+
+YOLOv6-nano achieves 35.0 mAP on COCO val2017 dataset with 1242 FPS on T4 using TensorRT FP16 for bs32 inference, and YOLOv6-s achieves 43.1 mAP on COCO val2017 dataset with 520 FPS on T4 using TensorRT FP16 for bs32 inference.
+
+YOLOv6 is composed of the following methods:
+
+- Hardware-friendly Design for Backbone and Neck
+- Efficient Decoupled Head with SIoU Loss
+
+
+## Coming soon
+
+- [ ] YOLOv6 m/l/x model.
+- [ ] Deployment for MNN/TNN/NCNN/CoreML...
+- [ ] Quantization tools
+
+
+## Quick Start
+
+### Install
+
+```shell
+git clone https://github.com/meituan/YOLOv6
+cd YOLOv6
+pip install -r requirements.txt
+```
+
+### Inference
+
+First, download a pretrained model from the YOLOv6 [release](https://github.com/meituan/YOLOv6/releases/tag/0.1.0)
+
+Second, run inference with `tools/infer.py`
+
+```shell
+python tools/infer.py --weights yolov6s.pt --source img.jpg / imgdir
+ yolov6n.pt
+```
+
+### Training
+
+Single GPU
+
+```shell
+python tools/train.py --batch 32 --conf configs/yolov6s.py --data data/coco.yaml --device 0
+ configs/yolov6n.py
+```
+
+Multi GPUs (DDP mode recommended)
+
+```shell
+python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --batch 256 --conf configs/yolov6s.py --data data/coco.yaml --device 0,1,2,3,4,5,6,7
+ configs/yolov6n.py
+```
+
+- conf: select config file to specify network/optimizer/hyperparameters
+- data: prepare [COCO](http://cocodataset.org) dataset and specify dataset paths in data.yaml
+
+
+### Evaluation
+
+Reproduce mAP on COCO val2017 dataset
+
+```shell
+python tools/eval.py --data data/coco.yaml --batch 32 --weights yolov6s.pt --task val
+ yolov6n.pt
+```
+
+
+### Deployment
+
+* [ONNX](./deploy/ONNX)
+* [OpenVINO](./deploy/OpenVINO)
+
+### Tutorials
+
+* [Train custom data](./docs/Train_custom_data.md)
+* [Test speed](./docs/Test_speed.md)
+
+
+
+## Benchmark
+
+
+| Model | Size | mAPval0.5:0.95 | SpeedV100
fp16 b32
(ms) | SpeedV100
fp32 b32
(ms) | SpeedT4
trt fp16 b1
(fps) | SpeedT4
trt fp16 b32
(fps) | Params
(M) | Flops
(G) | +| :-------------- | ----------- | :----------------------- | :------------------------------------ | :------------------------------------ | ---------------------------------------- | ----------------------------------------- | --------------- | -------------- | +| [**YOLOv6-n**](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6n.pt) | 416
640 | 30.8
35.0 | 0.3
0.5 | 0.4
0.7 | 1100
788 | 2716
1242 | 4.3
4.3 | 4.7
11.1 | +| [**YOLOv6-tiny**](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6t.pt) | 640 | 41.3 | 0.9 | 1.5 | 425 | 602 | 15.0 | 36.7 | +| [**YOLOv6-s**](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6s.pt) | 640 | 43.1 | 1.0 | 1.7 | 373 | 520 | 17.2 | 44.2 | + + +- Comparisons of the mAP and speed of different object detectors are tested on [COCO val2017](https://cocodataset.org/#download) dataset. +- Refer to [Test speed](./docs/Test_speed.md) tutorial to reproduce the speed results of YOLOv6. +- Params and Flops of YOLOv6 are estimated on deployed model. +- Speed results of other methods are tested in our environment using official codebase and model if not found from the corresponding official release. diff --git a/app.py b/app.py new file mode 100644 index 0000000000000000000000000000000000000000..4c659eef11bdf2e6247670981875d5b9dbf6fdcc --- /dev/null +++ b/app.py @@ -0,0 +1,34 @@ +import gradio as gr +import torch +from PIL import Image +import subprocess +import os +import PIL +from pathlib import Path +import uuid + +# Images +torch.hub.download_url_to_file('https://miro.medium.com/max/1400/1*EYFejGUjvjPcc4PZTwoufw.jpeg', '1*EYFejGUjvjPcc4PZTwoufw.jpeg') +torch.hub.download_url_to_file('https://production-media.paperswithcode.com/tasks/ezgif-frame-001_OZzxdny.jpg', 'ezgif-frame-001_OZzxdny.jpg') +torch.hub.download_url_to_file('https://favtutor.com/resources/images/uploads/Social_Distancing_Covid_19__1.jpg', 'Social_Distancing_Covid_19__1.jpg') +torch.hub.download_url_to_file('https://nkcf.org/wp-content/uploads/2017/11/people.jpg', 'people.jpg') + +def yolo(im): + file_name = str(uuid.uuid4()) + im.save(f'{file_name}.jpg') + os.system(f"python tools/infer.py --weights yolov6s.pt --source {str(file_name)}.jpg --project ''") + img = PIL.Image.open(f"exp/{file_name}.jpg") + os.remove(f"exp/{file_name}.jpg") + os.remove(f'{file_name}.jpg') + return img + +inputs = gr.inputs.Image(type='pil', label="Original Image") +outputs = gr.outputs.Image(type="pil", label="Output Image") + +title = "YOLOv6 - Demo" +description = "YOLOv6 is a single-stage object detection framework dedicated to industrial applications, with hardware-friendly efficient design and high performance. Here is a quick Gradio Demo for testing YOLOv6s model. More details from https://github.com/meituan/YOLOv6 " +article = "
YOLOv6-nano achieves 35.0 mAP on COCO val2017 dataset with 1242 FPS on T4 using TensorRT FP16 for bs32 inference, and YOLOv6-s achieves 43.1 mAP on COCO val2017 dataset with 520 FPS on T4 using TensorRT FP16 for bs32 inference. More information at https://github.com/meituan/YOLOv6
" + +examples = [['1*EYFejGUjvjPcc4PZTwoufw.jpeg'], ['ezgif-frame-001_OZzxdny.jpg'], ['Social_Distancing_Covid_19__1.jpg'], ['people.jpg']] + +gr.Interface(yolo, inputs, outputs, title=title, description=description, article=article, examples=examples, analytics_enabled = True, enable_queue=True).launch(inline=False, share=False, debug=False) \ No newline at end of file diff --git a/assets/picture.png b/assets/picture.png new file mode 100644 index 0000000000000000000000000000000000000000..602357b368e1bf4a24530a9f1c765289e3d81fab Binary files /dev/null and b/assets/picture.png differ diff --git a/configs/yolov6_tiny.py b/configs/yolov6_tiny.py new file mode 100644 index 0000000000000000000000000000000000000000..be455de25fb5779671e7adff1df1521b350d1d7f --- /dev/null +++ b/configs/yolov6_tiny.py @@ -0,0 +1,53 @@ +# YOLOv6t model +model = dict( + type='YOLOv6t', + pretrained=None, + depth_multiple=0.25, + width_multiple=0.50, + backbone=dict( + type='EfficientRep', + num_repeats=[1, 6, 12, 18, 6], + out_channels=[64, 128, 256, 512, 1024], + ), + neck=dict( + type='RepPAN', + num_repeats=[12, 12, 12, 12], + out_channels=[256, 128, 128, 256, 256, 512], + ), + head=dict( + type='EffiDeHead', + in_channels=[128, 256, 512], + num_layers=3, + begin_indices=24, + anchors=1, + out_indices=[17, 20, 23], + strides=[8, 16, 32], + iou_type='ciou' + ) +) + +solver = dict( + optim='SGD', + lr_scheduler='Cosine', + lr0=0.01, + lrf=0.01, + momentum=0.937, + weight_decay=0.0005, + warmup_epochs=3.0, + warmup_momentum=0.8, + warmup_bias_lr=0.1 +) + +data_aug = dict( + hsv_h=0.015, + hsv_s=0.7, + hsv_v=0.4, + degrees=0.0, + translate=0.1, + scale=0.5, + shear=0.0, + flipud=0.0, + fliplr=0.5, + mosaic=1.0, + mixup=0.0, +) diff --git a/configs/yolov6_tiny_finetune.py b/configs/yolov6_tiny_finetune.py new file mode 100644 index 0000000000000000000000000000000000000000..d751eff06a4ebcc82afe24c8abfef2f73dcfb76e --- /dev/null +++ b/configs/yolov6_tiny_finetune.py @@ -0,0 +1,53 @@ +# YOLOv6t model +model = dict( + type='YOLOv6t', + pretrained='./weights/yolov6t.pt', + depth_multiple=0.25, + width_multiple=0.50, + backbone=dict( + type='EfficientRep', + num_repeats=[1, 6, 12, 18, 6], + out_channels=[64, 128, 256, 512, 1024], + ), + neck=dict( + type='RepPAN', + num_repeats=[12, 12, 12, 12], + out_channels=[256, 128, 128, 256, 256, 512], + ), + head=dict( + type='EffiDeHead', + in_channels=[128, 256, 512], + num_layers=3, + begin_indices=24, + anchors=1, + out_indices=[17, 20, 23], + strides=[8, 16, 32], + iou_type='ciou' + ) +) + +solver = dict( + optim='SGD', + lr_scheduler='Cosine', + lr0=0.0032, + lrf=0.12, + momentum=0.843, + weight_decay=0.00036, + warmup_epochs=2.0, + warmup_momentum=0.5, + warmup_bias_lr=0.05 +) + +data_aug = dict( + hsv_h=0.0138, + hsv_s=0.664, + hsv_v=0.464, + degrees=0.373, + translate=0.245, + scale=0.898, + shear=0.602, + flipud=0.00856, + fliplr=0.5, + mosaic=1.0, + mixup=0.243, +) diff --git a/configs/yolov6n.py b/configs/yolov6n.py new file mode 100644 index 0000000000000000000000000000000000000000..40b6e0c4acda75f5784f5920353810fcb87aabd0 --- /dev/null +++ b/configs/yolov6n.py @@ -0,0 +1,53 @@ +# YOLOv6n model +model = dict( + type='YOLOv6n', + pretrained=None, + depth_multiple=0.33, + width_multiple=0.25, + backbone=dict( + type='EfficientRep', + num_repeats=[1, 6, 12, 18, 6], + out_channels=[64, 128, 256, 512, 1024], + ), + neck=dict( + type='RepPAN', + num_repeats=[12, 12, 12, 12], + out_channels=[256, 128, 128, 256, 256, 512], + ), + head=dict( + type='EffiDeHead', + in_channels=[128, 256, 512], + num_layers=3, + begin_indices=24, + anchors=1, + out_indices=[17, 20, 23], + strides=[8, 16, 32], + iou_type='ciou' + ) +) + +solver = dict( + optim='SGD', + lr_scheduler='Cosine', + lr0=0.01, + lrf=0.01, + momentum=0.937, + weight_decay=0.0005, + warmup_epochs=3.0, + warmup_momentum=0.8, + warmup_bias_lr=0.1 +) + +data_aug = dict( + hsv_h=0.015, + hsv_s=0.7, + hsv_v=0.4, + degrees=0.0, + translate=0.1, + scale=0.5, + shear=0.0, + flipud=0.0, + fliplr=0.5, + mosaic=1.0, + mixup=0.0, +) diff --git a/configs/yolov6n_finetune.py b/configs/yolov6n_finetune.py new file mode 100644 index 0000000000000000000000000000000000000000..7d1fab5a2c4946eb9fa1986b210af8ad98a5700c --- /dev/null +++ b/configs/yolov6n_finetune.py @@ -0,0 +1,53 @@ +# YOLOv6n model +model = dict( + type='YOLOv6n', + pretrained='./weights/yolov6n.pt', + depth_multiple=0.33, + width_multiple=0.25, + backbone=dict( + type='EfficientRep', + num_repeats=[1, 6, 12, 18, 6], + out_channels=[64, 128, 256, 512, 1024], + ), + neck=dict( + type='RepPAN', + num_repeats=[12, 12, 12, 12], + out_channels=[256, 128, 128, 256, 256, 512], + ), + head=dict( + type='EffiDeHead', + in_channels=[128, 256, 512], + num_layers=3, + begin_indices=24, + anchors=1, + out_indices=[17, 20, 23], + strides=[8, 16, 32], + iou_type='ciou' + ) +) + +solver = dict( + optim='SGD', + lr_scheduler='Cosine', + lr0=0.0032, + lrf=0.12, + momentum=0.843, + weight_decay=0.00036, + warmup_epochs=2.0, + warmup_momentum=0.5, + warmup_bias_lr=0.05 +) + +data_aug = dict( + hsv_h=0.0138, + hsv_s=0.664, + hsv_v=0.464, + degrees=0.373, + translate=0.245, + scale=0.898, + shear=0.602, + flipud=0.00856, + fliplr=0.5, + mosaic=1.0, + mixup=0.243 +) diff --git a/configs/yolov6s.py b/configs/yolov6s.py new file mode 100644 index 0000000000000000000000000000000000000000..8b281bf612fd4d309a2fd174f936c40f06451bba --- /dev/null +++ b/configs/yolov6s.py @@ -0,0 +1,53 @@ +# YOLOv6s model +model = dict( + type='YOLOv6s', + pretrained=None, + depth_multiple=0.33, + width_multiple=0.50, + backbone=dict( + type='EfficientRep', + num_repeats=[1, 6, 12, 18, 6], + out_channels=[64, 128, 256, 512, 1024], + ), + neck=dict( + type='RepPAN', + num_repeats=[12, 12, 12, 12], + out_channels=[256, 128, 128, 256, 256, 512], + ), + head=dict( + type='EffiDeHead', + in_channels=[128, 256, 512], + num_layers=3, + begin_indices=24, + anchors=1, + out_indices=[17, 20, 23], + strides=[8, 16, 32], + iou_type='siou' + ) +) + +solver = dict( + optim='SGD', + lr_scheduler='Cosine', + lr0=0.01, + lrf=0.01, + momentum=0.937, + weight_decay=0.0005, + warmup_epochs=3.0, + warmup_momentum=0.8, + warmup_bias_lr=0.1 +) + +data_aug = dict( + hsv_h=0.015, + hsv_s=0.7, + hsv_v=0.4, + degrees=0.0, + translate=0.1, + scale=0.5, + shear=0.0, + flipud=0.0, + fliplr=0.5, + mosaic=1.0, + mixup=0.0, +) diff --git a/configs/yolov6s_finetune.py b/configs/yolov6s_finetune.py new file mode 100644 index 0000000000000000000000000000000000000000..66e6600dc1565ce1281d880c37a88bec11196511 --- /dev/null +++ b/configs/yolov6s_finetune.py @@ -0,0 +1,53 @@ +# YOLOv6s model +model = dict( + type='YOLOv6s', + pretrained='./weights/yolov6s.pt', + depth_multiple=0.33, + width_multiple=0.50, + backbone=dict( + type='EfficientRep', + num_repeats=[1, 6, 12, 18, 6], + out_channels=[64, 128, 256, 512, 1024], + ), + neck=dict( + type='RepPAN', + num_repeats=[12, 12, 12, 12], + out_channels=[256, 128, 128, 256, 256, 512], + ), + head=dict( + type='EffiDeHead', + in_channels=[128, 256, 512], + num_layers=3, + begin_indices=24, + anchors=1, + out_indices=[17, 20, 23], + strides=[8, 16, 32], + iou_type='siou' + ) +) + +solver = dict( + optim='SGD', + lr_scheduler='Cosine', + lr0=0.0032, + lrf=0.12, + momentum=0.843, + weight_decay=0.00036, + warmup_epochs=2.0, + warmup_momentum=0.5, + warmup_bias_lr=0.05 +) + +data_aug = dict( + hsv_h=0.0138, + hsv_s=0.664, + hsv_v=0.464, + degrees=0.373, + translate=0.245, + scale=0.898, + shear=0.602, + flipud=0.00856, + fliplr=0.5, + mosaic=1.0, + mixup=0.243, +) diff --git a/data/coco.yaml b/data/coco.yaml new file mode 100644 index 0000000000000000000000000000000000000000..699551b91f1278c47e3724392c3b0501aaa949a4 --- /dev/null +++ b/data/coco.yaml @@ -0,0 +1,18 @@ +# COCO 2017 dataset http://cocodataset.org +train: ../coco/images/train2017 # 118287 images +val: ../coco/images/val2017 # 5000 images +test: ../coco/images/test2017 +anno_path: ../coco/annotations/instances_val2017.json +# number of classes +nc: 80 + +# class names +names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', + 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', + 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', + 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', + 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', + 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', + 'hair drier', 'toothbrush' ] diff --git a/data/images/image1.jpg b/data/images/image1.jpg new file mode 100644 index 0000000000000000000000000000000000000000..71a14e1b6d823eb34d8051d3dd1b7226c7834386 Binary files /dev/null and b/data/images/image1.jpg differ diff --git a/data/images/image2.jpg b/data/images/image2.jpg new file mode 100644 index 0000000000000000000000000000000000000000..6fb2b64124bb1a5c9306c302d1ac0cc452be0f17 Binary files /dev/null and b/data/images/image2.jpg differ diff --git a/data/images/image3.jpg b/data/images/image3.jpg new file mode 100644 index 0000000000000000000000000000000000000000..a11422270e965a30991ecbc65e42365b2667ee0f Binary files /dev/null and b/data/images/image3.jpg differ diff --git a/deploy/ONNX/README.md b/deploy/ONNX/README.md new file mode 100644 index 0000000000000000000000000000000000000000..5509bd0b0a2d4e0949bf9e4e931655469a07112c --- /dev/null +++ b/deploy/ONNX/README.md @@ -0,0 +1,17 @@ +## Export ONNX Model + +### Check requirements +```shell +pip install onnx>=1.10.0 +``` + +### Export script +```shell +python deploy/ONNX/export_onnx.py --weights yolov6s.pt --img 640 --batch 1 + +``` + +### Download +* [YOLOv6-nano](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6n.onnx) +* [YOLOv6-tiny](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6t.onnx) +* [YOLOv6-s](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6s.onnx) diff --git a/deploy/ONNX/export_onnx.py b/deploy/ONNX/export_onnx.py new file mode 100644 index 0000000000000000000000000000000000000000..6d5040ead9e186f4d6bcda01a6761c37312c631b --- /dev/null +++ b/deploy/ONNX/export_onnx.py @@ -0,0 +1,80 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import argparse +import time +import sys +import os +import torch +import torch.nn as nn +import onnx + +ROOT = os.getcwd() +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) + +from yolov6.models.yolo import * +from yolov6.models.effidehead import Detect +from yolov6.layers.common import * +from yolov6.utils.events import LOGGER +from yolov6.utils.checkpoint import load_checkpoint + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default='./yolov6s.pt', help='weights path') + parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size') # height, width + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--half', action='store_true', help='FP16 half-precision export') + parser.add_argument('--inplace', action='store_true', help='set Detect() inplace=True') + parser.add_argument('--device', default='0', help='cuda device, i.e. 0 or 0, 1, 2, 3 or cpu') + args = parser.parse_args() + args.img_size *= 2 if len(args.img_size) == 1 else 1 # expand + print(args) + t = time.time() + + # Check device + cuda = args.device != 'cpu' and torch.cuda.is_available() + device = torch.device('cuda:0' if cuda else 'cpu') + assert not (device.type == 'cpu' and args.half), '--half only compatible with GPU export, i.e. use --device 0' + # Load PyTorch model + model = load_checkpoint(args.weights, map_location=device, inplace=True, fuse=True) # load FP32 model + for layer in model.modules(): + if isinstance(layer, RepVGGBlock): + layer.switch_to_deploy() + + # Input + img = torch.zeros(args.batch_size, 3, *args.img_size).to(device) # image size(1,3,320,192) iDetection + + # Update model + if args.half: + img, model = img.half(), model.half() # to FP16 + model.eval() + for k, m in model.named_modules(): + if isinstance(m, Conv): # assign export-friendly activations + if isinstance(m.act, nn.SiLU): + m.act = SiLU() + elif isinstance(m, Detect): + m.inplace = args.inplace + + y = model(img) # dry run + + # ONNX export + try: + LOGGER.info('\nStarting to export ONNX...') + export_file = args.weights.replace('.pt', '.onnx') # filename + torch.onnx.export(model, img, export_file, verbose=False, opset_version=12, + training=torch.onnx.TrainingMode.EVAL, + do_constant_folding=True, + input_names=['image_arrays'], + output_names=['outputs'], + ) + + # Checks + onnx_model = onnx.load(export_file) # load onnx model + onnx.checker.check_model(onnx_model) # check onnx model + LOGGER.info(f'ONNX export success, saved as {export_file}') + except Exception as e: + LOGGER.info(f'ONNX export failure: {e}') + + # Finish + LOGGER.info('\nExport complete (%.2fs)' % (time.time() - t)) diff --git a/deploy/OpenVINO/README.md b/deploy/OpenVINO/README.md new file mode 100644 index 0000000000000000000000000000000000000000..76c25d011d59868d9f70ce5d9e7e7771273c3299 --- /dev/null +++ b/deploy/OpenVINO/README.md @@ -0,0 +1,18 @@ +## Export OpenVINO Model + +### Check requirements +```shell +pip install --upgrade pip +pip install openvino-dev +``` + +### Export script +```shell +python deploy/OpenVINO/export_openvino.py --weights yolov6s.pt --img 640 --batch 1 + +``` + +### Download +* [YOLOv6-nano](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6n_openvino.tar.gz) +* [YOLOv6-tiny](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6n_openvino.tar.gz) +* [YOLOv6-s](https://github.com/meituan/YOLOv6/releases/download/0.1.0/yolov6n_openvino.tar.gz) \ No newline at end of file diff --git a/deploy/OpenVINO/export_openvino.py b/deploy/OpenVINO/export_openvino.py new file mode 100644 index 0000000000000000000000000000000000000000..7b59ae0fc28dc9149568dab4cb7483185dd31dfe --- /dev/null +++ b/deploy/OpenVINO/export_openvino.py @@ -0,0 +1,92 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import argparse +import time +import sys +import os +import torch +import torch.nn as nn +import onnx +import subprocess + +ROOT = os.getcwd() +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) + +from yolov6.models.yolo import * +from yolov6.models.effidehead import Detect +from yolov6.layers.common import * +from yolov6.utils.events import LOGGER +from yolov6.utils.checkpoint import load_checkpoint + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default='./yolov6s.pt', help='weights path') + parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size') # height, width + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--half', action='store_true', help='FP16 half-precision export') + parser.add_argument('--inplace', action='store_true', help='set Detect() inplace=True') + parser.add_argument('--device', default='0', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + args = parser.parse_args() + args.img_size *= 2 if len(args.img_size) == 1 else 1 # expand + print(args) + t = time.time() + + # Check device + cuda = args.device != 'cpu' and torch.cuda.is_available() + device = torch.device('cuda:0' if cuda else 'cpu') + assert not (device.type == 'cpu' and args.half), '--half only compatible with GPU export, i.e. use --device 0' + # Load PyTorch model + model = load_checkpoint(args.weights, map_location=device, inplace=True, fuse=True) # load FP32 model + for layer in model.modules(): + if isinstance(layer, RepVGGBlock): + layer.switch_to_deploy() + + # Input + img = torch.zeros(args.batch_size, 3, *args.img_size).to(device) # image size(1,3,320,192) iDetection + + # Update model + if args.half: + img, model = img.half(), model.half() # to FP16 + model.eval() + for k, m in model.named_modules(): + if isinstance(m, Conv): # assign export-friendly activations + if isinstance(m.act, nn.SiLU): + m.act = SiLU() + elif isinstance(m, Detect): + m.inplace = args.inplace + + y = model(img) # dry run + + # ONNX export + try: + LOGGER.info('\nStarting to export ONNX...') + export_file = args.weights.replace('.pt', '.onnx') # filename + torch.onnx.export(model, img, export_file, verbose=False, opset_version=12, + training=torch.onnx.TrainingMode.EVAL, + do_constant_folding=True, + input_names=['image_arrays'], + output_names=['outputs'], + ) + + # Checks + onnx_model = onnx.load(export_file) # load onnx model + onnx.checker.check_model(onnx_model) # check onnx model + LOGGER.info(f'ONNX export success, saved as {export_file}') + except Exception as e: + LOGGER.info(f'ONNX export failure: {e}') + + # OpenVINO export + try: + LOGGER.info('\nStarting to export OpenVINO...') + import_file = args.weights.replace('.pt', '.onnx') + export_dir = str(import_file).replace('.onnx', '_openvino') + cmd = f"mo --input_model {import_file} --output_dir {export_dir} --data_type {'FP16' if args.half else 'FP32'}" + subprocess.check_output(cmd.split()) + LOGGER.info(f'OpenVINO export success, saved as {export_dir}') + except Exception as e: + LOGGER.info(f'OpenVINO export failure: {e}') + + # Finish + LOGGER.info('\nExport complete (%.2fs)' % (time.time() - t)) diff --git a/docs/About_naming_yolov6.md b/docs/About_naming_yolov6.md new file mode 100644 index 0000000000000000000000000000000000000000..1ab4a3b8b9a413bce3f904a6223d2864cd79ccb7 --- /dev/null +++ b/docs/About_naming_yolov6.md @@ -0,0 +1,12 @@ +# About the naming of YOLOv6 + +### WHY named YOLOv6 ? +The full name is actually MT-YOLOv6, which is called YOLOv6 for brevity. Our work is majorly inspired by the original idea of the one-stage YOLO detection algorithm and the implementation has leveraged various techniques and tricks of former relevant work . Therefore, we named the project YOLOv6 to pay tribute to the work of YOLO series. Furthermore, we have indeed adopted some novel method and made solid engineering improvements to dedicate the algorithm to industrial applications. +As for the project, we'll continue to improve and maintain it, contributing more values for industrial applications. + +P.S. We are contacting the authors of YOLO series about the naming of YOLOv6. + +Thanks for your attention！ + + + diff --git a/docs/Test_speed.md b/docs/Test_speed.md new file mode 100644 index 0000000000000000000000000000000000000000..526e0474db5cc415aef08459ce18bca442890db0 --- /dev/null +++ b/docs/Test_speed.md @@ -0,0 +1,41 @@ +# Test speed + +This guidence explains how to reproduce speed results of YOLOv6. For fair comparision, the speed results do not contain the time cost of data pre-processing and NMS post-processing. + +## 0. Prepare model + +Download the models you want to test from the latest release. + +## 1. Prepare testing environment + +Refer to README, install packages corresponding to CUDA, CUDNN and TensorRT version. + +Here, we use Torch1.8.0 inference on V100 and TensorRT 7.2 on T4. + +## 2. Reproduce speed + +#### 2.1 Torch Inference on V100 + +To get inference speed without TensorRT on V100, you can run the following command: + +```shell +python tools/eval.py --data data/coco.yaml --batch 32 --weights yolov6n.pt --task speed [--half] +``` + +- Speed results with batchsize = 1 are unstable in multiple runs, thus we do not provide the bs1 speed results. + +#### 2.2 TensorRT Inference on T4 + +To get inference speed with TensorRT in FP16 mode on T4, you can follow the steps below: + +First, export pytorch model as onnx format using the following command: + +```shell +python deploy/ONNX/export_onnx.py --weights yolov6n.pt --device 0 --batch [1 or 32] +``` + +Second, generate an inference trt engine and test speed using `trtexec`: + +``` +trtexec --onnx=yolov6n.onnx --workspace=1024 --avgRuns=1000 --inputIOFormats=fp16:chw --outputIOFormats=fp16:chw +``` diff --git a/docs/Train_custom_data.md b/docs/Train_custom_data.md new file mode 100644 index 0000000000000000000000000000000000000000..de67a37f50fa65a135a8e0f873ce1741066d9901 --- /dev/null +++ b/docs/Train_custom_data.md @@ -0,0 +1,129 @@ +# Train Custom Data + +This guidence explains how to train your own custom data with YOLOv6 ( take fine-tuning YOLOv6-s model for example). + +## 0. Before you start + +Clone this repo and follow README.md to install requirements in a Python3.8 environment. + + +## 1. Prepare your own dataset + +**Step 1** Prepare your own dataset with images. For labeling images, you can use tools like [Labelme](https://github.com/wkentaro/labelme). + +**Step 2** Generate label files in YOLO format. + +One image corresponds to one label file, and the label format example is presented as below. + +```json +# class_id center_x center_y bbox_width bbox_height +0 0.300926 0.617063 0.601852 0.765873 +1 0.575 0.319531 0.4 0.551562 +``` + +- Each row represents one object. +- Class id starts from `0`. +- Boundingbox coordinates must be in normalized `xywh` format (from 0 - 1). If your boxes are in pixels, divide `center_x` and `bbox_width` by image width, and `center_y` and `bbox_height` by image height. + +**Step 3** Organize directories. + +Organize your train and val images and label files according to the example below. + +```shell +# image directory +path/to/data/images/train/im0.jpg +path/to/data/images/val/im1.jpg +path/to/data/images/test/im2.jpg + +# label directory +path/to/data/labels/train/im0.txt +path/to/data/labels/val/im1.txt +path/to/data/labels/test/im2.txt +``` + +**Step 4** Create `dataset.yaml` in `$YOLOv6_DIR/data`. + +```yaml +train: path/to/data/images/train # train images +val: path/to/data/images/val # val images +test: path/to/data/images/test # test images (optional) + +# Classes +nc: 20 # number of classes +names: ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', + 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'] # class names + +``` + + +## 2. Create a config file + +We use a config file to specify the network structure and training setting, including optimizer and data augmentation hyperparameters. + +If you create a new config file, please put it under the configs directory. +Or just use the provided config file in `$YOLOV6_HOME/configs/*_finetune.py`. + +```python +## YOLOv6s Model config file +model = dict( + type='YOLOv6s', + pretrained='./weights/yolov6s.pt', # download pretrain model from YOLOv6 github if use pretrained model + depth_multiple = 0.33, + width_multiple = 0.50, + ... +) +solver=dict( + optim='SGD', + lr_scheduler='Cosine', + ... +) + +data_aug = dict( + hsv_h=0.015, + hsv_s=0.7, + hsv_v=0.4, + ... +) +``` + + + +## 3. Train + +Single GPU + +```shell +python tools/train.py --batch 256 --conf configs/yolov6s_finetune.py --data data/data.yaml --device 0 +``` + +Multi GPUs (DDP mode recommended) + +```shell +python -m torch.distributed.launch --nproc_per_node 4 tools/train.py --batch 256 --conf configs/yolov6s_finetune.py --data data/data.yaml --device 0,1,2,3 +``` + + + +## 4. Evaluation + +```shell +python tools/eval.py --data data/data.yaml --weights output_dir/name/weights/best_ckpt.pt --device 0 +``` + + + +## 5. Inference + +```shell +python tools/infer.py --weights output_dir/name/weights/best_ckpt.pt --source img.jpg --device 0 +``` + + + +## 6. Deployment + +Export as ONNX Format + +```shell +python deploy/ONNX/export_onnx.py --weights output_dir/name/weights/best_ckpt.pt --device 0 +``` diff --git a/packages.txt b/packages.txt new file mode 100644 index 0000000000000000000000000000000000000000..66cd8ee4a23a4dbf62845c96e27d6f7d64590bab --- /dev/null +++ b/packages.txt @@ -0,0 +1,5 @@ +ffmpeg +libsm6 +libxext6 -y +libgl1 +-y libgl1-mesa-glx \ No newline at end of file diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..304d48e935e5a6a724d74f02ab3341817254cd14 --- /dev/null +++ b/requirements.txt @@ -0,0 +1,15 @@ +# pip install -r requirements.txt +# python3.8 environment + +torch>=1.8.0 +torchvision>=0.9.0 +numpy>=1.18.5 +opencv-python>=4.1.2 +PyYAML>=5.3.1 +scipy>=1.4.1 +tqdm>=4.41.0 +addict>=2.4.0 +tensorboard>=2.7.0 +pycocotools>=2.0 +onnx>=1.10.0 # ONNX export +thop # FLOPs computation diff --git a/tools/eval.py b/tools/eval.py new file mode 100644 index 0000000000000000000000000000000000000000..79861ac2200a87fde94c9a8902dbf29f1ba421a3 --- /dev/null +++ b/tools/eval.py @@ -0,0 +1,86 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import argparse +import os +import sys +import torch + +ROOT = os.getcwd() +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) + +from yolov6.core.evaler import Evaler +from yolov6.utils.events import LOGGER + + +def get_args_parser(add_help=True): + parser = argparse.ArgumentParser(description='YOLOv6 PyTorch Evalating', add_help=add_help) + parser.add_argument('--data', type=str, default='./data/coco.yaml', help='dataset.yaml path') + parser.add_argument('--weights', type=str, default='./weights/yolov6s.pt', help='model.pt path(s)') + parser.add_argument('--batch-size', type=int, default=32, help='batch size') + parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)') + parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold') + parser.add_argument('--iou-thres', type=float, default=0.65, help='NMS IoU threshold') + parser.add_argument('--task', default='val', help='val, or speed') + parser.add_argument('--device', default='0', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--half', default=False, action='store_true', help='whether to use fp16 infer') + parser.add_argument('--save_dir', type=str, default='runs/val/exp', help='evaluation save dir') + args = parser.parse_args() + LOGGER.info(args) + return args + + +@torch.no_grad() +def run(data, + weights=None, + batch_size=32, + img_size=640, + conf_thres=0.001, + iou_thres=0.65, + task='val', + device='', + half=False, + model=None, + dataloader=None, + save_dir='', + ): + """ Run the evaluation process + + This function is the main process of evalutaion, supporting image file and dir containing images. + It has tasks of 'val', 'train' and 'speed'. Task 'train' processes the evaluation during training phase. + Task 'val' processes the evaluation purely and return the mAP of model.pt. Task 'speed' precesses the + evaluation of inference speed of model.pt. + + """ + + # task + Evaler.check_task(task) + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + # reload thres/device/half/data according task + conf_thres, iou_thres = Evaler.reload_thres(conf_thres, iou_thres, task) + device = Evaler.reload_device(device, model, task) + half = device.type != 'cpu' and half + data = Evaler.reload_dataset(data) if isinstance(data, str) else data + + # init + val = Evaler(data, batch_size, img_size, conf_thres, \ + iou_thres, device, half, save_dir) + model = val.init_model(model, weights, task) + dataloader = val.init_data(dataloader, task) + + # eval + model.eval() + pred_result = val.predict_model(model, dataloader, task) + eval_result = val.eval_model(pred_result, model, dataloader, task) + return eval_result + + +def main(args): + run(**vars(args)) + + +if __name__ == "__main__": + args = get_args_parser() + main(args) diff --git a/tools/infer.py b/tools/infer.py new file mode 100644 index 0000000000000000000000000000000000000000..89841b0af94d49523e37a840c5f8ec9d39f0acd2 --- /dev/null +++ b/tools/infer.py @@ -0,0 +1,108 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import argparse +import os +import sys +import os.path as osp + +import torch + +ROOT = os.getcwd() +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) + +from yolov6.utils.events import LOGGER +from yolov6.core.inferer import Inferer + + +def get_args_parser(add_help=True): + parser = argparse.ArgumentParser(description='YOLOv6 PyTorch Inference.', add_help=add_help) + parser.add_argument('--weights', type=str, default='weights/yolov6s.pt', help='model path(s) for inference.') + parser.add_argument('--source', type=str, default='data/images', help='the source path, e.g. image-file/dir.') + parser.add_argument('--yaml', type=str, default='data/coco.yaml', help='data yaml file.') + parser.add_argument('--img-size', type=int, default=640, help='the image-size(h,w) in inference size.') + parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold for inference.') + parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold for inference.') + parser.add_argument('--max-det', type=int, default=1000, help='maximal inferences per image.') + parser.add_argument('--device', default='0', help='device to run our model i.e. 0 or 0,1,2,3 or cpu.') + parser.add_argument('--save-txt', action='store_true', help='save results to *.txt.') + parser.add_argument('--save-img', action='store_false', help='save visuallized inference results.') + parser.add_argument('--classes', nargs='+', type=int, help='filter by classes, e.g. --classes 0, or --classes 0 2 3.') + parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS.') + parser.add_argument('--project', default='runs/inference', help='save inference results to project/name.') + parser.add_argument('--name', default='exp', help='save inference results to project/name.') + parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels.') + parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences.') + parser.add_argument('--half', action='store_true', help='whether to use FP16 half-precision inference.') + + args = parser.parse_args() + LOGGER.info(args) + return args + +@torch.no_grad() +def run(weights=osp.join(ROOT, 'yolov6s.pt'), + source=osp.join(ROOT, 'data/images'), + yaml=None, + img_size=640, + conf_thres=0.25, + iou_thres=0.45, + max_det=1000, + device='', + save_txt=False, + save_img=True, + classes=None, + agnostic_nms=False, + project=osp.join(ROOT, 'runs/inference'), + name='exp', + hide_labels=False, + hide_conf=False, + half=False, + ): + """ Inference process + + This function is the main process of inference, supporting image files or dirs containing images. + + Args: + weights: The path of model.pt, e.g. yolov6s.pt + source: Source path, supporting image files or dirs containing images. + yaml: Data yaml file, . + img_size: Inference image-size, e.g. 640 + conf_thres: Confidence threshold in inference, e.g. 0.25 + iou_thres: NMS IOU threshold in inference, e.g. 0.45 + max_det: Maximal detections per image, e.g. 1000 + device: Cuda device, e.e. 0, or 0,1,2,3 or cpu + save_txt: Save results to *.txt + save_img: Save visualized inference results + classes: Filter by class: --class 0, or --class 0 2 3 + agnostic_nms: Class-agnostic NMS + project: Save results to project/name + name: Save results to project/name, e.g. 'exp' + line_thickness: Bounding box thickness (pixels), e.g. 3 + hide_labels: Hide labels, e.g. False + hide_conf: Hide confidences + half: Use FP16 half-precision inference, e.g. False + """ + # create save dir + save_dir = osp.join(project, name) + if (save_img or save_txt) and not osp.exists(save_dir): + os.makedirs(save_dir) + else: + LOGGER.warning('Save directory already existed') + if save_txt: + os.mkdir(osp.join(save_dir, 'labels')) + + # Inference + inferer = Inferer(source, weights, device, yaml, img_size, half) + inferer.infer(conf_thres, iou_thres, classes, agnostic_nms, max_det, save_dir, save_txt, save_img, hide_labels, hide_conf) + + if save_txt or save_img: + LOGGER.info(f"Results saved to {save_dir}") + + +def main(args): + run(**vars(args)) + + +if __name__ == "__main__": + args = get_args_parser() + main(args) diff --git a/tools/quantization/mnn/README.md b/tools/quantization/mnn/README.md new file mode 100644 index 0000000000000000000000000000000000000000..12c3c0415352060b7e5c8f437730a9c6e35dfbef --- /dev/null +++ b/tools/quantization/mnn/README.md @@ -0,0 +1 @@ +# Coming soon \ No newline at end of file diff --git a/tools/quantization/tensorrt/post_training/Calibrator.py b/tools/quantization/tensorrt/post_training/Calibrator.py new file mode 100644 index 0000000000000000000000000000000000000000..e73e4187ba19f3e27496171ae00296ad8dc0dc79 --- /dev/null +++ b/tools/quantization/tensorrt/post_training/Calibrator.py @@ -0,0 +1,210 @@ +# +# Modified by Meituan +# 2022.6.24 +# + +# Copyright 2019 NVIDIA Corporation +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import sys +import glob +import random +import logging +import cv2 + +import numpy as np +from PIL import Image +import tensorrt as trt +import pycuda.driver as cuda +import pycuda.autoinit + +logging.basicConfig(level=logging.DEBUG, + format="%(asctime)s - %(name)s - %(levelname)s - %(message)s", + datefmt="%Y-%m-%d %H:%M:%S") +logger = logging.getLogger(__name__) + +def preprocess_yolov6(image, channels=3, height=224, width=224): + """Pre-processing for YOLOv6-based Object Detection Models + + Parameters + ---------- + image: PIL.Image + The image resulting from PIL.Image.open(filename) to preprocess + channels: int + The number of channels the image has (Usually 1 or 3) + height: int + The desired height of the image (usually 640) + width: int + The desired width of the image (usually 640) + + Returns + ------- + img_data: numpy array + The preprocessed image data in the form of a numpy array + + """ + # Get the image in CHW format + resized_image = image.resize((width, height), Image.BILINEAR) + img_data = np.asarray(resized_image).astype(np.float32) + + if len(img_data.shape) == 2: + # For images without a channel dimension, we stack + img_data = np.stack([img_data] * 3) + logger.debug("Received grayscale image. Reshaped to {:}".format(img_data.shape)) + else: + img_data = img_data.transpose([2, 0, 1]) + + mean_vec = np.array([0.0, 0.0, 0.0]) + stddev_vec = np.array([1.0, 1.0, 1.0]) + assert img_data.shape[0] == channels + + for i in range(img_data.shape[0]): + # Scale each pixel to [0, 1] and normalize per channel. + img_data[i, :, :] = (img_data[i, :, :] / 255.0 - mean_vec[i]) / stddev_vec[i] + + return img_data + +def get_int8_calibrator(calib_cache, calib_data, max_calib_size, calib_batch_size): + # Use calibration cache if it exists + if os.path.exists(calib_cache): + logger.info("Skipping calibration files, using calibration cache: {:}".format(calib_cache)) + calib_files = [] + # Use calibration files from validation dataset if no cache exists + else: + if not calib_data: + raise ValueError("ERROR: Int8 mode requested, but no calibration data provided. Please provide --calibration-data /path/to/calibration/files") + + calib_files = get_calibration_files(calib_data, max_calib_size) + + # Choose pre-processing function for INT8 calibration + preprocess_func = preprocess_yolov6 + + int8_calibrator = ImageCalibrator(calibration_files=calib_files, + batch_size=calib_batch_size, + cache_file=calib_cache) + return int8_calibrator + + +def get_calibration_files(calibration_data, max_calibration_size=None, allowed_extensions=(".jpeg", ".jpg", ".png")): + """Returns a list of all filenames ending with `allowed_extensions` found in the `calibration_data` directory. + + Parameters + ---------- + calibration_data: str + Path to directory containing desired files. + max_calibration_size: int + Max number of files to use for calibration. If calibration_data contains more than this number, + a random sample of size max_calibration_size will be returned instead. If None, all samples will be used. + + Returns + ------- + calibration_files: List[str] + List of filenames contained in the `calibration_data` directory ending with `allowed_extensions`. + """ + + logger.info("Collecting calibration files from: {:}".format(calibration_data)) + calibration_files = [path for path in glob.iglob(os.path.join(calibration_data, "**"), recursive=True) + if os.path.isfile(path) and path.lower().endswith(allowed_extensions)] + logger.info("Number of Calibration Files found: {:}".format(len(calibration_files))) + + if len(calibration_files) == 0: + raise Exception("ERROR: Calibration data path [{:}] contains no files!".format(calibration_data)) + + if max_calibration_size: + if len(calibration_files) > max_calibration_size: + logger.warning("Capping number of calibration images to max_calibration_size: {:}".format(max_calibration_size)) + random.seed(42) # Set seed for reproducibility + calibration_files = random.sample(calibration_files, max_calibration_size) + + return calibration_files + + +# https://docs.nvidia.com/deeplearning/sdk/tensorrt-api/python_api/infer/Int8/EntropyCalibrator2.html +class ImageCalibrator(trt.IInt8EntropyCalibrator2): + """INT8 Calibrator Class for Imagenet-based Image Classification Models. + + Parameters + ---------- + calibration_files: List[str] + List of image filenames to use for INT8 Calibration + batch_size: int + Number of images to pass through in one batch during calibration + input_shape: Tuple[int] + Tuple of integers defining the shape of input to the model (Default: (3, 224, 224)) + cache_file: str + Name of file to read/write calibration cache from/to. + preprocess_func: function -> numpy.ndarray + Pre-processing function to run on calibration data. This should match the pre-processing + done at inference time. In general, this function should return a numpy array of + shape `input_shape`. + """ + + def __init__(self, calibration_files=[], batch_size=32, input_shape=(3, 224, 224), + cache_file="calibration.cache", use_cv2=False): + super().__init__() + self.input_shape = input_shape + self.cache_file = cache_file + self.batch_size = batch_size + self.batch = np.zeros((self.batch_size, *self.input_shape), dtype=np.float32) + self.device_input = cuda.mem_alloc(self.batch.nbytes) + + self.files = calibration_files + self.use_cv2 = use_cv2 + # Pad the list so it is a multiple of batch_size + if len(self.files) % self.batch_size != 0: + logger.info("Padding # calibration files to be a multiple of batch_size {:}".format(self.batch_size)) + self.files += calibration_files[(len(calibration_files) % self.batch_size):self.batch_size] + + self.batches = self.load_batches() + self.preprocess_func = preprocess_yolov6 + + def load_batches(self): + # Populates a persistent self.batch buffer with images. + for index in range(0, len(self.files), self.batch_size): + for offset in range(self.batch_size): + if self.use_cv2: + image = cv2.imread(self.files[index + offset]) + else: + image = Image.open(self.files[index + offset]) + self.batch[offset] = self.preprocess_func(image, *self.input_shape) + logger.info("Calibration images pre-processed: {:}/{:}".format(index+self.batch_size, len(self.files))) + yield self.batch + + def get_batch_size(self): + return self.batch_size + + def get_batch(self, names): + try: + # Assume self.batches is a generator that provides batch data. + batch = next(self.batches) + # Assume that self.device_input is a device buffer allocated by the constructor. + cuda.memcpy_htod(self.device_input, batch) + return [int(self.device_input)] + except StopIteration: + # When we're out of batches, we return either [] or None. + # This signals to TensorRT that there is no calibration data remaining. + return None + + def read_calibration_cache(self): + # If there is a cache, use it instead of calibrating again. Otherwise, implicitly return None. + if os.path.exists(self.cache_file): + with open(self.cache_file, "rb") as f: + logger.info("Using calibration cache to save time: {:}".format(self.cache_file)) + return f.read() + + def write_calibration_cache(self, cache): + with open(self.cache_file, "wb") as f: + logger.info("Caching calibration data for future use: {:}".format(self.cache_file)) + f.write(cache) diff --git a/tools/quantization/tensorrt/post_training/LICENSE b/tools/quantization/tensorrt/post_training/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..bfbde9d70bda22f4809e95fd12079d820f94db0a --- /dev/null +++ b/tools/quantization/tensorrt/post_training/LICENSE @@ -0,0 +1,192 @@ + + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + Copyright 2020 NVIDIA Corporation + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. + diff --git a/tools/quantization/tensorrt/post_training/README.md b/tools/quantization/tensorrt/post_training/README.md new file mode 100644 index 0000000000000000000000000000000000000000..b50ad780682558b57554490cef1f9653dcd4ac8d --- /dev/null +++ b/tools/quantization/tensorrt/post_training/README.md @@ -0,0 +1,83 @@ +# ONNX -> TensorRT INT8 +These scripts were last tested using the +[NGC TensorRT Container Version 20.06-py3](https://ngc.nvidia.com/catalog/containers/nvidia:tensorrt). +You can see the corresponding framework versions for this container [here](https://docs.nvidia.com/deeplearning/sdk/tensorrt-container-release-notes/rel_20.06.html#rel_20.06). + +## Quickstart + +> **NOTE**: This INT8 example is only valid for **fixed-shape** ONNX models at the moment. +> +INT8 Calibration on **dynamic-shape** models is now supported, however this example has not been updated +to reflect that yet. For more details on INT8 Calibration for **dynamic-shape** models, please +see the [documentation](https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#int8-calib-dynamic-shapes). + +### 1. Convert ONNX model to TensorRT INT8 + +See `./onnx_to_tensorrt.py -h` for full list of command line arguments. + +```bash +./onnx_to_tensorrt.py --explicit-batch \ + --onnx resnet50/model.onnx \ + --fp16 \ + --int8 \ + --calibration-cache="caches/yolov6.cache" \ + -o resnet50.int8.engine +``` + +See the [INT8 Calibration](#int8-calibration) section below for details on calibration +using your own model or different data, where you don't have an existing calibration cache +or want to create a new one. + +## INT8 Calibration + +See [ImagenetCalibrator.py](ImagenetCalibrator.py) for a reference implementation +of TensorRT's [IInt8EntropyCalibrator2](https://docs.nvidia.com/deeplearning/sdk/tensorrt-api/python_api/infer/Int8/EntropyCalibrator2.html). + +This class can be tweaked to work for other kinds of models, inputs, etc. + +In the [Quickstart](#quickstart) section above, we made use of a pre-existing cache, +[caches/yolov6.cache](caches/yolov6.cache), to save time for the sake of an example. + +However, to calibrate using different data or a different model, you can do so with the `--calibration-data` argument. + +* This requires that you've mounted a dataset, such as Imagenet, to use for calibration. + * Add something like `-v /imagenet:/imagenet` to your Docker command in Step (1) + to mount a dataset found locally at `/imagenet`. +* You can specify your own `preprocess_func` by defining it inside of `ImageCalibrator.py` + +```bash +# Path to dataset to use for calibration. +# **Not necessary if you already have a calibration cache from a previous run. +CALIBRATION_DATA="/imagenet" + +# Truncate calibration images to a random sample of this amount if more are found. +# **Not necessary if you already have a calibration cache from a previous run. +MAX_CALIBRATION_SIZE=512 + +# Calibration cache to be used instead of calibration data if it already exists, +# or the cache will be created from the calibration data if it doesn't exist. +CACHE_FILENAME="caches/yolov6.cache" + +# Path to ONNX model +ONNX_MODEL="model/yolov6.onnx" + +# Path to write TensorRT engine to +OUTPUT="yolov6.int8.engine" + +# Creates an int8 engine from your ONNX model, creating ${CACHE_FILENAME} based +# on your ${CALIBRATION_DATA}, unless ${CACHE_FILENAME} already exists, then +# it will use simply use that instead. +python3 onnx_to_tensorrt.py --fp16 --int8 -v \ + --max_calibration_size=${MAX_CALIBRATION_SIZE} \ + --calibration-data=${CALIBRATION_DATA} \ + --calibration-cache=${CACHE_FILENAME} \ + --preprocess_func=${PREPROCESS_FUNC} \ + --explicit-batch \ + --onnx ${ONNX_MODEL} -o ${OUTPUT} + +``` + +### Pre-processing + +In order to calibrate your model correctly, you should `pre-process` your data the same way +that you would during inference. \ No newline at end of file diff --git a/tools/quantization/tensorrt/post_training/onnx_to_tensorrt.py b/tools/quantization/tensorrt/post_training/onnx_to_tensorrt.py new file mode 100644 index 0000000000000000000000000000000000000000..4e52d401a3a9da6723a19c63e812d18b6884f38f --- /dev/null +++ b/tools/quantization/tensorrt/post_training/onnx_to_tensorrt.py @@ -0,0 +1,220 @@ +#!