File size: 6,684 Bytes
3eca424 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 |
# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates.
from detectron2.config import CfgNode as CN
def add_mask_former_default_config(cfg):
# data config
# select the dataset mapper
cfg.INPUT.DATASET_MAPPER_NAME = "mask_former_semantic"
# Color augmentation
cfg.INPUT.COLOR_AUG_SSD = False
# We retry random cropping until no single category in semantic segmentation GT occupies more
# than `SINGLE_CATEGORY_MAX_AREA` part of the crop.
cfg.INPUT.CROP.SINGLE_CATEGORY_MAX_AREA = 1.0
# Pad image and segmentation GT in dataset mapper.
cfg.INPUT.SIZE_DIVISIBILITY = -1
# solver config
# weight decay on embedding
cfg.SOLVER.WEIGHT_DECAY_EMBED = 0.0
# optimizer
cfg.SOLVER.OPTIMIZER = "ADAMW"
cfg.SOLVER.BACKBONE_MULTIPLIER = 0.1
# mask_former model config
cfg.MODEL.MASK_FORMER = CN()
# loss
cfg.MODEL.MASK_FORMER.DEEP_SUPERVISION = True
cfg.MODEL.MASK_FORMER.NO_OBJECT_WEIGHT = 0.1
cfg.MODEL.MASK_FORMER.CLASS_WEIGHT = 1.0
cfg.MODEL.MASK_FORMER.DICE_WEIGHT = 1.0
cfg.MODEL.MASK_FORMER.MASK_WEIGHT = 20.0
# transformer config
cfg.MODEL.MASK_FORMER.NHEADS = 8
cfg.MODEL.MASK_FORMER.DROPOUT = 0.1
cfg.MODEL.MASK_FORMER.DIM_FEEDFORWARD = 2048
cfg.MODEL.MASK_FORMER.ENC_LAYERS = 0
cfg.MODEL.MASK_FORMER.DEC_LAYERS = 6
cfg.MODEL.MASK_FORMER.PRE_NORM = False
cfg.MODEL.MASK_FORMER.HIDDEN_DIM = 256
cfg.MODEL.MASK_FORMER.NUM_OBJECT_QUERIES = 100
cfg.MODEL.MASK_FORMER.TRANSFORMER_IN_FEATURE = "res5"
cfg.MODEL.MASK_FORMER.ENFORCE_INPUT_PROJ = False
# mask_former inference config
cfg.MODEL.MASK_FORMER.TEST = CN()
cfg.MODEL.MASK_FORMER.TEST.SEMANTIC_ON = False
cfg.MODEL.MASK_FORMER.TEST.INSTANCE_ON = False
cfg.MODEL.MASK_FORMER.TEST.PANOPTIC_ON = False
cfg.MODEL.MASK_FORMER.TEST.OBJECT_MASK_THRESHOLD = 0.0
cfg.MODEL.MASK_FORMER.TEST.OVERLAP_THRESHOLD = 0.0
cfg.MODEL.MASK_FORMER.TEST.SEM_SEG_POSTPROCESSING_BEFORE_INFERENCE = False
# you can use this config to override
cfg.MODEL.MASK_FORMER.SIZE_DIVISIBILITY = 32
# pixel decoder config
cfg.MODEL.SEM_SEG_HEAD.MASK_DIM = 256
# adding transformer in pixel decoder
cfg.MODEL.SEM_SEG_HEAD.TRANSFORMER_ENC_LAYERS = 0
# pixel decoder
cfg.MODEL.SEM_SEG_HEAD.PIXEL_DECODER_NAME = "BasePixelDecoder"
# swin transformer backbone
cfg.MODEL.SWIN = CN()
cfg.MODEL.SWIN.PRETRAIN_IMG_SIZE = 224
cfg.MODEL.SWIN.PATCH_SIZE = 4
cfg.MODEL.SWIN.EMBED_DIM = 96
cfg.MODEL.SWIN.DEPTHS = [2, 2, 6, 2]
cfg.MODEL.SWIN.NUM_HEADS = [3, 6, 12, 24]
cfg.MODEL.SWIN.WINDOW_SIZE = 7
cfg.MODEL.SWIN.MLP_RATIO = 4.0
cfg.MODEL.SWIN.QKV_BIAS = True
cfg.MODEL.SWIN.QK_SCALE = None
cfg.MODEL.SWIN.DROP_RATE = 0.0
cfg.MODEL.SWIN.ATTN_DROP_RATE = 0.0
cfg.MODEL.SWIN.DROP_PATH_RATE = 0.3
cfg.MODEL.SWIN.APE = False
cfg.MODEL.SWIN.PATCH_NORM = True
cfg.MODEL.SWIN.OUT_FEATURES = ["res2", "res3", "res4", "res5"]
# NOTE: maskformer2 extra configs
# transformer module
cfg.MODEL.MASK_FORMER.TRANSFORMER_DECODER_NAME = "MultiScaleMaskedTransformerDecoder"
# LSJ aug
cfg.INPUT.IMAGE_SIZE = 1024
cfg.INPUT.MIN_SCALE = 0.1
cfg.INPUT.MAX_SCALE = 2.0
# MSDeformAttn encoder configs
cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES = ["res3", "res4", "res5"]
cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_N_POINTS = 4
cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_N_HEADS = 8
# point loss configs
# Number of points sampled during training for a mask point head.
cfg.MODEL.MASK_FORMER.TRAIN_NUM_POINTS = 112 * 112
# Oversampling parameter for PointRend point sampling during training. Parameter `k` in the
# original paper.
cfg.MODEL.MASK_FORMER.OVERSAMPLE_RATIO = 3.0
# Importance sampling parameter for PointRend point sampling during training. Parametr `beta` in
# the original paper.
cfg.MODEL.MASK_FORMER.IMPORTANCE_SAMPLE_RATIO = 0.75
def add_our_config(cfg):
cfg.ORACLE = False
cfg.PSEUDO = False
cfg.PSEUDO_WITH_PRIOR = True
cfg.PSEUDO_REJECT_THRESHOLD = 0.0
cfg.TEST.SLIDING_WINDOW = False
cfg.TEST.SLIDING_TILE_SIZE = 224
cfg.TEST.SLIDING_OVERLAP = 2 / 3.0
cfg.PSEUDO_FLAG_NAME = "trainable_flag"
cfg.SOLVER.TEST_IMS_PER_BATCH = 1
cfg.DATASETS.SAMPLE_PER_CLASS = -1
cfg.DATASETS.SAMPLE_SEED = 0
cfg.TEST.OPTIM = CN()
cfg.TEST.OPTIM.LR = 0.001
cfg.INPUT.TASK_NAME = ["semantic segmentation."]
# whether to use dense crf
cfg.TEST.DENSE_CRF = False
# embedding head
cfg.MODEL.SEM_SEG_HEAD.EMBEDDING_DIM = 512
cfg.MODEL.SEM_SEG_HEAD.EMBED_HIDDEN_DIM = 1024
cfg.MODEL.SEM_SEG_HEAD.EMBED_LAYERS = 2
# clip_adapter
cfg.MODEL.CLIP_ADAPTER = CN()
cfg.MODEL.CLIP_ADAPTER.PROMPT_LEARNER = "learnable"
# for predefined
cfg.MODEL.CLIP_ADAPTER.PREDEFINED_PROMPT_TEMPLATES = ["a sculpture of a {}."]
# for learnable prompt
cfg.MODEL.CLIP_ADAPTER.PROMPT_DIM = 512
cfg.MODEL.CLIP_ADAPTER.PROMPT_SHAPE = (16, 0)
cfg.MODEL.CLIP_ADAPTER.TASK_PROMPT_SHAPE = 8
cfg.MODEL.CLIP_ADAPTER.PROMPT_CHECKPOINT = ""
cfg.MODEL.CLIP_ADAPTER.CLIP_MODEL_NAME = "ViT-B/16"
cfg.MODEL.CLIP_ADAPTER.MASK_FILL = "mean"
cfg.MODEL.CLIP_ADAPTER.MASK_EXPAND_RATIO = 1.0
cfg.MODEL.CLIP_ADAPTER.MASK_THR = 0.5
cfg.MODEL.CLIP_ADAPTER.MASK_MATTING = False
cfg.MODEL.CLIP_ADAPTER.REGION_RESIZED = True
cfg.MODEL.CLIP_ADAPTER.CLIP_ENSEMBLE = True
cfg.MODEL.CLIP_ADAPTER.CLIP_ENSEMBLE_WEIGHT = 0.8
#
cfg.MODEL.CLIP_ADAPTER.SEPERATE_ADAPTER = False
cfg.MODEL.CLIP_ADAPTER.REGION_CLIP_ADAPTER = CN()
cfg.MODEL.CLIP_ADAPTER.REGION_CLIP_ADAPTER.CLIP_MODEL_NAME = "ViT-B/16"
cfg.MODEL.CLIP_ADAPTER.REGION_CLIP_ADAPTER.PROMPT_LEARNER = "predefined"
# for predefined
cfg.MODEL.CLIP_ADAPTER.REGION_CLIP_ADAPTER.PREDEFINED_PROMPT_TEMPLATES = [
"a photo of a {}."
]
# for learnable prompt
cfg.MODEL.CLIP_ADAPTER.REGION_CLIP_ADAPTER.PROMPT_DIM = 512
cfg.MODEL.CLIP_ADAPTER.REGION_CLIP_ADAPTER.PROMPT_SHAPE = (16, 0)
cfg.MODEL.CLIP_ADAPTER.REGION_CLIP_ADAPTER.PROMPT_CHECKPOINT = ""
cfg.MODEL.SEM_SEG_HEAD.EMB_SIZE = 256
cfg.MODEL.SEM_SEG_HEAD.EMBED_DIM = 2048
cfg.MODEL.SEM_SEG_HEAD.NUM_HEADS = 8
cfg.MODEL.SEM_SEG_HEAD.USE_LAYER_SCALE = True
# wandb
cfg.WANDB = CN()
cfg.WANDB.PROJECT = "zero_shot_seg"
cfg.WANDB.NAME = None
def add_mask_former_config(cfg):
"""
Add config for MASK_FORMER.
"""
add_mask_former_default_config(cfg)
add_our_config(cfg)
|