxes(augmented_inputs, tfms) + # merge all detected boxes to obtain final predictions for boxes + merged_instances = self._merge_detections(all_boxes, all_scores, all_classes, orig_shape) + + if self.cfg.MODEL.MASK_ON: + # Use the detected boxes to obtain masks + augmented_instances = self._rescale_detected_boxes( + augmented_inputs, merged_instances, tfms + ) + # run forward on the detected boxes + outputs = self._batch_inference(augmented_inputs, augmented_instances) + # Delete now useless variables to avoid being out of memory + del augmented_inputs, augmented_instances + # average the predictions + merged_instances.pred_masks = self._reduce_pred_masks(outputs, tfms) + merged_instances = detector_postprocess(merged_instances, *orig_shape) + return {"instances": merged_instances} + else: + return {"instances": merged_instances} + + def _get_augmented_inputs(self, input): + augmented_inputs = self.tta_mapper(input) + tfms = [x.pop("transforms") for x in augmented_inputs] + return augmented_inputs, tfms + + def _get_augmented_boxes(self, augmented_inputs, tfms): + # 1: forward with all augmented images + outputs = self._batch_inference(augmented_inputs) + # 2: union the results + all_boxes = [] + all_scores = [] + all_classes = [] + for output, tfm in zip(outputs, tfms): + # Need to inverse the transforms on boxes, to obtain results on original image + pred_boxes = output.pred_boxes.tensor + original_pred_boxes = tfm.inverse().apply_box(pred_boxes.cpu().numpy()) + all_boxes.append(torch.from_numpy(original_pred_boxes).to(pred_boxes.device)) + + all_scores.extend(output.scores) + all_classes.extend(output.pred_classes) + all_boxes = torch.cat(all_boxes, dim=0) + return all_boxes, all_scores, all_classes + + def _merge_detections(self, all_boxes, all_scores, all_classes, shape_hw): + # select from the union of all results + num_boxes = len(all_boxes) + num_classes = self.cfg.MODEL.ROI_HEADS.NUM_CLASSES + # +1 because fast_rcnn_inference expects background scores as well + all_scores_2d = torch.zeros(num_boxes, num_classes + 1, device=all_boxes.device) + for idx, cls, score in zip(count(), all_classes, all_scores): + all_scores_2d[idx, cls] = score + + merged_instances, _ = fast_rcnn_inference_single_image( + all_boxes, + all_scores_2d, + shape_hw, + 1e-8, + self.cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST, + self.cfg.TEST.DETECTIONS_PER_IMAGE, + ) + + return merged_instances + + def _rescale_detected_boxes(self, augmented_inputs, merged_instances, tfms): + augmented_instances = [] + for input, tfm in zip(augmented_inputs, tfms): + # Transform the target box to the augmented image's coordinate space + pred_boxes = merged_instances.pred_boxes.tensor.cpu().numpy() + pred_boxes = torch.from_numpy(tfm.apply_box(pred_boxes)) + + aug_instances = Instances( + image_size=input["image"].shape[1:3], + pred_boxes=Boxes(pred_boxes), + pred_classes=merged_instances.pred_classes, + scores=merged_instances.scores, + ) + augmented_instances.append(aug_instances) + return augmented_instances + + def _reduce_pred_masks(self, outputs, tfms): + # Should apply inverse transforms on masks. + # We assume only resize & flip are used. pred_masks is a scale-invariant + # representation, so we handle flip specially + for output, tfm in zip(outputs, tfms): + if any(isinstance(t, HFlipTransform) for t in tfm.transforms): + output.pred_masks = output.pred_masks.flip(dims=[3]) + all_pred_masks = torch.stack([o.pred_masks for o in outputs], dim=0) + avg_pred_masks = torch.mean(all_pred_masks, dim=0) + return avg_pred_masks diff --git a/detectron2/modeling/text_encoder/__init__.py b/detectron2/modeling/text_encoder/__init__.py new file mode 100755 index 0000000000000000000000000000000000000000..e09753c06e7cd77d8df3bee03b04ae9f85ce80bb --- /dev/null +++ b/detectron2/modeling/text_encoder/__init__.py @@ -0,0 +1,9 @@ +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from .build import build_lang_encoder as build_text_encoder +from .build import build_tokenizer + +from .transformer import * +from .hf_model import * diff --git a/detectron2/modeling/text_encoder/__pycache__/__init__.cpython-38.pyc b/detectron2/modeling/text_encoder/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6a19b6aff2f82801eee0411b36038f9b9e66e048 Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/__init__.cpython-38.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/__init__.cpython-39.pyc b/detectron2/modeling/text_encoder/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..84f58c0b001757ddeef65f1368da451539ba289e Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/build.cpython-38.pyc b/detectron2/modeling/text_encoder/__pycache__/build.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..79e699ee02f59628f22b3ec08feed1fac16e6bf2 Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/build.cpython-38.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/build.cpython-39.pyc b/detectron2/modeling/text_encoder/__pycache__/build.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ecf435540ebc289f3bf1b35c8ab40053c1f6a3f5 Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/build.cpython-39.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/hf_model.cpython-38.pyc b/detectron2/modeling/text_encoder/__pycache__/hf_model.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..222d878ebf486ddc86ef253950b82e0aacd90d09 Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/hf_model.cpython-38.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/hf_model.cpython-39.pyc b/detectron2/modeling/text_encoder/__pycache__/hf_model.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..578e2b2c2551b86d456a1840635a8f38c8bb82bf Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/hf_model.cpython-39.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/registry.cpython-38.pyc b/detectron2/modeling/text_encoder/__pycache__/registry.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..84aee5e3d73b586cab182622fa4ace4d72681228 Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/registry.cpython-38.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/registry.cpython-39.pyc b/detectron2/modeling/text_encoder/__pycache__/registry.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..acc6e946d873982b3e4955eb74f4e8dae82200b3 Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/registry.cpython-39.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/transformer.cpython-38.pyc b/detectron2/modeling/text_encoder/__pycache__/transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6261fa5e84f31b3130b6720ceead41e07b3db058 Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/transformer.cpython-38.pyc differ diff --git a/detectron2/modeling/text_encoder/__pycache__/transformer.cpython-39.pyc b/detectron2/modeling/text_encoder/__pycache__/transformer.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c734b6a47013e5df44d09e68130e9d7fb3de9568 Binary files /dev/null and b/detectron2/modeling/text_encoder/__pycache__/transformer.cpython-39.pyc differ diff --git a/detectron2/modeling/text_encoder/build.py b/detectron2/modeling/text_encoder/build.py new file mode 100755 index 0000000000000000000000000000000000000000..21717b73146f2be5fa823e5bd8f4dd0b144d188c --- /dev/null +++ b/detectron2/modeling/text_encoder/build.py @@ -0,0 +1,31 @@ +import os + +from transformers import CLIPTokenizer +from transformers import AutoTokenizer + +from .registry import lang_encoders +from .registry import is_lang_encoder + + +def build_lang_encoder(config_encoder, tokenizer, verbose, **kwargs): + model_name = config_encoder['NAME'] + + if not is_lang_encoder(model_name): + raise ValueError(f'Unknown model: {model_name}') + + return lang_encoders(model_name)(config_encoder, tokenizer, verbose, **kwargs) + + +def build_tokenizer(config_encoder): + tokenizer = None + os.environ['TOKENIZERS_PARALLELISM'] = 'true' + if config_encoder['TOKENIZER'] == 'clip': + pretrained_tokenizer = config_encoder.get( + 'PRETRAINED_TOKENIZER', 'openai/clip-vit-base-patch32' + ) + tokenizer = CLIPTokenizer.from_pretrained(pretrained_tokenizer) + tokenizer.add_special_tokens({'cls_token': tokenizer.eos_token}) + else: + tokenizer = AutoTokenizer.from_pretrained(config_encoder['TOKENIZER']) + + return tokenizer diff --git a/detectron2/modeling/text_encoder/hf_model.py b/detectron2/modeling/text_encoder/hf_model.py new file mode 100644 index 0000000000000000000000000000000000000000..588aa4c5f2408905d3d080d8679abc47d8d22c25 --- /dev/null +++ b/detectron2/modeling/text_encoder/hf_model.py @@ -0,0 +1,27 @@ +import logging + +from transformers import AutoConfig +from transformers import AutoModel + +from .registry import register_lang_encoder + +logger = logging.getLogger(__name__) + + +@register_lang_encoder +def lang_encoder(config_encoder, tokenizer, verbose, **kwargs): + + hf_model = None + if config_encoder['LOAD_PRETRAINED']: + hf_model = AutoModel.from_pretrained(config_encoder['HF_MODEL']) + else: + hf_config = AutoConfig.from_pretrained(config_encoder['HF_MODEL']) + + if 'CONFIG_OVERRIDE' in config_encoder: + logger.warning(f'Override config: {config_encoder["CONFIG_OVERRIDE"]}') + hf_config.update(config_encoder['CONFIG_OVERRIDE']) + + logger.info(f'HF model config: {hf_config}') + hf_model = AutoModel.from_config(hf_config) + + return hf_model diff --git a/detectron2/modeling/text_encoder/registry.py b/detectron2/modeling/text_encoder/registry.py new file mode 100755 index 0000000000000000000000000000000000000000..8991272a6e2294ea86eee338cf61d87e4123f724 --- /dev/null +++ b/detectron2/modeling/text_encoder/registry.py @@ -0,0 +1,18 @@ +_lang_encoders = {} + + +def register_lang_encoder(fn): + module_name_split = fn.__module__.split('.') + model_name = module_name_split[-1] + + _lang_encoders[model_name] = fn + + return fn + + +def lang_encoders(model_name): + return _lang_encoders[model_name] + + +def is_lang_encoder(model_name): + return model_name in _lang_encoders diff --git a/detectron2/modeling/text_encoder/transformer.py b/detectron2/modeling/text_encoder/transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..25e00ca7a0b5899869857d1e1d4e8afc91665f8c --- /dev/null +++ b/detectron2/modeling/text_encoder/transformer.py @@ -0,0 +1,194 @@ +from collections import OrderedDict +from typing import Tuple, Union +import logging +import os + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from timm.models.layers import DropPath, trunc_normal_ + +from .registry import register_lang_encoder + +logger = logging.getLogger(__name__) + +class LayerNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-12): + """Construct a layernorm module in the TF style (epsilon inside the square root). + """ + super(LayerNorm, self).__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.bias = nn.Parameter(torch.zeros(hidden_size)) + self.variance_epsilon = eps + + def forward(self, x): + pdtype = x.dtype + x = x.float() + u = x.mean(-1, keepdim=True) + s = (x - u).pow(2).mean(-1, keepdim=True) + x = (x - u) / torch.sqrt(s + self.variance_epsilon) + return self.weight * x.to(pdtype) + self.bias + + +class QuickGELU(nn.Module): + def forward(self, x: torch.Tensor): + return x * torch.sigmoid(1.702 * x) + + +class ResidualAttentionBlock(nn.Module): + def __init__(self, + d_model: int, + n_head: int, + attn_mask: torch.Tensor = None, + drop_path: float = 0.0): + super().__init__() + + self.attn = nn.MultiheadAttention(d_model, n_head) + self.ln_1 = LayerNorm(d_model) + self.mlp = nn.Sequential(OrderedDict([ + ("c_fc", nn.Linear(d_model, d_model * 4)), + ("gelu", QuickGELU()), + ("c_proj", nn.Linear(d_model * 4, d_model)) + ])) + self.ln_2 = LayerNorm(d_model) + self.attn_mask = attn_mask + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + + def attention(self, x: torch.Tensor, key_padding_mask: torch.Tensor = None): + self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) \ + if self.attn_mask is not None else None + + + return self.attn( + x, x, x, + key_padding_mask=key_padding_mask, + need_weights=False, + attn_mask=self.attn_mask + )[0] + + def forward(self, x: torch.Tensor, key_padding_mask: torch.Tensor = None): + x = x + self.drop_path(self.attention(self.ln_1(x), key_padding_mask=key_padding_mask)) + x = x + self.drop_path(self.mlp(self.ln_2(x))) + return x + + +class Transformer(nn.Module): + def __init__(self, + context_length: int, + vocab_size: int, + width: int, + layers: int, + heads: int, + drop_path: float = 0.0, + autogressive: bool =True): + super().__init__() + + self.token_embedding = nn.Embedding(vocab_size, width) + + self.context_length = context_length + self.positional_embedding = nn.Parameter( + torch.empty(self.context_length, width) + ) + + self.width = width + self.layers = layers + self.autogressive = autogressive + attn_mask = self.build_attention_mask() if autogressive else None + dpr = [x.item() for x in torch.linspace(0, drop_path, layers)] # stochastic depth decay rule + self.resblocks = nn.ModuleList( + [ + ResidualAttentionBlock(width, heads, attn_mask, dpr[i]) + for i in range(layers) + ] + ) + + self.ln_final = LayerNorm(width) + + trunc_normal_(self.positional_embedding, std=.02) + # nn.init.normal_(self.token_embedding, std=.02) + trunc_normal_(self.token_embedding.weight, std=.02) + self.apply(self._init_weights) + + @property + def dim_out(self): + return self.width + + def build_attention_mask(self): + # lazily create causal attention mask, with full attention between the vision tokens + # pytorch uses additive attention mask; fill with -inf + mask = torch.empty(self.context_length, self.context_length) + mask.fill_(float("-inf")) + mask.triu_(1) # zero out the lower diagonal + return mask + + def _init_weights(self, m): + if isinstance(m, (nn.Linear, nn.Conv2d)): + logger.info('=> init weight of Linear/Conv2d from trunc norm') + trunc_normal_(m.weight, std=0.02) + if m.bias is not None: + logger.info('=> init bias of Linear/Conv2d to zeros') + nn.init.constant_(m.bias, 0) + elif isinstance(m, (nn.LayerNorm, nn.BatchNorm2d)): + nn.init.constant_(m.bias, 0) + + def load_pretrained(self, pretrained='', pretrained_layers=[], verbose=True): + if os.path.isfile(pretrained): + pretrained_dict = torch.load(pretrained, map_location='cpu') + logging.info(f'=> loading pretrained model {pretrained}') + model_dict = self.state_dict() + pretrained_dict = { + k: v for k, v in pretrained_dict.items() + if k in model_dict.keys() + } + need_init_state_dict = {} + for k, v in pretrained_dict.items(): + need_init = ( + k.split('.')[0] in pretrained_layers + or pretrained_layers[0] == '*' + ) + if need_init: + if verbose: + logging.info(f'=> init {k} from {pretrained}') + + need_init_state_dict[k] = v + self.load_state_dict(need_init_state_dict, strict=False) + + + @torch.jit.ignore + def no_weight_decay(self): + return { + 'positional_embedding', + 'token_embedding', + } + + def forward(self, input_ids, attention_mask=None): + key_padding_mask = (input_ids == 0) if not self.autogressive else None + x = self.token_embedding(input_ids) # [batch_size, n_ctx, d_model] + x = x + self.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + for block in self.resblocks: + x = block(x, key_padding_mask) + x = x.permute(1, 0, 2) # LND -> NLD + + x = self.ln_final(x) + + return {'last_hidden_state': x} + + +@register_lang_encoder +def lang_encoder(config_encoder, tokenizer, verbose, **kwargs): + transformer = Transformer( + context_length=config_encoder['CONTEXT_LENGTH'], + vocab_size=tokenizer.vocab_size, + width=config_encoder['WIDTH'], + layers=config_encoder['LAYERS'], + heads=config_encoder['HEADS'], + autogressive=config_encoder.get('AUTOGRESSIVE', True) + ) + + if config_encoder['LOAD_PRETRAINED']: + transformer.load_pretrained() + + return transformer diff --git a/detectron2/projects/README.md b/detectron2/projects/README.md new file mode 100644 index 0000000000000000000000000000000000000000..95afe7ff8c8a9bd2f56621fcc3c1bdac11c256a9 --- /dev/null +++ b/detectron2/projects/README.md @@ -0,0 +1,2 @@ + +Projects live in the [`projects` directory](../../projects) under the root of this repository, but not here. diff --git a/detectron2/projects/__init__.py b/detectron2/projects/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a68207db4ee3c2578e1042b00b3071a946b7adea --- /dev/null +++ b/detectron2/projects/__init__.py @@ -0,0 +1,31 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import importlib +from pathlib import Path + +_PROJECTS = { + "point_rend": "PointRend", + "deeplab": "DeepLab", + "panoptic_deeplab": "Panoptic-DeepLab", +} +_PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent / "projects" + +if _PROJECT_ROOT.is_dir(): + # This is true only for in-place installation (pip install -e, setup.py develop), + # where setup(package_dir=) does not work: https://github.com/pypa/setuptools/issues/230 + + class _D2ProjectsFinder(importlib.abc.MetaPathFinder): + def find_spec(self, name, path, target=None): + if not name.startswith("detectron2.projects."): + return + project_name = name.split(".")[-1] + project_dir = _PROJECTS.get(project_name) + if not project_dir: + return + target_file = _PROJECT_ROOT / f"{project_dir}/{project_name}/__init__.py" + if not target_file.is_file(): + return + return importlib.util.spec_from_file_location(name, target_file) + + import sys + + sys.meta_path.append(_D2ProjectsFinder()) diff --git a/detectron2/solver/__init__.py b/detectron2/solver/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..9a2dbd35bb24f0d4a979bc8f304142376d87e7ec --- /dev/null +++ b/detectron2/solver/__init__.py @@ -0,0 +1,5 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .build import build_lr_scheduler, build_optimizer, get_default_optimizer_params +from .lr_scheduler import WarmupCosineLR, WarmupMultiStepLR, LRMultiplier, WarmupParamScheduler + +__all__ = [k for k in globals().keys() if not k.startswith("_")] diff --git a/detectron2/solver/__pycache__/__init__.cpython-39.pyc b/detectron2/solver/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ad859888bbb7340d5fb13eb6c13e3b34c0298a9b Binary files /dev/null and b/detectron2/solver/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/solver/__pycache__/build.cpython-39.pyc b/detectron2/solver/__pycache__/build.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..806c0d0a01ab0f4b8035e57923eda9bf17afccea Binary files /dev/null and b/detectron2/solver/__pycache__/build.cpython-39.pyc differ diff --git a/detectron2/solver/__pycache__/lr_scheduler.cpython-39.pyc b/detectron2/solver/__pycache__/lr_scheduler.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e3f9828303383b21647572e1167c4ae807404f96 Binary files /dev/null and b/detectron2/solver/__pycache__/lr_scheduler.cpython-39.pyc differ diff --git a/detectron2/solver/build.py b/detectron2/solver/build.py new file mode 100644 index 0000000000000000000000000000000000000000..c01c0e0c0f88089dbd45eb06b0bdbe4f37343fbe --- /dev/null +++ b/detectron2/solver/build.py @@ -0,0 +1,252 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import itertools +import logging +from enum import Enum +from typing import Any, Callable, Dict, Iterable, List, Optional, Set, Type, Union +import torch +from fvcore.common.param_scheduler import CosineParamScheduler, MultiStepParamScheduler + +from detectron2.config import CfgNode + +from .lr_scheduler import LRMultiplier, WarmupParamScheduler + +_GradientClipperInput = Union[torch.Tensor, Iterable[torch.Tensor]] +_GradientClipper = Callable[[_GradientClipperInput], None] + + +class GradientClipType(Enum): + VALUE = "value" + NORM = "norm" + + +def _create_gradient_clipper(cfg: CfgNode) -> _GradientClipper: + """ + Creates gradient clipping closure to clip by value or by norm, + according to the provided config. + """ + cfg = copy.deepcopy(cfg) + + def clip_grad_norm(p: _GradientClipperInput): + torch.nn.utils.clip_grad_norm_(p, cfg.CLIP_VALUE, cfg.NORM_TYPE) + + def clip_grad_value(p: _GradientClipperInput): + torch.nn.utils.clip_grad_value_(p, cfg.CLIP_VALUE) + + _GRADIENT_CLIP_TYPE_TO_CLIPPER = { + GradientClipType.VALUE: clip_grad_value, + GradientClipType.NORM: clip_grad_norm, + } + return _GRADIENT_CLIP_TYPE_TO_CLIPPER[GradientClipType(cfg.CLIP_TYPE)] + + +def _generate_optimizer_class_with_gradient_clipping( + optimizer: Type[torch.optim.Optimizer], + *, + per_param_clipper: Optional[_GradientClipper] = None, + global_clipper: Optional[_GradientClipper] = None, +) -> Type[torch.optim.Optimizer]: + """ + Dynamically creates a new type that inherits the type of a given instance + and overrides the `step` method to add gradient clipping + """ + assert ( + per_param_clipper is None or global_clipper is None + ), "Not allowed to use both per-parameter clipping and global clipping" + + def optimizer_wgc_step(self, closure=None): + if per_param_clipper is not None: + for group in self.param_groups: + for p in group["params"]: + per_param_clipper(p) + else: + # global clipper for future use with detr + # (https://github.com/facebookresearch/detr/pull/287) + all_params = itertools.chain(*[g["params"] for g in self.param_groups]) + global_clipper(all_params) + super(type(self), self).step(closure) + + OptimizerWithGradientClip = type( + optimizer.__name__ + "WithGradientClip", + (optimizer,), + {"step": optimizer_wgc_step}, + ) + return OptimizerWithGradientClip + + +def maybe_add_gradient_clipping( + cfg: CfgNode, optimizer: Type[torch.optim.Optimizer] +) -> Type[torch.optim.Optimizer]: + """ + If gradient clipping is enabled through config options, wraps the existing + optimizer type to become a new dynamically created class OptimizerWithGradientClip + that inherits the given optimizer and overrides the `step` method to + include gradient clipping. + + Args: + cfg: CfgNode, configuration options + optimizer: type. A subclass of torch.optim.Optimizer + + Return: + type: either the input `optimizer` (if gradient clipping is disabled), or + a subclass of it with gradient clipping included in the `step` method. + """ + if not cfg.SOLVER.CLIP_GRADIENTS.ENABLED: + return optimizer + if isinstance(optimizer, torch.optim.Optimizer): + optimizer_type = type(optimizer) + else: + assert issubclass(optimizer, torch.optim.Optimizer), optimizer + optimizer_type = optimizer + + grad_clipper = _create_gradient_clipper(cfg.SOLVER.CLIP_GRADIENTS) + OptimizerWithGradientClip = _generate_optimizer_class_with_gradient_clipping( + optimizer_type, per_param_clipper=grad_clipper + ) + if isinstance(optimizer, torch.optim.Optimizer): + optimizer.__class__ = OptimizerWithGradientClip # a bit hacky, not recommended + return optimizer + else: + return OptimizerWithGradientClip + + +def build_optimizer(cfg: CfgNode, model: torch.nn.Module) -> torch.optim.Optimizer: + """ + Build an optimizer from config. + """ + params = get_default_optimizer_params( + model, + base_lr=cfg.SOLVER.BASE_LR, + weight_decay_norm=cfg.SOLVER.WEIGHT_DECAY_NORM, + bias_lr_factor=cfg.SOLVER.BIAS_LR_FACTOR, + weight_decay_bias=cfg.SOLVER.WEIGHT_DECAY_BIAS, + ) + return maybe_add_gradient_clipping(cfg, torch.optim.SGD)( + params, + lr=cfg.SOLVER.BASE_LR, + momentum=cfg.SOLVER.MOMENTUM, + nesterov=cfg.SOLVER.NESTEROV, + weight_decay=cfg.SOLVER.WEIGHT_DECAY, + ) + + +def get_default_optimizer_params( + model: torch.nn.Module, + base_lr: Optional[float] = None, + weight_decay: Optional[float] = None, + weight_decay_norm: Optional[float] = None, + bias_lr_factor: Optional[float] = 1.0, + weight_decay_bias: Optional[float] = None, + overrides: Optional[Dict[str, Dict[str, float]]] = None, +): + """ + Get default param list for optimizer, with support for a few types of + overrides. If no overrides needed, this is equivalent to `model.parameters()`. + + Args: + base_lr: lr for every group by default. Can be omitted to use the one in optimizer. + weight_decay: weight decay for every group by default. Can be omitted to use the one + in optimizer. + weight_decay_norm: override weight decay for params in normalization layers + bias_lr_factor: multiplier of lr for bias parameters. + weight_decay_bias: override weight decay for bias parameters + overrides: if not `None`, provides values for optimizer hyperparameters + (LR, weight decay) for module parameters with a given name; e.g. + ``{"embedding": {"lr": 0.01, "weight_decay": 0.1}}`` will set the LR and + weight decay values for all module parameters named `embedding`. + + For common detection models, ``weight_decay_norm`` is the only option + needed to be set. ``bias_lr_factor,weight_decay_bias`` are legacy settings + from Detectron1 that are not found useful. + + Example: + :: + torch.optim.SGD(get_default_optimizer_params(model, weight_decay_norm=0), + lr=0.01, weight_decay=1e-4, momentum=0.9) + """ + if overrides is None: + overrides = {} + defaults = {} + if base_lr is not None: + defaults["lr"] = base_lr + if weight_decay is not None: + defaults["weight_decay"] = weight_decay + bias_overrides = {} + if bias_lr_factor is not None and bias_lr_factor != 1.0: + # NOTE: unlike Detectron v1, we now by default make bias hyperparameters + # exactly the same as regular weights. + if base_lr is None: + raise ValueError("bias_lr_factor requires base_lr") + bias_overrides["lr"] = base_lr * bias_lr_factor + if weight_decay_bias is not None: + bias_overrides["weight_decay"] = weight_decay_bias + if len(bias_overrides): + if "bias" in overrides: + raise ValueError("Conflicting overrides for 'bias'") + overrides["bias"] = bias_overrides + + norm_module_types = ( + torch.nn.BatchNorm1d, + torch.nn.BatchNorm2d, + torch.nn.BatchNorm3d, + torch.nn.SyncBatchNorm, + # NaiveSyncBatchNorm inherits from BatchNorm2d + torch.nn.GroupNorm, + torch.nn.InstanceNorm1d, + torch.nn.InstanceNorm2d, + torch.nn.InstanceNorm3d, + torch.nn.LayerNorm, + torch.nn.LocalResponseNorm, + ) + params: List[Dict[str, Any]] = [] + memo: Set[torch.nn.parameter.Parameter] = set() + for module in model.modules(): + for module_param_name, value in module.named_parameters(recurse=False): + if not value.requires_grad: + continue + # Avoid duplicating parameters + if value in memo: + continue + memo.add(value) + + hyperparams = copy.copy(defaults) + if isinstance(module, norm_module_types) and weight_decay_norm is not None: + hyperparams["weight_decay"] = weight_decay_norm + hyperparams.update(overrides.get(module_param_name, {})) + params.append({"params": [value], **hyperparams}) + return params + + +def build_lr_scheduler( + cfg: CfgNode, optimizer: torch.optim.Optimizer +) -> torch.optim.lr_scheduler._LRScheduler: + """ + Build a LR scheduler from config. + """ + name = cfg.SOLVER.LR_SCHEDULER_NAME + + if name == "WarmupMultiStepLR": + steps = [x for x in cfg.SOLVER.STEPS if x <= cfg.SOLVER.MAX_ITER] + if len(steps) != len(cfg.SOLVER.STEPS): + logger = logging.getLogger(__name__) + logger.warning( + "SOLVER.STEPS contains values larger than SOLVER.MAX_ITER. " + "These values will be ignored." + ) + sched = MultiStepParamScheduler( + values=[cfg.SOLVER.GAMMA ** k for k in range(len(steps) + 1)], + milestones=steps, + num_updates=cfg.SOLVER.MAX_ITER, + ) + elif name == "WarmupCosineLR": + sched = CosineParamScheduler(1, 0) + else: + raise ValueError("Unknown LR scheduler: {}".format(name)) + + sched = WarmupParamScheduler( + sched, + cfg.SOLVER.WARMUP_FACTOR, + min(cfg.SOLVER.WARMUP_ITERS / cfg.SOLVER.MAX_ITER, 1.0), + cfg.SOLVER.WARMUP_METHOD, + ) + return LRMultiplier(optimizer, multiplier=sched, max_iter=cfg.SOLVER.MAX_ITER) diff --git a/detectron2/solver/lr_scheduler.py b/detectron2/solver/lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..8803e87b9e60cffdbe048c97c282d353191ae4c8 --- /dev/null +++ b/detectron2/solver/lr_scheduler.py @@ -0,0 +1,238 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +import math +from bisect import bisect_right +from typing import List +import torch +from fvcore.common.param_scheduler import ( + CompositeParamScheduler, + ConstantParamScheduler, + LinearParamScheduler, + ParamScheduler, +) + +logger = logging.getLogger(__name__) + + +class WarmupParamScheduler(CompositeParamScheduler): + """ + Add an initial warmup stage to another scheduler. + """ + + def __init__( + self, + scheduler: ParamScheduler, + warmup_factor: float, + warmup_length: float, + warmup_method: str = "linear", + ): + """ + Args: + scheduler: warmup will be added at the beginning of this scheduler + warmup_factor: the factor w.r.t the initial value of ``scheduler``, e.g. 0.001 + warmup_length: the relative length (in [0, 1]) of warmup steps w.r.t the entire + training, e.g. 0.01 + warmup_method: one of "linear" or "constant" + """ + end_value = scheduler(warmup_length) # the value to reach when warmup ends + start_value = warmup_factor * scheduler(0.0) + if warmup_method == "constant": + warmup = ConstantParamScheduler(start_value) + elif warmup_method == "linear": + warmup = LinearParamScheduler(start_value, end_value) + else: + raise ValueError("Unknown warmup method: {}".format(warmup_method)) + super().