/usr/bin/env python3 + +# +# Modified by Meituan +# 2022.6.24 +# + +# Copyright 2019 NVIDIA Corporation +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import sys +import glob +import math +import logging +import argparse + +import tensorrt as trt +#sys.path.remove('/opt/ros/kinetic/lib/python2.7/dist-packages') + +TRT_LOGGER = trt.Logger() +logging.basicConfig(level=logging.DEBUG, + format="%(asctime)s - %(name)s - %(levelname)s - %(message)s", + datefmt="%Y-%m-%d %H:%M:%S") +logger = logging.getLogger(__name__) + + +def add_profiles(config, inputs, opt_profiles): + logger.debug("=== Optimization Profiles ===") + for i, profile in enumerate(opt_profiles): + for inp in inputs: + _min, _opt, _max = profile.get_shape(inp.name) + logger.debug("{} - OptProfile {} - Min {} Opt {} Max {}".format(inp.name, i, _min, _opt, _max)) + config.add_optimization_profile(profile) + + +def mark_outputs(network): + # Mark last layer's outputs if not already marked + # NOTE: This may not be correct in all cases + last_layer = network.get_layer(network.num_layers-1) + if not last_layer.num_outputs: + logger.error("Last layer contains no outputs.") + return + + for i in range(last_layer.num_outputs): + network.mark_output(last_layer.get_output(i)) + + +def check_network(network): + if not network.num_outputs: + logger.warning("No output nodes found, marking last layer's outputs as network outputs. Correct this if wrong.") + mark_outputs(network) + + inputs = [network.get_input(i) for i in range(network.num_inputs)] + outputs = [network.get_output(i) for i in range(network.num_outputs)] + max_len = max([len(inp.name) for inp in inputs] + [len(out.name) for out in outputs]) + + logger.debug("=== Network Description ===") + for i, inp in enumerate(inputs): + logger.debug("Input {0} | Name: {1:{2}} | Shape: {3}".format(i, inp.name, max_len, inp.shape)) + for i, out in enumerate(outputs): + logger.debug("Output {0} | Name: {1:{2}} | Shape: {3}".format(i, out.name, max_len, out.shape)) + + +def get_batch_sizes(max_batch_size): + # Returns powers of 2, up to and including max_batch_size + max_exponent = math.log2(max_batch_size) + for i in range(int(max_exponent)+1): + batch_size = 2**i + yield batch_size + + if max_batch_size != batch_size: + yield max_batch_size + + +# TODO: This only covers dynamic shape for batch size, not dynamic shape for other dimensions +def create_optimization_profiles(builder, inputs, batch_sizes=[1,8,16,32,64]): + # Check if all inputs are fixed explicit batch to create a single profile and avoid duplicates + if all([inp.shape[0] > -1 for inp in inputs]): + profile = builder.create_optimization_profile() + for inp in inputs: + fbs, shape = inp.shape[0], inp.shape[1:] + profile.set_shape(inp.name, min=(fbs, *shape), opt=(fbs, *shape), max=(fbs, *shape)) + return [profile] + + # Otherwise for mixed fixed+dynamic explicit batch inputs, create several profiles + profiles = {} + for bs in batch_sizes: + if not profiles.get(bs): + profiles[bs] = builder.create_optimization_profile() + + for inp in inputs: + shape = inp.shape[1:] + # Check if fixed explicit batch + if inp.shape[0] > -1: + bs = inp.shape[0] + + profiles[bs].set_shape(inp.name, min=(bs, *shape), opt=(bs, *shape), max=(bs, *shape)) + + return list(profiles.values()) + +def main(): + parser = argparse.ArgumentParser(description="Creates a TensorRT engine from the provided ONNX file.\n") + parser.add_argument("--onnx", required=True, help="The ONNX model file to convert to TensorRT") + parser.add_argument("-o", "--output", type=str, default="model.engine", help="The path at which to write the engine") + parser.add_argument("-b", "--max-batch-size", type=int, help="The max batch size for the TensorRT engine input") + parser.add_argument("-v", "--verbosity", action="count", help="Verbosity for logging. (None) for ERROR, (-v) for INFO/WARNING/ERROR, (-vv) for VERBOSE.") + parser.add_argument("--explicit-batch", action='store_true', help="Set trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH.") + parser.add_argument("--explicit-precision", action='store_true', help="Set trt.NetworkDefinitionCreationFlag.EXPLICIT_PRECISION.") + parser.add_argument("--gpu-fallback", action='store_true', help="Set trt.BuilderFlag.GPU_FALLBACK.") + parser.add_argument("--refittable", action='store_true', help="Set trt.BuilderFlag.REFIT.") + parser.add_argument("--debug", action='store_true', help="Set trt.BuilderFlag.DEBUG.") + parser.add_argument("--strict-types", action='store_true', help="Set trt.BuilderFlag.STRICT_TYPES.") + parser.add_argument("--fp16", action="store_true", help="Attempt to use FP16 kernels when possible.") + parser.add_argument("--int8", action="store_true", help="Attempt to use INT8 kernels when possible. This should generally be used in addition to the --fp16 flag. \ + ONLY SUPPORTS RESNET-LIKE MODELS SUCH AS RESNET50/VGG16/INCEPTION/etc.") + parser.add_argument("--calibration-cache", help="(INT8 ONLY) The path to read/write from calibration cache.", default="calibration.cache") + parser.add_argument("--calibration-data", help="(INT8 ONLY) The directory containing {*.jpg, *.jpeg, *.png} files to use for calibration. (ex: Imagenet Validation Set)", default=None) + parser.add_argument("--calibration-batch-size", help="(INT8 ONLY) The batch size to use during calibration.", type=int, default=128) + parser.add_argument("--max-calibration-size", help="(INT8 ONLY) The max number of data to calibrate on from --calibration-data.", type=int, default=2048) + parser.add_argument("-s", "--simple", action="store_true", help="Use SimpleCalibrator with random data instead of ImagenetCalibrator for INT8 calibration.") + args, _ = parser.parse_known_args() + + print(args) + + # Adjust logging verbosity + if args.verbosity is None: + TRT_LOGGER.min_severity = trt.Logger.Severity.ERROR + # -v + elif args.verbosity == 1: + TRT_LOGGER.min_severity = trt.Logger.Severity.INFO + # -vv + else: + TRT_LOGGER.min_severity = trt.Logger.Severity.VERBOSE + logger.info("TRT_LOGGER Verbosity: {:}".format(TRT_LOGGER.min_severity)) + + # Network flags + network_flags = 0 + if args.explicit_batch: + network_flags |= 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH) + if args.explicit_precision: + network_flags |= 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_PRECISION) + + builder_flag_map = { + 'gpu_fallback': trt.BuilderFlag.GPU_FALLBACK, + 'refittable': trt.BuilderFlag.REFIT, + 'debug': trt.BuilderFlag.DEBUG, + 'strict_types': trt.BuilderFlag.STRICT_TYPES, + 'fp16': trt.BuilderFlag.FP16, + 'int8': trt.BuilderFlag.INT8, + } + + # Building engine + with trt.Builder(TRT_LOGGER) as builder, \ + builder.create_network(network_flags) as network, \ + builder.create_builder_config() as config, \ + trt.OnnxParser(network, TRT_LOGGER) as parser: + + config.max_workspace_size = 2**30 # 1GiB + + # Set Builder Config Flags + for flag in builder_flag_map: + if getattr(args, flag): + logger.info("Setting {}".format(builder_flag_map[flag])) + config.set_flag(builder_flag_map[flag]) + + # Fill network atrributes with information by parsing model + with open(args.onnx, "rb") as f: + if not parser.parse(f.read()): + print('ERROR: Failed to parse the ONNX file: {}'.format(args.onnx)) + for error in range(parser.num_errors): + print(parser.get_error(error)) + sys.exit(1) + + # Display network info and check certain properties + check_network(network) + + if args.explicit_batch: + # Add optimization profiles + batch_sizes = [1, 8, 16, 32, 64] + inputs = [network.get_input(i) for i in range(network.num_inputs)] + opt_profiles = create_optimization_profiles(builder, inputs, batch_sizes) + add_profiles(config, inputs, opt_profiles) + # Implicit Batch Network + else: + builder.max_batch_size = args.max_batch_size + opt_profiles = [] + + # Precision flags + if args.fp16 and not builder.platform_has_fast_fp16: + logger.warning("FP16 not supported on this platform.") + + if args.int8 and not builder.platform_has_fast_int8: + logger.warning("INT8 not supported on this platform.") + + if args.int8: + from Calibrator import ImageCalibrator, get_int8_calibrator # local module + config.int8_calibrator = get_int8_calibrator(args.calibration_cache, + args.calibration_data, + args.max_calibration_size, + args.calibration_batch_size) + + logger.info("Building Engine...") + with builder.build_engine(network, config) as engine, open(args.output, "wb") as f: + logger.info("Serializing engine to file: {:}".format(args.output)) + f.write(engine.serialize()) + +if __name__ == "__main__": + main() diff --git a/tools/quantization/tensorrt/post_training/quant.sh b/tools/quantization/tensorrt/post_training/quant.sh new file mode 100644 index 0000000000000000000000000000000000000000..9a66cc58178226a35eacd2ca23db47ab70acf78d --- /dev/null +++ b/tools/quantization/tensorrt/post_training/quant.sh @@ -0,0 +1,23 @@ +# Path to ONNX model +# ex: ../yolov6.onnx +ONNX_MODEL=$1 + +# Path to dataset to use for calibration. +# **Not necessary if you already have a calibration cache from a previous run. +CALIBRATION_DATA=$2 + +# Path to Cache file to Serving +# ex: ./caches/demo.cache +CACHE_FILENAME=$3 + +# Path to write TensorRT engine to +OUTPUT=$4 + +# Creates an int8 engine from your ONNX model, creating ${CACHE_FILENAME} based +# on your ${CALIBRATION_DATA}, unless ${CACHE_FILENAME} already exists, then +# it will use simply use that instead. +python3 onnx_to_tensorrt.py --fp16 --int8 -v \ + --calibration-data=${CALIBRATION_DATA} \ + --calibration-cache=${CACHE_FILENAME} \ + --explicit-batch \ + --onnx ${ONNX_MODEL} -o ${OUTPUT} diff --git a/tools/quantization/tensorrt/requirements.txt b/tools/quantization/tensorrt/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..36d9384213a3c7fa9c7df0e0b9cf5c164b66cdca --- /dev/null +++ b/tools/quantization/tensorrt/requirements.txt @@ -0,0 +1,7 @@ +# pip install -r requirements.txt +# python3.8 environment + +tensorrt # TensorRT 8.0+ +pycuda==2020.1 # CUDA 11.0 +nvidia-pyindex +pytorch-quantization \ No newline at end of file diff --git a/tools/quantization/tensorrt/training_aware/QAT_quantizer.py b/tools/quantization/tensorrt/training_aware/QAT_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..4621aa426fd0ba7875717a1442ab07a74cc49b30 --- /dev/null +++ b/tools/quantization/tensorrt/training_aware/QAT_quantizer.py @@ -0,0 +1,39 @@ +# +# QAT_quantizer.py +# YOLOv6 +# +# Created by Meituan on 2022/06/24. +# Copyright © 2022 +# + +from absl import logging +from pytorch_quantization import nn as quant_nn +from pytorch_quantization import quant_modules + +# Call this function before defining the model +def tensorrt_official_qat(): + # Quantization Aware Training is based on Straight Through Estimator (STE) derivative approximation. + # It is some time known as “quantization aware training”. + + # PyTorch-Quantization is a toolkit for training and evaluating PyTorch models with simulated quantization. + # Quantization can be added to the model automatically, or manually, allowing the model to be tuned for accuracy and performance. + # Quantization is compatible with NVIDIAs high performance integer kernels which leverage integer Tensor Cores. + # The quantized model can be exported to ONNX and imported by TensorRT 8.0 and later. + # https://github.com/NVIDIA/TensorRT/blob/main/tools/pytorch-quantization/examples/finetune_quant_resnet50.ipynb + + # The example to export the + # model.eval() + # quant_nn.TensorQuantizer.use_fb_fake_quant = True # We have to shift to pytorch's fake quant ops before exporting the model to ONNX + # opset_version = 13 + + # Export ONNX for multiple batch sizes + # print("Creating ONNX file: " + onnx_filename) + # dummy_input = torch.randn(batch_onnx, 3, 224, 224, device='cuda') #TODO: switch input dims by model + # torch.onnx.export(model, dummy_input, onnx_filename, verbose=False, opset_version=opset_version, enable_onnx_checker=False, do_constant_folding=True) + try: + quant_modules.initialize() + except NameError: + logging.info("initialzation error for quant_modules") + +# def QAT_quantizer(): +# coming soon \ No newline at end of file diff --git a/tools/train.py b/tools/train.py new file mode 100644 index 0000000000000000000000000000000000000000..927d997bd8c6e572e43bcd8e137928ceab0f9afb --- /dev/null +++ b/tools/train.py @@ -0,0 +1,87 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import argparse +import os +import os.path as osp +import torch +import torch.distributed as dist +import sys + +ROOT = os.getcwd() +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) + +from yolov6.core.engine import Trainer +from yolov6.utils.config import Config +from yolov6.utils.events import LOGGER, save_yaml +from yolov6.utils.envs import get_envs, select_device, set_random_seed + + +def get_args_parser(add_help=True): + parser = argparse.ArgumentParser(description='YOLOv6 PyTorch Training', add_help=add_help) + parser.add_argument('--data-path', default='./data/coco.yaml', type=str, help='dataset path') + parser.add_argument('--conf-file', default='./configs/yolov6s.py', type=str, help='experiment description file') + parser.add_argument('--img-size', type=int, default=640, help='train, val image size (pixels)') + parser.add_argument('--batch-size', default=32, type=int, help='total batch size for all GPUs') + parser.add_argument('--epochs', default=400, type=int, help='number of total epochs to run') + parser.add_argument('--workers', default=8, type=int, help='number of data loading workers (default: 8)') + parser.add_argument('--device', default='0', type=str, help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--noval', action='store_true', help='only evaluate in final epoch') + parser.add_argument('--check-images', action='store_true', help='check images when initializing datasets') + parser.add_argument('--check-labels', action='store_true', help='check label files when initializing datasets') + parser.add_argument('--output-dir', default='./runs/train', type=str, help='path to save outputs') + parser.add_argument('--name', default='exp', type=str, help='experiment name, save to output_dir/name') + parser.add_argument('--dist_url', type=str, default="tcp://127.0.0.1:8888") + parser.add_argument('--gpu_count', type=int, default=0) + parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify') + + return parser + + +def check_and_init(args): + '''check config files and device, and initialize ''' + + # check files + args.save_dir = osp.join(args.output_dir, args.name) + os.makedirs(args.save_dir, exist_ok=True) + cfg = Config.fromfile(args.conf_file) + + # check device + device = select_device(args.device) + + # set random seed + set_random_seed(1+args.rank, deterministic=(args.rank == -1)) + + # save args + save_yaml(vars(args), osp.join(args.save_dir, 'args.yaml')) + + return cfg, device + + +def main(args): + '''main function of training''' + # Setup + args.rank, args.local_rank, args.world_size = get_envs() + LOGGER.info(f'training args are: {args}\n') + cfg, device = check_and_init(args) + + if args.local_rank != -1: # if DDP mode + torch.cuda.set_device(args.local_rank) + device = torch.device('cuda', args.local_rank) + LOGGER.info('Initializing process group... ') + dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo", \ + init_method=args.dist_url, rank=args.local_rank, world_size=args.world_size) + + # Start + trainer = Trainer(args, cfg, device) + trainer.train() + + # End + if args.world_size > 1 and args.rank == 0: + LOGGER.info('Destroying process group... ') + dist.destroy_process_group() + + +if __name__ == '__main__': + args = get_args_parser().parse_args() + main(args) diff --git a/yolov6/core/engine.py b/yolov6/core/engine.py new file mode 100644 index 0000000000000000000000000000000000000000..927523ed8accff0dd6410a5e1fcc529f9247d16a --- /dev/null +++ b/yolov6/core/engine.py @@ -0,0 +1,262 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import os +import time +from copy import deepcopy +import os.path as osp + +from tqdm import tqdm + +import numpy as np +import torch +from torch.cuda import amp +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter + +import tools.eval as eval +from yolov6.data.data_load import create_dataloader +from yolov6.models.yolo import build_model +from yolov6.models.loss import ComputeLoss +from yolov6.utils.events import LOGGER, NCOLS, load_yaml, write_tblog +from yolov6.utils.ema import ModelEMA, de_parallel +from yolov6.utils.checkpoint import load_state_dict, save_checkpoint, strip_optimizer +from yolov6.solver.build import build_optimizer, build_lr_scheduler + + +class Trainer: + def __init__(self, args, cfg, device): + self.args = args + self.cfg = cfg + self.device = device + + self.rank = args.rank + self.local_rank = args.local_rank + self.world_size = args.world_size + self.main_process = self.rank in [-1, 0] + self.save_dir = args.save_dir + # get data loader + self.data_dict = load_yaml(args.data_path) + self.num_classes = self.data_dict['nc'] + self.train_loader, self.val_loader = self.get_data_loader(args, cfg, self.data_dict) + # get model and optimizer + model = self.get_model(args, cfg, self.num_classes, device) + self.optimizer = self.get_optimizer(args, cfg, model) + self.scheduler, self.lf = self.get_lr_scheduler(args, cfg, self.optimizer) + self.ema = ModelEMA(model) if self.main_process else None + self.model = self.parallel_model(args, model, device) + self.model.nc, self.model.names = self.data_dict['nc'], self.data_dict['names'] + # tensorboard + self.tblogger = SummaryWriter(self.save_dir) if self.main_process else None + + self.start_epoch = 0 + self.max_epoch = args.epochs + self.max_stepnum = len(self.train_loader) + self.batch_size = args.batch_size + self.img_size = args.img_size + + # Training Process + def train(self): + try: + self.train_before_loop() + for self.epoch in range(self.start_epoch, self.max_epoch): + self.train_in_loop() + + except Exception as _: + LOGGER.error('ERROR in training loop or eval/save model.') + raise + finally: + self.train_after_loop() + + # Training loop for each epoch + def train_in_loop(self): + try: + self.prepare_for_steps() + for self.step, self.batch_data in self.pbar: + self.train_in_steps() + self.print_details() + except Exception as _: + LOGGER.error('ERROR in training steps.') + raise + try: + self.eval_and_save() + except Exception as _: + LOGGER.error('ERROR in evaluate and save model.') + raise + + # Training loop for batchdata + def train_in_steps(self): + images, targets = self.prepro_data(self.batch_data, self.device) + # forward + with amp.autocast(enabled=self.device != 'cpu'): + preds = self.model(images) + total_loss, loss_items = self.compute_loss(preds, targets) + if self.rank != -1: + total_loss *= self.world_size + # backward + self.scaler.scale(total_loss).backward() + self.loss_items = loss_items + self.update_optimizer() + + def eval_and_save(self): + epoch_sub = self.max_epoch - self.epoch + val_period = 20 if epoch_sub > 100 else 1 # to fasten training time, evaluate in every 20 epochs for the early stage. + is_val_epoch = (not self.args.noval or (epoch_sub == 1)) and (self.epoch % val_period == 0) + if self.main_process: + self.ema.update_attr(self.model, include=['nc', 'names', 'stride']) # update attributes for ema model + if is_val_epoch: + self.eval_model() + self.ap = self.evaluate_results[0] * 0.1 + self.evaluate_results[1] * 0.9 + self.best_ap = max(self.ap, self.best_ap) + # save ckpt + ckpt = { + 'model': deepcopy(de_parallel(self.model)).half(), + 'ema': deepcopy(self.ema.ema).half(), + 'updates': self.ema.updates, + 'optimizer': self.optimizer.state_dict(), + 'epoch': self.epoch, + } + + save_ckpt_dir = osp.join(self.save_dir, 'weights') + save_checkpoint(ckpt, (is_val_epoch) and (self.ap == self.best_ap), save_ckpt_dir, model_name='last_ckpt') + del ckpt + # log for tensorboard + write_tblog(self.tblogger, self.epoch, self.evaluate_results, self.mean_loss) + + def eval_model(self): + results = eval.run(self.data_dict, + batch_size=self.batch_size // self.world_size * 2, + img_size=self.img_size, + model=self.ema.ema, + dataloader=self.val_loader, + save_dir=self.save_dir, + task='train') + + LOGGER.info(f"Epoch: {self.epoch} | mAP@0.5: {results[0]} | mAP@0.50:0.95: {results[1]}") + self.evaluate_results = results[:2] + + def train_before_loop(self): + LOGGER.info('Training start...') + self.start_time = time.time() + self.warmup_stepnum = max(round(self.cfg.solver.warmup_epochs * self.max_stepnum), 1000) + self.scheduler.last_epoch = self.start_epoch - 1 + self.last_opt_step = -1 + self.scaler = amp.GradScaler(enabled=self.device != 'cpu') + + self.best_ap, self.ap = 0.0, 0.0 + self.evaluate_results = (0, 0) # AP50, AP50_95 + self.compute_loss = ComputeLoss(iou_type=self.cfg.model.head.iou_type) + + def prepare_for_steps(self): + if self.epoch > self.start_epoch: + self.scheduler.step() + self.model.train() + if self.rank != -1: + self.train_loader.sampler.set_epoch(self.epoch) + self.mean_loss = torch.zeros(4, device=self.device) + self.optimizer.zero_grad() + + LOGGER.info(('\n' + '%10s' * 5) % ('Epoch', 'iou_loss', 'l1_loss', 'obj_loss', 'cls_loss')) + self.pbar = enumerate(self.train_loader) + if self.main_process: + self.pbar = tqdm(self.pbar, total=self.max_stepnum, ncols=NCOLS, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') + + # Print loss after each steps + def print_details(self): + if self.main_process: + self.mean_loss = (self.mean_loss * self.step + self.loss_items) / (self.step + 1) + self.pbar.set_description(('%10s' + '%10.4g' * 4) % (f'{self.epoch}/{self.max_epoch - 1}', \ + *(self.mean_loss))) + + # Empty cache if training finished + def train_after_loop(self): + if self.main_process: + LOGGER.info(f'\nTraining completed in {(time.time() - self.start_time) / 3600:.3f} hours.') + save_ckpt_dir = osp.join(self.save_dir, 'weights') + strip_optimizer(save_ckpt_dir) # strip optimizers for saved pt model + if self.device != 'cpu': + torch.cuda.empty_cache() + + def update_optimizer(self): + curr_step = self.step + self.max_stepnum * self.epoch + self.accumulate = max(1, round(64 / self.batch_size)) + if curr_step <= self.warmup_stepnum: + self.accumulate = max(1, np.interp(curr_step, [0, self.warmup_stepnum], [1, 64 / self.batch_size]).round()) + for k, param in enumerate(self.optimizer.param_groups): + warmup_bias_lr = self.cfg.solver.warmup_bias_lr if k == 2 else 0.0 + param['lr'] = np.interp(curr_step, [0, self.warmup_stepnum], [warmup_bias_lr, param['initial_lr'] * self.lf(self.epoch)]) + if 'momentum' in param: + param['momentum'] = np.interp(curr_step, [0, self.warmup_stepnum], [self.cfg.solver.warmup_momentum, self.cfg.solver.momentum]) + if curr_step - self.last_opt_step >= self.accumulate: + self.scaler.step(self.optimizer) + self.scaler.update() + self.optimizer.zero_grad() + if self.ema: + self.ema.update(self.model) + self.last_opt_step = curr_step + + @staticmethod + def get_data_loader(args, cfg, data_dict): + train_path, val_path = data_dict['train'], data_dict['val'] + # check data + nc = int(data_dict['nc']) + class_names = data_dict['names'] + assert len(class_names) == nc, f'the length of class names does not match the number of classes defined' + grid_size = max(int(max(cfg.model.head.strides)), 32) + # create train dataloader + train_loader = create_dataloader(train_path, args.img_size, args.batch_size // args.world_size, grid_size, + hyp=dict(cfg.data_aug), augment=True, rect=False, rank=args.local_rank, + workers=args.workers, shuffle=True, check_images=args.check_images, + check_labels=args.check_labels, class_names=class_names, task='train')[0] + # create val dataloader + val_loader = None + if args.rank in [-1, 0]: + val_loader = create_dataloader(val_path, args.img_size, args.batch_size // args.world_size * 2, grid_size, + hyp=dict(cfg.data_aug), rect=True, rank=-1, pad=0.5, + workers=args.workers, check_images=args.check_images, + check_labels=args.check_labels, class_names=class_names, task='val')[0] + + return train_loader, val_loader + + @staticmethod + def prepro_data(batch_data, device): + images = batch_data[0].to(device, non_blocking=True).float() / 255 + targets = batch_data[1].to(device) + return images, targets + + @staticmethod + def get_model(args, cfg, nc, device): + model = build_model(cfg, nc, device) + weights = cfg.model.pretrained + if weights: # finetune if pretrained model is set + LOGGER.info(f'Loading state_dict from {weights} for fine-tuning...') + model = load_state_dict(weights, model, map_location=device) + LOGGER.info('Model: {}'.format(model)) + return model + + @staticmethod + def parallel_model(args, model, device): + # If DP mode + dp_mode = device.type != 'cpu' and args.rank == -1 + if dp_mode and torch.cuda.device_count() > 1: + LOGGER.warning('WARNING: DP not recommended, use DDP instead.\n') + model = torch.nn.DataParallel(model) + + # If DDP mode + ddp_mode = device.type != 'cpu' and args.rank != -1 + if ddp_mode: + model = DDP(model, device_ids=[args.local_rank], output_device=args.local_rank) + + return model + + @staticmethod + def get_optimizer(args, cfg, model): + accumulate = max(1, round(64 / args.batch_size)) + cfg.solver.weight_decay *= args.batch_size * accumulate / 64 + optimizer = build_optimizer(cfg, model) + return optimizer + + @staticmethod + def get_lr_scheduler(args, cfg, optimizer): + epochs = args.epochs + lr_scheduler, lf = build_lr_scheduler(cfg, optimizer, epochs) + return lr_scheduler, lf diff --git a/yolov6/core/evaler.py b/yolov6/core/evaler.py new file mode 100644 index 0000000000000000000000000000000000000000..ef617177b6bce4c120bd6eb34bfa5353bb53cc20 --- /dev/null +++ b/yolov6/core/evaler.py @@ -0,0 +1,258 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import os +from tqdm import tqdm +import numpy as np +import json +import torch +import yaml +from pathlib import Path + +from pycocotools.coco import COCO +from pycocotools.cocoeval import COCOeval + +from yolov6.data.data_load import create_dataloader +from yolov6.utils.events import LOGGER, NCOLS +from yolov6.utils.nms import non_max_suppression +from yolov6.utils.checkpoint import load_checkpoint +from yolov6.utils.torch_utils import time_sync, get_model_info + +''' + +python tools/eval.py --task 'train'/'val'/'speed' + +''' + + +class Evaler: + def __init__(self, + data, + batch_size=32, + img_size=640, + conf_thres=0.001, + iou_thres=0.65, + device='', + half=True, + save_dir=''): + self.data = data + self.batch_size = batch_size + self.img_size = img_size + self.conf_thres = conf_thres + self.iou_thres = iou_thres + self.device = device + self.half = half + self.save_dir = save_dir + + def init_model(self, model, weights, task): + if task != 'train': + model = load_checkpoint(weights, map_location=self.device) + self.stride = int(model.stride.max()) + if