__init__( + [warmup, scheduler], + interval_scaling=["rescaled", "fixed"], + lengths=[warmup_length, 1 - warmup_length], + ) + + +class LRMultiplier(torch.optim.lr_scheduler._LRScheduler): + """ + A LRScheduler which uses fvcore :class:`ParamScheduler` to multiply the + learning rate of each param in the optimizer. + Every step, the learning rate of each parameter becomes its initial value + multiplied by the output of the given :class:`ParamScheduler`. + + The absolute learning rate value of each parameter can be different. + This scheduler can be used as long as the relative scale among them do + not change during training. + + Examples: + :: + LRMultiplier( + opt, + WarmupParamScheduler( + MultiStepParamScheduler( + [1, 0.1, 0.01], + milestones=[60000, 80000], + num_updates=90000, + ), 0.001, 100 / 90000 + ), + max_iter=90000 + ) + """ + + # NOTES: in the most general case, every LR can use its own scheduler. + # Supporting this requires interaction with the optimizer when its parameter + # group is initialized. For example, classyvision implements its own optimizer + # that allows different schedulers for every parameter group. + # To avoid this complexity, we use this class to support the most common cases + # where the relative scale among all LRs stay unchanged during training. In this + # case we only need a total of one scheduler that defines the relative LR multiplier. + + def __init__( + self, + optimizer: torch.optim.Optimizer, + multiplier: ParamScheduler, + max_iter: int, + last_iter: int = -1, + ): + """ + Args: + optimizer, last_iter: See ``torch.optim.lr_scheduler._LRScheduler``. + ``last_iter`` is the same as ``last_epoch``. + multiplier: a fvcore ParamScheduler that defines the multiplier on + every LR of the optimizer + max_iter: the total number of training iterations + """ + if not isinstance(multiplier, ParamScheduler): + raise ValueError( + "_LRMultiplier(multiplier=) must be an instance of fvcore " + f"ParamScheduler. Got {multiplier} instead." + ) + self._multiplier = multiplier + self._max_iter = max_iter + super().__init__(optimizer, last_epoch=last_iter) + + def state_dict(self): + # fvcore schedulers are stateless. Only keep pytorch scheduler states + return {"base_lrs": self.base_lrs, "last_epoch": self.last_epoch} + + def get_lr(self) -> List[float]: + multiplier = self._multiplier(self.last_epoch / self._max_iter) + return [base_lr * multiplier for base_lr in self.base_lrs] + + +""" +Content below is no longer needed! +""" + +# NOTE: PyTorch's LR scheduler interface uses names that assume the LR changes +# only on epoch boundaries. We typically use iteration based schedules instead. +# As a result, "epoch" (e.g., as in self.last_epoch) should be understood to mean +# "iteration" instead. + +# FIXME: ideally this would be achieved with a CombinedLRScheduler, separating +# MultiStepLR with WarmupLR but the current LRScheduler design doesn't allow it. + + +class WarmupMultiStepLR(torch.optim.lr_scheduler._LRScheduler): + def __init__( + self, + optimizer: torch.optim.Optimizer, + milestones: List[int], + gamma: float = 0.1, + warmup_factor: float = 0.001, + warmup_iters: int = 1000, + warmup_method: str = "linear", + last_epoch: int = -1, + ): + logger.warning( + "WarmupMultiStepLR is deprecated! Use LRMultipilier with fvcore ParamScheduler instead!" + ) + if not list(milestones) == sorted(milestones): + raise ValueError( + "Milestones should be a list of" " increasing integers. Got {}", milestones + ) + self.milestones = milestones + self.gamma = gamma + self.warmup_factor = warmup_factor + self.warmup_iters = warmup_iters + self.warmup_method = warmup_method + super().__init__(optimizer, last_epoch) + + def get_lr(self) -> List[float]: + warmup_factor = _get_warmup_factor_at_iter( + self.warmup_method, self.last_epoch, self.warmup_iters, self.warmup_factor + ) + return [ + base_lr * warmup_factor * self.gamma ** bisect_right(self.milestones, self.last_epoch) + for base_lr in self.base_lrs + ] + + def _compute_values(self) -> List[float]: + # The new interface + return self.get_lr() + + +class WarmupCosineLR(torch.optim.lr_scheduler._LRScheduler): + def __init__( + self, + optimizer: torch.optim.Optimizer, + max_iters: int, + warmup_factor: float = 0.001, + warmup_iters: int = 1000, + warmup_method: str = "linear", + last_epoch: int = -1, + ): + logger.warning( + "WarmupCosineLR is deprecated! Use LRMultipilier with fvcore ParamScheduler instead!" + ) + self.max_iters = max_iters + self.warmup_factor = warmup_factor + self.warmup_iters = warmup_iters + self.warmup_method = warmup_method + super().__init__(optimizer, last_epoch) + + def get_lr(self) -> List[float]: + warmup_factor = _get_warmup_factor_at_iter( + self.warmup_method, self.last_epoch, self.warmup_iters, self.warmup_factor + ) + # Different definitions of half-cosine with warmup are possible. For + # simplicity we multiply the standard half-cosine schedule by the warmup + # factor. An alternative is to start the period of the cosine at warmup_iters + # instead of at 0. In the case that warmup_iters << max_iters the two are + # very close to each other. + return [ + base_lr + * warmup_factor + * 0.5 + * (1.0 + math.cos(math.pi * self.last_epoch / self.max_iters)) + for base_lr in self.base_lrs + ] + + def _compute_values(self) -> List[float]: + # The new interface + return self.get_lr() + + +def _get_warmup_factor_at_iter( + method: str, iter: int, warmup_iters: int, warmup_factor: float +) -> float: + """ + Return the learning rate warmup factor at a specific iteration. + See :paper:`ImageNet in 1h` for more details. + + Args: + method (str): warmup method; either "constant" or "linear". + iter (int): iteration at which to calculate the warmup factor. + warmup_iters (int): the number of warmup iterations. + warmup_factor (float): the base warmup factor (the meaning changes according + to the method used). + + Returns: + float: the effective warmup factor at the given iteration. + """ + if iter >= warmup_iters: + return 1.0 + + if method == "constant": + return warmup_factor + elif method == "linear": + alpha = iter / warmup_iters + return warmup_factor * (1 - alpha) + alpha + else: + raise ValueError("Unknown warmup method: {}".format(method)) diff --git a/detectron2/structures/__init__.py b/detectron2/structures/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..117e55bc05e6bb5cfb4d5f1178f3da4928d064af --- /dev/null +++ b/detectron2/structures/__init__.py @@ -0,0 +1,17 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from .boxes import Boxes, BoxMode, pairwise_iou, pairwise_ioa +from .image_list import ImageList + +from .instances import Instances +from .keypoints import Keypoints, heatmaps_to_keypoints +from .masks import BitMasks, PolygonMasks, polygons_to_bitmask, ROIMasks +from .rotated_boxes import RotatedBoxes +from .rotated_boxes import pairwise_iou as pairwise_iou_rotated + +__all__ = [k for k in globals().keys() if not k.startswith("_")] + + +from detectron2.utils.env import fixup_module_metadata + +fixup_module_metadata(__name__, globals(), __all__) +del fixup_module_metadata diff --git a/detectron2/structures/__pycache__/__init__.cpython-39.pyc b/detectron2/structures/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..39d22247fef96861591b19fd0137b00be579fd71 Binary files /dev/null and b/detectron2/structures/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/structures/__pycache__/boxes.cpython-39.pyc b/detectron2/structures/__pycache__/boxes.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..5a0bc8d9219b76781da6f371056a4af6538907f2 Binary files /dev/null and b/detectron2/structures/__pycache__/boxes.cpython-39.pyc differ diff --git a/detectron2/structures/__pycache__/image_list.cpython-39.pyc b/detectron2/structures/__pycache__/image_list.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..01fae675cc4f4c282ace2a1f7bedbe1f93b91dfe Binary files /dev/null and b/detectron2/structures/__pycache__/image_list.cpython-39.pyc differ diff --git a/detectron2/structures/__pycache__/instances.cpython-39.pyc b/detectron2/structures/__pycache__/instances.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7d424ff72b955bdc221055c3451e26cf554e8760 Binary files /dev/null and b/detectron2/structures/__pycache__/instances.cpython-39.pyc differ diff --git a/detectron2/structures/__pycache__/keypoints.cpython-39.pyc b/detectron2/structures/__pycache__/keypoints.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8548c2c1f52fb437ce10cb00637405fd75c209b7 Binary files /dev/null and b/detectron2/structures/__pycache__/keypoints.cpython-39.pyc differ diff --git a/detectron2/structures/__pycache__/masks.cpython-39.pyc b/detectron2/structures/__pycache__/masks.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8869b3572c11d85bb5b40ef2f1c5875b1ef61b8c Binary files /dev/null and b/detectron2/structures/__pycache__/masks.cpython-39.pyc differ diff --git a/detectron2/structures/__pycache__/rotated_boxes.cpython-39.pyc b/detectron2/structures/__pycache__/rotated_boxes.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..c9fb1ba110008808910e8a5d47b56ac3a163391b Binary files /dev/null and b/detectron2/structures/__pycache__/rotated_boxes.cpython-39.pyc differ diff --git a/detectron2/structures/__pycache__/tsv_file.cpython-39.pyc b/detectron2/structures/__pycache__/tsv_file.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b08b599f1469afc8c00c1920e12b8a78188bb0ac Binary files /dev/null and b/detectron2/structures/__pycache__/tsv_file.cpython-39.pyc differ diff --git a/detectron2/structures/boxes.py b/detectron2/structures/boxes.py new file mode 100644 index 0000000000000000000000000000000000000000..6d8762d60a873c6b6daa42e9e7fcac41eda32fec --- /dev/null +++ b/detectron2/structures/boxes.py @@ -0,0 +1,416 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import math +import numpy as np +from enum import IntEnum, unique +from typing import List, Tuple, Union +import torch +from torch import device + +from detectron2.utils.env import TORCH_VERSION + +_RawBoxType = Union[List[float], Tuple[float, ...], torch.Tensor, np.ndarray] + + +if TORCH_VERSION < (1, 8): + _maybe_jit_unused = torch.jit.unused +else: + + def _maybe_jit_unused(x): + return x + + +@unique +class BoxMode(IntEnum): + """ + Enum of different ways to represent a box. + """ + + XYXY_ABS = 0 + """ + (x0, y0, x1, y1) in absolute floating points coordinates. + The coordinates in range [0, width or height]. + """ + XYWH_ABS = 1 + """ + (x0, y0, w, h) in absolute floating points coordinates. + """ + XYXY_REL = 2 + """ + Not yet supported! + (x0, y0, x1, y1) in range [0, 1]. They are relative to the size of the image. + """ + XYWH_REL = 3 + """ + Not yet supported! + (x0, y0, w, h) in range [0, 1]. They are relative to the size of the image. + """ + XYWHA_ABS = 4 + """ + (xc, yc, w, h, a) in absolute floating points coordinates. + (xc, yc) is the center of the rotated box, and the angle a is in degrees ccw. + """ + + @staticmethod + def convert(box: _RawBoxType, from_mode: "BoxMode", to_mode: "BoxMode") -> _RawBoxType: + """ + Args: + box: can be a k-tuple, k-list or an Nxk array/tensor, where k = 4 or 5 + from_mode, to_mode (BoxMode) + + Returns: + The converted box of the same type. + """ + if from_mode == to_mode: + return box + + original_type = type(box) + is_numpy = isinstance(box, np.ndarray) + single_box = isinstance(box, (list, tuple)) + if single_box: + assert len(box) == 4 or len(box) == 5, ( + "BoxMode.convert takes either a k-tuple/list or an Nxk array/tensor," + " where k == 4 or 5" + ) + arr = torch.tensor(box)[None, :] + else: + # avoid modifying the input box + if is_numpy: + arr = torch.from_numpy(np.asarray(box)).clone() + else: + arr = box.clone() + + assert to_mode not in [BoxMode.XYXY_REL, BoxMode.XYWH_REL] and from_mode not in [ + BoxMode.XYXY_REL, + BoxMode.XYWH_REL, + ], "Relative mode not yet supported!" + + if from_mode == BoxMode.XYWHA_ABS and to_mode == BoxMode.XYXY_ABS: + assert ( + arr.shape[-1] == 5 + ), "The last dimension of input shape must be 5 for XYWHA format" + original_dtype = arr.dtype + arr = arr.double() + + w = arr[:, 2] + h = arr[:, 3] + a = arr[:, 4] + c = torch.abs(torch.cos(a * math.pi / 180.0)) + s = torch.abs(torch.sin(a * math.pi / 180.0)) + # This basically computes the horizontal bounding rectangle of the rotated box + new_w = c * w + s * h + new_h = c * h + s * w + + # convert center to top-left corner + arr[:, 0] -= new_w / 2.0 + arr[:, 1] -= new_h / 2.0 + # bottom-right corner + arr[:, 2] = arr[:, 0] + new_w + arr[:, 3] = arr[:, 1] + new_h + + arr = arr[:, :4].to(dtype=original_dtype) + elif from_mode == BoxMode.XYWH_ABS and to_mode == BoxMode.XYWHA_ABS: + original_dtype = arr.dtype + arr = arr.double() + arr[:, 0] += arr[:, 2] / 2.0 + arr[:, 1] += arr[:, 3] / 2.0 + angles = torch.zeros((arr.shape[0], 1), dtype=arr.dtype) + arr = torch.cat((arr, angles), axis=1).to(dtype=original_dtype) + else: + if to_mode == BoxMode.XYXY_ABS and from_mode == BoxMode.XYWH_ABS: + arr[:, 2] += arr[:, 0] + arr[:, 3] += arr[:, 1] + elif from_mode == BoxMode.XYXY_ABS and to_mode == BoxMode.XYWH_ABS: + arr[:, 2] -= arr[:, 0] + arr[:, 3] -= arr[:, 1] + else: + raise NotImplementedError( + "Conversion from BoxMode {} to {} is not supported yet".format( + from_mode, to_mode + ) + ) + + if single_box: + return original_type(arr.flatten().tolist()) + if is_numpy: + return arr.numpy() + else: + return arr + + +class Boxes: + """ + This structure stores a list of boxes as a Nx4 torch.Tensor. + It supports some common methods about boxes + (`area`, `clip`, `nonempty`, etc), + and also behaves like a Tensor + (support indexing, `to(device)`, `.device`, and iteration over all boxes) + + Attributes: + tensor (torch.Tensor): float matrix of Nx4. Each row is (x1, y1, x2, y2). + """ + + def __init__(self, tensor: torch.Tensor): + """ + Args: + tensor (Tensor[float]): a Nx4 matrix. Each row is (x1, y1, x2, y2). + """ + device = tensor.device if isinstance(tensor, torch.Tensor) else torch.device("cpu") + tensor = torch.as_tensor(tensor, dtype=torch.float32, device=device) + if tensor.numel() == 0: + # Use reshape, so we don't end up creating a new tensor that does not depend on + # the inputs (and consequently confuses jit) + tensor = tensor.reshape((-1, 4)).to(dtype=torch.float32, device=device) + assert tensor.dim() == 2 and tensor.size(-1) == 4, tensor.size() + + self.tensor = tensor + + def clone(self) -> "Boxes": + """ + Clone the Boxes. + + Returns: + Boxes + """ + return Boxes(self.tensor.clone()) + + @_maybe_jit_unused + def to(self, device: torch.device): + # Boxes are assumed float32 and does not support to(dtype) + return Boxes(self.tensor.to(device=device)) + + def area(self) -> torch.Tensor: + """ + Computes the area of all the boxes. + + Returns: + torch.Tensor: a vector with areas of each box. + """ + box = self.tensor + area = (box[:, 2] - box[:, 0]) * (box[:, 3] - box[:, 1]) + return area + + def clip(self, box_size: Tuple[int, int]) -> None: + """ + Clip (in place) the boxes by limiting x coordinates to the range [0, width] + and y coordinates to the range [0, height]. + + Args: + box_size (height, width): The clipping box's size. + """ + assert torch.isfinite(self.tensor).all(), "Box tensor contains infinite or NaN!" + h, w = box_size + x1 = self.tensor[:, 0].clamp(min=0, max=w) + y1 = self.tensor[:, 1].clamp(min=0, max=h) + x2 = self.tensor[:, 2].clamp(min=0, max=w) + y2 = self.tensor[:, 3].clamp(min=0, max=h) + self.tensor = torch.stack((x1, y1, x2, y2), dim=-1) + + def nonempty(self, threshold: float = 0.0) -> torch.Tensor: + """ + Find boxes that are non-empty. + A box is considered empty, if either of its side is no larger than threshold. + + Returns: + Tensor: + a binary vector which represents whether each box is empty + (False) or non-empty (True). + """ + box = self.tensor + widths = box[:, 2] - box[:, 0] + heights = box[:, 3] - box[:, 1] + keep = (widths > threshold) & (heights > threshold) + return keep + + def __getitem__(self, item) -> "Boxes": + """ + Args: + item: int, slice, or a BoolTensor + + Returns: + Boxes: Create a new :class:`Boxes` by indexing. + + The following usage are allowed: + + 1. `new_boxes = boxes[3]`: return a `Boxes` which contains only one box. + 2. `new_boxes = boxes[2:10]`: return a slice of boxes. + 3. `new_boxes = boxes[vector]`, where vector is a torch.BoolTensor + with `length = len(boxes)`. Nonzero elements in the vector will be selected. + + Note that the returned Boxes might share storage with this Boxes, + subject to Pytorch's indexing semantics. + """ + if isinstance(item, int): + return Boxes(self.tensor[item].view(1, -1)) + b = self.tensor[item] + assert b.dim() == 2, "Indexing on Boxes with {} failed to return a matrix!".format(item) + return Boxes(b) + + def __len__(self) -> int: + return self.tensor.shape[0] + + def __repr__(self) -> str: + return "Boxes(" + str(self.tensor) + ")" + + def inside_box(self, box_size: Tuple[int, int], boundary_threshold: int = 0) -> torch.Tensor: + """ + Args: + box_size (height, width): Size of the reference box. + boundary_threshold (int): Boxes that extend beyond the reference box + boundary by more than boundary_threshold are considered "outside". + + Returns: + a binary vector, indicating whether each box is inside the reference box. + """ + height, width = box_size + inds_inside = ( + (self.tensor[..., 0] >= -boundary_threshold) + & (self.tensor[..., 1] >= -boundary_threshold) + & (self.tensor[..., 2] < width + boundary_threshold) + & (self.tensor[..., 3] < height + boundary_threshold) + ) + return inds_inside + + def get_centers(self) -> torch.Tensor: + """ + Returns: + The box centers in a Nx2 array of (x, y). + """ + return (self.tensor[:, :2] + self.tensor[:, 2:]) / 2 + + def scale(self, scale_x: float, scale_y: float) -> None: + """ + Scale the box with horizontal and vertical scaling factors + """ + self.tensor[:, 0::2] *= scale_x + self.tensor[:, 1::2] *= scale_y + + @classmethod + @_maybe_jit_unused + def cat(cls, boxes_list: List["Boxes"]) -> "Boxes": + """ + Concatenates a list of Boxes into a single Boxes + + Arguments: + boxes_list (list[Boxes]) + + Returns: + Boxes: the concatenated Boxes + """ + assert isinstance(boxes_list, (list, tuple)) + if len(boxes_list) == 0: + return cls(torch.empty(0)) + assert all([isinstance(box, Boxes) for box in boxes_list]) + + # use torch.cat (v.s. layers.cat) so the returned boxes never share storage with input + cat_boxes = cls(torch.cat([b.tensor for b in boxes_list], dim=0)) + return cat_boxes + + @property + def device(self) -> device: + return self.tensor.device + + # type "Iterator[torch.Tensor]", yield, and iter() not supported by torchscript + # https://github.com/pytorch/pytorch/issues/18627 + @torch.jit.unused + def __iter__(self): + """ + Yield a box as a Tensor of shape (4,) at a time. + """ + yield from self.tensor + + +def pairwise_intersection(boxes1: Boxes, boxes2: Boxes) -> torch.Tensor: + """ + Given two lists of boxes of size N and M, + compute the intersection area between __all__ N x M pairs of boxes. + The box order must be (xmin, ymin, xmax, ymax) + + Args: + boxes1,boxes2 (Boxes): two `Boxes`. Contains N & M boxes, respectively. + + Returns: + Tensor: intersection, sized [N,M]. + """ + boxes1, boxes2 = boxes1.tensor, boxes2.tensor + width_height = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) - torch.max( + boxes1[:, None, :2], boxes2[:, :2] + ) # [N,M,2] + + width_height.clamp_(min=0) # [N,M,2] + intersection = width_height.prod(dim=2) # [N,M] + return intersection + + +# implementation from https://github.com/kuangliu/torchcv/blob/master/torchcv/utils/box.py +# with slight modifications +def pairwise_iou(boxes1: Boxes, boxes2: Boxes) -> torch.Tensor: + """ + Given two lists of boxes of size N and M, compute the IoU + (intersection over union) between **all** N x M pairs of boxes. + The box order must be (xmin, ymin, xmax, ymax). + + Args: + boxes1,boxes2 (Boxes): two `Boxes`. Contains N & M boxes, respectively. + + Returns: + Tensor: IoU, sized [N,M]. + """ + area1 = boxes1.area() # [N] + area2 = boxes2.area() # [M] + inter = pairwise_intersection(boxes1, boxes2) + + # handle empty boxes + iou = torch.where( + inter > 0, + inter / (area1[:, None] + area2 - inter), + torch.zeros(1, dtype=inter.dtype, device=inter.device), + ) + return iou + + +def pairwise_ioa(boxes1: Boxes, boxes2: Boxes) -> torch.Tensor: + """ + Similar to :func:`pariwise_iou` but compute the IoA (intersection over boxes2 area). + + Args: + boxes1,boxes2 (Boxes): two `Boxes`. Contains N & M boxes, respectively. + + Returns: + Tensor: IoA, sized [N,M]. + """ + area2 = boxes2.area() # [M] + inter = pairwise_intersection(boxes1, boxes2) + + # handle empty boxes + ioa = torch.where( + inter > 0, inter / area2, torch.zeros(1, dtype=inter.dtype, device=inter.device) + ) + return ioa + + +def matched_boxlist_iou(boxes1: Boxes, boxes2: Boxes) -> torch.Tensor: + """ + Compute pairwise intersection over union (IOU) of two sets of matched + boxes. The box order must be (xmin, ymin, xmax, ymax). + Similar to boxlist_iou, but computes only diagonal elements of the matrix + + Args: + boxes1: (Boxes) bounding boxes, sized [N,4]. + boxes2: (Boxes) bounding boxes, sized [N,4]. + Returns: + Tensor: iou, sized [N]. + """ + assert len(boxes1) == len( + boxes2 + ), "boxlists should have the same" "number of entries, got {}, {}".format( + len(boxes1), len(boxes2) + ) + area1 = boxes1.area() # [N] + area2 = boxes2.area() # [N] + box1, box2 = boxes1.tensor, boxes2.tensor + lt = torch.max(box1[:, :2], box2[:, :2]) # [N,2] + rb = torch.min(box1[:, 2:], box2[:, 2:]) # [N,2] + wh = (rb - lt).clamp(min=0) # [N,2] + inter = wh[:, 0] * wh[:, 1] # [N] + iou = inter / (area1 + area2 - inter) # [N] + return iou diff --git a/detectron2/structures/image_list.py b/detectron2/structures/image_list.py new file mode 100644 index 0000000000000000000000000000000000000000..26e6e49c55e27120ab26b6107cebb6c885f81c38 --- /dev/null +++ b/detectron2/structures/image_list.py @@ -0,0 +1,124 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from __future__ import division +from typing import Any, List, Tuple +import torch +from torch import device +from torch.nn import functional as F + +from detectron2.utils.env import TORCH_VERSION + + +def _as_tensor(x: Tuple[int, int]) -> torch.Tensor: + """ + An equivalent of `torch.as_tensor`, but works under tracing if input + is a list of tensor. `torch.as_tensor` will record a constant in tracing, + but this function will use `torch.stack` instead. + """ + if torch.jit.is_scripting(): + return torch.as_tensor(x) + if isinstance(x, (list, tuple)) and all([isinstance(t, torch.Tensor) for t in x]): + return torch.stack(x) + return torch.as_tensor(x) + + +class ImageList(object): + """ + Structure that holds a list of images (of possibly + varying sizes) as a single tensor. + This works by padding the images to the same size, + and storing in a field the original sizes of each image + + Attributes: + image_sizes (list[tuple[int, int]]): each tuple is (h, w). + During tracing, it becomes list[Tensor] instead. + """ + + def __init__(self, tensor: torch.Tensor, image_sizes: List[Tuple[int, int]]): + """ + Arguments: + tensor (Tensor): of shape (N, H, W) or (N, C_1, ..., C_K, H, W) where K >= 1 + image_sizes (list[tuple[int, int]]): Each tuple is (h, w). It can + be smaller than (H, W) due to padding. + """ + self.tensor = tensor + self.image_sizes = image_sizes + + def __len__(self) -> int: + return len(self.image_sizes) + + def __getitem__(self, idx) -> torch.Tensor: + """ + Access the individual image in its original size. + + Args: + idx: int or slice + + Returns: + Tensor: an image of shape (H, W) or (C_1, ..., C_K, H, W) where K >= 1 + """ + size = self.image_sizes[idx] + return self.tensor[idx, ..., : size[0], : size[1]] + + @torch.jit.unused + def to(self, *args: Any, **kwargs: Any) -> "ImageList": + cast_tensor = self.tensor.to(*args, **kwargs) + return ImageList(cast_tensor, self.image_sizes) + + @property + def device(self) -> device: + return self.tensor.device + + @staticmethod + def from_tensors( + tensors: List[torch.Tensor], size_divisibility: int = 0, pad_value: float = 0.0 + ) -> "ImageList": + """ + Args: + tensors: a tuple or list of `torch.Tensor`, each of shape (Hi, Wi) or + (C_1, ..., C_K, Hi, Wi) where K >= 1. The Tensors will be padded + to the same shape with `pad_value`. + size_divisibility (int): If `size_divisibility > 0`, add padding to ensure + the common height and width is divisible by `size_divisibility`. + This depends on the model and many models need a divisibility of 32. + pad_value (float): value to pad + + Returns: + an `ImageList`. + """ + assert len(tensors) > 0 + assert isinstance(tensors, (tuple, list)) + for t in tensors: + assert isinstance(t, torch.Tensor), type(t) + assert t.shape[:-2] == tensors[0].shape[:-2], t.shape + + image_sizes = [(im.shape[-2], im.shape[-1]) for im in tensors] + image_sizes_tensor = [_as_tensor(x) for x in image_sizes] + max_size = torch.stack(image_sizes_tensor).max(0).values + + if size_divisibility > 1: + stride = size_divisibility + # the last two dims are H,W, both subject to divisibility requirement + max_size = (max_size + (stride - 1)) // stride * stride + + # handle weirdness of scripting and tracing ... + if torch.jit.is_scripting(): + max_size: List[int] = max_size.to(dtype=torch.long).tolist() + else: + # https://github.com/pytorch/pytorch/issues/42448 + if TORCH_VERSION >= (1, 7) and torch.jit.is_tracing(): + image_sizes = image_sizes_tensor + + if len(tensors) == 1: + # This seems slightly (2%) faster. + # TODO: check whether it's faster for multiple images as well + image_size = image_sizes[0] + padding_size = [0, max_size[-1] - image_size[1], 0, max_size[-2] - image_size[0]] + batched_imgs = F.pad(tensors[0], padding_size, value=pad_value).unsqueeze_(0) + else: + # max_size can be a tensor in tracing mode, therefore convert to list + batch_shape = [len(tensors)] + list(tensors[0].shape[:-2]) + list(max_size) + batched_imgs = tensors[0].new_full(batch_shape, pad_value) + for img, pad_img in zip(tensors, batched_imgs): + pad_img[..., : img.shape[-2], : img.shape[-1]].copy_(img) + + return ImageList(batched_imgs.contiguous(), image_sizes) diff --git a/detectron2/structures/instances.py b/detectron2/structures/instances.py new file mode 100644 index 0000000000000000000000000000000000000000..e6bc832796b1a71dfa3ce6c06735ad02acb7a482 --- /dev/null +++ b/detectron2/structures/instances.py @@ -0,0 +1,191 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import itertools +from typing import Any, Dict, List, Tuple, Union +import torch + + +class Instances: + """ + This class represents a list of instances in an image. + It stores the attributes of instances (e.g., boxes, masks, labels, scores) as "fields". + All fields must have the same ``__len__`` which is the number of instances. + + All other (non-field) attributes of this class are considered private: + they must start with '_' and are not modifiable by a user. + + Some basic usage: + + 1. Set/get/check a field: + + .. code-block:: python + + instances.gt_boxes = Boxes(...) + print(instances.pred_masks) # a tensor of shape (N, H, W) + print('gt_masks' in instances) + + 2. ``len(instances)`` returns the number of instances + 3. Indexing: ``instances[indices]`` will apply the indexing on all the fields + and returns a new :class:`Instances`. + Typically, ``indices`` is a integer vector of indices, + or a binary mask of length ``num_instances`` + + .. code-block:: python + + category_3_detections = instances[instances.pred_classes == 3] + confident_detections = instances[instances.scores > 0.9] + """ + + def __init__(self, image_size: Tuple[int, int], **kwargs: Any): + """ + Args: + image_size (height, width): the spatial size of the image. + kwargs: fields to add to this `Instances`. + """ + self._image_size = image_size + self._fields: Dict[str, Any] = {} + for k, v in kwargs.items(): + self.set(k, v) + + @property + def image_size(self) -> Tuple[int, int]: + """ + Returns: + tuple: height, width + """ + return self._image_size + + def __setattr__(self, name: str, val: Any) -> None: + if name.startswith("_"): + super().__setattr__(name, val) + else: + self.set(name, val) + + def __getattr__(self, name: str) -> Any: + if name == "_fields" or name not in self._fields: + raise AttributeError("Cannot find field '{}' in the given Instances!".format(name)) + return self._fields[name] + + def set(self, name: str, value: Any) -> None: + """ + Set the field named `name` to `value`. + The length of `value` must be the number of instances, + and must agree with other existing fields in this object. + """ + data_len = len(value) + if len(self._fields): + assert ( + len(self) == data_len + ), "Adding a field of length {} to a Instances of length {}".format(data_len, len(self)) + self._fields[name] = value + + def has(self, name: str) -> bool: + """ + Returns: + bool: whether the field called `name` exists. + """ + return name in self._fields + + def remove(self, name: str) -> None: + """ + Remove the field called `name`. + """ + del self._fields[name] + + def get(self, name: str) -> Any: + """ + Returns the field called `name`. + """ + return self._fields[name] + + def get_fields(self) -> Dict[str, Any]: + """ + Returns: + dict: a dict which maps names (str) to data of the fields + + Modifying the returned dict will modify this instance. + """ + return self._fields + + # Tensor-like methods + def to(self, *args: Any, **kwargs: Any) -> "Instances": + """ + Returns: + Instances: all fields are called with a `to(device)`, if the field has this method. + """ + ret = Instances(self._image_size) + for k, v in self._fields.items(): + if hasattr(v, "to"): + v = v.to(*args, **kwargs) + ret.set(k, v) + return ret + + def __getitem__(self, item: Union[int, slice, torch.BoolTensor]) -> "Instances": + """ + Args: + item: an index-like object and will be used to index all the fields. + + Returns: + If `item` is a string, return the data in the corresponding field. + Otherwise, returns an `Instances` where all fields are indexed by `item`. + """ + if type(item) == int: + if item >= len(self) or item < -len(self): + raise IndexError("Instances index out of range!") + else: + item = slice(item, None, len(self)) + + ret = Instances(self._image_size) + for k, v in self._fields.items(): + ret.set(k, v[item]) + return ret + + def __len__(self) -> int: + for v in self._fields.values(): + # use __len__ because len() has to be int and is not friendly to tracing + return v.