self.device.type != 'cpu': + model(torch.zeros(1, 3, self.img_size, self.img_size).to(self.device).type_as(next(model.parameters()))) + # switch to deploy + from yolov6.layers.common import RepVGGBlock + for layer in model.modules(): + if isinstance(layer, RepVGGBlock): + layer.switch_to_deploy() + LOGGER.info("Switch model to deploy modality.") + LOGGER.info("Model Summary: {}".format(get_model_info(model, self.img_size))) + model.half() if self.half else model.float() + return model + + def init_data(self, dataloader, task): + '''Initialize dataloader. + Returns a dataloader for task val or speed. + ''' + self.is_coco = isinstance(self.data.get('val'), str) and 'coco' in self.data['val'] # COCO dataset + self.ids = self.coco80_to_coco91_class() if self.is_coco else list(range(1000)) + if task != 'train': + pad = 0.0 if task == 'speed' else 0.5 + dataloader = create_dataloader(self.data[task if task in ('train', 'val', 'test') else 'val'], + self.img_size, self.batch_size, self.stride, pad=pad, rect=True, + class_names=self.data['names'], task=task)[0] + return dataloader + + def predict_model(self, model, dataloader, task): + '''Model prediction + Predicts the whole dataset and gets the prediced results and inference time. + ''' + self.speed_result = torch.zeros(4, device=self.device) + pred_results = [] + pbar = tqdm(dataloader, desc="Inferencing model in val datasets.", ncols=NCOLS) + for imgs, targets, paths, shapes in pbar: + # pre-process + t1 = time_sync() + imgs = imgs.to(self.device, non_blocking=True) + imgs = imgs.half() if self.half else imgs.float() + imgs /= 255 + self.speed_result[1] += time_sync() - t1 # pre-process time + + # Inference + t2 = time_sync() + outputs = model(imgs) + self.speed_result[2] += time_sync() - t2 # inference time + + # post-process + t3 = time_sync() + outputs = non_max_suppression(outputs, self.conf_thres, self.iou_thres, multi_label=True) + self.speed_result[3] += time_sync() - t3 # post-process time + self.speed_result[0] += len(outputs) + + # save result + pred_results.extend(self.convert_to_coco_format(outputs, imgs, paths, shapes, self.ids)) + return pred_results + + def eval_model(self, pred_results, model, dataloader, task): + '''Evaluate models + For task speed, this function only evaluates the speed of model and outputs inference time. + For task val, this function evalutates the speed and mAP by pycocotools, and returns + inference time and mAP value. + ''' + LOGGER.info(f'\nEvaluating speed.') + self.eval_speed(task) + + LOGGER.info(f'\nEvaluating mAP by pycocotools.') + if task != 'speed' and len(pred_results): + if 'anno_path' in self.data: + anno_json = self.data['anno_path'] + else: + # generated coco format labels in dataset initialization + dataset_root = os.path.dirname(os.path.dirname(self.data['val'])) + base_name = os.path.basename(self.data['val']) + anno_json = os.path.join(dataset_root, 'annotations', f'instances_{base_name}.json') + pred_json = os.path.join(self.save_dir, "predictions.json") + LOGGER.info(f'Saving {pred_json}...') + with open(pred_json, 'w') as f: + json.dump(pred_results, f) + + anno = COCO(anno_json) + pred = anno.loadRes(pred_json) + cocoEval = COCOeval(anno, pred, 'bbox') + if self.is_coco: + imgIds = [int(os.path.basename(x).split(".")[0]) + for x in dataloader.dataset.img_paths] + cocoEval.params.imgIds = imgIds + cocoEval.evaluate() + cocoEval.accumulate() + cocoEval.summarize() + map, map50 = cocoEval.stats[:2] # update results (mAP@0.5:0.95, mAP@0.5) + # Return results + model.float() # for training + if task != 'train': + LOGGER.info(f"Results saved to {self.save_dir}") + return (map50, map) + return (0.0, 0.0) + + def eval_speed(self, task): + '''Evaluate model inference speed.''' + if task != 'train': + n_samples = self.speed_result[0].item() + pre_time, inf_time, nms_time = 1000 * self.speed_result[1:].cpu().numpy() / n_samples + for n, v in zip(["pre-process", "inference", "NMS"],[pre_time, inf_time, nms_time]): + LOGGER.info("Average {} time: {:.2f} ms".format(n, v)) + + def box_convert(self, x): + # Convert boxes with shape [n, 4] from [x1, y1, x2, y2] to [x, y, w, h] where x1y1=top-left, x2y2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center + y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center + y[:, 2] = x[:, 2] - x[:, 0] # width + y[:, 3] = x[:, 3] - x[:, 1] # height + return y + + def scale_coords(self, img1_shape, coords, img0_shape, ratio_pad=None): + # Rescale coords (xyxy) from img1_shape to img0_shape + if ratio_pad is None: # calculate from img0_shape + gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1]) # gain = old / new + pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2 # wh padding + else: + gain = ratio_pad[0][0] + pad = ratio_pad[1] + + coords[:, [0, 2]] -= pad[0] # x padding + coords[:, [1, 3]] -= pad[1] # y padding + coords[:, :4] /= gain + if isinstance(coords, torch.Tensor): # faster individually + coords[:, 0].clamp_(0, img0_shape[1]) # x1 + coords[:, 1].clamp_(0, img0_shape[0]) # y1 + coords[:, 2].clamp_(0, img0_shape[1]) # x2 + coords[:, 3].clamp_(0, img0_shape[0]) # y2 + else: # np.array (faster grouped) + coords[:, [0, 2]] = coords[:, [0, 2]].clip(0, img0_shape[1]) # x1, x2 + coords[:, [1, 3]] = coords[:, [1, 3]].clip(0, img0_shape[0]) # y1, y2 + return coords + + def convert_to_coco_format(self, outputs, imgs, paths, shapes, ids): + pred_results = [] + for i, pred in enumerate(outputs): + if len(pred) == 0: + continue + path, shape = Path(paths[i]), shapes[i][0] + self.scale_coords(imgs[i].shape[1:], pred[:, :4], shape, shapes[i][1]) + image_id = int(path.stem) if path.stem.isnumeric() else path.stem + bboxes = self.box_convert(pred[:, 0:4]) + bboxes[:, :2] -= bboxes[:, 2:] / 2 + cls = pred[:, 5] + scores = pred[:, 4] + for ind in range(pred.shape[0]): + category_id = ids[int(cls[ind])] + bbox = [round(x, 3) for x in bboxes[ind].tolist()] + score = round(scores[ind].item(), 5) + pred_data = { + "image_id": image_id, + "category_id": category_id, + "bbox": bbox, + "score": score + } + pred_results.append(pred_data) + return pred_results + + @staticmethod + def check_task(task): + if task not in ['train','val','speed']: + raise Exception("task argument error: only support 'train' / 'val' / 'speed' task.") + + @staticmethod + def reload_thres(conf_thres, iou_thres, task): + '''Sets conf and iou threshold for task val/speed''' + if task != 'train': + if task == 'val': + conf_thres = 0.001 + if task == 'speed': + conf_thres = 0.25 + iou_thres = 0.45 + return conf_thres, iou_thres + + @staticmethod + def reload_device(device, model, task): + # device = 'cpu' or '0' or '0,1,2,3' + if task == 'train': + device = next(model.parameters()).device + else: + if device == 'cpu': + os.environ['CUDA_VISIBLE_DEVICES'] = '-1' + elif device: + os.environ['CUDA_VISIBLE_DEVICES'] = device + assert torch.cuda.is_available() + cuda = device != 'cpu' and torch.cuda.is_available() + device = torch.device('cuda:0' if cuda else 'cpu') + return device + + @staticmethod + def reload_dataset(data): + with open(data, errors='ignore') as yaml_file: + data = yaml.safe_load(yaml_file) + val = data.get('val') + if not os.path.exists(val): + raise Exception('Dataset not found.') + return data + + @staticmethod + def coco80_to_coco91_class(): # converts 80-index (val2014) to 91-index (paper) + # https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/ + x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, + 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, + 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, + 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, + 80, 81, 82, 84, 85, 86, 87, 88, 89, 90] + return x diff --git a/yolov6/core/inferer.py b/yolov6/core/inferer.py new file mode 100644 index 0000000000000000000000000000000000000000..d4aee34440a0fa798da02476393a6df648598da3 --- /dev/null +++ b/yolov6/core/inferer.py @@ -0,0 +1,196 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import os +import os.path as osp +import math + +from tqdm import tqdm + +import numpy as np +import cv2 +import torch +from PIL import ImageFont + +from yolov6.utils.events import LOGGER, load_yaml + +from yolov6.layers.common import DetectBackend +from yolov6.data.data_augment import letterbox +from yolov6.utils.nms import non_max_suppression + + +class Inferer: + def __init__(self, source, weights, device, yaml, img_size, half): + import glob + from yolov6.data.datasets import IMG_FORMATS + + self.__dict__.update(locals()) + + # Init model + self.device = device + self.img_size = img_size + cuda = self.device != 'cpu' and torch.cuda.is_available() + self.device = torch.device('cuda:0' if cuda else 'cpu') + self.model = DetectBackend(weights, device=self.device) + self.stride = self.model.stride + self.class_names = load_yaml(yaml)['names'] + self.img_size = self.check_img_size(self.img_size, s=self.stride) # check image size + + # Half precision + if half & (self.device.type != 'cpu'): + self.model.model.half() + else: + self.model.model.float() + half = False + + if self.device.type != 'cpu': + self.model(torch.zeros(1, 3, *self.img_size).to(self.device).type_as(next(self.model.model.parameters()))) # warmup + + # Load data + if os.path.isdir(source): + img_paths = sorted(glob.glob(os.path.join(source, '*.*'))) # dir + elif os.path.isfile(source): + img_paths = [source] # files + else: + raise Exception(f'Invalid path: {source}') + self.img_paths = [img_path for img_path in img_paths if img_path.split('.')[-1].lower() in IMG_FORMATS] + + def infer(self, conf_thres, iou_thres, classes, agnostic_nms, max_det, save_dir, save_txt, save_img, hide_labels, hide_conf): + ''' Model Inference and results visualization ''' + + for img_path in tqdm(self.img_paths): + img, img_src = self.precess_image(img_path, self.img_size, self.stride, self.half) + img = img.to(self.device) + if len(img.shape) == 3: + img = img[None] + # expand for batch dim + pred_results = self.model(img) + det = non_max_suppression(pred_results, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det)[0] + + save_path = osp.join(save_dir, osp.basename(img_path)) # im.jpg + txt_path = osp.join(save_dir, 'labels', osp.basename(img_path).split('.')[0]) + + gn = torch.tensor(img_src.shape)[[1, 0, 1, 0]] # normalization gain whwh + img_ori = img_src + + # check image and font + assert img_ori.data.contiguous, 'Image needs to be contiguous. Please apply to input images with np.ascontiguousarray(im).' + self.font_check() + + if len(det): + det[:, :4] = self.rescale(img.shape[2:], det[:, :4], img_src.shape).round() + + for *xyxy, conf, cls in reversed(det): + if save_txt: # Write to file + xywh = (self.box_convert(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh + line = (cls, *xywh, conf) + with open(txt_path + '.txt', 'a') as f: + f.write(('%g ' * len(line)).rstrip() % line + '\n') + + if save_img: + class_num = int(cls) # integer class + label = None if hide_labels else (self.class_names[class_num] if hide_conf else f'{self.class_names[class_num]} {conf:.2f}') + + self.plot_box_and_label(img_ori, max(round(sum(img_ori.shape) / 2 * 0.003), 2), xyxy, label, color=self.generate_colors(class_num, True)) + + img_src = np.asarray(img_ori) + + # Save results (image with detections) + if save_img: + cv2.imwrite(save_path, img_src) + + @staticmethod + def precess_image(path, img_size, stride, half): + '''Process image before image inference.''' + try: + img_src = cv2.imread(path) + assert img_src is not None, f'Invalid image: {path}' + except Exception as e: + LOGGER.Warning(e) + image = letterbox(img_src, img_size, stride=stride)[0] + + # Convert + image = image.