__len__() + raise NotImplementedError("Empty Instances does not support __len__!") + + def __iter__(self): + raise NotImplementedError("`Instances` object is not iterable!") + + @staticmethod + def cat(instance_lists: List["Instances"]) -> "Instances": + """ + Args: + instance_lists (list[Instances]) + + Returns: + Instances + """ + assert all(isinstance(i, Instances) for i in instance_lists) + assert len(instance_lists) > 0 + if len(instance_lists) == 1: + return instance_lists[0] + + image_size = instance_lists[0].image_size + for i in instance_lists[1:]: + assert i.image_size == image_size + ret = Instances(image_size) + for k in instance_lists[0]._fields.keys(): + values = [i.get(k) for i in instance_lists] + v0 = values[0] + if isinstance(v0, torch.Tensor): + values = torch.cat(values, dim=0) + elif isinstance(v0, list): + values = list(itertools.chain(*values)) + elif hasattr(type(v0), "cat"): + values = type(v0).cat(values) + else: + raise ValueError("Unsupported type {} for concatenation".format(type(v0))) + ret.set(k, values) + return ret + + def __str__(self) -> str: + s = self.__class__.__name__ + "(" + s += "num_instances={}, ".format(len(self)) + s += "image_height={}, ".format(self._image_size[0]) + s += "image_width={}, ".format(self._image_size[1]) + s += "fields=[{}])".format(", ".join((f"{k}: {v}" for k, v in self._fields.items()))) + return s + + __repr__ = __str__ diff --git a/detectron2/structures/keypoints.py b/detectron2/structures/keypoints.py new file mode 100644 index 0000000000000000000000000000000000000000..3d956a2d57e30be18ccef1fd3cf201d5ba3d8ab4 --- /dev/null +++ b/detectron2/structures/keypoints.py @@ -0,0 +1,230 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import numpy as np +from typing import Any, List, Tuple, Union +import torch +from torch.nn import functional as F + +from detectron2.utils.env import TORCH_VERSION + +if TORCH_VERSION < (1, 8): + + def script_if_tracing(fn): + return fn + + +else: + script_if_tracing = torch.jit.script_if_tracing + + +class Keypoints: + """ + Stores keypoint **annotation** data. GT Instances have a `gt_keypoints` property + containing the x,y location and visibility flag of each keypoint. This tensor has shape + (N, K, 3) where N is the number of instances and K is the number of keypoints per instance. + + The visibility flag follows the COCO format and must be one of three integers: + + * v=0: not labeled (in which case x=y=0) + * v=1: labeled but not visible + * v=2: labeled and visible + """ + + def __init__(self, keypoints: Union[torch.Tensor, np.ndarray, List[List[float]]]): + """ + Arguments: + keypoints: A Tensor, numpy array, or list of the x, y, and visibility of each keypoint. + The shape should be (N, K, 3) where N is the number of + instances, and K is the number of keypoints per instance. + """ + device = keypoints.device if isinstance(keypoints, torch.Tensor) else torch.device("cpu") + keypoints = torch.as_tensor(keypoints, dtype=torch.float32, device=device) + assert keypoints.dim() == 3 and keypoints.shape[2] == 3, keypoints.shape + self.tensor = keypoints + + def __len__(self) -> int: + return self.tensor.size(0) + + def to(self, *args: Any, **kwargs: Any) -> "Keypoints": + return type(self)(self.tensor.to(*args, **kwargs)) + + @property + def device(self) -> torch.device: + return self.tensor.device + + def to_heatmap(self, boxes: torch.Tensor, heatmap_size: int) -> torch.Tensor: + """ + Convert keypoint annotations to a heatmap of one-hot labels for training, + as described in :paper:`Mask R-CNN`. + + Arguments: + boxes: Nx4 tensor, the boxes to draw the keypoints to + + Returns: + heatmaps: + A tensor of shape (N, K), each element is integer spatial label + in the range [0, heatmap_size**2 - 1] for each keypoint in the input. + valid: + A tensor of shape (N, K) containing whether each keypoint is in the roi or not. + """ + return _keypoints_to_heatmap(self.tensor, boxes, heatmap_size) + + def __getitem__(self, item: Union[int, slice, torch.BoolTensor]) -> "Keypoints": + """ + Create a new `Keypoints` by indexing on this `Keypoints`. + + The following usage are allowed: + + 1. `new_kpts = kpts[3]`: return a `Keypoints` which contains only one instance. + 2. `new_kpts = kpts[2:10]`: return a slice of key points. + 3. `new_kpts = kpts[vector]`, where vector is a torch.ByteTensor + with `length = len(kpts)`. Nonzero elements in the vector will be selected. + + Note that the returned Keypoints might share storage with this Keypoints, + subject to Pytorch's indexing semantics. + """ + if isinstance(item, int): + return Keypoints([self.tensor[item]]) + return Keypoints(self.tensor[item]) + + def __repr__(self) -> str: + s = self.__class__.__name__ + "(" + s += "num_instances={})".format(len(self.tensor)) + return s + + +# TODO make this nicer, this is a direct translation from C2 (but removing the inner loop) +def _keypoints_to_heatmap( + keypoints: torch.Tensor, rois: torch.Tensor, heatmap_size: int +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Encode keypoint locations into a target heatmap for use in SoftmaxWithLoss across space. + + Maps keypoints from the half-open interval [x1, x2) on continuous image coordinates to the + closed interval [0, heatmap_size - 1] on discrete image coordinates. We use the + continuous-discrete conversion from Heckbert 1990 ("What is the coordinate of a pixel?"): + d = floor(c) and c = d + 0.5, where d is a discrete coordinate and c is a continuous coordinate. + + Arguments: + keypoints: tensor of keypoint locations in of shape (N, K, 3). + rois: Nx4 tensor of rois in xyxy format + heatmap_size: integer side length of square heatmap. + + Returns: + heatmaps: A tensor of shape (N, K) containing an integer spatial label + in the range [0, heatmap_size**2 - 1] for each keypoint in the input. + valid: A tensor of shape (N, K) containing whether each keypoint is in + the roi or not. + """ + + if rois.numel() == 0: + return rois.new().long(), rois.new().long() + offset_x = rois[:, 0] + offset_y = rois[:, 1] + scale_x = heatmap_size / (rois[:, 2] - rois[:, 0]) + scale_y = heatmap_size / (rois[:, 3] - rois[:, 1]) + + offset_x = offset_x[:, None] + offset_y = offset_y[:, None] + scale_x = scale_x[:, None] + scale_y = scale_y[:, None] + + x = keypoints[..., 0] + y = keypoints[..., 1] + + x_boundary_inds = x == rois[:, 2][:, None] + y_boundary_inds = y == rois[:, 3][:, None] + + x = (x - offset_x) * scale_x + x = x.floor().long() + y = (y - offset_y) * scale_y + y = y.floor().long() + + x[x_boundary_inds] = heatmap_size - 1 + y[y_boundary_inds] = heatmap_size - 1 + + valid_loc = (x >= 0) & (y >= 0) & (x < heatmap_size) & (y < heatmap_size) + vis = keypoints[..., 2] > 0 + valid = (valid_loc & vis).long() + + lin_ind = y * heatmap_size + x + heatmaps = lin_ind * valid + + return heatmaps, valid + + +@script_if_tracing +def heatmaps_to_keypoints(maps: torch.Tensor, rois: torch.Tensor) -> torch.Tensor: + """ + Extract predicted keypoint locations from heatmaps. + + Args: + maps (Tensor): (#ROIs, #keypoints, POOL_H, POOL_W). The predicted heatmap of logits for + each ROI and each keypoint. + rois (Tensor): (#ROIs, 4). The box of each ROI. + + Returns: + Tensor of shape (#ROIs, #keypoints, 4) with the last dimension corresponding to + (x, y, logit, score) for each keypoint. + + When converting discrete pixel indices in an NxN image to a continuous keypoint coordinate, + we maintain consistency with :meth:`Keypoints.to_heatmap` by using the conversion from + Heckbert 1990: c = d + 0.5, where d is a discrete coordinate and c is a continuous coordinate. + """ + # The decorator use of torch.no_grad() was not supported by torchscript. + # https://github.com/pytorch/pytorch/issues/44768 + maps = maps.detach() + rois = rois.detach() + + offset_x = rois[:, 0] + offset_y = rois[:, 1] + + widths = (rois[:, 2] - rois[:, 0]).clamp(min=1) + heights = (rois[:, 3] - rois[:, 1]).clamp(min=1) + widths_ceil = widths.ceil() + heights_ceil = heights.ceil() + + num_rois, num_keypoints = maps.shape[:2] + xy_preds = maps.new_zeros(rois.shape[0], num_keypoints, 4) + + width_corrections = widths / widths_ceil + height_corrections = heights / heights_ceil + + keypoints_idx = torch.arange(num_keypoints, device=maps.device) + + for i in range(num_rois): + outsize = (int(heights_ceil[i]), int(widths_ceil[i])) + roi_map = F.interpolate( + maps[[i]], size=outsize, mode="bicubic", align_corners=False + ).squeeze( + 0 + ) # #keypoints x H x W + + # softmax over the spatial region + max_score, _ = roi_map.view(num_keypoints, -1).max(1) + max_score = max_score.view(num_keypoints, 1, 1) + tmp_full_resolution = (roi_map - max_score).exp_() + tmp_pool_resolution = (maps[i] - max_score).exp_() + # Produce scores over the region H x W, but normalize with POOL_H x POOL_W, + # so that the scores of objects of different absolute sizes will be more comparable + roi_map_scores = tmp_full_resolution / tmp_pool_resolution.sum((1, 2), keepdim=True) + + w = roi_map.shape[2] + pos = roi_map.view(num_keypoints, -1).argmax(1) + + x_int = pos % w + y_int = (pos - x_int) // w + + assert ( + roi_map_scores[keypoints_idx, y_int, x_int] + == roi_map_scores.view(num_keypoints, -1).max(1)[0] + ).all() + + x = (x_int.float() + 0.5) * width_corrections[i] + y = (y_int.float() + 0.5) * height_corrections[i] + + xy_preds[i, :, 0] = x + offset_x[i] + xy_preds[i, :, 1] = y + offset_y[i] + xy_preds[i, :, 2] = roi_map[keypoints_idx, y_int, x_int] + xy_preds[i, :, 3] = roi_map_scores[keypoints_idx, y_int, x_int] + + return xy_preds diff --git a/detectron2/structures/masks.py b/detectron2/structures/masks.py new file mode 100644 index 0000000000000000000000000000000000000000..3513a38dbdf55b10d5107209a81d16c36975423e --- /dev/null +++ b/detectron2/structures/masks.py @@ -0,0 +1,527 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import itertools +import numpy as np +from typing import Any, Iterator, List, Union +import pycocotools.mask as mask_util +import torch +from torch import device + +from detectron2.layers.roi_align import ROIAlign +from detectron2.utils.memory import retry_if_cuda_oom + +from .boxes import Boxes + + +def polygon_area(x, y): + # Using the shoelace formula + # https://stackoverflow.com/questions/24467972/calculate-area-of-polygon-given-x-y-coordinates + return 0.5 * np.abs(np.dot(x, np.roll(y, 1)) - np.dot(y, np.roll(x, 1))) + + +def polygons_to_bitmask(polygons: List[np.ndarray], height: int, width: int) -> np.ndarray: + """ + Args: + polygons (list[ndarray]): each array has shape (Nx2,) + height, width (int) + + Returns: + ndarray: a bool mask of shape (height, width) + """ + assert len(polygons) > 0, "COCOAPI does not support empty polygons" + rles = mask_util.frPyObjects(polygons, height, width) + rle = mask_util.merge(rles) + return mask_util.decode(rle).astype(np.bool) + + +def rasterize_polygons_within_box( + polygons: List[np.ndarray], box: np.ndarray, mask_size: int +) -> torch.Tensor: + """ + Rasterize the polygons into a mask image and + crop the mask content in the given box. + The cropped mask is resized to (mask_size, mask_size). + + This function is used when generating training targets for mask head in Mask R-CNN. + Given original ground-truth masks for an image, new ground-truth mask + training targets in the size of `mask_size x mask_size` + must be provided for each predicted box. This function will be called to + produce such targets. + + Args: + polygons (list[ndarray[float]]): a list of polygons, which represents an instance. + box: 4-element numpy array + mask_size (int): + + Returns: + Tensor: BoolTensor of shape (mask_size, mask_size) + """ + # 1. Shift the polygons w.r.t the boxes + w, h = box[2] - box[0], box[3] - box[1] + + polygons = copy.deepcopy(polygons) + for p in polygons: + p[0::2] = p[0::2] - box[0] + p[1::2] = p[1::2] - box[1] + + # 2. Rescale the polygons to the new box size + # max() to avoid division by small number + ratio_h = mask_size / max(h, 0.1) + ratio_w = mask_size / max(w, 0.1) + + if ratio_h == ratio_w: + for p in polygons: + p *= ratio_h + else: + for p in polygons: + p[0::2] *= ratio_w + p[1::2] *= ratio_h + + # 3. Rasterize the polygons with coco api + mask = polygons_to_bitmask(polygons, mask_size, mask_size) + mask = torch.from_numpy(mask) + return mask + + +class BitMasks: + """ + This class stores the segmentation masks for all objects in one image, in + the form of bitmaps. + + Attributes: + tensor: bool Tensor of N,H,W, representing N instances in the image. + """ + + def __init__(self, tensor: Union[torch.Tensor, np.ndarray]): + """ + Args: + tensor: bool Tensor of N,H,W, representing N instances in the image. + """ + device = tensor.device if isinstance(tensor, torch.Tensor) else torch.device("cpu") + tensor = torch.as_tensor(tensor, dtype=torch.bool, device=device) + assert tensor.dim() == 3, tensor.size() + self.image_size = tensor.shape[1:] + self.tensor = tensor + + @torch.jit.unused + def to(self, *args: Any, **kwargs: Any) -> "BitMasks": + return BitMasks(self.tensor.to(*args, **kwargs)) + + @property + def device(self) -> torch.device: + return self.tensor.device + + @torch.jit.unused + def __getitem__(self, item: Union[int, slice, torch.BoolTensor]) -> "BitMasks": + """ + Returns: + BitMasks: Create a new :class:`BitMasks` by indexing. + + The following usage are allowed: + + 1. `new_masks = masks[3]`: return a `BitMasks` which contains only one mask. + 2. `new_masks = masks[2:10]`: return a slice of masks. + 3. `new_masks = masks[vector]`, where vector is a torch.BoolTensor + with `length = len(masks)`. Nonzero elements in the vector will be selected. + + Note that the returned object might share storage with this object, + subject to Pytorch's indexing semantics. + """ + if isinstance(item, int): + return BitMasks(self.tensor[item].view(1, -1)) + m = self.tensor[item] + assert m.dim() == 3, "Indexing on BitMasks with {} returns a tensor with shape {}!".format( + item, m.shape + ) + return BitMasks(m) + + @torch.jit.unused + def __iter__(self) -> torch.Tensor: + yield from self.tensor + + @torch.jit.unused + def __repr__(self) -> str: + s = self.__class__.__name__ + "(" + s += "num_instances={})".format(len(self.tensor)) + return s + + def __len__(self) -> int: + return self.tensor.shape[0] + + def nonempty(self) -> torch.Tensor: + """ + Find masks that are non-empty. + + Returns: + Tensor: a BoolTensor which represents + whether each mask is empty (False) or non-empty (True). + """ + return self.tensor.flatten(1).any(dim=1) + + @staticmethod + def from_polygon_masks( + polygon_masks: Union["PolygonMasks", List[List[np.ndarray]]], height: int, width: int + ) -> "BitMasks": + """ + Args: + polygon_masks (list[list[ndarray]] or PolygonMasks) + height, width (int) + """ + if isinstance(polygon_masks, PolygonMasks): + polygon_masks = polygon_masks.polygons + masks = [polygons_to_bitmask(p, height, width) for p in polygon_masks] + return BitMasks(torch.stack([torch.from_numpy(x) for x in masks])) + + @staticmethod + def from_roi_masks(roi_masks: "ROIMasks", height: int, width: int) -> "BitMasks": + """ + Args: + roi_masks: + height, width (int): + """ + return roi_masks.to_bitmasks(height, width) + + def crop_and_resize(self, boxes: torch.Tensor, mask_size: int) -> torch.Tensor: + """ + Crop each bitmask by the given box, and resize results to (mask_size, mask_size). + This can be used to prepare training targets for Mask R-CNN. + It has less reconstruction error compared to rasterization with polygons. + However we observe no difference in accuracy, + but BitMasks requires more memory to store all the masks. + + Args: + boxes (Tensor): Nx4 tensor storing the boxes for each mask + mask_size (int): the size of the rasterized mask. + + Returns: + Tensor: + A bool tensor of shape (N, mask_size, mask_size), where + N is the number of predicted boxes for this image. + """ + assert len(boxes) == len(self), "{} != {}".format(len(boxes), len(self)) + device = self.tensor.device + + batch_inds = torch.arange(len(boxes), device=device).to(dtype=boxes.dtype)[:, None] + rois = torch.cat([batch_inds, boxes], dim=1) # Nx5 + + bit_masks = self.tensor.to(dtype=torch.float32) + rois = rois.to(device=device) + output = ( + ROIAlign((mask_size, mask_size), 1.0, 0, aligned=True) + .forward(bit_masks[:, None, :, :], rois) + .squeeze(1) + ) + output = output >= 0.5 + return output + + def get_bounding_boxes(self) -> Boxes: + """ + Returns: + Boxes: tight bounding boxes around bitmasks. + If a mask is empty, it's bounding box will be all zero. + """ + boxes = torch.zeros(self.tensor.shape[0], 4, dtype=torch.float32) + x_any = torch.any(self.tensor, dim=1) + y_any = torch.any(self.tensor, dim=2) + for idx in range(self.tensor.shape[0]): + x = torch.where(x_any[idx, :])[0] + y = torch.where(y_any[idx, :])[0] + if len(x) > 0 and len(y) > 0: + boxes[idx, :] = torch.as_tensor( + [x[0], y[0], x[-1] + 1, y[-1] + 1], dtype=torch.float32 + ) + return Boxes(boxes) + + @staticmethod + def cat(bitmasks_list: List["BitMasks"]) -> "BitMasks": + """ + Concatenates a list of BitMasks into a single BitMasks + + Arguments: + bitmasks_list (list[BitMasks]) + + Returns: + BitMasks: the concatenated BitMasks + """ + assert isinstance(bitmasks_list, (list, tuple)) + assert len(bitmasks_list) > 0 + assert all(isinstance(bitmask, BitMasks) for bitmask in bitmasks_list) + + cat_bitmasks = type(bitmasks_list[0])(torch.cat([bm.tensor for bm in bitmasks_list], dim=0)) + return cat_bitmasks + + +class PolygonMasks: + """ + This class stores the segmentation masks for all objects in one image, in the form of polygons. + + Attributes: + polygons: list[list[ndarray]]. Each ndarray is a float64 vector representing a polygon. + """ + + def __init__(self, polygons: List[List[Union[torch.Tensor, np.ndarray]]]): + """ + Arguments: + polygons (list[list[np.ndarray]]): The first + level of the list correspond to individual instances, + the second level to all the polygons that compose the + instance, and the third level to the polygon coordinates. + The third level array should have the format of + [x0, y0, x1, y1, ..., xn, yn] (n >= 3). + """ + if not isinstance(polygons, list): + raise ValueError( + "Cannot create PolygonMasks: Expect a list of list of polygons per image. " + "Got '{}' instead.".format(type(polygons)) + ) + + def _make_array(t: Union[torch.Tensor, np.ndarray]) -> np.ndarray: + # Use float64 for higher precision, because why not? + # Always put polygons on CPU (self.to is a no-op) since they + # are supposed to be small tensors. + # May need to change this assumption if GPU placement becomes useful + if isinstance(t, torch.Tensor): + t = t.cpu().numpy() + return np.asarray(t).astype("float64") + + def process_polygons( + polygons_per_instance: List[Union[torch.Tensor, np.ndarray]] + ) -> List[np.ndarray]: + if not isinstance(polygons_per_instance, list): + raise ValueError( + "Cannot create polygons: Expect a list of polygons per instance. " + "Got '{}' instead.".format(type(polygons_per_instance)) + ) + # transform each polygon to a numpy array + polygons_per_instance = [_make_array(p) for p in polygons_per_instance] + for polygon in polygons_per_instance: + if len(polygon) % 2 != 0 or len(polygon) < 6: + raise ValueError(f"Cannot create a polygon from {len(polygon)} coordinates.") + return polygons_per_instance + + self.polygons: List[List[np.ndarray]] = [ + process_polygons(polygons_per_instance) for polygons_per_instance in polygons + ] + + def to(self, *args: Any, **kwargs: Any) -> "PolygonMasks": + return self + + @property + def device(self) -> torch.device: + return torch.device("cpu") + + def get_bounding_boxes(self) -> Boxes: + """ + Returns: + Boxes: tight bounding boxes around polygon masks. + """ + boxes = torch.zeros(len(self.polygons), 4, dtype=torch.float32) + for idx, polygons_per_instance in enumerate(self.polygons): + minxy = torch.as_tensor([float("inf"), float("inf")], dtype=torch.float32) + maxxy = torch.zeros(2, dtype=torch.float32) + for polygon in polygons_per_instance: + coords = torch.from_numpy(polygon).view(-1, 2).to(dtype=torch.float32) + minxy = torch.min(minxy, torch.min(coords, dim=0).values) + maxxy = torch.max(maxxy, torch.max(coords, dim=0).values) + boxes[idx, :2] = minxy + boxes[idx, 2:] = maxxy + return Boxes(boxes) + + def nonempty(self) -> torch.Tensor: + """ + Find masks that are non-empty. + + Returns: + Tensor: + a BoolTensor which represents whether each mask is empty (False) or not (True). + """ + keep = [1 if len(polygon) > 0 else 0 for polygon in self.polygons] + return torch.from_numpy(np.asarray(keep, dtype=np.bool)) + + def __getitem__(self, item: Union[int, slice, List[int], torch.BoolTensor]) -> "PolygonMasks": + """ + Support indexing over the instances and return a `PolygonMasks` object. + `item` can be: + + 1. An integer. It will return an object with only one instance. + 2. A slice. It will return an object with the selected instances. + 3. A list[int]. It will return an object with the selected instances, + correpsonding to the indices in the list. + 4. A vector mask of type BoolTensor, whose length is num_instances. + It will return an object with the instances whose mask is nonzero. + """ + if isinstance(item, int): + selected_polygons = [self.polygons[item]] + elif isinstance(item, slice): + selected_polygons = self.polygons[item] + elif isinstance(item, list): + selected_polygons = [self.polygons[i] for i in item] + elif isinstance(item, torch.Tensor): + # Polygons is a list, so we have to move the indices back to CPU. + if item.dtype == torch.bool: + assert item.dim() == 1, item.shape + item = item.nonzero().squeeze(1).cpu().numpy().tolist() + elif item.dtype in [torch.int32, torch.int64]: + item = item.cpu().numpy().tolist() + else: + raise ValueError("Unsupported tensor dtype={} for indexing!".format(item.dtype)) + selected_polygons = [self.polygons[i] for i in item] + return PolygonMasks(selected_polygons) + + def __iter__(self) -> Iterator[List[np.ndarray]]: + """ + Yields: + list[ndarray]: the polygons for one instance. + Each Tensor is a float64 vector representing a polygon. + """ + return iter(self.polygons) + + def __repr__(self) -> str: + s = self.__class__.__name__ + "(" + s += "num_instances={})".format(len(self.polygons)) + return s + + def __len__(self) -> int: + return len(self.polygons) + + def crop_and_resize(self, boxes: torch.Tensor, mask_size: int) -> torch.Tensor: + """ + Crop each mask by the given box, and resize results to (mask_size, mask_size). + This can be used to prepare training targets for Mask R-CNN. + + Args: + boxes (Tensor): Nx4 tensor storing the boxes for each mask + mask_size (int): the size of the rasterized mask. + + Returns: + Tensor: A bool tensor of shape (N, mask_size, mask_size), where + N is the number of predicted boxes for this image. + """ + assert len(boxes) == len(self), "{} != {}".format(len(boxes), len(self)) + + device = boxes.device + # Put boxes on the CPU, as the polygon representation is not efficient GPU-wise + # (several small tensors for representing a single instance mask) + boxes = boxes.to(torch.device("cpu")) + + results = [ + rasterize_polygons_within_box(poly, box.numpy(), mask_size) + for poly, box in zip(self.polygons, boxes) + ] + """ + poly: list[list[float]], the polygons for one instance + box: a tensor of shape (4,) + """ + if len(results) == 0: + return torch.empty(0, mask_size, mask_size, dtype=torch.bool, device=device) + return torch.stack(results, dim=0).to(device=device) + + def area(self): + """ + Computes area of the mask. + Only works with Polygons, using the shoelace formula: + https://stackoverflow.com/questions/24467972/calculate-area-of-polygon-given-x-y-coordinates + + Returns: + Tensor: a vector, area for each instance + """ + + area = [] + for polygons_per_instance in self.polygons: + area_per_instance = 0 + for p in polygons_per_instance: + area_per_instance += polygon_area(p[0::2], p[1::2]) + area.append(area_per_instance) + + return torch.tensor(area) + + @staticmethod + def cat(polymasks_list: List["PolygonMasks"]) -> "PolygonMasks": + """ + Concatenates a list of PolygonMasks into a single PolygonMasks + + Arguments: + polymasks_list (list[PolygonMasks]) + + Returns: + PolygonMasks: the concatenated PolygonMasks + """ + assert isinstance(polymasks_list, (list, tuple)) + assert len(polymasks_list) > 0 + assert all(isinstance(polymask, PolygonMasks) for polymask in polymasks_list) + + cat_polymasks = type(polymasks_list[0])( + list(itertools.chain.from_iterable(pm.polygons for pm in polymasks_list)) + ) + return cat_polymasks + + +class ROIMasks: + """ + Represent masks by N smaller masks defined in some ROIs. Once ROI boxes are given, + full-image bitmask can be obtained by "pasting" the mask on the region defined + by the corresponding ROI box. + """ + + def __init__(self, tensor: torch.Tensor): + """ + Args: + tensor: (N, M, M) mask tensor that defines the mask within each ROI. + """ + if tensor.dim() != 3: + raise ValueError("ROIMasks must take a masks of 3 dimension.") + self.tensor = tensor + + def to(self, device: torch.device) -> "ROIMasks": + return ROIMasks(self.tensor.to(device)) + + @property + def device(self) -> device: + return self.tensor.device + + def __len__(self): + return self.tensor.shape[0] + + def __getitem__(self, item) -> "ROIMasks": + """ + Returns: + ROIMasks: Create a new :class:`ROIMasks` by indexing. + + The following usage are allowed: + + 1. `new_masks = masks[2:10]`: return a slice of masks. + 2. `new_masks = masks[vector]`, where vector is a torch.BoolTensor + with `length = len(masks)`. Nonzero elements in the vector will be selected. + + Note that the returned object might share storage with this object, + subject to Pytorch's indexing semantics. + """ + t = self.tensor[item] + if t.dim() != 3: + raise ValueError( + f"Indexing on ROIMasks with {item} returns a tensor with shape {t.shape}!" + ) + return ROIMasks(t) + + @torch.jit.unused + def __repr__(self) -> str: + s = self.__class__.__name__ + "(" + s += "num_instances={})".format(len(self.tensor)) + return s + + @torch.jit.unused + def to_bitmasks(self, boxes: torch.Tensor, height, width, threshold=0.5): + """ + Args: + + """ + from detectron2.layers import paste_masks_in_image + + paste = retry_if_cuda_oom(paste_masks_in_image) + bitmasks = paste( + self.tensor, + boxes, + (height, width), + threshold=threshold, + ) + return BitMasks(bitmasks) diff --git a/detectron2/structures/rotated_boxes.py b/detectron2/structures/rotated_boxes.py new file mode 100644 index 0000000000000000000000000000000000000000..8f48b40560f2f409b20d87bb1ff448bf44e090d2 --- /dev/null +++ b/detectron2/structures/rotated_boxes.py @@ -0,0 +1,505 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import math +from typing import List, Tuple +import torch + +from detectron2.layers.rotated_boxes import pairwise_iou_rotated + +from .boxes import Boxes, _maybe_jit_unused + + +class RotatedBoxes(Boxes): + """ + This structure stores a list of rotated boxes as a Nx5 torch.Tensor. + It supports some common methods about boxes + (`area`, `clip`, `nonempty`, etc), + and also behaves like a Tensor + (support indexing, `to(device)`, `.device`, and iteration over all boxes) + """ + + def __init__(self, tensor: torch.Tensor): + """ + Args: + tensor (Tensor[float]): a Nx5 matrix. Each row is + (x_center, y_center, width, height, angle), + in which angle is represented in degrees. + While there's no strict range restriction for it, + the recommended principal range is between [-180, 180) degrees. + + Assume we have a horizontal box B = (x_center, y_center, width, height), + where width is along the x-axis and height is along the y-axis. + The rotated box B_rot (x_center, y_center, width, height, angle) + can be seen as: + + 1. When angle == 0: + B_rot == B + 2. When angle > 0: + B_rot is obtained by rotating B w.r.t its center by :math:`|angle|` degrees CCW; + 3. When angle < 0: + B_rot is obtained by rotating B w.r.t its center by :math:`|angle|` degrees CW. + + Mathematically, since the right-handed coordinate system for image space + is (y, x), where y is top->down and x is left->right, the 4 vertices of the + rotated rectangle :math:`(yr_i, xr_i)` (i = 1, 2, 3, 4) can be obtained from + the vertices of the horizontal rectangle :math:`(y_i, x_i)` (i = 1, 2, 3, 4) + in the following way (:math:`\\theta = angle*\\pi/180` is the angle in radians, + :math:`(y_c, x_c)` is the center of the rectangle): + + .. math:: + + yr_i = \\cos(\\theta) (y_i - y_c) - \\sin(\\theta) (x_i - x_c) + y_c, + + xr_i = \\sin(\\theta) (y_i - y_c) + \\cos(\\theta) (x_i - x_c) + x_c, + + which is the standard rigid-body rotation transformation. + + Intuitively, the angle is + (1) the rotation angle from y-axis in image space + to the height vector (top->down in the box's local coordinate system) + of the box in CCW, and + (2) the rotation angle from x-axis in image space + to the width vector (left->right in the box's local coordinate system) + of the box in CCW. + + More intuitively, consider the following horizontal box ABCD represented + in (x1, y1, x2, y2): (3, 2, 7, 4), + covering the [3, 7] x [2, 4] region of the continuous coordinate system + which looks like this: + + .. code:: none + + O--------> x + | + | A---B + | | | + | D---C + | + v y + + Note that each capital letter represents one 0-dimensional geometric point + instead of a 'square pixel' here. + + In the example above, using (x, y) to represent a point we have: + + .. math:: + + O = (0, 0), A = (3, 2), B = (7, 2), C = (7, 4), D = (3, 4) + + We name vector AB = vector DC as the width vector in box's local coordinate system, and + vector AD = vector BC as the height vector in box's local coordinate system. Initially, + when angle = 0 degree, they're aligned with the positive directions of x-axis and y-axis + in the image space, respectively. + + For better illustration, we denote the center of the box as E, + + .. code:: none + + O--------> x + | + | A---B + | | E | + | D---C + | + v y + + where the center E = ((3+7)/2, (2+4)/2) = (5, 3). + + Also, + + .. math:: + + width = |AB| = |CD| = 7 - 3 = 4, + height = |AD| = |BC| = 4 - 2 = 2. + + Therefore, the corresponding representation for the same shape in rotated box in + (x_center, y_center, width, height, angle) format is: + + (5, 3, 4, 2, 0), + + Now, let's consider (5, 3, 4, 2, 90), which is rotated by 90 degrees + CCW (counter-clockwise) by definition. It looks like this: + + .. code:: none + + O--------> x + | B-C + | | | + | |E| + | | | + | A-D + v y + + The center E is still located at the same point (5, 3), while the vertices + ABCD are rotated by 90 degrees CCW with regard to E: + A = (4, 5), B = (4, 1), C = (6, 1), D = (6, 5) + + Here, 90 degrees can be seen as the CCW angle to rotate from y-axis to + vector AD or vector BC (the top->down height vector in box's local coordinate system), + or the CCW angle to rotate from x-axis to vector AB or vector DC (the left->right + width vector in box's local coordinate system). + + .. math:: + + width = |AB| = |CD| = 5 - 1 = 4, + height = |AD| = |BC| = 6 - 4 = 2. + + Next, how about (5, 3, 4, 2, -90), which is rotated by 90 degrees CW (clockwise) + by definition? It looks like this: + + .. code:: none + + O--------> x + | D-A + | | | + | |E| + | | | + | C-B + v y + + The center E is still located at the same point (5, 3), while the vertices + ABCD are rotated by 90 degrees CW with regard to E: + A = (6, 1), B = (6, 5), C = (4, 5), D = (4, 1) + + .. math:: + + width = |AB| = |CD| = 5 - 1 = 4, + height = |AD| = |BC| = 6 - 4 = 2. + + This covers exactly the same region as (5, 3, 4, 2, 90) does, and their IoU + will be 1. However, these two will generate different RoI Pooling results and + should not be treated as an identical box. + + On the other hand, it's easy to see that (X, Y, W, H, A) is identical to + (X, Y, W, H, A+360N), for any integer N. For example (5, 3, 4, 2, 270) would be + identical to (5, 3, 4, 2, -90), because rotating the shape 270 degrees CCW is + equivalent to rotating the same shape 90 degrees CW. + + We could rotate further to get (5, 3, 4, 2, 180), or (5, 3, 4, 2, -180): + + .. code:: none + + O--------> x + | + | C---D + | | E | + | B---A + | + v y + + .. math:: + + A = (7, 4), B = (3, 4), C = (3, 2), D = (7, 2), + + width = |AB| = |CD| = 7 - 3 = 4, + height = |AD| = |BC| = 4 - 2 = 2. + + Finally, this is a very inaccurate (heavily quantized) illustration of + how (5, 3, 4, 2, 60) looks like in case anyone wonders: + + .. code:: none + + O--------> x + | B\ + | / C + | /E / + | A / + | `D + v y + + It's still a rectangle with center of (5, 3), width of 4 and height of 2, + but its angle (and thus orientation) is somewhere between + (5, 3, 4, 2, 0) and (5, 3, 4, 2, 90). + """ + device = tensor.device if isinstance(tensor, torch.Tensor) else torch.device("cpu") + tensor = torch.as_tensor(tensor, dtype=torch.float32, device=device) + if tensor.numel() == 0: + # Use reshape, so we don't end up creating a new tensor that does not depend on + # the inputs (and consequently confuses jit) + tensor = tensor.reshape((0, 5)).to(dtype=torch.float32, device=device) + assert tensor.dim() == 2 and tensor.size(-1) == 5, tensor.size() + + self.tensor = tensor + + def clone(self) -> "RotatedBoxes": + """ + Clone the RotatedBoxes. + + Returns: + RotatedBoxes + """ + return RotatedBoxes(self.tensor.clone()) + + @_maybe_jit_unused + def to(self, device: torch.device): + # Boxes are assumed float32 and does not support to(dtype) + return RotatedBoxes(self.tensor.to(device=device)) + + def area(self) -> torch.Tensor: + """ + Computes the area of all the boxes. + + Returns: + torch.Tensor: a vector with areas of each box. + """ + box = self.tensor + area = box[:, 2] * box[:, 3] + return area + + def normalize_angles(self) -> None: + """ + Restrict angles to the range of [-180, 180) degrees + """ + self.tensor[:, 4] = (self.tensor[:, 4] + 180.0) % 360.0 - 180.0 + + def clip(self, box_size: Tuple[int, int], clip_angle_threshold: float = 1.0) -> None: + """ + Clip (in place) the boxes by limiting x coordinates to the range [0, width] + and y coordinates to the range [0, height]. + + For RRPN: + Only clip boxes that are almost horizontal with a tolerance of + clip_angle_threshold to maintain backward compatibility. + + Rotated boxes beyond this threshold are not clipped for two reasons: + + 1. There are potentially multiple ways to clip a rotated box to make it + fit within the image. + 2. It's tricky to make the entire rectangular box fit within the image + and still be able to not leave out pixels of interest. + + Therefore we rely on ops like RoIAlignRotated to safely handle this. + + Args: + box_size (height, width): The clipping box's size. + clip_angle_threshold: + Iff. abs(normalized(angle)) <= clip_angle_threshold (in degrees), + we do the clipping as horizontal boxes. + """ + h, w = box_size + + # normalize angles to be within (-180, 180] degrees + self.normalize_angles() + + idx = torch.where(torch.abs(self.tensor[:, 4]) <= clip_angle_threshold)[0] + + # convert to (x1, y1, x2, y2) + x1 = self.tensor[idx, 0] - self.tensor[idx, 2] / 2.0 + y1 = self.tensor[idx, 1] - self.tensor[idx, 3] / 2.0 + x2 = self.tensor[idx, 0] + self.tensor[idx, 2] / 2.0 + y2 = self.tensor[idx, 1] + self.tensor[idx, 3] / 2.0 + + # clip + x1.clamp_(min=0, max=w) + y1.clamp_(min=0, max=h) + x2.clamp_(min=0, max=w) + y2.clamp_(min=0, max=h) + + # convert back to (xc, yc, w, h) + self.tensor[idx, 0] = (x1 + x2) / 2.0 + self.tensor[idx, 1] = (y1 + y2) / 2.0 + # make sure widths and heights do not increase due to numerical errors + self.tensor[idx, 2] = torch.min(self.tensor[idx, 2], x2 - x1) + self.tensor[idx, 3] = torch.min(self.tensor[idx, 3], y2 - y1) + + def nonempty(self, threshold: float = 0.0) -> torch.Tensor: + """ + Find boxes that are non-empty. + A box is considered empty, if either of its side is no larger than threshold. + + Returns: + Tensor: a binary vector which represents + whether each box is empty (False) or non-empty (True). + """ + box = self.tensor + widths = box[:, 2] + heights = box[:, 3] + keep = (widths > threshold) & (heights > threshold) + return keep + + def __getitem__(self, item) -> "RotatedBoxes": + """ + Returns: + RotatedBoxes: Create a new :class:`RotatedBoxes` by indexing. + + The following usage are allowed: + + 1. `new_boxes = boxes[3]`: return a `RotatedBoxes` which contains only one box. + 2. `new_boxes = boxes[2:10]`: return a slice of boxes. + 3. `new_boxes = boxes[vector]`, where vector is a torch.ByteTensor + with `length = len(boxes)`. Nonzero elements in the vector will be selected. + + Note that the returned RotatedBoxes might share storage with this RotatedBoxes, + subject to Pytorch's indexing semantics. + """ + if isinstance(item, int): + return RotatedBoxes(self.tensor[item].view(1, -1)) + b = self.tensor[item] + assert b.dim() == 2, "Indexing on RotatedBoxes with {} failed to return a matrix!".format( + item + ) + return RotatedBoxes(b) + + def __len__(self) -> int: + return self.tensor.shape[0] + + def __repr__(self) -> str: + return "RotatedBoxes(" + str(self.tensor) + ")" + + def inside_box(self, box_size: Tuple[int, int], boundary_threshold: int = 0) -> torch.Tensor: + """ + Args: + box_size (height, width): Size of the reference box covering + [0, width] x [0, height] + boundary_threshold (int): Boxes that extend beyond the reference box + boundary by more than boundary_threshold are considered "outside". + + For RRPN, it might not be necessary to call this function since it's common + for rotated box to extend to outside of the image boundaries + (the clip function only clips the near-horizontal boxes) + + Returns: + a binary vector, indicating whether each box is inside the reference box. + """ + height, width = box_size + + cnt_x = self.tensor[..., 0] + cnt_y = self.tensor[..., 1] + half_w = self.tensor[..., 2] / 2.0 + half_h = self.tensor[..., 3] / 2.0 + a = self.tensor[..., 4] + c = torch.abs(torch.cos(a * math.pi / 180.0)) + s = torch.abs(torch.sin(a * math.pi / 180.0)) + # This basically computes the horizontal bounding rectangle of the rotated box + max_rect_dx = c * half_w + s * half_h + max_rect_dy = c * half_h + s * half_w + + inds_inside = ( + (cnt_x - max_rect_dx >= -boundary_threshold) + & (cnt_y - max_rect_dy >= -boundary_threshold) + & (cnt_x + max_rect_dx < width + boundary_threshold) + & (cnt_y + max_rect_dy < height + boundary_threshold) + ) + + return inds_inside + + def get_centers(self) -> torch.Tensor: + """ + Returns: + The box centers in a Nx2 array of (x, y). + """ + return self.tensor[:, :2] + + def scale(self, scale_x: float, scale_y: float) -> None: + """ + Scale the rotated box with horizontal and vertical scaling factors + Note: when scale_factor_x != scale_factor_y, + the rotated box does not preserve the rectangular shape when the angle + is not a multiple of 90 degrees under resize transformation. + Instead, the shape is a parallelogram (that has skew) + Here we make an approximation by fitting a rotated rectangle to the parallelogram. + """ + self.tensor[:, 0] *= scale_x + self.tensor[:, 1] *= scale_y + theta = self.tensor[:, 4] * math.pi / 180.0 + c = torch.cos(theta) + s = torch.sin(theta) + + # In image space, y is top->down and x is left->right + # Consider the local coordintate system for the rotated box, + # where the box center is located at (0, 0), and the four vertices ABCD are + # A(-w / 2, -h / 2), B(w / 2, -h / 2), C(w / 2, h / 2), D(-w / 2, h / 2) + # the midpoint of the left edge AD of the rotated box E is: + # E = (A+D)/2 = (-w / 2, 0) + # the midpoint of the top edge AB of the rotated box F is: + # F(0, -h / 2) + # To get the old coordinates in the global system, apply the rotation transformation + # (Note: the right-handed coordinate system for image space is yOx): + # (old_x, old_y) = (s * y + c * x, c * y - s * x) + # E(old) = (s * 0 + c * (-w/2), c * 0 - s * (-w/2)) = (-c * w / 2, s * w / 2) + # F(old) = (s * (-h / 2) + c * 0, c * (-h / 2) - s * 0) = (-s * h / 2, -c * h / 2) + # After applying the scaling factor (sfx, sfy): + # E(new) = (-sfx * c * w / 2, sfy * s * w / 2) + # F(new) = (-sfx * s * h / 2, -sfy * c * h / 2) + # The new width after scaling tranformation becomes: + + # w(new) = |E(new) - O| * 2 + # = sqrt[(sfx * c * w / 2)^2 + (sfy * s * w / 2)^2] * 2 + # = sqrt[(sfx * c)^2 + (sfy * s)^2] * w + # i.e., scale_factor_w = sqrt[(sfx * c)^2 + (sfy * s)^2] + # + # For example, + # when angle = 0 or 180, |c| = 1, s = 0, scale_factor_w == scale_factor_x; + # when |angle| = 90, c = 0, |s| = 1, scale_factor_w == scale_factor_y + self.tensor[:, 2] *= torch.sqrt((scale_x * c) ** 2 + (scale_y * s) ** 2) + + # h(new) = |F(new) - O| * 2 + # = sqrt[(sfx * s * h / 2)^2 + (sfy * c * h / 2)^2] * 2 + # = sqrt[(sfx * s)^2 + (sfy * c)^2] * h + # i.e., scale_factor_h = sqrt[(sfx * s)^2 + (sfy * c)^2] + # + # For example, + # when angle = 0 or 180, |c| = 1, s = 0, scale_factor_h == scale_factor_y; + # when |angle| = 90, c = 0, |s| = 1, scale_factor_h == scale_factor_x + self.tensor[:, 3] *= torch.sqrt((scale_x * s) ** 2 + (scale_y * c) ** 2) + + # The angle is the rotation angle from y-axis in image space to the height + # vector (top->down in the box's local coordinate system) of the box in CCW. + # + # angle(new) = angle_yOx(O - F(new)) + # = angle_yOx( (sfx * s * h / 2, sfy * c * h / 2) ) + # = atan2(sfx * s * h / 2, sfy * c * h / 2) + # = atan2(sfx * s, sfy * c) + # + # For example, + # when sfx == sfy, angle(new) == atan2(s, c) == angle(old) + self.tensor[:, 4] = torch.atan2(scale_x * s, scale_y * c) * 180 / math.pi + + @classmethod + @_maybe_jit_unused + def cat(cls, boxes_list: List["RotatedBoxes"]) -> "RotatedBoxes": + """ + Concatenates a list of RotatedBoxes into a single RotatedBoxes + + Arguments: + boxes_list (list[RotatedBoxes]) + + Returns: + RotatedBoxes: the concatenated RotatedBoxes + """ + assert isinstance(boxes_list, (list, tuple)) + if len(boxes_list) == 0: + return cls(torch.empty(0)) + assert all([isinstance(box, RotatedBoxes) for box in boxes_list]) + + # use torch.cat (v.s. layers.cat) so the returned boxes never share storage with input + cat_boxes = cls(torch.cat([b.tensor for b in boxes_list], dim=0)) + return cat_boxes + + @property + def device(self) -> torch.device: + return self.tensor.device + + @torch.jit.unused + def __iter__(self): + """ + Yield a box as a Tensor of shape (5,) at a time. + """ + yield from self.tensor + + +def pairwise_iou(boxes1: RotatedBoxes, boxes2: RotatedBoxes) -> None: + """ + Given two lists of rotated boxes of size N and M, + compute the IoU (intersection over union) + between **all** N x M pairs of boxes. + The box order must be (x_center, y_center, width, height, angle). + + Args: + boxes1, boxes2 (RotatedBoxes): + two `RotatedBoxes`. Contains N & M rotated boxes, respectively. + + Returns: + Tensor: IoU, sized [N,M]. + """ + + return pairwise_iou_rotated(boxes1.tensor, boxes2.tensor) diff --git a/detectron2/structures/tsv_file.py b/detectron2/structures/tsv_file.py new file mode 100644 index 0000000000000000000000000000000000000000..207affe2c42020954925db07a33f8fe6688ed498 --- /dev/null +++ b/detectron2/structures/tsv_file.py @@ -0,0 +1,352 @@ +import logging +import os +import json +import os.path as op +import numpy as np +from typing import List, Union +from collections import OrderedDict + +def generate_lineidx(filein, idxout): + idxout_tmp = idxout + '.tmp' + with open(filein, 'r') as tsvin, open(idxout_tmp,'w') as tsvout: + fsize = os.fstat(tsvin.fileno()).st_size + fpos = 0 + while fpos!=fsize: + tsvout.write(str(fpos)+"\n") + tsvin.readline() + fpos = tsvin.tell() + os.rename(idxout_tmp, idxout) + + +def read_to_character(fp, c): + result = [] + while True: + s = fp.read(32) + assert s != '' + if c in s: + result.append(s[: s.index(c)]) + break + else: + result.append(s) + return ''.join(result) + + +class TSVFile(object): + def __init__(self, tsv_file, generate_lineidx=False): + self.tsv_file = tsv_file + self.lineidx = op.splitext(tsv_file)[0] + '.lineidx' + self._fp = None + self._lineidx = None + # the process always keeps the process which opens the file. + # If the pid is not equal to the currrent pid, we will re-open the file. + self.pid = None + # generate lineidx if not exist + if not op.isfile(self.lineidx) and generate_lineidx: + generate_lineidx(self.tsv_file, self.lineidx) + + def __del__(self): + if self._fp: + self._fp.close() + + def __str__(self): + return "TSVFile(tsv_file='{}')".format(self.tsv_file) + + def __repr__(self): + return str(self) + + def num_rows(self): + self._ensure_lineidx_loaded() + return len(self._lineidx) + + def seek(self, idx): + self._ensure_tsv_opened() + self._ensure_lineidx_loaded() + try: + pos = self._lineidx[idx] + except: + logging.info('{}-{}'.format(self.tsv_file, idx)) + raise + self._fp.seek(pos) + return [s.strip() for s in self._fp.readline().split('\t')] + + def seek_first_column(self, idx): + self._ensure_tsv_opened() + self._ensure_lineidx_loaded() + pos = self._lineidx[idx] + self._fp.seek(pos) + return read_to_character(self._fp, '\t') + + def get_key(self, idx): + return self.seek_first_column(idx) + + def __getitem__(self, index): + return self.seek(index) + + def __len__(self): + return self.num_rows() + + def _ensure_lineidx_loaded(self): + if self._lineidx is None: + # print('loading lineidx: {}'.format(self.lineidx)) + + with open(self.lineidx, 'r') as fp: + self._lineidx = [int(i.strip()) for i in fp.readlines()] + + def _ensure_tsv_opened(self): + if self._fp is None: + self._fp = open(self.tsv_file, 'r') + self.pid = os.getpid() + + if self.pid != os.getpid(): + # print('re-open {} because the process id changed'.format(self.tsv_file)) + self._fp = open(self.tsv_file, 'r') + self.pid = os.getpid() + +class TSVFileNew(object): + def __init__(self, + tsv_file: str, + if_generate_lineidx: bool = False, + lineidx: str = None, + class_selector: List[str] = None): + self.tsv_file = tsv_file + self.lineidx = op.splitext(tsv_file)[0] + '.lineidx' \ + if not lineidx else lineidx + self.linelist = op.splitext(tsv_file)[0] + '.linelist' + self.chunks = op.splitext(tsv_file)[0] + '.chunks' + self._fp = None + self._lineidx = None + self._sample_indices = None + self._class_boundaries = None + self._class_selector = class_selector + # the process always keeps the process which opens the file. + # If the pid is not equal to the currrent pid, we will re-open the file. + self.pid = None + # generate lineidx if not exist + if not op.isfile(self.lineidx) and if_generate_lineidx: + generate_lineidx(self.tsv_file, self.lineidx) + + def __del__(self): + if self._fp: + self._fp.close() + + def __str__(self): + return "TSVFile(tsv_file='{}')".format(self.tsv_file) + + def __repr__(self): + return str(self) + + def get_class_boundaries(self): + return self._class_boundaries + + def num_rows(self): + self._ensure_lineidx_loaded() + return len(self._sample_indices) + + def seek(self, idx: int): + self._ensure_tsv_opened() + self._ensure_lineidx_loaded() + try: + pos = self._lineidx[self._sample_indices[idx]] + except: + logging.info('=> {}-{}'.format(self.tsv_file, idx)) + raise + self._fp.seek(pos) + return [s.strip() for s in self._fp.readline().split('\t')] + + def seek_first_column(self, idx: int): + self._ensure_tsv_opened() + self._ensure_lineidx_loaded() + pos = self._lineidx[idx] + self._fp.seek(pos) + return read_to_character(self._fp, '\t') + + def get_key(self, idx: int): + return self.seek_first_column(idx) + + def __getitem__(self, index: int): + return self.seek(index) + + def __len__(self): + return self.num_rows() + + def _ensure_lineidx_loaded(self): + if self._lineidx is None: + # print('=> loading lineidx: {}'.format(self.lineidx)) + with open(self.lineidx, 'r') as fp: + lines = fp.readlines() + lines = [line.strip() for line in lines] + self._lineidx = [int(line) for line in lines] + # except: + # print("error in loading lineidx file {}, regenerate it".format(self.lineidx)) + # generate_lineidx(self.tsv_file, self.lineidx) + # with open(self.lineidx, 'r') as fp: + # lines = fp.readlines() + # lines = [line.strip() for line in lines] + # self._lineidx = [int(line) for line in lines] + # read the line list if exists + linelist = None + if op.isfile(self.linelist): + with open(self.linelist, 'r') as fp: + linelist = sorted( + [ + int(line.strip()) + for line in fp.readlines() + ] + ) + if op.isfile(self.chunks) and self._class_selector: + self._sample_indices = [] + self._class_boundaries = [] + class_boundaries = json.load(open(self.chunks, 'r')) + for class_name, boundary in class_boundaries.items(): + start = len(self._sample_indices) + if class_name in self._class_selector: + for idx in range(boundary[0], boundary[1] + 1): + # NOTE: potentially slow when linelist is long, try to speed it up + if linelist and idx not in linelist: + continue + self._sample_indices.append(idx) + end = len(self._sample_indices) + self._class_boundaries.append((start, end)) + else: + if linelist: + self._sample_indices = linelist + else: + self._sample_indices = list(range(len(self._lineidx))) + + def _ensure_tsv_opened(self): + if self._fp is None: + self._fp = open(self.tsv_file, 'r') + self.pid = os.getpid() + + if self.pid != os.getpid(): + logging.debug('=> re-open {} because the process id changed'.format(self.tsv_file)) + self._fp = open(self.tsv_file, 'r') + self.pid = os.getpid() + +class LRU(OrderedDict): + """Limit size, evicting the least recently looked-up key when full. + https://docs.python.org/3/library/collections.html#collections.OrderedDict + """ + + def __init__(self, maxsize=4, *args, **kwds): + self.maxsize = maxsize + super().__init__(*args, **kwds) + + def __getitem__(self, key): + value = super().__getitem__(key) + self.move_to_end(key) + return value + + def __setitem__(self, key, value): + if key in self: + self.move_to_end(key) + super().__setitem__(key, value) + if len(self) > self.maxsize: + oldest = next(iter(self)) + del self[oldest] + + +class CompositeTSVFile: + def __init__(self, + file_list: Union[str, list], + root: str = '.', + class_selector: List[str] = None): + if isinstance(file_list, str): + self.file_list = load_list_file(file_list) + else: + assert isinstance(file_list, list) + self.file_list = file_list + + self.root = root + self.cache = LRU() + self.tsvs = None + self.chunk_sizes = None + self.accum_chunk_sizes = None + self._class_selector = class_selector + self._class_boundaries = None + self.initialized = False + self.initialize() + + def get_key(self, index: int): + idx_source, idx_row = self._calc_chunk_idx_row(index) + k = self.tsvs[idx_source].get_key(idx_row) + return '_'.join([self.file_list[idx_source], k]) + + def get_class_boundaries(self): + return self._class_boundaries + + def get_chunk_size(self): + return self.chunk_sizes + + def num_rows(self): + return sum(self.chunk_sizes) + + def _calc_chunk_idx_row(self, index: int): + idx_chunk = 0 + idx_row = index + while index >= self.accum_chunk_sizes[idx_chunk]: + idx_chunk += 1 + idx_row = index - self.accum_chunk_sizes[idx_chunk-1] + return idx_chunk, idx_row + + + def __getitem__(self, index: int): + idx_source, idx_row = self._calc_chunk_idx_row(index) + if idx_source not in self.cache: + self.cache[idx_source] = TSVFileNew( + op.join(self.root, self.file_list[idx_source]), + class_selector=self._class_selector + ) + return self.cache[idx_source].seek(idx_row) + + def __len__(self): + return sum(self.chunk_sizes) + + def initialize(self): + """ + this function has to be called in init function if cache_policy is + enabled. Thus, let's always call it in init funciton to make it simple. + """ + if self.initialized: + return + tsvs = [ + TSVFileNew( + op.join(self.root, f), + class_selector=self._class_selector + ) for f in self.file_list + ] + logging.info("Calculating chunk sizes ...") + self.chunk_sizes = [len(tsv) for tsv in tsvs] + + self.accum_chunk_sizes = [0] + for size in self.chunk_sizes: + self.accum_chunk_sizes += [self.accum_chunk_sizes[-1] + size] + self.accum_chunk_sizes = self.accum_chunk_sizes[1:] + + if ( + self._class_selector + and all([tsv.get_class_boundaries() for tsv in tsvs]) + ): + """ + Note: When using CompositeTSVFile, make sure that the classes contained in each + tsv file do not overlap. Otherwise, the class boundaries won't be correct. + """ + self._class_boundaries = [] + offset = 0 + for tsv in tsvs: + boundaries = tsv.get_class_boundaries() + for bound in boundaries: + self._class_boundaries.append((bound[0] + offset, bound[1] + offset)) + offset += len(tsv) + # NOTE: in current setting, get_key is not used during training, so we remove tsvs for saving memory cost + del tsvs + self.initialized = True + + +def load_list_file(fname: str) -> List[str]: + with open(fname, 'r') as fp: + lines = fp.readlines() + result = [line.strip() for line in lines] + if len(result) > 0 and result[-1] == '': + result = result[:-1] + return result diff --git a/detectron2/utils/README.md b/detectron2/utils/README.md new file mode 100644 index 0000000000000000000000000000000000000000..9765b24a730b77556104187ac3ef5439ab0859fd --- /dev/null +++ b/detectron2/utils/README.md @@ -0,0 +1,5 @@ +# Utility functions + +This folder contain utility functions that are not used in the +core library, but are useful for building models or training +code using the config system. diff --git a/detectron2/utils/__init__.py b/detectron2/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..9020c2df23e2af280b7bb168b996ae9eaf312eb8 --- /dev/null +++ b/detectron2/utils/__init__.py @@ -0,0 +1 @@ +# Copyright (c) Facebook, Inc. and its affiliates. diff --git a/detectron2/utils/__pycache__/__init__.cpython-39.pyc b/detectron2/utils/__pycache__/__init__.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2b08dc92a52b1fc1646e8d7d5d4dfe21dfd47677 Binary files /dev/null and b/detectron2/utils/__pycache__/__init__.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/collect_env.cpython-39.pyc b/detectron2/utils/__pycache__/collect_env.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..1aec0dabfe883f85a132ce35c53eb9fcbefcf5ce Binary files /dev/null and b/detectron2/utils/__pycache__/collect_env.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/colormap.cpython-39.pyc b/detectron2/utils/__pycache__/colormap.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..126cba7a23b6d8dbcbf35fa2a7f6a0c0153e28aa Binary files /dev/null and b/detectron2/utils/__pycache__/colormap.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/comm.cpython-39.pyc b/detectron2/utils/__pycache__/comm.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..17585cc904adf357554a4f214cf64ee3dc82698c Binary files /dev/null and b/detectron2/utils/__pycache__/comm.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/env.cpython-39.pyc b/detectron2/utils/__pycache__/env.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..ce8ba13886aaa358ae68f3760fccca6677ab2b14 Binary files /dev/null and b/detectron2/utils/__pycache__/env.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/events.cpython-39.pyc b/detectron2/utils/__pycache__/events.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7ab7ffb96652f3a34f720809127f24dcccf64831 Binary files /dev/null and b/detectron2/utils/__pycache__/events.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/file_io.cpython-39.pyc b/detectron2/utils/__pycache__/file_io.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b20ccc76a96f8469f910f529220e472ec0d5ceb1 Binary files /dev/null and b/detectron2/utils/__pycache__/file_io.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/logger.cpython-39.pyc b/detectron2/utils/__pycache__/logger.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d98e8e49a91592ce7009ee4ce262d8712eac10d6 Binary files /dev/null and b/detectron2/utils/__pycache__/logger.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/memory.cpython-39.pyc b/detectron2/utils/__pycache__/memory.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..250a740c1aa6183787c7d7fc47c9f679ba2a6c1e Binary files /dev/null and b/detectron2/utils/__pycache__/memory.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/registry.cpython-39.pyc b/detectron2/utils/__pycache__/registry.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a1091ab262cdd1e5138e0914b66dc29254ee81c3 Binary files /dev/null and b/detectron2/utils/__pycache__/registry.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/serialize.cpython-39.pyc b/detectron2/utils/__pycache__/serialize.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e50beb6529bb232c680b0a89ddd8b84a0198afaf Binary files /dev/null and b/detectron2/utils/__pycache__/serialize.cpython-39.pyc differ diff --git a/detectron2/utils/__pycache__/visualizer.cpython-39.pyc b/detectron2/utils/__pycache__/visualizer.cpython-39.pyc new file mode 100644 index 0000000000000000000000000000000000000000..da2191724e9ac68b1e2d3d0a5f3ffa0cee815090 Binary files /dev/null and b/detectron2/utils/__pycache__/visualizer.cpython-39.pyc differ diff --git a/detectron2/utils/analysis.py b/detectron2/utils/analysis.py new file mode 100644 index 0000000000000000000000000000000000000000..51b453cabb2f369e470296d468955432446de1a5 --- /dev/null +++ b/detectron2/utils/analysis.py @@ -0,0 +1,152 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# -*- coding: utf-8 -*- + +import typing +import fvcore +from fvcore.nn import activation_count, flop_count, parameter_count, parameter_count_table +from torch import nn + +from detectron2.export import TracingAdapter + +__all__ = [ + "activation_count_operators", + "flop_count_operators", + "parameter_count_table", + "parameter_count", +] + +FLOPS_MODE = "flops" +ACTIVATIONS_MODE = "activations" + + +# Some extra ops to ignore from counting, including elementwise and reduction ops +_IGNORED_OPS = { + "aten::add", + "aten::add_", + "aten::argmax", + "aten::argsort", + "aten::batch_norm", + "aten::constant_pad_nd", + "aten::div", + "aten::div_", + "aten::exp", + "aten::log2", + "aten::max_pool2d", + "aten::meshgrid", + "aten::mul", + "aten::mul_", + "aten::neg", + "aten::nonzero_numpy", + "aten::reciprocal", + "aten::rsub", + "aten::sigmoid", + "aten::sigmoid_", + "aten::softmax", + "aten::sort", + "aten::sqrt", + "aten::sub", + "torchvision::nms", # TODO estimate flop for nms +} + + +class FlopCountAnalysis(fvcore.nn.FlopCountAnalysis): + """ + Same as :class:`fvcore.nn.FlopCountAnalysis`, but supports detectron2 models. + """ + + def __init__(self, model, inputs): + """ + Args: + model (nn.Module): + inputs (Any): inputs of the given model. Does not have to be tuple of tensors. + """ + wrapper = TracingAdapter(model, inputs, allow_non_tensor=True) + super().__init__(wrapper, wrapper.flattened_inputs) + self.set_op_handle(**{k: None for k in _IGNORED_OPS}) + + +def flop_count_operators(model: nn.Module, inputs: list) -> typing.DefaultDict[str, float]: + """ + Implement operator-level flops counting using jit. + This is a wrapper of :func:`fvcore.nn.flop_count` and adds supports for standard + detection models in detectron2. + Please use :class:`FlopCountAnalysis` for more advanced functionalities. + + Note: + The function runs the input through the model to compute flops. + The flops of a detection model is often input-dependent, for example, + the flops of box & mask head depends on the number of proposals & + the number of detected objects. + Therefore, the flops counting using a single input may not accurately + reflect the computation cost of a model. It's recommended to average + across a number of inputs. + + Args: + model: a detectron2 model that takes `list[dict]` as input. + inputs (list[dict]): inputs to model, in detectron2's standard format. + Only "image" key will be used. + supported_ops (dict[str, Handle]): see documentation of :func:`fvcore.nn.flop_count` + + Returns: + Counter: Gflop count per operator + """ + old_train = model.training + model.eval() + ret = FlopCountAnalysis(model, inputs).by_operator() + model.train(old_train) + return {k: v / 1e9 for k, v in ret.items()} + + +def activation_count_operators( + model: nn.Module, inputs: list, **kwargs +) -> typing.DefaultDict[str, float]: + """ + Implement operator-level activations counting using jit. + This is a wrapper of fvcore.nn.activation_count, that supports standard detection models + in detectron2. + + Note: + The function runs the input through the model to compute activations. + The activations of a detection model is often input-dependent, for example, + the activations of box & mask head depends on the number of proposals & + the number of detected objects. + + Args: + model: a detectron2 model that takes `list[dict]` as input. + inputs (list[dict]): inputs to model, in detectron2's standard format. + Only "image" key will be used. + + Returns: + Counter: activation count per operator + """ + return _wrapper_count_operators(model=model, inputs=inputs, mode=ACTIVATIONS_MODE, **kwargs) + + +def _wrapper_count_operators( + model: nn.Module, inputs: list, mode: str, **kwargs +) -> typing.DefaultDict[str, float]: + # ignore some ops + supported_ops = {k: lambda *args, **kwargs: {} for k in _IGNORED_OPS} + supported_ops.update(kwargs.pop("supported_ops", {})) + kwargs["supported_ops"] = supported_ops + + assert len(inputs) == 1, "Please use batch size=1" + tensor_input = inputs[0]["image"] + inputs = [{"image": tensor_input}] # remove other keys, in case there are any + + old_train = model.training + if isinstance(model, (nn.parallel.distributed.DistributedDataParallel, nn.DataParallel)): + model = model.module + wrapper = TracingAdapter(model, inputs) + wrapper.eval() + if mode == FLOPS_MODE: + ret = flop_count(wrapper, (tensor_input,), **kwargs) + elif mode == ACTIVATIONS_MODE: + ret = activation_count(wrapper, (tensor_input,), **kwargs) + else: + raise NotImplementedError("Count for mode {} is not supported yet.".format(mode)) + # compatible with change in fvcore + if isinstance(ret, tuple): + ret = ret[0] + model.train(old_train) + return ret diff --git a/detectron2/utils/collect_env.py b/detectron2/utils/collect_env.py new file mode 100644 index 0000000000000000000000000000000000000000..a2359d332124b49024c5ae59fe3a4a51a92d181b --- /dev/null +++ b/detectron2/utils/collect_env.py @@ -0,0 +1,211 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import importlib +import numpy as np +import os +import re +import subprocess +import sys +from collections import defaultdict +import PIL +import torch +import torchvision +from tabulate import tabulate + +__all__ = ["collect_env_info"] + + +def collect_torch_env(): + try: + import torch.__config__ + + return torch.__config__.show() + except ImportError: + # compatible with older versions of pytorch + from torch.utils.collect_env import get_pretty_env_info + + return get_pretty_env_info() + + +def get_env_module(): + var_name = "DETECTRON2_ENV_MODULE" + return var_name, os.environ.get(var_name, "") + + +def detect_compute_compatibility(CUDA_HOME, so_file): + try: + cuobjdump = os.path.join(CUDA_HOME, "bin", "cuobjdump") + if os.path.isfile(cuobjdump): + output = subprocess.check_output( + "'{}' --list-elf '{}'".format(cuobjdump, so_file), shell=True + ) + output = output.decode("utf-8").strip().split("\n") + arch = [] + for line in output: + line = re.findall(r"\.sm_([0-9]*)\.", line)[0] + arch.append(".".join(line)) + arch = sorted(set(arch)) + return ", ".join(arch) + else: + return so_file + "; cannot find cuobjdump" + except Exception: + # unhandled failure + return so_file + + +def collect_env_info(): + has_gpu = torch.cuda.is_available() # true for both CUDA & ROCM + torch_version = torch.__version__ + + # NOTE that CUDA_HOME/ROCM_HOME could be None even when CUDA runtime libs are functional + from torch.utils.cpp_extension import CUDA_HOME, ROCM_HOME + + has_rocm = False + if (getattr(torch.version, "hip", None) is not None) and (ROCM_HOME is not None): + has_rocm = True + has_cuda = has_gpu and (not has_rocm) + + data = [] + data.append(("sys.platform", sys.platform)) # check-template.yml depends on it + data.append(("Python", sys.version.replace("\n", ""))) + data.append(("numpy", np.__version__)) + + try: + import detectron2 # noqa + + data.append( + ("detectron2", detectron2.__version__ + " @" + os.path.dirname(detectron2.__file__)) + ) + except ImportError: + data.append(("detectron2", "failed to import")) + except AttributeError: + data.append(("detectron2", "imported a wrong installation")) + + try: + import detectron2._C as _C + except ImportError as e: + data.append(("detectron2._C", f"not built correctly: {e}")) + + # print system compilers when extension fails to build + if sys.platform != "win32": # don't know what to do for windows + try: + # this is how torch/utils/cpp_extensions.py choose compiler + cxx = os.environ.get("CXX", "c++") + cxx = subprocess.check_output("'{}' --version".format(cxx), shell=True) + cxx = cxx.decode("utf-8").strip().split("\n")[0] + except subprocess.SubprocessError: + cxx = "Not found" + data.append(("Compiler ($CXX)", cxx)) + + if has_cuda and CUDA_HOME is not None: + try: + nvcc = os.path.join(CUDA_HOME, "bin", "nvcc") + nvcc = subprocess.check_output("'{}' -V".format(nvcc), shell=True) + nvcc = nvcc.decode("utf-8").strip().split("\n")[-1] + except subprocess.SubprocessError: + nvcc = "Not found" + data.append(("CUDA compiler", nvcc)) + if has_cuda and sys.platform != "win32": + try: + so_file = importlib.util.find_spec("detectron2._C").origin + except (ImportError, AttributeError): + pass + else: + data.append( + ("detectron2 arch flags", detect_compute_compatibility(CUDA_HOME, so_file)) + ) + else: + # print compilers that are used to build extension + data.append(("Compiler", _C.get_compiler_version())) + data.append(("CUDA compiler", _C.get_cuda_version())) # cuda or hip + if has_cuda and getattr(_C, "has_cuda", lambda: True)(): + data.append( + ("detectron2 arch flags", detect_compute_compatibility(CUDA_HOME, _C.__file__)) + ) + + data.append(get_env_module()) + data.append(("PyTorch", torch_version + " @" + os.path.dirname(torch.__file__))) + data.append(("PyTorch debug build", torch.version.debug)) + + data.append(("GPU available", has_gpu)) + if has_gpu: + devices = defaultdict(list) + for k in range(torch.cuda.device_count()): + cap = ".".join((str(x) for x in torch.cuda.get_device_capability(k))) + name = torch.cuda.get_device_name(k) + f" (arch={cap})" + devices[name].append(str(k)) + for name, devids in devices.items(): + data.append(("GPU " + ",".join(devids), name)) + + if has_rocm: + msg = " - invalid!" if not (ROCM_HOME and os.path.isdir(ROCM_HOME)) else "" + data.append(("ROCM_HOME", str(ROCM_HOME) + msg)) + else: + msg = " - invalid!" if not (CUDA_HOME and os.path.isdir(CUDA_HOME)) else "" + data.append(("CUDA_HOME", str(CUDA_HOME) + msg)) + + cuda_arch_list = os.environ.get("TORCH_CUDA_ARCH_LIST", None) + if cuda_arch_list: + data.append(("TORCH_CUDA_ARCH_LIST", cuda_arch_list)) + data.append(("Pillow", PIL.__version__)) + + try: + data.append( + ( + "torchvision", + str(torchvision.__version__) + " @" + os.path.dirname(torchvision.__file__), + ) + ) + if has_cuda: + try: + torchvision_C = importlib.util.find_spec("torchvision._C").origin + msg = detect_compute_compatibility(CUDA_HOME, torchvision_C) + data.append(("torchvision arch flags", msg)) + except (ImportError, AttributeError): + data.append(("torchvision._C", "Not found")) + except AttributeError: + data.append(("torchvision", "unknown")) + + try: + import fvcore + + data.append(("fvcore", fvcore.__version__)) + except (ImportError, AttributeError): + pass + + try: + import iopath + + data.append(("iopath", iopath.__version__)) + except (ImportError, AttributeError): + pass + + try: + import cv2 + + data.append(("cv2", cv2.__version__)) + except (ImportError, AttributeError): + data.append(("cv2", "Not found")) + env_str = tabulate(data) + "\n" + env_str += collect_torch_env() + return env_str + + +if __name__ == "__main__": + try: + from detectron2.utils.collect_env import collect_env_info as f + + print(f()) + except ImportError: + print(collect_env_info()) + + if torch.cuda.is_available(): + for k in range(torch.cuda.device_count()): + device = f"cuda:{k}" + try: + x = torch.tensor([1, 2.0], dtype=torch.float32) + x = x.to(device) + except Exception as e: + print( + f"Unable to copy tensor to device={device}: {e}. " + "Your CUDA environment is broken." + ) diff --git a/detectron2/utils/colormap.py b/detectron2/utils/colormap.py new file mode 100644 index 0000000000000000000000000000000000000000..150ccc372262ec4de0b36db66a303cae9495e67f --- /dev/null +++ b/detectron2/utils/colormap.py @@ -0,0 +1,140 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +""" +An awesome colormap for really neat visualizations. +Copied from Detectron, and removed gray colors. +""" + +import numpy as np + +__all__ = ["colormap", "random_color"] + +# fmt: off +# RGB: +_COLORS = np.array( + [ + 0.000, 0.447, 0.741, + 0.850, 0.325, 0.098, + 0.929, 0.694, 0.125, + 0.494, 0.184, 0.556, + 0.466, 0.674, 0.188, + 0.301, 0.745, 0.933, + 0.635, 0.078, 0.184, + 0.300, 0.300, 0.300, + 0.600, 0.600, 0.600, + 1.000, 0.000, 0.000, + 1.000, 0.500, 0.000, + 0.749, 0.749, 0.000, + 0.000, 1.000, 0.000, + 0.000, 0.000, 1.000, + 0.667, 0.000, 1.000, + 0.333, 0.333, 0.000, + 0.333, 0.667, 0.000, + 0.333, 1.000, 0.000, + 0.667, 0.333, 0.000, + 0.667, 0.667, 0.000, + 0.667, 1.000, 0.000, + 1.000, 0.333, 0.000, + 1.000, 0.667, 0.000, + 1.000, 1.000, 0.000, + 0.000, 0.333, 0.500, + 0.000, 0.667, 0.500, + 0.000, 1.000, 0.500, + 0.333, 0.000, 0.500, + 0.333, 0.333, 0.500, + 0.333, 0.667, 0.500, + 0.333, 1.000, 0.500, + 0.667, 0.000, 0.500, + 0.667, 0.333, 0.500, + 0.667, 0.667, 0.500, + 0.667, 1.000, 0.500, + 1.000, 0.000, 0.500, + 1.000, 0.333, 0.500, + 1.000, 0.667, 0.500, + 1.000, 1.000, 0.500, + 0.000, 0.333, 1.000, + 0.000, 0.667, 1.000, + 0.000, 1.000, 1.000, + 0.333, 0.000, 1.000, + 0.333, 0.333, 1.000, + 0.333, 0.667, 1.000, + 0.333, 1.000, 1.000, + 0.667, 0.000, 1.000, + 0.667, 0.333, 1.000, + 0.667, 0.667, 1.000, + 0.667, 1.000, 1.000, + 1.000, 0.000, 1.000, + 1.000, 0.333, 1.000, + 1.000, 0.667, 1.000, + 0.333, 0.000, 0.000, + 0.500, 0.000, 0.000, + 0.667, 0.000, 0.000, + 0.833, 0.000, 0.000, + 1.000, 0.000, 0.000, + 0.000, 0.167, 0.000, + 0.000, 0.333, 0.000, + 0.000, 0.500, 0.000, + 0.000, 0.667, 0.000, + 0.000, 0.833, 0.000, + 0.000, 1.000, 0.000, + 0.000, 0.000, 0.167, + 0.000, 0.000, 0.333, + 0.000, 0.000, 0.500, + 0.000, 0.000, 0.667, + 0.000, 0.000, 0.833, + 0.000, 0.000, 1.000, + 0.000, 0.000, 0.000, + 0.143, 0.143, 0.143, + 0.857, 0.857, 0.857, + 1.000, 1.000, 1.000 + ] +).astype(np.float32).reshape(-1, 3) +# fmt: on + + +def colormap(rgb=False, maximum=255): + """ + Args: + rgb (bool): whether to return RGB colors or BGR colors. + maximum (int): either 255 or 1 + + Returns: + ndarray: a float32 array of Nx3 colors, in range [0, 255] or [0, 1] + """ + assert maximum in [255, 1], maximum + c = _COLORS * maximum + if not rgb: + c = c[:, ::-1] + return c + + +def random_color(rgb=False, maximum=255): + """ + Args: + rgb (bool): whether to return RGB colors or BGR colors. + maximum (int): either 255 or 1 + + Returns: + ndarray: a vector of 3 numbers + """ + idx = np.random.randint(0, len(_COLORS)) + ret = _COLORS[idx] * maximum + if not rgb: + ret = ret[::-1] + return ret + + +if __name__ == "__main__": + import cv2 + + size = 100 + H, W = 10, 10 + canvas = np.random.rand(H * size, W * size, 3).astype("float32") + for h in range(H): + for w in range(W): + idx = h * W + w + if idx >= len(_COLORS): + break + canvas[h * size : (h + 1) * size, w * size : (w + 1) * size] = _COLORS[idx] + cv2.imshow("a", canvas) + cv2.waitKey(0) diff --git a/detectron2/utils/comm.py b/detectron2/utils/comm.py new file mode 100644 index 0000000000000000000000000000000000000000..867788958f95bfa23a12fc1f144a422cd1c9e7c7 --- /dev/null +++ b/detectron2/utils/comm.py @@ -0,0 +1,355 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +""" +This file contains primitives for multi-gpu communication. +This is useful when doing distributed training. +""" + +import functools +import logging +import numpy as np +import pickle +import torch +import torch.distributed as dist +import diffdist +from torch import nn +from torch.nn import functional as F + +_LOCAL_PROCESS_GROUP = None +""" +A torch process group which only includes processes that on the same machine as the current process. +This variable is set when processes are spawned by `launch()` in "engine/launch.py". +""" + + +def get_world_size() -> int: + if not dist.is_available(): + return 1 + if not dist.is_initialized(): + return 1 + return dist.get_world_size() + + +def get_rank() -> int: + if not dist.is_available(): + return 0 + if not dist.is_initialized(): + return 0 + return dist.get_rank() + + +def get_local_rank() -> int: + """ + Returns: + The rank of the current process within the local (per-machine) process group. + """ + if not dist.is_available(): + return 0 + if not dist.is_initialized(): + return 0 + assert _LOCAL_PROCESS_GROUP is not None + return dist.get_rank(group=_LOCAL_PROCESS_GROUP) + + +def get_local_size() -> int: + """ + Returns: + The size of the per-machine process group, + i.e. the number of processes per machine. + """ + if not dist.is_available(): + return 1 + if not dist.is_initialized(): + return 1 + return dist.get_world_size(group=_LOCAL_PROCESS_GROUP) + + +def is_main_process() -> bool: + return get_rank() == 0 + + +def synchronize(): + """ + Helper function to synchronize (barrier) among all processes when + using distributed training + """ + if not dist.is_available(): + return + if not dist.is_initialized(): + return + world_size = dist.get_world_size() + if world_size == 1: + return + dist.barrier() + + +@functools.lru_cache() +def _get_global_gloo_group(): + """ + Return a process group based on gloo backend, containing all the ranks + The result is cached. + """ + if dist.get_backend() == "nccl": + return dist.new_group(backend="gloo") + else: + return dist.group.WORLD + + +def _serialize_to_tensor(data, group): + backend = dist.get_backend(group) + assert backend in ["gloo", "nccl"] + device = torch.device("cpu" if backend == "gloo" else "cuda") + + buffer = pickle.dumps(data) + if len(buffer) > 1024 ** 3: + logger = logging.getLogger(__name__) + logger.warning( + "Rank {} trying to all-gather {:.2f} GB of data on device {}".format( + get_rank(), len(buffer) / (1024 ** 3), device + ) + ) + storage = torch.ByteStorage.from_buffer(buffer) + tensor = torch.ByteTensor(storage).to(device=device) + return tensor + + +def _pad_to_largest_tensor(tensor, group): + """ + Returns: + list[int]: size of the tensor, on each rank + Tensor: padded tensor that has the max size + """ + world_size = dist.get_world_size(group=group) + assert ( + world_size >= 1 + ), "comm.gather/all_gather must be called from ranks within the given group!" + local_size = torch.tensor([tensor.numel()], dtype=torch.int64, device=tensor.device) + size_list = [ + torch.zeros([1], dtype=torch.int64, device=tensor.device) for _ in range(world_size) + ] + dist.all_gather(size_list, local_size, group=group) + size_list = [int(size.item()) for size in size_list] + + max_size = max(size_list) + + # we pad the tensor because torch all_gather does not support + # gathering tensors of different shapes + if local_size != max_size: + padding = torch.zeros((max_size - local_size,), dtype=torch.uint8, device=tensor.device) + tensor = torch.cat((tensor, padding), dim=0) + return size_list, tensor + + +def all_gather(data, group=None): + """ + Run all_gather on arbitrary picklable data (not necessarily tensors). + + Args: + data: any picklable object + group: a torch process group. By default, will use a group which + contains all ranks on gloo backend. + + Returns: + list[data]: list of data gathered from each rank + """ + if get_world_size() == 1: + return [data] + if group is None: + group = _get_global_gloo_group() + if dist.get_world_size(group) == 1: + return [data] + + tensor = _serialize_to_tensor(data, group) + + size_list, tensor = _pad_to_largest_tensor(tensor, group) + max_size = max(size_list) + + # receiving Tensor from all ranks + tensor_list = [ + torch.empty((max_size,), dtype=torch.uint8, device=tensor.device) for _ in size_list + ] + dist.all_gather(tensor_list, tensor, group=group) + + data_list = [] + for size, tensor in zip(size_list, tensor_list): + buffer = tensor.cpu().numpy().tobytes()[:size] + data_list.append(pickle.loads(buffer)) + + return data_list + + +def gather(data, dst=0, group=None): + """ + Run gather on arbitrary picklable data (not necessarily tensors). + + Args: + data: any picklable object + dst (int): destination rank + group: a torch process group. By default, will use a group which + contains all ranks on gloo backend. + + Returns: + list[data]: on dst, a list of data gathered from each rank. Otherwise, + an empty list. + """ + if get_world_size() == 1: + return [data] + if group is None: + group = _get_global_gloo_group() + if dist.get_world_size(group=group) == 1: + return [data] + rank = dist.get_rank(group=group) + + tensor = _serialize_to_tensor(data, group) + size_list, tensor = _pad_to_largest_tensor(tensor, group) + + # receiving Tensor from all ranks + if rank == dst: + max_size = max(size_list) + tensor_list = [ + torch.empty((max_size,), dtype=torch.uint8, device=tensor.device) for _ in size_list + ] + dist.gather(tensor, tensor_list, dst=dst, group=group) + + data_list = [] + for size, tensor in zip(size_list, tensor_list): + buffer = tensor.cpu().numpy().tobytes()[:size] + data_list.append(pickle.loads(buffer)) + return data_list + else: + dist.gather(tensor, [], dst=dst, group=group) + return [] + + +def shared_random_seed(): + """ + Returns: + int: a random number that is the same across all workers. + If workers need a shared RNG, they can use this shared seed to + create one. + + All workers must call this function, otherwise it will deadlock. + """ + ints = np.random.randint(2 ** 31) + all_ints = all_gather(ints) + return all_ints[0] + + +def reduce_dict(input_dict, average=True): + """ + Reduce the values in the dictionary from all processes so that process with rank + 0 has the reduced results. + + Args: + input_dict (dict): inputs to be reduced. All the values must be scalar CUDA Tensor. + average (bool): whether to do average or sum + + Returns: + a dict with the same keys as input_dict, after reduction. + """ + world_size = get_world_size() + if world_size < 2: + return input_dict + with torch.no_grad(): + names = [] + values = [] + # sort the keys so that they are consistent across processes + for k in sorted(input_dict.keys()): + names.append(k) + values.append(input_dict[k]) + values = torch.stack(values, dim=0) + dist.reduce(values, dst=0) + if dist.get_rank() == 0 and average: + # only main process gets accumulated, so only divide by + # world_size in this case + values /= world_size + reduced_dict = {k: v for k, v in zip(names, values)} + return reduced_dict + +def gather_tensors(tensor, method=""): + """ + Performs all_gather operation on the provided tensors. + *** Warning ***: torch.distributed.all_gather has no gradient. + """ + world_size = get_world_size() + rank = get_rank() + + if world_size <= 1: + return tensor, tensor.shape[0] + + batch_size = torch.tensor(tensor.shape[0], device=tensor.device) + batch_size_full = [torch.zeros_like(batch_size) + for _ in range(world_size)] + dist.all_gather(batch_size_full, batch_size) + + # cutting all data to min batch size across all GPUs + min_bs = min([bs.item() for bs in batch_size_full]) + if min_bs < batch_size: + tensor = tensor[:min_bs] + + if "svd" in method: + # curently, svd does not support half-precision + # convert tenosr back to full-precision + with torch.cuda.amp.autocast(enabled=False): + U, Sig, V = torch.svd_lowrank(tensor.cpu(), q=int(method.split("_")[1])) + # gather U + Us = _gather_tensor(U.to(tensor.device), world_size) # N x B x LR + Sigs = _gather_tensor(torch.diag(Sig.to(tensor.device)), world_size) # N x LR x LR + Vs = _gather_tensor(V.to(tensor.device).T, world_size) # N x LR x D + # perform batch mm + # outputs = [] + # for k in range(Us.shape[0]): + # temp = torch.mm(Us[k], Sigs[k]) # B x LR + # output = torch.mm(temp, Vs[k]) # B x D + # outputs.append(output) + # outputs[rank] = tensor + # output = torch.cat(outputs, 0) + + output = torch.bmm(torch.bmm(Us, Sigs), Vs) + output[rank] = tensor + output = output.view(-1, output.shape[-1]) + elif "pca" in method: + raise NotImplementedError + else: + tensors_gather = [ + torch.ones_like(tensor) + for _ in range(world_size) + ] + # dist.all_gather(tensors_gather, tensor, async_op=False) + # need to do this to restore propagation of the gradients + # tensors_gather[rank] = tensor + tensors_gather = diffdist.functional.all_gather(tensors_gather, tensor) + output = torch.cat(tensors_gather, dim=0) + return output, min_bs + +class SoftTargetCrossEntropy(nn.Module): + def __init__(self): + super(SoftTargetCrossEntropy, self).__init__() + + def forward(self, x, target, dim=-1): + loss = torch.sum(-target * F.log_softmax(x, dim=dim), dim=dim) / (torch.sum(target, dim=dim) + 1e-6) + return loss.mean() + +class MILCrossEntropy(nn.Module): + """ + Multi-instance learning loss + """ + def __init__(self): + super(MILCrossEntropy, self).__init__() + + def forward(self, x, target, dim=-1, weights=None, avg_positives=False): + # for numerical stability + logits_max, _ = torch.max(x, dim=1, keepdim=True) + logits = x - logits_max.detach() + exp_logits = torch.exp(logits) + + # get non-zero entries off-diagonal + # identity = torch.eye(target.shape[0]).type_as(target) + # laplacian = 1 - (target - identity) + probs = exp_logits / (exp_logits).sum(dim=dim, keepdim=True) + if avg_positives: # average the logits over positive targets + loss = -torch.log(torch.sum(target * probs, dim=dim) / (torch.sum(target, dim=dim) + 1e-6)) + else: # sum the logits over positive targets + loss = -torch.log(torch.sum(target * probs, dim=dim)) + if weights is not None: + return (loss * weights).mean() + return loss.mean() \ No newline at end of file diff --git a/detectron2/utils/env.py b/detectron2/utils/env.py new file mode 100644 index 0000000000000000000000000000000000000000..40634c17c73273ac8927632be164f466cfe7d1fa --- /dev/null +++ b/detectron2/utils/env.py @@ -0,0 +1,170 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import importlib +import importlib.util +import logging +import numpy as np +import os +import random +import sys +from datetime import datetime +import torch + +__all__ = ["seed_all_rng"] + + +TORCH_VERSION = tuple(int(x) for x in torch.__version__.split(".")[:2]) +""" +PyTorch version as a tuple of 2 ints. Useful for comparison. +""" + + +DOC_BUILDING = os.getenv("_DOC_BUILDING", False) # set in docs/conf.py +""" +Whether we're building documentation. +""" + + +def seed_all_rng(seed=None): + """ + Set the random seed for the RNG in torch, numpy and python. + + Args: + seed (int): if None, will use a strong random seed. + """ + if seed is None: + seed = ( + os.getpid() + + int(datetime.now().strftime("%S%f")) + + int.from_bytes(os.urandom(2), "big") + ) + logger = logging.getLogger(__name__) + logger.info("Using a generated random seed {}".format(seed)) + np.random.seed(seed) + torch.manual_seed(seed) + random.seed(seed) + os.environ["PYTHONHASHSEED"] = str(seed) + + +# from https://stackoverflow.com/questions/67631/how-to-import-a-module-given-the-full-path +def _import_file(module_name, file_path, make_importable=False): + spec = importlib.util.spec_from_file_location(module_name, file_path) + module = importlib.util.module_from_spec(spec) + spec.loader.exec_module(module) + if make_importable: + sys.modules[module_name] = module + return module + + +def _configure_libraries(): + """ + Configurations for some libraries. + """ + # An environment option to disable `import cv2` globally, + # in case it leads to negative performance impact + disable_cv2 = int(os.environ.get("DETECTRON2_DISABLE_CV2", False)) + if disable_cv2: + sys.modules["cv2"] = None + else: + # Disable opencl in opencv since its interaction with cuda often has negative effects + # This envvar is supported after OpenCV 3.4.0 + os.environ["OPENCV_OPENCL_RUNTIME"] = "disabled" + try: + import cv2 + + if int(cv2.__version__.split(".")[0]) >= 3: + cv2.ocl.setUseOpenCL(False) + except ModuleNotFoundError: + # Other types of ImportError, if happened, should not be ignored. + # Because a failed opencv import could mess up address space + # https://github.com/skvark/opencv-python/issues/381 + pass + + def get_version(module, digit=2): + return tuple(map(int, module.__version__.split(".")[:digit])) + + # fmt: off + assert get_version(torch) >= (1, 4), "Requires torch>=1.4" + import fvcore + assert get_version(fvcore, 3) >= (0, 1, 2), "Requires fvcore>=0.1.2" + import yaml + assert get_version(yaml) >= (5, 1), "Requires pyyaml>=5.1" + # fmt: on + + +_ENV_SETUP_DONE = False + + +def setup_environment(): + """Perform environment setup work. The default setup is a no-op, but this + function allows the user to specify a Python source file or a module in + the $DETECTRON2_ENV_MODULE environment variable, that performs + custom setup work that may be necessary to their computing environment. + """ + global _ENV_SETUP_DONE + if _ENV_SETUP_DONE: + return + _ENV_SETUP_DONE = True + + _configure_libraries() + + custom_module_path = os.environ.get("DETECTRON2_ENV_MODULE") + + if custom_module_path: + setup_custom_environment(custom_module_path) + else: + # The default setup is a no-op + pass + + +def setup_custom_environment(custom_module): + """ + Load custom environment setup by importing a Python source file or a + module, and run the setup function. + """ + if custom_module.endswith(".py"): + module = _import_file("detectron2.utils.env.custom_module", custom_module) + else: + module = importlib.import_module(custom_module) + assert hasattr(module, "setup_environment") and callable(module.setup_environment), ( + "Custom environment module defined in {} does not have the " + "required callable attribute 'setup_environment'." + ).format(custom_module) + module.setup_environment() + + +def fixup_module_metadata(module_name, namespace, keys=None): + """ + Fix the __qualname__ of module members to be their exported api name, so + when they are referenced in docs, sphinx can find them. Reference: + https://github.com/python-trio/trio/blob/6754c74eacfad9cc5c92d5c24727a2f3b620624e/trio/_util.py#L216-L241 + """ + if not DOC_BUILDING: + return + seen_ids = set() + + def fix_one(qualname, name, obj): + # avoid infinite recursion (relevant when using + # typing.Generic, for example) + if id(obj) in seen_ids: + return + seen_ids.add(id(obj)) + + mod = getattr(obj, "__module__", None) + if mod is not None and (mod.startswith(module_name) or mod.startswith("fvcore.")): + obj.__module__ = module_name + # Modules, unlike everything else in Python, put fully-qualitied + # names into their __name__ attribute. We check for "." to avoid + # rewriting these. + if hasattr(obj, "__name__") and "." not in obj.__name__: + obj.__name__ = name + obj.__qualname__ = qualname + if isinstance(obj, type): + for attr_name, attr_value in obj.__dict__.items(): + fix_one(objname + "." + attr_name, attr_name, attr_value) + + if keys is None: + keys = namespace.keys() + for objname in keys: + if not objname.startswith("_"): + obj = namespace[objname] + fix_one(objname, objname, obj) diff --git a/detectron2/utils/events.py b/detectron2/utils/events.py new file mode 100644 index 0000000000000000000000000000000000000000..5dee954bdd6ad7dc5ea999562d1d2b03c3a520d9 --- /dev/null +++ b/detectron2/utils/events.py @@ -0,0 +1,486 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import datetime +import json +import logging +import os +import time +from collections import defaultdict +from contextlib import contextmanager +from typing import Optional +import torch +from fvcore.common.history_buffer import HistoryBuffer + +from detectron2.utils.file_io import PathManager + +__all__ = [ + "get_event_storage", + "JSONWriter", + "TensorboardXWriter", + "CommonMetricPrinter", + "EventStorage", +] + +_CURRENT_STORAGE_STACK = [] + + +def get_event_storage(): + """ + Returns: + The :class:`EventStorage` object that's currently being used. + Throws an error if no :class:`EventStorage` is currently enabled. + """ + assert len( + _CURRENT_STORAGE_STACK + ), "get_event_storage() has to be called inside a 'with EventStorage(...)' context!" + return _CURRENT_STORAGE_STACK[-1] + + +class EventWriter: + """ + Base class for writers that obtain events from :class:`EventStorage` and process them. + """ + + def write(self): + raise NotImplementedError + + def close(self): + pass + + +class JSONWriter(EventWriter): + """ + Write scalars to a json file. + + It saves scalars as one json per line (instead of a big json) for easy parsing. + + Examples parsing such a json file: + :: + $ cat metrics.json | jq -s '.[0:2]' + [ + { + "data_time": 0.008433341979980469, + "iteration": 19, + "loss": 1.9228371381759644, + "loss_box_reg": 0.050025828182697296, + "loss_classifier": 0.5316952466964722, + "loss_mask": 0.7236229181289673, + "loss_rpn_box": 0.0856662318110466, + "loss_rpn_cls": 0.48198649287223816, + "lr": 0.007173333333333333, + "time": 0.25401854515075684 + }, + { + "data_time": 0.007216215133666992, + "iteration": 39, + "loss": 1.282649278640747, + "loss_box_reg": 0.06222952902317047, + "loss_classifier": 0.30682939291000366, + "loss_mask": 0.6970193982124329, + "loss_rpn_box": 0.038663312792778015, + "loss_rpn_cls": 0.1471673548221588, + "lr": 0.007706666666666667, + "time": 0.2490077018737793 + } + ] + + $ cat metrics.json | jq '.loss_mask' + 0.7126231789588928 + 0.689423680305481 + 0.6776131987571716 + ... + + """ + + def __init__(self, json_file, window_size=20): + """ + Args: + json_file (str): path to the json file. New data will be appended if the file exists. + window_size (int): the window size of median smoothing for the scalars whose + `smoothing_hint` are True. + """ + self._file_handle = PathManager.open(json_file, "a") + self._window_size = window_size + self._last_write = -1 + + def write(self): + storage = get_event_storage() + to_save = defaultdict(dict) + + for k, (v, iter) in storage.latest_with_smoothing_hint(self._window_size).items(): + # keep scalars that have not been written + if iter <= self._last_write: + continue + to_save[iter][k] = v + if len(to_save): + all_iters = sorted(to_save.keys()) + self._last_write = max(all_iters) + + for itr, scalars_per_iter in to_save.items(): + scalars_per_iter["iteration"] = itr + self._file_handle.write(json.dumps(scalars_per_iter, sort_keys=True) + "\n") + self._file_handle.flush() + try: + os.fsync(self._file_handle.fileno()) + except AttributeError: + pass + + def close(self): + self._file_handle.close() + + +class TensorboardXWriter(EventWriter): + """ + Write all scalars to a tensorboard file. + """ + + def __init__(self, log_dir: str, window_size: int = 20, **kwargs): + """ + Args: + log_dir (str): the directory to save the output events + window_size (int): the scalars will be median-smoothed by this window size + + kwargs: other arguments passed to `torch.utils.tensorboard.SummaryWriter(...)` + """ + self._window_size = window_size + from torch.utils.tensorboard import SummaryWriter + + self._writer = SummaryWriter(log_dir, **kwargs) + self._last_write = -1 + + def write(self): + storage = get_event_storage() + new_last_write = self._last_write + for k, (v, iter) in storage.latest_with_smoothing_hint(self._window_size).items(): + if iter > self._last_write: + self._writer.add_scalar(k, v, iter) + new_last_write = max(new_last_write, iter) + self._last_write = new_last_write + + # storage.put_{image,histogram} is only meant to be used by + # tensorboard writer. So we access its internal fields directly from here. + if len(storage._vis_data) >= 1: + for img_name, img, step_num in storage._vis_data: + self._writer.add_image(img_name, img, step_num) + # Storage stores all image data and rely on this writer to clear them. + # As a result it assumes only one writer will use its image data. + # An alternative design is to let storage store limited recent + # data (e.g. only the most recent image) that all writers can access. + # In that case a writer may not see all image data if its period is long. + storage.clear_images() + + if len(storage._histograms) >= 1: + for params in storage._histograms: + self._writer.add_histogram_raw(**params) + storage.clear_histograms() + + def close(self): + if hasattr(self, "_writer"): # doesn't exist when the code fails at import + self._writer.close() + + +class CommonMetricPrinter(EventWriter): + """ + Print **common** metrics to the terminal, including + iteration time, ETA, memory, all losses, and the learning rate. + It also applies smoothing using a window of 20 elements. + + It's meant to print common metrics in common ways. + To print something in more customized ways, please implement a similar printer by yourself. + """ + + def __init__(self, max_iter: Optional[int] = None, window_size: int = 20): + """ + Args: + max_iter: the maximum number of iterations to train. + Used to compute ETA. If not given, ETA will not be printed. + window_size (int): the losses will be median-smoothed by this window size + """ + self.logger = logging.getLogger(__name__) + self._max_iter = max_iter + self._window_size = window_size + self._last_write = None # (step, time) of last call to write(). Used to compute ETA + + def _get_eta(self, storage) -> Optional[str]: + if self._max_iter is None: + return "" + iteration = storage.iter + try: + eta_seconds = storage.history("time").median(1000) * (self._max_iter - iteration - 1) + storage.put_scalar("eta_seconds", eta_seconds, smoothing_hint=False) + return str(datetime.timedelta(seconds=int(eta_seconds))) + except KeyError: + # estimate eta on our own - more noisy + eta_string = None + if self._last_write is not None: + estimate_iter_time = (time.perf_counter() - self._last_write[1]) / ( + iteration - self._last_write[0] + ) + eta_seconds = estimate_iter_time * (self._max_iter - iteration - 1) + eta_string = str(datetime.timedelta(seconds=int(eta_seconds))) + self._last_write = (iteration, time.perf_counter()) + return eta_string + + def write(self): + storage = get_event_storage() + iteration = storage.iter + if iteration == self._max_iter: + # This hook only reports training progress (loss, ETA, etc) but not other data, + # therefore do not write anything after training succeeds, even if this method + # is called. + return + + try: + data_time = storage.history("data_time").avg(20) + except KeyError: + # they may not exist in the first few iterations (due to warmup) + # or when SimpleTrainer is not used + data_time = None + try: + iter_time = storage.history("time").global_avg() + except KeyError: + iter_time = None + try: + lr = "{:.5g}".format(storage.history("lr").latest()) + except KeyError: + lr = "N/A" + + eta_string = self._get_eta(storage) + + if torch.cuda.is_available(): + max_mem_mb = torch.cuda.max_memory_allocated() / 1024.0 / 1024.0 + else: + max_mem_mb = None + + # NOTE: max_mem is parsed by grep in "dev/parse_results.sh" + self.logger.info( + " {eta}iter: {iter} {losses} {time}{data_time}lr: {lr} {memory}".format( + eta=f"eta: {eta_string} " if eta_string else "", + iter=iteration, + losses=" ".join( + [ + "{}: {:.4g}".format(k, v.median(self._window_size)) + for k, v in storage.histories().items() + if "loss" in k + ] + ), + time="time: {:.4f} ".format(iter_time) if iter_time is not None else "", + data_time="data_time: {:.4f} ".format(data_time) if data_time is not None else "", + lr=lr, + memory="max_mem: {:.0f}M".format(max_mem_mb) if max_mem_mb is not None else "", + ) + ) + + +class EventStorage: + """ + The user-facing class that provides metric storage functionalities. + + In the future we may add support for storing / logging other types of data if needed. + """ + + def __init__(self, start_iter=0): + """ + Args: + start_iter (int): the iteration number to start with + """ + self._history = defaultdict(HistoryBuffer) + self._smoothing_hints = {} + self._latest_scalars = {} + self._iter = start_iter + self._current_prefix = "" + self._vis_data = [] + self._histograms = [] + + def put_image(self, img_name, img_tensor): + """ + Add an `img_tensor` associated with `img_name`, to be shown on + tensorboard. + + Args: + img_name (str): The name of the image to put into tensorboard. + img_tensor (torch.Tensor or numpy.array): An `uint8` or `float` + Tensor of shape `[channel, height, width]` where `channel` is + 3. The image format should be RGB. The elements in img_tensor + can either have values in [0, 1] (float32) or [0, 255] (uint8). + The `img_tensor` will be visualized in tensorboard. + """ + self._vis_data.append((img_name, img_tensor, self._iter)) + + def put_scalar(self, name, value, smoothing_hint=True): + """ + Add a scalar `value` to the `HistoryBuffer` associated with `name`. + + Args: + smoothing_hint (bool): a 'hint' on whether this scalar is noisy and should be + smoothed when logged. The hint will be accessible through + :meth:`EventStorage.smoothing_hints`. A writer may ignore the hint + and apply custom smoothing rule. + + It defaults to True because most scalars we save need to be smoothed to + provide any useful signal. + """ + name = self._current_prefix + name + history = self._history[name] + value = float(value) + history.update(value, self._iter) + self._latest_scalars[name] = (value, self._iter) + + existing_hint = self._smoothing_hints.get(name) + if existing_hint is not None: + assert ( + existing_hint == smoothing_hint + ), "Scalar {} was put with a different smoothing_hint!".format(name) + else: + self._smoothing_hints[name] = smoothing_hint + + def put_scalars(self, *, smoothing_hint=True, **kwargs): + """ + Put multiple scalars from keyword arguments. + + Examples: + + storage.put_scalars(loss=my_loss, accuracy=my_accuracy, smoothing_hint=True) + """ + for k, v in kwargs.items(): + self.put_scalar(k, v, smoothing_hint=smoothing_hint) + + def put_histogram(self, hist_name, hist_tensor, bins=1000): + """ + Create a histogram from a tensor. + + Args: + hist_name (str): The name of the histogram to put into tensorboard. + hist_tensor (torch.Tensor): A Tensor of arbitrary shape to be converted + into a histogram. + bins (int): Number of histogram bins. + """ + ht_min, ht_max = hist_tensor.min().item(), hist_tensor.max().item() + + # Create a histogram with PyTorch + hist_counts = torch.histc(hist_tensor, bins=bins) + hist_edges = torch.linspace(start=ht_min, end=ht_max, steps=bins + 1, dtype=torch.float32) + + # Parameter for the add_histogram_raw function of SummaryWriter + hist_params = dict( + tag=hist_name, + min=ht_min, + max=ht_max, + num=len(hist_tensor), + sum=float(hist_tensor.sum()), + sum_squares=float(torch.sum(hist_tensor ** 2)), + bucket_limits=hist_edges[1:].tolist(), + bucket_counts=hist_counts.tolist(), + global_step=self._iter, + ) + self._histograms.append(hist_params) + + def history(self, name): + """ + Returns: + HistoryBuffer: the scalar history for name + """ + ret = self._history.get(name, None) + if ret is None: + raise KeyError("No history metric available for {}!".format(name)) + return ret + + def histories(self): + """ + Returns: + dict[name -> HistoryBuffer]: the HistoryBuffer for all scalars + """ + return self._history + + def latest(self): + """ + Returns: + dict[str -> (float, int)]: mapping from the name of each scalar to the most + recent value and the iteration number its added. + """ + return self._latest_scalars + + def latest_with_smoothing_hint(self, window_size=20): + """ + Similar to :meth:`latest`, but the returned values + are either the un-smoothed original latest value, + or a median of the given window_size, + depend on whether the smoothing_hint is True. + + This provides a default behavior that other writers can use. + """ + result = {} + for k, (v, itr) in self._latest_scalars.items(): + result[k] = ( + self._history[k].median(window_size) if self._smoothing_hints[k] else v, + itr, + ) + return result + + def smoothing_hints(self): + """ + Returns: + dict[name -> bool]: the user-provided hint on whether the scalar + is noisy and needs smoothing. + """ + return self._smoothing_hints + + def step(self): + """ + User should either: (1) Call this function to increment storage.iter when needed. Or + (2) Set `storage.iter` to the correct iteration number before each iteration. + + The storage will then be able to associate the new data with an iteration number. + """ + self._iter += 1 + + @property + def iter(self): + """ + Returns: + int: The current iteration number. When used together with a trainer, + this is ensured to be the same as trainer.iter. + """ + return self._iter + + @iter.setter + def iter(self, val): + self._iter = int(val) + + @property + def iteration(self): + # for backward compatibility + return self._iter + + def __enter__(self): + _CURRENT_STORAGE_STACK.append(self) + return self + + def __exit__(self, exc_type, exc_val, exc_tb): + assert _CURRENT_STORAGE_STACK[-1] == self + _CURRENT_STORAGE_STACK.pop() + + @contextmanager + def name_scope(self, name): + """ + Yields: + A context within which all the events added to this storage + will be prefixed by the name scope. + """ + old_prefix = self._current_prefix + self._current_prefix = name.rstrip("/") + "/" + yield + self._current_prefix = old_prefix + + def clear_images(self): + """ + Delete all the stored images for visualization. This should be called + after images are written to tensorboard. + """ + self._vis_data = [] + + def clear_histograms(self): + """ + Delete all the stored histograms for visualization. + This should be called after histograms are written to tensorboard. + """ + self._histograms = [] diff --git a/detectron2/utils/file_io.py b/detectron2/utils/file_io.py new file mode 100644 index 0000000000000000000000000000000000000000..46ee4ec31d04eee77976ff3edbbf84762a3409ed --- /dev/null +++ b/detectron2/utils/file_io.py @@ -0,0 +1,37 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from iopath.common.file_io import HTTPURLHandler, OneDrivePathHandler, PathHandler +from iopath.common.file_io import PathManager as PathManagerBase + +__all__ = ["PathManager", "PathHandler"] + + +PathManager = PathManagerBase() +""" +This is a detectron2 project-specific PathManager. +We try to stay away from global PathManager in fvcore as it +introduces potential conflicts among other libraries. +""" + + +class Detectron2Handler(PathHandler): + """ + Resolve anything that's hosted under detectron2's namespace. + """ + + PREFIX = "detectron2://" + S3_DETECTRON2_PREFIX = "https://dl.fbaipublicfiles.com/detectron2/" + + def _get_supported_prefixes(self): + return [self.PREFIX] + + def _get_local_path(self, path, **kwargs): + name = path[len(self.PREFIX) :] + return PathManager.get_local_path(self.S3_DETECTRON2_PREFIX + name, **kwargs) + + def _open(self, path, mode="r", **kwargs): + return PathManager.open(self._get_local_path(path), mode, **kwargs) + + +PathManager.register_handler(HTTPURLHandler()) +PathManager.register_handler(OneDrivePathHandler()) +PathManager.register_handler(Detectron2Handler()) diff --git a/detectron2/utils/logger.py b/detectron2/utils/logger.py new file mode 100644 index 0000000000000000000000000000000000000000..7c7890f8bec5db44098fe1a38d26eb13231f7063 --- /dev/null +++ b/detectron2/utils/logger.py @@ -0,0 +1,237 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import atexit +import functools +import logging +import os +import sys +import time +from collections import Counter +import torch +from tabulate import tabulate +from termcolor import colored + +from detectron2.utils.file_io import PathManager + +__all__ = ["setup_logger", "log_first_n", "log_every_n", "log_every_n_seconds"] + + +class _ColorfulFormatter(logging.Formatter): + def __init__(self, *args, **kwargs): + self._root_name = kwargs.pop("root_name") + "." + self._abbrev_name = kwargs.pop("abbrev_name", "") + if len(self._abbrev_name): + self._abbrev_name = self._abbrev_name + "." + super(_ColorfulFormatter, self).__init__(*args, **kwargs) + + def formatMessage(self, record): + record.name = record.name.replace(self._root_name, self._abbrev_name) + log = super(_ColorfulFormatter, self).formatMessage(record) + if record.levelno == logging.WARNING: + prefix = colored("WARNING", "red", attrs=["blink"]) + elif record.levelno == logging.ERROR or record.levelno == logging.CRITICAL: + prefix = colored("ERROR", "red", attrs=["blink", "underline"]) + else: + return log + return prefix + " " + log + + +@functools.lru_cache() # so that calling setup_logger multiple times won't add many handlers +def setup_logger( + output=None, distributed_rank=0, *, color=True, name="detectron2", abbrev_name=None +): + """ + Initialize the detectron2 logger and set its verbosity level to "DEBUG". + + Args: + output (str): a file name or a directory to save log. If None, will not save log file. + If ends with ".txt" or ".log", assumed to be a file name. + Otherwise, logs will be saved to `output/log.txt`. + name (str): the root module name of this logger + abbrev_name (str): an abbreviation of the module, to avoid long names in logs. + Set to "" to not log the root module in logs. + By default, will abbreviate "detectron2" to "d2" and leave other + modules unchanged. + + Returns: + logging.Logger: a logger + """ + logger = logging.getLogger(name) + logger.setLevel(logging.DEBUG) + logger.propagate = False + + if abbrev_name is None: + abbrev_name = "d2" if name == "detectron2" else name + + plain_formatter = logging.Formatter( + "[%(asctime)s] %(name)s %(levelname)s: %(message)s", datefmt="%m/%d %H:%M:%S" + ) + # stdout logging: master only + if distributed_rank == 0: + ch = logging.StreamHandler(stream=sys.stdout) + ch.setLevel(logging.DEBUG) + if color: + formatter = _ColorfulFormatter( + colored("[%(asctime)s %(name)s]: ", "green") + "%(message)s", + datefmt="%m/%d %H:%M:%S", + root_name=name, + abbrev_name=str(abbrev_name), + ) + else: + formatter = plain_formatter + ch.setFormatter(formatter) + logger.addHandler(ch) + + # file logging: all workers + if output is not None: + if output.endswith(".txt") or output.endswith(".log"): + filename = output + else: + filename = os.path.join(output, "log.txt") + if distributed_rank > 0: + filename = filename + ".rank{}".format(distributed_rank) + PathManager.mkdirs(os.path.dirname(filename)) + + fh = logging.StreamHandler(_cached_log_stream(filename)) + fh.setLevel(logging.DEBUG) + fh.setFormatter(plain_formatter) + logger.addHandler(fh) + + return logger + + +# cache the opened file object, so that different calls to `setup_logger` +# with the same file name can safely write to the same file. +@functools.lru_cache(maxsize=None) +def _cached_log_stream(filename): + # use 1K buffer if writing to cloud storage + io = PathManager.open(filename, "a", buffering=1024 if "://" in filename else -1) + atexit.register(io.close) + return io + + +""" +Below are some other convenient logging methods. +They are mainly adopted from +https://github.com/abseil/abseil-py/blob/master/absl/logging/__init__.py +""" + + +def _find_caller(): + """ + Returns: + str: module name of the caller + tuple: a hashable key to be used to identify different callers + """ + frame = sys._getframe(2) + while frame: + code = frame.f_code + if os.path.join("utils", "logger.") not in code.co_filename: + mod_name = frame.f_globals["__name__"] + if mod_name == "__main__": + mod_name = "detectron2" + return mod_name, (code.co_filename, frame.f_lineno, code.co_name) + frame = frame.f_back + + +_LOG_COUNTER = Counter() +_LOG_TIMER = {} + + +def log_first_n(lvl, msg, n=1, *, name=None, key="caller"): + """ + Log only for the first n times. + + Args: + lvl (int): the logging level + msg (str): + n (int): + name (str): name of the logger to use. Will use the caller's module by default. + key (str or tuple[str]): the string(s) can be one of "caller" or + "message", which defines how to identify duplicated logs. + For example, if called with `n=1, key="caller"`, this function + will only log the first call from the same caller, regardless of + the message content. + If called with `n=1, key="message"`, this function will log the + same content only once, even if they are called from different places. + If called with `n=1, key=("caller", "message")`, this function + will not log only if the same caller has logged the same message before. + """ + if isinstance(key, str): + key = (key,) + assert len(key) > 0 + + caller_module, caller_key = _find_caller() + hash_key = () + if "caller" in key: + hash_key = hash_key + caller_key + if "message" in key: + hash_key = hash_key + (msg,) + + _LOG_COUNTER[hash_key] += 1 + if _LOG_COUNTER[hash_key] <= n: + logging.getLogger(name or caller_module).log(lvl, msg) + + +def log_every_n(lvl, msg, n=1, *, name=None): + """ + Log once per n times. + + Args: + lvl (int): the logging level + msg (str): + n (int): + name (str): name of the logger to use. Will use the caller's module by default. + """ + caller_module, key = _find_caller() + _LOG_COUNTER[key] += 1 + if n == 1 or _LOG_COUNTER[key] % n == 1: + logging.getLogger(name or caller_module).log(lvl, msg) + + +def log_every_n_seconds(lvl, msg, n=1, *, name=None): + """ + Log no more than once per n seconds. + + Args: + lvl (int): the logging level + msg (str): + n (int): + name (str): name of the logger to use. Will use the caller's module by default. + """ + caller_module, key = _find_caller() + last_logged = _LOG_TIMER.get(key, None) + current_time = time.time() + if last_logged is None or current_time - last_logged >= n: + logging.getLogger(name or caller_module).log(lvl, msg) + _LOG_TIMER[key] = current_time + + +def create_small_table(small_dict): + """ + Create a small table using the keys of small_dict as headers. This is only + suitable for small dictionaries. + + Args: + small_dict (dict): a result dictionary of only a few items. + + Returns: + str: the table as a string. + """ + keys, values = tuple(zip(*small_dict.items())) + table = tabulate( + [values], + headers=keys, + tablefmt="pipe", + floatfmt=".3f", + stralign="center", + numalign="center", + ) + return table + + +def _log_api_usage(identifier: str): + """ + Internal function used to log the usage of different detectron2 components + inside facebook's infra. + """ + torch._C._log_api_usage_once("detectron2." + identifier) diff --git a/detectron2/utils/memory.py b/detectron2/utils/memory.py new file mode 100644 index 0000000000000000000000000000000000000000..bd494780b9dbbd1571688cd270bb9b53d113c13e --- /dev/null +++ b/detectron2/utils/memory.py @@ -0,0 +1,84 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +import logging +from contextlib import contextmanager +from functools import wraps +import torch + +__all__ = ["retry_if_cuda_oom"] + + +@contextmanager +def _ignore_torch_cuda_oom(): + """ + A context which ignores CUDA OOM exception from pytorch. + """ + try: + yield + except RuntimeError as e: + # NOTE: the string may change? + if "CUDA out of memory. " in str(e): + pass + else: + raise + + +def retry_if_cuda_oom(func): + """ + Makes a function retry itself after encountering + pytorch's CUDA OOM error. + It will first retry after calling `torch.cuda.empty_cache()`. + + If that still fails, it will then retry by trying to convert inputs to CPUs. + In this case, it expects the function to dispatch to CPU implementation. + The return values may become CPU tensors as well and it's user's + responsibility to convert it back to CUDA tensor if needed. + + Args: + func: a stateless callable that takes tensor-like objects as arguments + + Returns: + a callable which retries `func` if OOM is encountered. + + Examples: + :: + output = retry_if_cuda_oom(some_torch_function)(input1, input2) + # output may be on CPU even if inputs are on GPU + + Note: + 1. When converting inputs to CPU, it will only look at each argument and check + if it has `.device` and `.to` for conversion. Nested structures of tensors + are not supported. + + 2. Since the function might be called more than once, it has to be + stateless. + """ + + def maybe_to_cpu(x): + try: + like_gpu_tensor = x.device.type == "cuda" and hasattr(x, "to") + except AttributeError: + like_gpu_tensor = False + if like_gpu_tensor: + return x.to(device="cpu") + else: + return x + + @wraps(func) + def wrapped(*args, **kwargs): + with _ignore_torch_cuda_oom(): + return func(*args, **kwargs) + + # Clear cache and retry + torch.cuda.empty_cache() + with _ignore_torch_cuda_oom(): + return func(*args, **kwargs) + + # Try on CPU. This slows down the code significantly, therefore print a notice. + logger = logging.getLogger(__name__) + logger.info("Attempting to copy inputs of {} to CPU due to CUDA OOM".format(str(func))) + new_args = (maybe_to_cpu(x) for x in args) + new_kwargs = {k: maybe_to_cpu(v) for k, v in kwargs.items()} + return func(*new_args, **new_kwargs) + + return wrapped diff --git a/detectron2/utils/registry.py b/detectron2/utils/registry.py new file mode 100644 index 0000000000000000000000000000000000000000..4b01e9007c2578a7b5ae555c926cc06c8a3010f9 --- /dev/null +++ b/detectron2/utils/registry.py @@ -0,0 +1,60 @@ +# Copyright (c) Facebook, Inc. and its affiliates. + +from typing import Any +import pydoc +from fvcore.common.registry import Registry # for backward compatibility. + +""" +``Registry`` and `locate` provide ways to map a string (typically found +in config files) to callable objects. +""" + +__all__ = ["Registry", "locate"] + + +def _convert_target_to_string(t: Any) -> str: + """ + Inverse of ``locate()``. + + Args: + t: any object with ``__module__`` and ``__qualname__`` + """ + module, qualname = t.__module__, t.__qualname__ + + # Compress the path to this object, e.g. ``module.submodule._impl.class`` + # may become ``module.submodule.class``, if the later also resolves to the same + # object. This simplifies the string, and also is less affected by moving the + # class implementation. + module_parts = module.split(".") + for k in range(1, len(module_parts)): + prefix = ".".join(module_parts[:k]) + candidate = f"{prefix}.{qualname}" + try: + if locate(candidate) is t: + return candidate + except ImportError: + pass + return f"{module}.{qualname}" + + +def locate(name: str) -> Any: + """ + Locate and return an object ``x`` using an input string ``{x.__module__}.{x.__qualname__}``, + such as "module.submodule.class_name". + + Raise Exception if it cannot be found. + """ + obj = pydoc.locate(name) + + # Some cases (e.g. torch.optim.sgd.SGD) not handled correctly + # by pydoc.locate. Try a private function from hydra. + if obj is None: + try: + # from hydra.utils import get_method - will print many errors + from hydra.utils import _locate + except ImportError as e: + raise ImportError(f"Cannot dynamically locate object {name}!") from e + else: + obj = _locate(name) # it raises if fails + + return obj diff --git a/detectron2/utils/serialize.py b/detectron2/utils/serialize.py new file mode 100644 index 0000000000000000000000000000000000000000..96bb153ec82117d062ad4849237d41d9877e7f9c --- /dev/null +++ b/detectron2/utils/serialize.py @@ -0,0 +1,29 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import cloudpickle + + +class PicklableWrapper(object): + """ + Wrap an object to make it more picklable, note that it uses + heavy weight serialization libraries that are slower than pickle. + It's best to use it only on closures (which are usually not picklable). + + This is a simplified version of + https://github.com/joblib/joblib/blob/master/joblib/externals/loky/cloudpickle_wrapper.py + """ + + def __init__(self, obj): + self._obj = obj + + def __reduce__(self): + s = cloudpickle.dumps(self._obj) + return cloudpickle.loads, (s,) + + def __call__(self, *args, **kwargs): + return self._obj(*args, **kwargs) + + def __getattr__(self, attr): + # Ensure that the wrapped object can be used seamlessly as the previous object. + if attr not in ["_obj"]: + return getattr(self._obj, attr) + return getattr(self, attr) diff --git a/detectron2/utils/testing.py b/detectron2/utils/testing.py new file mode 100644 index 0000000000000000000000000000000000000000..a2e94076e79625d4853f86164e690d1720455e0f --- /dev/null +++ b/detectron2/utils/testing.py @@ -0,0 +1,132 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import io +import numpy as np +import torch + +from detectron2 import model_zoo +from detectron2.data import DatasetCatalog +from detectron2.data.detection_utils import read_image +from detectron2.modeling import build_model +from detectron2.structures import Boxes, Instances, ROIMasks +from detectron2.utils.file_io import PathManager + + +""" +Internal utilities for tests. Don't use except for writing tests. +""" + + +def get_model_no_weights(config_path): + """ + Like model_zoo.get, but do not load any weights (even pretrained) + """ + cfg = model_zoo.get_config(config_path) + if not torch.cuda.is_available(): + cfg.MODEL.DEVICE = "cpu" + return build_model(cfg) + + +def random_boxes(num_boxes, max_coord=100, device="cpu"): + """ + Create a random Nx4 boxes tensor, with coordinates < max_coord. + """ + boxes = torch.rand(num_boxes, 4, device=device) * (max_coord * 0.5) + boxes.clamp_(min=1.0) # tiny boxes cause numerical instability in box regression + # Note: the implementation of this function in torchvision is: + # boxes[:, 2:] += torch.rand(N, 2) * 100 + # but it does not guarantee non-negative widths/heights constraints: + # boxes[:, 2] >= boxes[:, 0] and boxes[:, 3] >= boxes[:, 1]: + boxes[:, 2:] += boxes[:, :2] + return boxes + + +def get_sample_coco_image(tensor=True): + """ + Args: + tensor (bool): if True, returns 3xHxW tensor. + else, returns a HxWx3 numpy array. + + Returns: + an image, in BGR color. + """ + try: + file_name = DatasetCatalog.get("coco_2017_val_100")[0]["file_name"] + if not PathManager.exists(file_name): + raise FileNotFoundError() + except IOError: + # for public CI to run + file_name = "http://images.cocodataset.org/train2017/000000000009.jpg" + ret = read_image(file_name, format="BGR") + if tensor: + ret = torch.from_numpy(np.ascontiguousarray(ret.transpose(2, 0, 1))) + return ret + + +def convert_scripted_instances(instances): + """ + Convert a scripted Instances object to a regular :class:`Instances` object + """ + ret = Instances(instances.image_size) + for name in instances._field_names: + val = getattr(instances, "_" + name, None) + if val is not None: + ret.set(name, val) + return ret + + +def assert_instances_allclose(input, other, *, rtol=1e-5, msg="", size_as_tensor=False): + """ + Args: + input, other (Instances): + size_as_tensor: compare image_size of the Instances as tensors (instead of tuples). + Useful for comparing outputs of tracing. + """ + if not isinstance(input, Instances): + input = convert_scripted_instances(input) + if not isinstance(other, Instances): + other = convert_scripted_instances(other) + + if not msg: + msg = "Two Instances are different! " + else: + msg = msg.rstrip() + " " + + size_error_msg = msg + f"image_size is {input.image_size} vs. {other.image_size}!" + if size_as_tensor: + assert torch.equal( + torch.tensor(input.image_size), torch.tensor(other.image_size) + ), size_error_msg + else: + assert input.image_size == other.image_size, size_error_msg + fields = sorted(input.get_fields().keys()) + fields_other = sorted(other.get_fields().keys()) + assert fields == fields_other, msg + f"Fields are {fields} vs {fields_other}!" + + for f in fields: + val1, val2 = input.get(f), other.get(f) + if isinstance(val1, (Boxes, ROIMasks)): + # boxes in the range of O(100) and can have a larger tolerance + assert torch.allclose(val1.tensor, val2.tensor, atol=100 * rtol), ( + msg + f"Field {f} differs too much!" + ) + elif isinstance(val1, torch.Tensor): + if val1.dtype.is_floating_point: + mag = torch.abs(val1).max().cpu().item() + assert torch.allclose(val1, val2, atol=mag * rtol), ( + msg + f"Field {f} differs too much!" + ) + else: + assert torch.equal(val1, val2), msg + f"Field {f} is different!" + else: + raise ValueError(f"Don't know how to compare type {type(val1)}") + + +def reload_script_model(module): + """ + Save a jit module and load it back. + Similar to the `getExportImportCopy` function in torch/testing/ + """ + buffer = io.BytesIO() + torch.jit.save(module, buffer) + buffer.seek(0) + return torch.jit.load(buffer) diff --git a/detectron2/utils/video_visualizer.py b/detectron2/utils/video_visualizer.py new file mode 100644 index 0000000000000000000000000000000000000000..4287f7cf3152b40c30218c5dbdc41873388db591 --- /dev/null +++ b/detectron2/utils/video_visualizer.py @@ -0,0 +1,235 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import numpy as np +import pycocotools.mask as mask_util + +from detectron2.utils.visualizer import ( + ColorMode, + Visualizer, + _create_text_labels, + _PanopticPrediction, +) + +from .colormap import random_color + + +class _DetectedInstance: + """ + Used to store data about detected objects in video frame, + in order to transfer color to objects in the future frames. + + Attributes: + label (int): + bbox (tuple[float]): + mask_rle (dict): + color (tuple[float]): RGB colors in range (0, 1) + ttl (int): time-to-live for the instance. For example, if ttl=2, + the instance color can be transferred to objects in the next two frames. + """ + + __slots__ = ["label", "bbox", "mask_rle", "color", "ttl"] + + def __init__(self, label, bbox, mask_rle, color, ttl): + self.label = label + self.bbox = bbox + self.mask_rle = mask_rle + self.color = color + self.ttl = ttl + + +class VideoVisualizer: + def __init__(self, metadata, instance_mode=ColorMode.IMAGE): + """ + Args: + metadata (MetadataCatalog): image metadata. + """ + self.metadata = metadata + self._old_instances = [] + assert instance_mode in [ + ColorMode.IMAGE, + ColorMode.IMAGE_BW, + ], "Other mode not supported yet." + self._instance_mode = instance_mode + + def draw_instance_predictions(self, frame, predictions): + """ + Draw instance-level prediction results on an image. + + Args: + frame (ndarray): an RGB image of shape (H, W, C), in the range [0, 255]. + predictions (Instances): the output of an instance detection/segmentation + model. Following fields will be used to draw: + "pred_boxes", "pred_classes", "scores", "pred_masks" (or "pred_masks_rle"). + + Returns: + output (VisImage): image object with visualizations. + """ + frame_visualizer = Visualizer(frame, self.metadata) + num_instances = len(predictions) + if num_instances == 0: + return frame_visualizer.output + + boxes = predictions.pred_boxes.tensor.numpy() if predictions.has("pred_boxes") else None + scores = predictions.scores if predictions.has("scores") else None + classes = predictions.pred_classes.numpy() if predictions.has("pred_classes") else None + keypoints = predictions.pred_keypoints if predictions.has("pred_keypoints") else None + + if predictions.has("pred_masks"): + masks = predictions.pred_masks + # mask IOU is not yet enabled + # masks_rles = mask_util.encode(np.asarray(masks.permute(1, 2, 0), order="F")) + # assert len(masks_rles) == num_instances + else: + masks = None + + detected = [ + _DetectedInstance(classes[i], boxes[i], mask_rle=None, color=None, ttl=8) + for i in range(num_instances) + ] + colors = self._assign_colors(detected) + + labels = _create_text_labels(classes, scores, self.metadata.get("thing_classes", None)) + + if self._instance_mode == ColorMode.IMAGE_BW: + # any() returns uint8 tensor + frame_visualizer.output.img = frame_visualizer._create_grayscale_image( + (masks.any(dim=0) > 0).numpy() if masks is not None else None + ) + alpha = 0.3 + else: + alpha = 0.5 + + frame_visualizer.overlay_instances( + boxes=None if masks is not None else boxes, # boxes are a bit distracting + masks=masks, + labels=labels, + keypoints=keypoints, + assigned_colors=colors, + alpha=alpha, + ) + + return frame_visualizer.output + + def draw_sem_seg(self, frame, sem_seg, area_threshold=None): + """ + Args: + sem_seg (ndarray or Tensor): semantic segmentation of shape (H, W), + each value is the integer label. + area_threshold (Optional[int]): only draw segmentations larger than the threshold + """ + # don't need to do anything special + frame_visualizer = Visualizer(frame, self.metadata) + frame_visualizer.draw_sem_seg(sem_seg, area_threshold=None) + return frame_visualizer.output + + def draw_panoptic_seg_predictions( + self, frame, panoptic_seg, segments_info, area_threshold=None, alpha=0.5 + ): + frame_visualizer = Visualizer(frame, self.metadata) + pred = _PanopticPrediction(panoptic_seg, segments_info, self.metadata) + + if self._instance_mode == ColorMode.IMAGE_BW: + frame_visualizer.output.img = frame_visualizer._create_grayscale_image( + pred.non_empty_mask() + ) + + # draw mask for all semantic segments first i.e. "stuff" + for mask, sinfo in pred.semantic_masks(): + category_idx = sinfo["category_id"] + try: + mask_color = [x / 255 for x in self.metadata.stuff_colors[category_idx]] + except AttributeError: + mask_color = None + + frame_visualizer.draw_binary_mask( + mask, + color=mask_color, + text=self.metadata.stuff_classes[category_idx], + alpha=alpha, + area_threshold=area_threshold, + ) + + all_instances = list(pred.instance_masks()) + if len(all_instances) == 0: + return frame_visualizer.output + # draw mask for all instances second + masks, sinfo = list(zip(*all_instances)) + num_instances = len(masks) + masks_rles = mask_util.encode( + np.asarray(np.asarray(masks).transpose(1, 2, 0), dtype=np.uint8, order="F") + ) + assert len(masks_rles) == num_instances + + category_ids = [x["category_id"] for x in sinfo] + detected = [ + _DetectedInstance(category_ids[i], bbox=None, mask_rle=masks_rles[i], color=None, ttl=8) + for i in range(num_instances) + ] + colors = self._assign_colors(detected) + labels = [self.metadata.thing_classes[k] for k in category_ids] + + frame_visualizer.overlay_instances( + boxes=None, + masks=masks, + labels=labels, + keypoints=None, + assigned_colors=colors, + alpha=alpha, + ) + return frame_visualizer.output + + def _assign_colors(self, instances): + """ + Naive tracking heuristics to assign same color to the same instance, + will update the internal state of tracked instances. + + Returns: + list[tuple[float]]: list of colors. + """ + + # Compute iou with either boxes or masks: + is_crowd = np.zeros((len(instances),), dtype=np.bool) + if instances[0].bbox is None: + assert instances[0].mask_rle is not None + # use mask iou only when box iou is None + # because box seems good enough + rles_old = [x.mask_rle for x in self._old_instances] + rles_new = [x.mask_rle for x in instances] + ious = mask_util.iou(rles_old, rles_new, is_crowd) + threshold = 0.5 + else: + boxes_old = [x.bbox for x in self._old_instances] + boxes_new = [x.bbox for x in instances] + ious = mask_util.iou(boxes_old, boxes_new, is_crowd) + threshold = 0.6 + if len(ious) == 0: + ious = np.zeros((len(self._old_instances), len(instances)), dtype="float32") + + # Only allow matching instances of the same label: + for old_idx, old in enumerate(self._old_instances): + for new_idx, new in enumerate(instances): + if old.label != new.label: + ious[old_idx, new_idx] = 0 + + matched_new_per_old = np.asarray(ious).argmax(axis=1) + max_iou_per_old = np.asarray(ious).max(axis=1) + + # Try to find match for each old instance: + extra_instances = [] + for idx, inst in enumerate(self._old_instances): + if max_iou_per_old[idx] > threshold: + newidx = matched_new_per_old[idx] + if instances[newidx].color is None: + instances[newidx].color = inst.color + continue + # If an old instance does not match any new instances, + # keep it for the next frame in case it is just missed by the detector + inst.ttl -= 1 + if inst.ttl > 0: + extra_instances.append(inst) + + # Assign random color to newly-detected instances: + for inst in instances: + if inst.color is None: + inst.color = random_color(rgb=True, maximum=1) + self._old_instances = instances[:] + extra_instances + return [d.color for d in instances] diff --git a/detectron2/utils/visualizer.py b/detectron2/utils/visualizer.py new file mode 100644 index 0000000000000000000000000000000000000000..2540102c48229ba9db2b667d94bc1f85d0131114 --- /dev/null +++ b/detectron2/utils/visualizer.py @@ -0,0 +1,1219 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import colorsys +import logging +import math +import numpy as np +from enum import Enum, unique +import cv2 +import matplotlib as mpl +import matplotlib.colors as mplc +import matplotlib.figure as mplfigure +import pycocotools.mask as mask_util +import torch +from matplotlib.backends.backend_agg import FigureCanvasAgg +from PIL import Image + +from detectron2.data import MetadataCatalog +from detectron2.structures import BitMasks, Boxes, BoxMode, Keypoints, PolygonMasks, RotatedBoxes +from detectron2.utils.file_io import PathManager + +from .colormap import random_color + +logger = logging.getLogger(__name__) + +__all__ = ["ColorMode", "VisImage", "Visualizer"] + + +_SMALL_OBJECT_AREA_THRESH = 1000 +_LARGE_MASK_AREA_THRESH = 120000 +_OFF_WHITE = (1.0, 1.0, 240.0 / 255) +_BLACK = (0, 0, 0) +_RED = (1.0, 0, 0) + +_KEYPOINT_THRESHOLD = 0.05 + + +@unique +class ColorMode(Enum): + """ + Enum of different color modes to use for instance visualizations. + """ + + IMAGE = 0 + """ + Picks a random color for every instance and overlay segmentations with low opacity. + """ + SEGMENTATION = 1 + """ + Let instances of the same category have similar colors + (from metadata.thing_colors), and overlay them with + high opacity. This provides more attention on the quality of segmentation. + """ + IMAGE_BW = 2 + """ + Same as IMAGE, but convert all areas without masks to gray-scale. + Only available for drawing per-instance mask predictions. + """ + + +class GenericMask: + """ + Attribute: + polygons (list[ndarray]): list[ndarray]: polygons for this mask. + Each ndarray has format [x, y, x, y, ...] + mask (ndarray): a binary mask + """ + + def __init__(self, mask_or_polygons, height, width): + self._mask = self._polygons = self._has_holes = None + self.height = height + self.width = width + + m = mask_or_polygons + if isinstance(m, dict): + # RLEs + assert "counts" in m and "size" in m + if isinstance(m["counts"], list): # uncompressed RLEs + h, w = m["size"] + assert h == height and w == width + m = mask_util.frPyObjects(m, h, w) + self._mask = mask_util.decode(m)[:, :] + return + + if isinstance(m, list): # list[ndarray] + self._polygons = [np.asarray(x).reshape(-1) for x in m] + return + + if isinstance(m, np.ndarray): # assumed to be a binary mask + assert m.shape[1] != 2, m.shape + assert m.shape == (height, width), m.shape + self._mask = m.astype("uint8") + return + + raise ValueError("GenericMask cannot handle object {} of type '{}'".format(m, type(m))) + + @property + def mask(self): + if self._mask is None: + self._mask = self.polygons_to_mask(self._polygons) + return self._mask + + @property + def polygons(self): + if self._polygons is None: + self._polygons, self._has_holes = self.mask_to_polygons(self._mask) + return self._polygons + + @property + def has_holes(self): + if self._has_holes is None: + if self._mask is not None: + self._polygons, self._has_holes = self.mask_to_polygons(self._mask) + else: + self._has_holes = False # if original format is polygon, does not have holes + return self._has_holes + + def mask_to_polygons(self, mask): + # cv2.RETR_CCOMP flag retrieves all the contours and arranges them to a 2-level + # hierarchy. External contours (boundary) of the object are placed in hierarchy-1. + # Internal contours (holes) are placed in hierarchy-2. + # cv2.CHAIN_APPROX_NONE flag gets vertices of polygons from contours. + mask = np.ascontiguousarray(mask) # some versions of cv2 does not support incontiguous arr + res = cv2.findContours(mask.astype("uint8"), cv2.RETR_CCOMP, cv2.CHAIN_APPROX_NONE) + hierarchy = res[-1] + if hierarchy is None: # empty mask + return [], False + has_holes = (hierarchy.reshape(-1, 4)[:, 3] >= 0).sum() > 0 + res = res[-2] + res = [x.flatten() for x in res] + # These coordinates from OpenCV are integers in range [0, W-1 or H-1]. + # We add 0.5 to turn them into real-value coordinate space. A better solution + # would be to first +0.5 and then dilate the returned polygon by 0.5. + res = [x + 0.5 for x in res if len(x) >= 6] + return res, has_holes + + def polygons_to_mask(self, polygons): + rle = mask_util.frPyObjects(polygons, self.height, self.width) + rle = mask_util.merge(rle) + return mask_util.decode(rle)[:, :] + + def area(self): + return self.mask.sum() + + def bbox(self): + p = mask_util.frPyObjects(self.polygons, self.height, self.width) + p = mask_util.merge(p) + bbox = mask_util.toBbox(p) + bbox[2] += bbox[0] + bbox[3] += bbox[1] + return bbox + + +class _PanopticPrediction: + """ + Unify different panoptic annotation/prediction formats + """ + + def __init__(self, panoptic_seg, segments_info, metadata=None): + if segments_info is None: + assert metadata is not None + # If "segments_info" is None, we assume "panoptic_img" is a + # H*W int32 image storing the panoptic_id in the format of + # category_id * label_divisor + instance_id. We reserve -1 for + # VOID label. + label_divisor = metadata.label_divisor + segments_info = [] + for panoptic_label in np.unique(panoptic_seg.numpy()): + if panoptic_label == -1: + # VOID region. + continue + pred_class = panoptic_label // label_divisor + isthing = pred_class in metadata.thing_dataset_id_to_contiguous_id.values() + segments_info.append( + { + "id": int(panoptic_label), + "category_id": int(pred_class), + "isthing": bool(isthing), + } + ) + del metadata + + self._seg = panoptic_seg + + self._sinfo = {s["id"]: s for s in segments_info} # seg id -> seg info + segment_ids, areas = torch.unique(panoptic_seg, sorted=True, return_counts=True) + areas = areas.numpy() + sorted_idxs = np.argsort(-areas) + self._seg_ids, self._seg_areas = segment_ids[sorted_idxs], areas[sorted_idxs] + self._seg_ids = self._seg_ids.tolist() + for sid, area in zip(self._seg_ids, self._seg_areas): + if sid in self._sinfo: + self._sinfo[sid]["area"] = float(area) + + def non_empty_mask(self): + """ + Returns: + (H, W) array, a mask for all pixels that have a prediction + """ + empty_ids = [] + for id in self._seg_ids: + if id not in self._sinfo: + empty_ids.append(id) + if len(empty_ids) == 0: + return np.zeros(self._seg.shape, dtype=np.uint8) + assert ( + len(empty_ids) == 1 + ), ">1 ids corresponds to no labels. This is currently not supported" + return (self._seg != empty_ids[0]).numpy().astype(np.bool) + + def semantic_masks(self): + for sid in self._seg_ids: + sinfo = self._sinfo.get(sid) + if sinfo is None or sinfo["isthing"]: + # Some pixels (e.g. id 0 in PanopticFPN) have no instance or semantic predictions. + continue + yield (self._seg == sid).numpy().astype(np.bool), sinfo + + def instance_masks(self): + for sid in self._seg_ids: + sinfo = self._sinfo.get(sid) + if sinfo is None or not sinfo["isthing"]: + continue + mask = (self._seg == sid).numpy().astype(np.bool) + if mask.sum() > 0: + yield mask, sinfo + + +def _create_text_labels(classes, scores, class_names, is_crowd=None): + """ + Args: + classes (list[int] or None): + scores (list[float] or None): + class_names (list[str] or None): + is_crowd (list[bool] or None): + + Returns: + list[str] or None + """ + labels = None + if classes is not None: + if class_names is not None and len(class_names) > 0: + labels = [class_names[i] for i in classes] + else: + labels = [str(i) for i in classes] + if scores is not None: + if labels is None: + labels = ["{:.0f}%".format(s * 100) for s in scores] + else: + labels = ["{} {:.0f}%".format(l, s * 100) for l, s in zip(labels, scores)] + if labels is not None and is_crowd is not None: + labels = [l + ("|crowd" if crowd else "") for l, crowd in zip(labels, is_crowd)] + return labels + + +class VisImage: + def __init__(self, img, scale=1.0): + """ + Args: + img (ndarray): an RGB image of shape (H, W, 3). + scale (float): scale the input image + """ + self.img = img + self.scale = scale + self.width, self.height = img.shape[1], img.shape[0] + self._setup_figure(img) + + def _setup_figure(self, img): + """ + Args: + Same as in :meth:`__init__()`. + + Returns: + fig (matplotlib.pyplot.figure): top level container for all the image plot elements. + ax (matplotlib.pyplot.Axes): contains figure elements and sets the coordinate system. + """ + fig = mplfigure.Figure(frameon=False) + self.dpi = fig.get_dpi() + # add a small 1e-2 to avoid precision lost due to matplotlib's truncation + # (https://github.com/matplotlib/matplotlib/issues/15363) + fig.set_size_inches( + (self.width * self.scale + 1e-2) / self.dpi, + (self.height * self.scale + 1e-2) / self.dpi, + ) + self.canvas = FigureCanvasAgg(fig) + # self.canvas = mpl.backends.backend_cairo.FigureCanvasCairo(fig) + ax = fig.add_axes([0.0, 0.0, 1.0, 1.0]) + ax.axis("off") + # Need to imshow this first so that other patches can be drawn on top + ax.imshow(img, extent=(0, self.width, self.height, 0), interpolation="nearest") + + self.fig = fig + self.ax = ax + + def save(self, filepath): + """ + Args: + filepath (str): a string that contains the absolute path, including the file name, where + the visualized image will be saved. + """ + self.fig.savefig(filepath) + + def get_image(self): + """ + Returns: + ndarray: + the visualized image of shape (H, W, 3) (RGB) in uint8 type. + The shape is scaled w.r.t the input image using the given `scale` argument. + """ + canvas = self.canvas + s, (width, height) = canvas.print_to_buffer() + # buf = io.BytesIO() # works for cairo backend + # canvas.print_rgba(buf) + # width, height = self.width, self.height + # s = buf.getvalue() + + buffer = np.frombuffer(s, dtype="uint8") + + img_rgba = buffer.reshape(height, width, 4) + rgb, alpha = np.split(img_rgba, [3], axis=2) + return rgb.astype("uint8") + + +class Visualizer: + """ + Visualizer that draws data about detection/segmentation on images. + + It contains methods like `draw_{text,box,circle,line,binary_mask,polygon}` + that draw primitive objects to images, as well as high-level wrappers like + `draw_{instance_predictions,sem_seg,panoptic_seg_predictions,dataset_dict}` + that draw composite data in some pre-defined style. + + Note that the exact visualization style for the high-level wrappers are subject to change. + Style such as color, opacity, label contents, visibility of labels, or even the visibility + of objects themselves (e.g. when the object is too small) may change according + to different heuristics, as long as the results still look visually reasonable. + + To obtain a consistent style, you can implement custom drawing functions with the + abovementioned primitive methods instead. If you need more customized visualization + styles, you can process the data yourself following their format documented in + tutorials (:doc:`/tutorials/models`, :doc:`/tutorials/datasets`). This class does not + intend to satisfy everyone's preference on drawing styles. + + This visualizer focuses on high rendering quality rather than performance. It is not + designed to be used for real-time applications. + """ + + # TODO implement a fast, rasterized version using OpenCV + + def __init__(self, img_rgb, metadata=None, scale=1.0, instance_mode=ColorMode.IMAGE): + """ + Args: + img_rgb: a numpy array of shape (H, W, C), where H and W correspond to + the height and width of the image respectively. C is the number of + color channels. The image is required to be in RGB format since that + is a requirement of the Matplotlib library. The image is also expected + to be in the range [0, 255]. + metadata (Metadata): dataset metadata (e.g. class names and colors) + instance_mode (ColorMode): defines one of the pre-defined style for drawing + instances on an image. + """ + self.img = np.asarray(img_rgb).clip(0, 255).astype(np.uint8) + if metadata is None: + metadata = MetadataCatalog.get("__nonexist__") + self.metadata = metadata + self.output = VisImage(self.img, scale=scale) + self.cpu_device = torch.device("cpu") + + # too small texts are useless, therefore clamp to 9 + self._default_font_size = max( + np.sqrt(self.output.height * self.output.width) // 90, 10 // scale + ) + self._instance_mode = instance_mode + + def draw_instance_predictions(self, predictions): + """ + Draw instance-level prediction results on an image. + + Args: + predictions (Instances): the output of an instance detection/segmentation + model. Following fields will be used to draw: + "pred_boxes", "pred_classes", "scores", "pred_masks" (or "pred_masks_rle"). + + Returns: + output (VisImage): image object with visualizations. + """ + boxes = predictions.pred_boxes if predictions.has("pred_boxes") else None + scores = predictions.scores if predictions.has("scores") else None + classes = predictions.pred_classes.tolist() if predictions.has("pred_classes") else None + labels = _create_text_labels(classes, scores, self.metadata.get("thing_classes", None)) + keypoints = predictions.pred_keypoints if predictions.has("pred_keypoints") else None + + if predictions.has("pred_masks"): + masks = np.asarray(predictions.pred_masks) + masks = [GenericMask(x, self.output.height, self.output.width) for x in masks] + else: + masks = None + + if self._instance_mode == ColorMode.SEGMENTATION and self.metadata.get("thing_colors"): + colors = [ + self._jitter([x / 255 for x in self.metadata.thing_colors[c]]) for c in classes + ] + alpha = 0.8 + else: + colors = None + alpha = 0.5 + + if self._instance_mode == ColorMode.IMAGE_BW: + self.output.img = self._create_grayscale_image( + (predictions.pred_masks.any(dim=0) > 0).numpy() + if predictions.has("pred_masks") + else None + ) + alpha = 0.3 + + self.overlay_instances( + masks=masks, + boxes=boxes, + labels=labels, + keypoints=keypoints, + assigned_colors=colors, + alpha=alpha, + ) + return self.output + + def draw_sem_seg(self, sem_seg, area_threshold=None, alpha=0.8): + """ + Draw semantic segmentation predictions/labels. + + Args: + sem_seg (Tensor or ndarray): the segmentation of shape (H, W). + Each value is the integer label of the pixel. + area_threshold (int): segments with less than `area_threshold` are not drawn. + alpha (float): the larger it is, the more opaque the segmentations are. + + Returns: + output (VisImage): image object with visualizations. + """ + if isinstance(sem_seg, torch.Tensor): + sem_seg = sem_seg.numpy() + labels, areas = np.unique(sem_seg, return_counts=True) + sorted_idxs = np.argsort(-areas).tolist() + labels = labels[sorted_idxs] + for label in filter(lambda l: l < len(self.metadata.stuff_classes), labels): + try: + mask_color = [x / 255 for x in self.metadata.stuff_colors[label]] + except (AttributeError, IndexError): + mask_color = None + + binary_mask = (sem_seg == label).astype(np.uint8) + text = self.metadata.stuff_classes[label] + self.draw_binary_mask( + binary_mask, + color=mask_color, + edge_color=_OFF_WHITE, + text=text, + alpha=alpha, + area_threshold=area_threshold, + ) + return self.output + + def draw_panoptic_seg(self, panoptic_seg, segments_info, area_threshold=None, alpha=0.7): + """ + Draw panoptic prediction annotations or results. + + Args: + panoptic_seg (Tensor): of shape (height, width) where the values are ids for each + segment. + segments_info (list[dict] or None): Describe each segment in `panoptic_seg`. + If it is a ``list[dict]``, each dict contains keys "id", "category_id". + If None, category id of each pixel is computed by + ``pixel // metadata.label_divisor``. + area_threshold (int): stuff segments with less than `area_threshold` are not drawn. + + Returns: + output (VisImage): image object with visualizations. + """ + pred = _PanopticPrediction(panoptic_seg, segments_info, self.metadata) + + if self._instance_mode == ColorMode.IMAGE_BW: + self.output.img = self._create_grayscale_image(pred.non_empty_mask()) + + # draw mask for all semantic segments first i.e. "stuff" + for mask, sinfo in pred.semantic_masks(): + category_idx = sinfo["category_id"] + try: + mask_color = [x / 255 for x in self.metadata.stuff_colors[category_idx]] + except AttributeError: + mask_color = None + + text = self.metadata.stuff_classes[category_idx] + self.draw_binary_mask( + mask, + color=mask_color, + edge_color=_OFF_WHITE, + text=text, + alpha=alpha, + area_threshold=area_threshold, + ) + + # draw mask for all instances second + all_instances = list(pred.instance_masks()) + if len(all_instances) == 0: + return self.output + masks, sinfo = list(zip(*all_instances)) + category_ids = [x["category_id"] for x in sinfo] + + try: + scores = [x["score"] for x in sinfo] + except KeyError: + scores = None + labels = _create_text_labels( + category_ids, scores, self.metadata.thing_classes, [x.get("iscrowd", 0) for x in sinfo] + ) + + try: + colors = [ + self._jitter([x / 255 for x in self.metadata.thing_colors[c]]) for c in category_ids + ] + except AttributeError: + colors = None + self.overlay_instances(masks=masks, labels=labels, assigned_colors=colors, alpha=alpha) + + return self.output + + draw_panoptic_seg_predictions = draw_panoptic_seg # backward compatibility + + def draw_dataset_dict(self, dic): + """ + Draw annotations/segmentaions in Detectron2 Dataset format. + + Args: + dic (dict): annotation/segmentation data of one image, in Detectron2 Dataset format. + + Returns: + output (VisImage): image object with visualizations. + """ + annos = dic.get("annotations", None) + if annos: + if "segmentation" in annos[0]: + masks = [x["segmentation"] for x in annos] + else: + masks = None + if "keypoints" in annos[0]: + keypts = [x["keypoints"] for x in annos] + keypts = np.array(keypts).reshape(len(annos), -1, 3) + else: + keypts = None + + boxes = [ + BoxMode.convert(x["bbox"], x["bbox_mode"], BoxMode.XYXY_ABS) + if len(x["bbox"]) == 4 + else x["bbox"] + for x in annos + ] + + colors = None + category_ids = [x["category_id"] for x in annos] + if self._instance_mode == ColorMode.SEGMENTATION and self.metadata.get("thing_colors"): + colors = [ + self._jitter([x / 255 for x in self.metadata.thing_colors[c]]) + for c in category_ids + ] + names = self.metadata.get("thing_classes", None) + labels = _create_text_labels( + category_ids, + scores=None, + class_names=names, + is_crowd=[x.get("iscrowd", 0) for x in annos], + ) + self.overlay_instances( + labels=labels, boxes=boxes, masks=masks, keypoints=keypts, assigned_colors=colors + ) + + sem_seg = dic.get("sem_seg", None) + if sem_seg is None and "sem_seg_file_name" in dic: + with PathManager.open(dic["sem_seg_file_name"], "rb") as f: + sem_seg = Image.open(f) + sem_seg = np.asarray(sem_seg, dtype="uint8") + if sem_seg is not None: + self.draw_sem_seg(sem_seg, area_threshold=0, alpha=0.5) + + pan_seg = dic.get("pan_seg", None) + if pan_seg is None and "pan_seg_file_name" in dic: + with PathManager.open(dic["pan_seg_file_name"], "rb") as f: + pan_seg = Image.open(f) + pan_seg = np.asarray(pan_seg) + from panopticapi.utils import rgb2id + + pan_seg = rgb2id(pan_seg) + if pan_seg is not None: + segments_info = dic["segments_info"] + pan_seg = torch.tensor(pan_seg) + self.draw_panoptic_seg(pan_seg, segments_info, area_threshold=0, alpha=0.5) + return self.output + + def overlay_instances( + self, + *, + boxes=None, + labels=None, + masks=None, + keypoints=None, + assigned_colors=None, + alpha=0.5 + ): + """ + Args: + boxes (Boxes, RotatedBoxes or ndarray): either a :class:`Boxes`, + or an Nx4 numpy array of XYXY_ABS format for the N objects in a single image, + or a :class:`RotatedBoxes`, + or an Nx5 numpy array of (x_center, y_center, width, height, angle_degrees) format + for the N objects in a single image, + labels (list[str]): the text to be displayed for each instance. + masks (masks-like object): Supported types are: + + * :class:`detectron2.structures.PolygonMasks`, + :class:`detectron2.structures.BitMasks`. + * list[list[ndarray]]: contains the segmentation masks for all objects in one image. + The first level of the list corresponds to individual instances. The second + level to all the polygon that compose the instance, and the third level + to the polygon coordinates. The third level should have the format of + [x0, y0, x1, y1, ..., xn, yn] (n >= 3). + * list[ndarray]: each ndarray is a binary mask of shape (H, W). + * list[dict]: each dict is a COCO-style RLE. + keypoints (Keypoint or array like): an array-like object of shape (N, K, 3), + where the N is the number of instances and K is the number of keypoints. + The last dimension corresponds to (x, y, visibility or score). + assigned_colors (list[matplotlib.colors]): a list of colors, where each color + corresponds to each mask or box in the image. Refer to 'matplotlib.colors' + for full list of formats that the colors are accepted in. + + Returns: + output (VisImage): image object with visualizations. + """ + num_instances = 0 + if boxes is not None: + boxes = self._convert_boxes(boxes) + num_instances = len(boxes) + if masks is not None: + masks = self._convert_masks(masks) + if num_instances: + assert len(masks) == num_instances + else: + num_instances = len(masks) + if keypoints is not None: + if num_instances: + assert len(keypoints) == num_instances + else: + num_instances = len(keypoints) + keypoints = self._convert_keypoints(keypoints) + if labels is not None: + assert len(labels) == num_instances + if assigned_colors is None: + assigned_colors = [random_color(rgb=True, maximum=1) for _ in range(num_instances)] + if num_instances == 0: + return self.output + if boxes is not None and boxes.shape[1] == 5: + return self.overlay_rotated_instances( + boxes=boxes, labels=labels, assigned_colors=assigned_colors + ) + + # Display in largest to smallest order to reduce occlusion. + areas = None + if boxes is not None: + areas = np.prod(boxes[:, 2:] - boxes[:, :2], axis=1) + elif masks is not None: + areas = np.asarray([x.area() for x in masks]) + + if areas is not None: + sorted_idxs = np.argsort(-areas).tolist() + # Re-order overlapped instances in descending order. + boxes = boxes[sorted_idxs] if boxes is not None else None + labels = [labels[k] for k in sorted_idxs] if labels is not None else None + masks = [masks[idx] for idx in sorted_idxs] if masks is not None else None + assigned_colors = [assigned_colors[idx] for idx in sorted_idxs] + keypoints = keypoints[sorted_idxs] if keypoints is not None else None + + for i in range(num_instances): + color = assigned_colors[i] + if boxes is not None: + self.draw_box(boxes[i], edge_color=color) + + if masks is not None: + for segment in masks[i].polygons: + self.draw_polygon(segment.reshape(-1, 2), color, alpha=alpha) + + if labels is not None: + # first get a box + if boxes is not None: + x0, y0, x1, y1 = boxes[i] + text_pos = (x0, y0) # if drawing boxes, put text on the box corner. + horiz_align = "left" + elif masks is not None: + # skip small mask without polygon + if len(masks[i].polygons) == 0: + continue + + x0, y0, x1, y1 = masks[i].bbox() + + # draw text in the center (defined by median) when box is not drawn + # median is less sensitive to outliers. + text_pos = np.median(masks[i].mask.nonzero(), axis=1)[::-1] + horiz_align = "center" + else: + continue # drawing the box confidence for keypoints isn't very useful. + # for small objects, draw text at the side to avoid occlusion + instance_area = (y1 - y0) * (x1 - x0) + if ( + instance_area < _SMALL_OBJECT_AREA_THRESH * self.output.scale + or y1 - y0 < 40 * self.output.scale + ): + if y1 >= self.output.height - 5: + text_pos = (x1, y0) + else: + text_pos = (x0, y1) + + height_ratio = (y1 - y0) / np.sqrt(self.output.height * self.output.width) + lighter_color = self._change_color_brightness(color, brightness_factor=0.7) + font_size = ( + np.clip((height_ratio - 0.02) / 0.08 + 1, 1.2, 2) + * 0.5 + * self._default_font_size + ) + self.draw_text( + labels[i], + text_pos, + color=lighter_color, + horizontal_alignment=horiz_align, + font_size=font_size, + ) + + # draw keypoints + if keypoints is not None: + for keypoints_per_instance in keypoints: + self.draw_and_connect_keypoints(keypoints_per_instance) + + return self.output + + def overlay_rotated_instances(self, boxes=None, labels=None, assigned_colors=None): + """ + Args: + boxes (ndarray): an Nx5 numpy array of + (x_center, y_center, width, height, angle_degrees) format + for the N objects in a single image. + labels (list[str]): the text to be displayed for each instance. + assigned_colors (list[matplotlib.colors]): a list of colors, where each color + corresponds to each mask or box in the image. Refer to 'matplotlib.colors' + for full list of formats that the colors are accepted in. + + Returns: + output (VisImage): image object with visualizations. + """ + num_instances = len(boxes) + + if assigned_colors is None: + assigned_colors = [random_color(rgb=True, maximum=1) for _ in range(num_instances)] + if num_instances == 0: + return self.output + + # Display in largest to smallest order to reduce occlusion. + if boxes is not None: + areas = boxes[:, 2] * boxes[:, 3] + + sorted_idxs = np.argsort(-areas).tolist() + # Re-order overlapped instances in descending order. + boxes = boxes[sorted_idxs] + labels = [labels[k] for k in sorted_idxs] if labels is not None else None + colors = [assigned_colors[idx] for idx in sorted_idxs] + + for i in range(num_instances): + self.draw_rotated_box_with_label( + boxes[i], edge_color=colors[i], label=labels[i] if labels is not None else None + ) + + return self.output + + def draw_and_connect_keypoints(self, keypoints): + """ + Draws keypoints of an instance and follows the rules for keypoint connections + to draw lines between appropriate keypoints. This follows color heuristics for + line color. + + Args: + keypoints (Tensor): a tensor of shape (K, 3), where K is the number of keypoints + and the last dimension corresponds to (x, y, probability). + + Returns: + output (VisImage): image object with visualizations. + """ + visible = {} + keypoint_names = self.metadata.get("keypoint_names") + for idx, keypoint in enumerate(keypoints): + # draw keypoint + x, y, prob = keypoint + if prob > _KEYPOINT_THRESHOLD: + self.draw_circle((x, y), color=_RED) + if keypoint_names: + keypoint_name = keypoint_names[idx] + visible[keypoint_name] = (x, y) + + if self.metadata.get("keypoint_connection_rules"): + for kp0, kp1, color in self.metadata.keypoint_connection_rules: + if kp0 in visible and kp1 in visible: + x0, y0 = visible[kp0] + x1, y1 = visible[kp1] + color = tuple(x / 255.0 for x in color) + self.draw_line([x0, x1], [y0, y1], color=color) + + # draw lines from nose to mid-shoulder and mid-shoulder to mid-hip + # Note that this strategy is specific to person keypoints. + # For other keypoints, it should just do nothing + try: + ls_x, ls_y = visible["left_shoulder"] + rs_x, rs_y = visible["right_shoulder"] + mid_shoulder_x, mid_shoulder_y = (ls_x + rs_x) / 2, (ls_y + rs_y) / 2 + except KeyError: + pass + else: + # draw line from nose to mid-shoulder + nose_x, nose_y = visible.get("nose", (None, None)) + if nose_x is not None: + self.draw_line([nose_x, mid_shoulder_x], [nose_y, mid_shoulder_y], color=_RED) + + try: + # draw line from mid-shoulder to mid-hip + lh_x, lh_y = visible["left_hip"] + rh_x, rh_y = visible["right_hip"] + except KeyError: + pass + else: + mid_hip_x, mid_hip_y = (lh_x + rh_x) / 2, (lh_y + rh_y) / 2 + self.draw_line([mid_hip_x, mid_shoulder_x], [mid_hip_y, mid_shoulder_y], color=_RED) + return self.output + + """ + Primitive drawing functions: + """ + + def draw_text( + self, + text, + position, + *, + font_size=None, + color="g", + horizontal_alignment="center", + rotation=0 + ): + """ + Args: + text (str): class label + position (tuple): a tuple of the x and y coordinates to place text on image. + font_size (int, optional): font of the text. If not provided, a font size + proportional to the image width is calculated and used. + color: color of the text. Refer to `matplotlib.colors` for full list + of formats that are accepted. + horizontal_alignment (str): see `matplotlib.text.Text` + rotation: rotation angle in degrees CCW + + Returns: + output (VisImage): image object with text drawn. + """ + if not font_size: + font_size = self._default_font_size + + # since the text background is dark, we don't want the text to be dark + color = np.maximum(list(mplc.to_rgb(color)), 0.2) + color[np.argmax(color)] = max(0.8, np.max(color)) + + x, y = position + self.output.ax.text( + x, + y, + text, + size=font_size * self.output.scale, + family="sans-serif", + bbox={"facecolor": "black", "alpha": 0.8, "pad": 0.7, "edgecolor": "none"}, + verticalalignment="top", + horizontalalignment=horizontal_alignment, + color=color, + zorder=10, + rotation=rotation, + ) + return self.output + + def draw_box(self, box_coord, alpha=0.5, edge_color="g", line_style="-"): + """ + Args: + box_coord (tuple): a tuple containing x0, y0, x1, y1 coordinates, where x0 and y0 + are the coordinates of the image's top left corner. x1 and y1 are the + coordinates of the image's bottom right corner. + alpha (float): blending efficient. Smaller values lead to more transparent masks. + edge_color: color of the outline of the box. Refer to `matplotlib.colors` + for full list of formats that are accepted. + line_style (string): the string to use to create the outline of the boxes. + + Returns: + output (VisImage): image object with box drawn. + """ + x0, y0, x1, y1 = box_coord + width = x1 - x0 + height = y1 - y0 + + linewidth = max(self._default_font_size / 4, 1) + + self.output.ax.add_patch( + mpl.patches.Rectangle( + (x0, y0), + width, + height, + fill=False, + edgecolor=edge_color, + linewidth=linewidth * self.output.scale, + alpha=alpha, + linestyle=line_style, + ) + ) + return self.output + + def draw_rotated_box_with_label( + self, rotated_box, alpha=0.5, edge_color="g", line_style="-", label=None + ): + """ + Draw a rotated box with label on its top-left corner. + + Args: + rotated_box (tuple): a tuple containing (cnt_x, cnt_y, w, h, angle), + where cnt_x and cnt_y are the center coordinates of the box. + w and h are the width and height of the box. angle represents how + many degrees the box is rotated CCW with regard to the 0-degree box. + alpha (float): blending efficient. Smaller values lead to more transparent masks. + edge_color: color of the outline of the box. Refer to `matplotlib.colors` + for full list of formats that are accepted. + line_style (string): the string to use to create the outline of the boxes. + label (string): label for rotated box. It will not be rendered when set to None. + + Returns: + output (VisImage): image object with box drawn. + """ + cnt_x, cnt_y, w, h, angle = rotated_box + area = w * h + # use thinner lines when the box is small + linewidth = self._default_font_size / ( + 6 if area < _SMALL_OBJECT_AREA_THRESH * self.output.scale else 3 + ) + + theta = angle * math.pi / 180.0 + c = math.cos(theta) + s = math.sin(theta) + rect = [(-w / 2, h / 2), (-w / 2, -h / 2), (w / 2, -h / 2), (w / 2, h / 2)] + # x: left->right ; y: top->down + rotated_rect = [(s * yy + c * xx + cnt_x, c * yy - s * xx + cnt_y) for (xx, yy) in rect] + for k in range(4): + j = (k + 1) % 4 + self.draw_line( + [rotated_rect[k][0], rotated_rect[j][0]], + [rotated_rect[k][1], rotated_rect[j][1]], + color=edge_color, + linestyle="--" if k == 1 else line_style, + linewidth=linewidth, + ) + + if label is not None: + text_pos = rotated_rect[1] # topleft corner + + height_ratio = h / np.sqrt(self.output.height * self.output.width) + label_color = self._change_color_brightness(edge_color, brightness_factor=0.7) + font_size = ( + np.clip((height_ratio - 0.02) / 0.08 + 1, 1.2, 2) * 0.5 * self._default_font_size + ) + self.draw_text(label, text_pos, color=label_color, font_size=font_size, rotation=angle) + + return self.output + + def draw_circle(self, circle_coord, color, radius=3): + """ + Args: + circle_coord (list(int) or tuple(int)): contains the x and y coordinates + of the center of the circle. + color: color of the polygon. Refer to `matplotlib.colors` for a full list of + formats that are accepted. + radius (int): radius of the circle. + + Returns: + output (VisImage): image object with box drawn. + """ + x, y = circle_coord + self.output.ax.add_patch( + mpl.patches.Circle(circle_coord, radius=radius, fill=True, color=color) + ) + return self.output + + def draw_line(self, x_data, y_data, color, linestyle="-", linewidth=None): + """ + Args: + x_data (list[int]): a list containing x values of all the points being drawn. + Length of list should match the length of y_data. + y_data (list[int]): a list containing y values of all the points being drawn. + Length of list should match the length of x_data. + color: color of the line. Refer to `matplotlib.colors` for a full list of + formats that are accepted. + linestyle: style of the line. Refer to `matplotlib.lines.Line2D` + for a full list of formats that are accepted. + linewidth (float or None): width of the line. When it's None, + a default value will be computed and used. + + Returns: + output (VisImage): image object with line drawn. + """ + if linewidth is None: + linewidth = self._default_font_size / 3 + linewidth = max(linewidth, 1) + self.output.ax.add_line( + mpl.lines.Line2D( + x_data, + y_data, + linewidth=linewidth * self.output.scale, + color=color, + linestyle=linestyle, + ) + ) + return self.output + + def draw_binary_mask( + self, binary_mask, color=None, *, edge_color=None, text=None, alpha=0.5, area_threshold=0 + ): + """ + Args: + binary_mask (ndarray): numpy array of shape (H, W), where H is the image height and + W is the image width. Each value in the array is either a 0 or 1 value of uint8 + type. + color: color of the mask. Refer to `matplotlib.colors` for a full list of + formats that are accepted. If None, will pick a random color. + edge_color: color of the polygon edges. Refer to `matplotlib.colors` for a + full list of formats that are accepted. + text (str): if None, will be drawn in the object's center of mass. + alpha (float): blending efficient. Smaller values lead to more transparent masks. + area_threshold (float): a connected component small than this will not be shown. + + Returns: + output (VisImage): image object with mask drawn. + """ + if color is None: + color = random_color(rgb=True, maximum=1) + color = mplc.to_rgb(color) + + has_valid_segment = False + binary_mask = binary_mask.astype("uint8") # opencv needs uint8 + mask = GenericMask(binary_mask, self.output.height, self.output.width) + shape2d = (binary_mask.shape[0], binary_mask.shape[1]) + + if not mask.has_holes: + # draw polygons for regular masks + for segment in mask.polygons: + area = mask_util.area(mask_util.frPyObjects([segment], shape2d[0], shape2d[1])) + if area < (area_threshold or 0): + continue + has_valid_segment = True + segment = segment.reshape(-1, 2) + self.draw_polygon(segment, color=color, edge_color=edge_color, alpha=alpha) + else: + # TODO: Use Path/PathPatch to draw vector graphics: + # https://stackoverflow.com/questions/8919719/how-to-plot-a-complex-polygon + rgba = np.zeros(shape2d + (4,), dtype="float32") + rgba[:, :, :3] = color + rgba[:, :, 3] = (mask.mask == 1).astype("float32") * alpha + has_valid_segment = True + self.output.ax.imshow(rgba, extent=(0, self.output.width, self.output.height, 0)) + + if text is not None and has_valid_segment: + # TODO sometimes drawn on wrong objects. the heuristics here can improve. + lighter_color = self._change_color_brightness(color, brightness_factor=0.7) + _num_cc, cc_labels, stats, centroids = cv2.connectedComponentsWithStats(binary_mask, 8) + largest_component_id = np.argmax(stats[1:, -1]) + 1 + + # draw text on the largest component, as well as other very large components. + for cid in range(1, _num_cc): + if cid == largest_component_id or stats[cid, -1] > _LARGE_MASK_AREA_THRESH: + # median is more stable than centroid + # center = centroids[largest_component_id] + center = np.median((cc_labels == cid).nonzero(), axis=1)[::-1] + self.draw_text(text, center, color=lighter_color) + return self.output + + def draw_polygon(self, segment, color, edge_color=None, alpha=0.5): + """ + Args: + segment: numpy array of shape Nx2, containing all the points in the polygon. + color: color of the polygon. Refer to `matplotlib.colors` for a full list of + formats that are accepted. + edge_color: color of the polygon edges. Refer to `matplotlib.colors` for a + full list of formats that are accepted. If not provided, a darker shade + of the polygon color will be used instead. + alpha (float): blending efficient. Smaller values lead to more transparent masks. + + Returns: + output (VisImage): image object with polygon drawn. + """ + if edge_color is None: + # make edge color darker than the polygon color + if alpha > 0.8: + edge_color = self._change_color_brightness(color, brightness_factor=-0.7) + else: + edge_color = color + edge_color = mplc.to_rgb(edge_color) + (1,) + + polygon = mpl.patches.Polygon( + segment, + fill=True, + facecolor=mplc.to_rgb(color) + (alpha,), + edgecolor=edge_color, + linewidth=max(self._default_font_size // 15 * self.output.scale, 1), + ) + self.output.ax.add_patch(polygon) + return self.output + + """ + Internal methods: + """ + + def _jitter(self, color): + """ + Randomly modifies given color to produce a slightly different color than the color given. + + Args: + color (tuple[double]): a tuple of 3 elements, containing the RGB values of the color + picked. The values in the list are in the [0.0, 1.0] range. + + Returns: + jittered_color (tuple[double]): a tuple of 3 elements, containing the RGB values of the + color after being jittered. The values in the list are in the [0.0, 1.0] range. + """ + color = mplc.to_rgb(color) + vec = np.random.rand(3) + # better to do it in another color space + vec = vec / np.linalg.norm(vec) * 0.5 + res = np.clip(vec + color, 0, 1) + return tuple(res) + + def _create_grayscale_image(self, mask=None): + """ + Create a grayscale version of the original image. + The colors in masked area, if given, will be kept. + """ + img_bw = self.img.astype("f4").mean(axis=2) + img_bw = np.stack([img_bw] * 3, axis=2) + if mask is not None: + img_bw[mask] = self.img[mask] + return img_bw + + def _change_color_brightness(self, color, brightness_factor): + """ + Depending on the brightness_factor, gives a lighter or darker color i.e. a color with + less or more saturation than the original color. + + Args: + color: color of the polygon. Refer to `matplotlib.colors` for a full list of + formats that are accepted. + brightness_factor (float): a value in [-1.0, 1.0] range. A lightness factor of + 0 will correspond to no change, a factor in [-1.0, 0) range will result in + a darker color and a factor in (0, 1.0] range will result in a lighter color. + + Returns: + modified_color (tuple[double]): a tuple containing the RGB values of the + modified color. Each value in the tuple is in the [0.0, 1.0] range. + """ + assert brightness_factor >= -1.0 and brightness_factor <= 1.0 + color = mplc.to_rgb(color) + polygon_color = colorsys.rgb_to_hls(*mplc.to_rgb(color)) + modified_lightness = polygon_color[1] + (brightness_factor * polygon_color[1]) + modified_lightness = 0.0 if modified_lightness < 0.0 else modified_lightness + modified_lightness = 1.0 if modified_lightness > 1.0 else modified_lightness + modified_color = colorsys.hls_to_rgb(polygon_color[0], modified_lightness, polygon_color[2]) + return modified_color + + def _convert_boxes(self, boxes): + """ + Convert different format of boxes to an NxB array, where B = 4 or 5 is the box dimension. + """ + if isinstance(boxes, Boxes) or isinstance(boxes, RotatedBoxes): + return boxes.tensor.numpy() + else: + return np.asarray(boxes) + + def _convert_masks(self, masks_or_polygons): + """ + Convert different format of masks or polygons to a tuple of masks and polygons. + + Returns: + list[GenericMask]: + """ + + m = masks_or_polygons + if isinstance(m, PolygonMasks): + m = m.polygons + if isinstance(m, BitMasks): + m = m.tensor.numpy() + if isinstance(m, torch.Tensor): + m = m.numpy() + ret = [] + for x in m: + if isinstance(x, GenericMask): + ret.append(x) + else: + ret.append(GenericMask(x, self.output.height, self.output.width)) + return ret + + def _convert_keypoints(self, keypoints): + if isinstance(keypoints, Keypoints): + keypoints = keypoints.tensor + keypoints = np.asarray(keypoints) + return keypoints + + def get_output(self): + """ + Returns: + output (VisImage): the image output containing the visualizations added + to the image. + """ + return self.output diff --git a/install.sh b/install.sh new file mode 100644 index 0000000000000000000000000000000000000000..43abad3a5b87ff72b771bbbd07c5c68c3c427c7d --- /dev/null +++ b/install.sh @@ -0,0 +1,14 @@ +pip install dataclasses ftfy regex tqdm timm diffdist spacy +pip install git+https://github.com/lvis-dataset/lvis-api.git +python -m spacy download en_core_web_sm + +CURRENT_DIR=${PWD##*/} +cd ../ +python -m pip install -e $CURRENT_DIR +cd $CURRENT_DIR + +# for tsv loading +pip install Pillow==7.1.2 + +ln -s /mnt/data_storage/coco datasets/coco +ln -s /mnt/data_storage/lvis datasets/lvis \ No newline at end of file diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..d878706514a62cae374dbea4238288c5f7931852 --- /dev/null +++ b/requirements.txt @@ -0,0 +1,8 @@ +torch==1.10.1 +torchvision==0.11.2 +opencv-python-headless==4.5.3.56 +timm==0.4.12 +numpy +yacs +lvis +transformers \ No newline at end of file diff --git a/setup.cfg b/setup.cfg new file mode 100644 index 0000000000000000000000000000000000000000..2a1ccd4e54e57367b2c26325ec14fbc8d7b8b972 --- /dev/null +++ b/setup.cfg @@ -0,0 +1,26 @@ +[isort] +line_length=100 +multi_line_output=3 +include_trailing_comma=True +known_standard_library=numpy,setuptools,mock +skip=./datasets,docs +skip_glob=*/__init__.py,**/configs/**,tests/config/** +known_myself=detectron2 +known_third_party=fvcore,matplotlib,cv2,torch,torchvision,PIL,pycocotools,yacs,termcolor,cityscapesscripts,tabulate,tqdm,scipy,lvis,psutil,pkg_resources,caffe2,onnx,panopticapi,black,isort,av,iopath,omegaconf,hydra,yaml,pydoc,submitit,cloudpickle +no_lines_before=STDLIB,THIRDPARTY +sections=FUTURE,STDLIB,THIRDPARTY,myself,FIRSTPARTY,LOCALFOLDER +default_section=FIRSTPARTY + +[mypy] +python_version=3.6 +ignore_missing_imports = True +warn_unused_configs = True +disallow_untyped_defs = True +check_untyped_defs = True +warn_unused_ignores = True +warn_redundant_casts = True +show_column_numbers = True +follow_imports = silent +allow_redefinition = True +; Require all functions to be annotated +disallow_incomplete_defs = True diff --git a/setup.py b/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..16f1a522526a56831880c3c28be81782958b085d --- /dev/null +++ b/setup.py @@ -0,0 +1,247 @@ +#!/usr/bin/env python +# Copyright (c) Facebook, Inc. and its affiliates. + +import glob +import os +import shutil +from os import path +from setuptools import find_packages, setup +from typing import List +import torch +from torch.utils.cpp_extension import CUDA_HOME, CppExtension, CUDAExtension +from torch.utils.hipify import hipify_python + +torch_ver = [int(x) for x in torch.__version__.split(".")[:2]] +assert torch_ver >= [1, 6], "Requires PyTorch >= 1.6" + + +def get_version(): + init_py_path = path.join(path.abspath(path.dirname(__file__)), "detectron2", "__init__.py") + init_py = open(init_py_path, "r").readlines() + version_line = [l.strip() for l in init_py if l.startswith("__version__")][0] + version = version_line.split("=")[-1].strip().strip("'\"") + + # The following is used to build release packages. + # Users should never use it. + suffix = os.getenv("D2_VERSION_SUFFIX", "") + version = version + suffix + if os.getenv("BUILD_NIGHTLY", "0") == "1": + from datetime import datetime + + date_str = datetime.today().strftime("%y%m%d") + version = version + ".dev" + date_str + + new_init_py = [l for l in init_py if not l.startswith("__version__")] + new_init_py.append('__version__ = "{}"\n'.format(version)) + with open(init_py_path, "w") as f: + f.write("".join(new_init_py)) + return version + + +def get_extensions(): + this_dir = path.dirname(path.abspath(__file__)) + extensions_dir = path.join(this_dir, "detectron2", "layers", "csrc") + + main_source = path.join(extensions_dir, "vision.cpp") + sources = glob.glob(path.join(extensions_dir, "**", "*.cpp")) + + from torch.utils.cpp_extension import ROCM_HOME + + is_rocm_pytorch = ( + True if ((torch.version.hip is not None) and (ROCM_HOME is not None)) else False + ) + + hipify_ver = ( + [int(x) for x in torch.utils.hipify.__version__.split(".")] + if hasattr(torch.utils.hipify, "__version__") + else [0, 0, 0] + ) + + if is_rocm_pytorch and hipify_ver < [1, 0, 0]: # TODO not needed since pt1.8 + + # Earlier versions of hipification and extension modules were not + # transparent, i.e. would require an explicit call to hipify, and the + # hipification would introduce "hip" subdirectories, possibly changing + # the relationship between source and header files. + # This path is maintained for backwards compatibility. + + hipify_python.hipify( + project_directory=this_dir, + output_directory=this_dir, + includes="/detectron2/layers/csrc/*", + show_detailed=True, + is_pytorch_extension=True, + ) + + source_cuda = glob.glob(path.join(extensions_dir, "**", "hip", "*.hip")) + glob.glob( + path.join(extensions_dir, "hip", "*.hip") + ) + + shutil.copy( + "detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_utils.h", + "detectron2/layers/csrc/box_iou_rotated/hip/box_iou_rotated_utils.h", + ) + shutil.copy( + "detectron2/layers/csrc/deformable/deform_conv.h", + "detectron2/layers/csrc/deformable/hip/deform_conv.h", + ) + + sources = [main_source] + sources + sources = [ + s + for s in sources + if not is_rocm_pytorch or torch_ver < [1, 7] or not s.endswith("hip/vision.cpp") + ] + + else: + + # common code between cuda and rocm platforms, + # for hipify version [1,0,0] and later. + + source_cuda = glob.glob(path.join(extensions_dir, "**", "*.cu")) + glob.glob( + path.join(extensions_dir, "*.cu") + ) + + sources = [main_source] + sources + + extension = CppExtension + + extra_compile_args = {"cxx": []} + define_macros = [] + + if (torch.cuda.is_available() and ((CUDA_HOME is not None) or is_rocm_pytorch)) or os.getenv( + "FORCE_CUDA", "0" + ) == "1": + extension = CUDAExtension + sources += source_cuda + + if not is_rocm_pytorch: + define_macros += [("WITH_CUDA", None)] + extra_compile_args["nvcc"] = [ + "-O3", + "-DCUDA_HAS_FP16=1", + "-D__CUDA_NO_HALF_OPERATORS__", + "-D__CUDA_NO_HALF_CONVERSIONS__", + "-D__CUDA_NO_HALF2_OPERATORS__", + ] + else: + define_macros += [("WITH_HIP", None)] + extra_compile_args["nvcc"] = [] + + if torch_ver < [1, 7]: + # supported by https://github.com/pytorch/pytorch/pull/43931 + CC = os.environ.get("CC", None) + if CC is not None: + extra_compile_args["nvcc"].append("-ccbin={}".format(CC)) + + include_dirs = [extensions_dir] + + ext_modules = [ + extension( + "detectron2._C", + sources, + include_dirs=include_dirs, + define_macros=define_macros, + extra_compile_args=extra_compile_args, + ) + ] + + return ext_modules + + +def get_model_zoo_configs() -> List[str]: + """ + Return a list of configs to include in package for model zoo. Copy over these configs inside + detectron2/model_zoo. + """ + + # Use absolute paths while symlinking. + source_configs_dir = path.join(path.dirname(path.realpath(__file__)), "configs") + destination = path.join( + path.dirname(path.realpath(__file__)), "detectron2", "model_zoo", "configs" + ) + # Symlink the config directory inside package to have a cleaner pip install. + + # Remove stale symlink/directory from a previous build. + if path.exists(source_configs_dir): + if path.islink(destination): + os.unlink(destination) + elif path.isdir(destination): + shutil.rmtree(destination) + + if not path.exists(destination): + try: + os.symlink(source_configs_dir, destination) + except OSError: + # Fall back to copying if symlink fails: ex. on Windows. + shutil.copytree(source_configs_dir, destination) + + config_paths = glob.glob("configs/**/*.yaml", recursive=True) + glob.glob( + "configs/**/*.py", recursive=True + ) + return config_paths + + +# For projects that are relative small and provide features that are very close +# to detectron2's core functionalities, we install them under detectron2.projects +PROJECTS = { + # "detectron2.projects.point_rend": "projects/PointRend/point_rend", + # "detectron2.projects.deeplab": "projects/DeepLab/deeplab", + # "detectron2.projects.panoptic_deeplab": "projects/Panoptic-DeepLab/panoptic_deeplab", +} + +setup( + name="detectron2", + version=get_version(), + author="FAIR", + url="https://github.com/facebookresearch/detectron2", + description="Detectron2 is FAIR's next-generation research " + "platform for object detection and segmentation.", + packages=find_packages(exclude=("configs", "tests*")) + list(PROJECTS.keys()), + package_dir=PROJECTS, + package_data={"detectron2.model_zoo": get_model_zoo_configs()}, + python_requires=">=3.6", + install_requires=[ + # Do not add opencv here. Just like pytorch, user should install + # opencv themselves, preferrably by OS's package manager, or by + # choosing the proper pypi package name at https://github.com/skvark/opencv-python + "termcolor>=1.1", + "Pillow>=7.1", # or use pillow-simd for better performance + "yacs>=0.1.6", + "tabulate", + "cloudpickle", + "matplotlib", + "tqdm>4.29.0", + "tensorboard", + # Lock version of fvcore/iopath because they may have breaking changes + # NOTE: when updating fvcore/iopath version, make sure fvcore depends + # on compatible version of iopath. + "fvcore>=0.1.5,<0.1.6", # required like this to make it pip installable + "iopath>=0.1.7,<0.1.9", + "pycocotools>=2.0.2", # corresponds to https://github.com/ppwwyyxx/cocoapi + "future", # used by caffe2 + "pydot", # used to save caffe2 SVGs + "dataclasses; python_version<'3.7'", + "omegaconf>=2.1.0rc1", + "hydra-core>=1.1.0rc1", + "black==21.4b2", + # When adding to the list, may need to update docs/requirements.txt + # or add mock in docs/conf.py + ], + extras_require={ + "all": [ + "shapely", + "pygments>=2.2", + "psutil", + "panopticapi @ https://github.com/cocodataset/panopticapi/archive/master.zip", + ], + "dev": [ + "flake8==3.8.1", + "isort==4.3.21", + "flake8-bugbear", + "flake8-comprehensions", + ], + }, + ext_modules=get_extensions(), + cmdclass={"build_ext": torch.utils.cpp_extension.BuildExtension}, +)