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB + image = torch.from_numpy(np.ascontiguousarray(image)) + image = image.half() if half else image.float() # uint8 to fp16/32 + image /= 255 # 0 - 255 to 0.0 - 1.0 + + return image, img_src + + @staticmethod + def rescale(ori_shape, boxes, target_shape): + '''Rescale the output to the original image shape''' + ratio = min(ori_shape[0] / target_shape[0], ori_shape[1] / target_shape[1]) + padding = (ori_shape[1] - target_shape[1] * ratio) / 2, (ori_shape[0] - target_shape[0] * ratio) / 2 + + boxes[:, [0, 2]] -= padding[0] + boxes[:, [1, 3]] -= padding[1] + boxes[:, :4] /= ratio + + boxes[:, 0].clamp_(0, target_shape[1]) # x1 + boxes[:, 1].clamp_(0, target_shape[0]) # y1 + boxes[:, 2].clamp_(0, target_shape[1]) # x2 + boxes[:, 3].clamp_(0, target_shape[0]) # y2 + + return boxes + + def check_img_size(self, img_size, s=32, floor=0): + """Make sure image size is a multiple of stride s in each dimension, and return a new shape list of image.""" + if isinstance(img_size, int): # integer i.e. img_size=640 + new_size = max(self.make_divisible(img_size, int(s)), floor) + elif isinstance(img_size, list): # list i.e. img_size=[640, 480] + new_size = [max(self.make_divisible(x, int(s)), floor) for x in img_size] + else: + raise Exception(f"Unsupported type of img_size: {type(img_size)}") + + if new_size != img_size: + print(f'WARNING: --img-size {img_size} must be multiple of max stride {s}, updating to {new_size}') + return new_size if isinstance(img_size,list) else [new_size]*2 + + def make_divisible(self, x, divisor): + # Upward revision the value x to make it evenly divisible by the divisor. + return math.ceil(x / divisor) * divisor + + @staticmethod + def plot_box_and_label(image, lw, box, label='', color=(128, 128, 128), txt_color=(255, 255, 255)): + # Add one xyxy box to image with label + p1, p2 = (int(box[0]), int(box[1])), (int(box[2]), int(box[3])) + cv2.rectangle(image, p1, p2, color, thickness=lw, lineType=cv2.LINE_AA) + if label: + tf = max(lw - 1, 1) # font thickness + w, h = cv2.getTextSize(label, 0, fontScale=lw / 3, thickness=tf)[0] # text width, height + outside = p1[1] - h - 3 >= 0 # label fits outside box + p2 = p1[0] + w, p1[1] - h - 3 if outside else p1[1] + h + 3 + cv2.rectangle(image, p1, p2, color, -1, cv2.LINE_AA) # filled + cv2.putText(image, label, (p1[0], p1[1] - 2 if outside else p1[1] + h + 2), 0, lw / 3, txt_color, + thickness=tf, lineType=cv2.LINE_AA) + + @staticmethod + def font_check(font='./yolov6/utils/Arial.ttf', size=10): + # Return a PIL TrueType Font, downloading to CONFIG_DIR if necessary + assert osp.exists(font), f'font path not exists: {font}' + try: + return ImageFont.truetype(str(font) if font.exists() else font.name, size) + except Exception as e: # download if missing + return ImageFont.truetype(str(font), size) + + @staticmethod + def box_convert(x): + # Convert boxes with shape [n, 4] from [x1, y1, x2, y2] to [x, y, w, h] where x1y1=top-left, x2y2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center + y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center + y[:, 2] = x[:, 2] - x[:, 0] # width + y[:, 3] = x[:, 3] - x[:, 1] # height + return y + + @staticmethod + def generate_colors(i, bgr=False): + hex = ('FF3838', 'FF9D97', 'FF701F', 'FFB21D', 'CFD231', '48F90A', '92CC17', '3DDB86', '1A9334', '00D4BB', + '2C99A8', '00C2FF', '344593', '6473FF', '0018EC', '8438FF', '520085', 'CB38FF', 'FF95C8', 'FF37C7') + palette = [] + for iter in hex: + h = '#' + iter + palette.append(tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4))) + num = len(palette) + color = palette[int(i) % num] + return (color[2], color[1], color[0]) if bgr else color diff --git a/yolov6/data/data_augment.py b/yolov6/data/data_augment.py new file mode 100644 index 0000000000000000000000000000000000000000..e4acea61a71f5cb42efc667fbcfd0d66a525672e --- /dev/null +++ b/yolov6/data/data_augment.py @@ -0,0 +1,193 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +# This code is based on +# https://github.com/ultralytics/yolov5/blob/master/utils/dataloaders.py + +import math +import random + +import cv2 +import numpy as np + +def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5): + # HSV color-space augmentation + if hgain or sgain or vgain: + r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1 # random gains + hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV)) + dtype = im.dtype # uint8 + + x = np.arange(0, 256, dtype=r.dtype) + lut_hue = ((x * r[0]) % 180).astype(dtype) + lut_sat = np.clip(x * r[1], 0, 255).astype(dtype) + lut_val = np.clip(x * r[2], 0, 255).astype(dtype) + + im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val))) + cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im) # no return needed + + + +def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleup=True, stride=32): + # Resize and pad image while meeting stride-multiple constraints + shape = im.shape[:2] # current shape [height, width] + if isinstance(new_shape, int): + new_shape = (new_shape, new_shape) + + # Scale ratio (new / old) + r = min(new_shape[0] / shape[0], new_shape[1] / shape[1]) + if not scaleup: # only scale down, do not scale up (for better val mAP) + r = min(r, 1.0) + + # Compute padding + new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) + dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh padding + + if auto: # minimum rectangle + dw, dh = np.mod(dw, stride), np.mod(dh, stride) # wh padding + + dw /= 2 # divide padding into 2 sides + dh /= 2 + + if shape[::-1] != new_unpad: # resize + im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR) + top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1)) + left, right = int(round(dw - 0.1)), int(round(dw + 0.1)) + im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add border + return im, r, (dw, dh) + + +def mixup(im, labels, im2, labels2): + # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf + r = np.random.beta(32.0, 32.0) # mixup ratio, alpha=beta=32.0 + im = (im * r + im2 * (1 - r)).astype(np.uint8) + labels = np.concatenate((labels, labels2), 0) + return im, labels + + +def box_candidates(box1, box2, wh_thr=2, ar_thr=20, area_thr=0.1, eps=1e-16): # box1(4,n), box2(4,n) + # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio + w1, h1 = box1[2] - box1[0], box1[3] - box1[1] + w2, h2 = box2[2] - box2[0], box2[3] - box2[1] + ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps)) # aspect ratio + return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr) # candidates + + +def random_affine(img, labels=(), degrees=10, translate=.1, scale=.1, shear=10, + new_shape=(640,640)): + + n = len(labels) + height,width = new_shape + + M,s = get_transform_matrix(img.shape[:2],(height,width),degrees,scale,shear,translate) + if (M != np.eye(3)).any(): # image changed + img = cv2.warpAffine(img, M[:2], dsize=(width, height), borderValue=(114, 114, 114)) + + # Transform label coordinates + if n: + new = np.zeros((n, 4)) + + xy = np.ones((n * 4, 3)) + xy[:, :2] = labels[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2) # x1y1, x2y2, x1y2, x2y1 + xy = xy @ M.T # transform + xy = xy[:, :2].reshape(n, 8) # perspective rescale or affine + + # create new boxes + x = xy[:, [0, 2, 4, 6]] + y = xy[:, [1, 3, 5, 7]] + new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T + + # clip + new[:, [0, 2]] = new[:, [0, 2]].clip(0, width) + new[:, [1, 3]] = new[:, [1, 3]].clip(0, height) + + # filter candidates + i = box_candidates(box1=labels[:, 1:5].T * s, box2=new.T, area_thr=0.1) + labels = labels[i] + labels[:, 1:5] = new[i] + + return img, labels + + +def get_transform_matrix(img_shape,new_shape,degrees,scale,shear,translate): + new_height,new_width = new_shape + # Center + C = np.eye(3) + C[0, 2] = -img_shape[1] / 2 # x translation (pixels) + C[1, 2] = -img_shape[0] / 2 # y translation (pixels) + + # Rotation and Scale + R = np.eye(3) + a = random.uniform(-degrees, degrees) + # a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations + s = random.uniform(1 - scale, 1 + scale) + # s = 2 ** random.uniform(-scale, scale) + R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s) + + # Shear + S = np.eye(3) + S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg) + S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg) + + # Translation + T = np.eye(3) + T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * new_width # x translation (pixels) + T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * new_height # y transla ion (pixels) + + # Combined rotation matrix + M = T @ S @ R @ C # order of operations (right to left) is IMPORTANT + return M,s + + +def mosaic_augmentation(img_size, imgs, hs, ws, labels, hyp): + + assert len(imgs)==4, "Mosaic augmentaion of current version only supports 4 images." + + labels4 = [] + s = img_size + yc, xc = (int(random.uniform(s//2, 3*s//2)) for _ in range(2)) # mosaic center x, y + for i in range(len(imgs)): + # Load image + img, h, w = imgs[i],hs[i],ws[i] + # place img in img4 + if i == 0: # top left + img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles + x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image) + x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image) + elif i == 1: # top right + x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc + x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h + elif i == 2: # bottom left + x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h) + x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h) + elif i == 3: # bottom right + x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h) + x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h) + + img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax] + padw = x1a - x1b + padh = y1a - y1b + + # Labels + labels_per_img= labels[i].copy() + if labels_per_img.size: + boxes = np.copy(labels_per_img[:,1:]) + boxes[:, 0] = w * (labels_per_img[:, 1] - labels_per_img[:, 3] / 2) + padw # top left x + boxes[:, 1] = h * (labels_per_img[:, 2] - labels_per_img[:, 4] / 2) + padh # top left y + boxes[:, 2] = w * (labels_per_img[:, 1] + labels_per_img[:, 3] / 2) + padw # bottom right x + boxes[:, 3] = h * (labels_per_img[:, 2] + labels_per_img[:, 4] / 2) + padh # bottom right y + labels_per_img[:,1:] = boxes + + labels4.append(labels_per_img) + + # Concat/clip labels + labels4 = np.concatenate(labels4, 0) + for x in (labels4[:, 1:]): + np.clip(x, 0, 2 * s, out=x) + + # Augment + img4, labels4 = random_affine(img4, labels4, + degrees=hyp['degrees'], + translate=hyp['translate'], + scale=hyp['scale'], + shear=hyp['shear']) + + return img4, labels4 diff --git a/yolov6/data/data_load.py b/yolov6/data/data_load.py new file mode 100644 index 0000000000000000000000000000000000000000..e8e273caf3d4ce08c1de9cfdf9377a5c2abebbd1 --- /dev/null +++ b/yolov6/data/data_load.py @@ -0,0 +1,77 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +# This code is based on +# https://github.com/ultralytics/yolov5/blob/master/utils/dataloaders.py + +import os +from torch.utils.data import dataloader, distributed + +from .datasets import TrainValDataset +from yolov6.utils.events import LOGGER +from yolov6.utils.torch_utils import torch_distributed_zero_first + + +def create_dataloader(path, img_size, batch_size, stride, hyp=None, augment=False, check_images=False, check_labels=False, pad=0.0, rect=False, rank=-1, workers=8, shuffle=False,class_names=None, task='Train'): + '''Create general dataloader. + + Returns dataloader and dataset + ''' + if rect and shuffle: + LOGGER.warning('WARNING: --rect is incompatible with DataLoader shuffle, setting shuffle=False') + shuffle = False + with torch_distributed_zero_first(rank): + dataset = TrainValDataset(path, img_size, batch_size, + augment=augment, + hyp=hyp, + rect=rect, + check_images=check_images, + stride=int(stride), + pad=pad, + rank=rank, + class_names=class_names, + task=task) + + batch_size = min(batch_size, len(dataset)) + workers = min([os.cpu_count() // int(os.getenv('WORLD_SIZE', 1)), batch_size if batch_size > 1 else 0, workers]) # number of workers + sampler = None if rank == -1 else distributed.DistributedSampler(dataset, shuffle=shuffle) + return TrainValDataLoader(dataset, + batch_size=batch_size, + shuffle=shuffle and sampler is None, + num_workers=workers, + sampler=sampler, + pin_memory=True, + collate_fn=TrainValDataset.collate_fn), dataset + + +class TrainValDataLoader(dataloader.DataLoader): + """ Dataloader that reuses workers + + Uses same syntax as vanilla DataLoader + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + object.__setattr__(self, 'batch_sampler', _RepeatSampler(self.batch_sampler)) + self.iterator = super().__iter__() + + def __len__(self): + return len(self.batch_sampler.sampler) + + def __iter__(self): + for i in range(len(self)): + yield next(self.iterator) + + +class _RepeatSampler: + """ Sampler that repeats forever + + Args: + sampler (Sampler) + """ + + def __init__(self, sampler): + self.sampler = sampler + + def __iter__(self): + while True: + yield from iter(self.sampler) diff --git a/yolov6/data/datasets.py b/yolov6/data/datasets.py new file mode 100644 index 0000000000000000000000000000000000000000..7627d608245ec5765b910dfb251419502ebdd32f --- /dev/null +++ b/yolov6/data/datasets.py @@ -0,0 +1,533 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- + +import glob +import os +import os.path as osp +import random +import json +import time + +from multiprocessing.pool import Pool + +import cv2 +import numpy as np +import torch +from PIL import ExifTags, Image, ImageOps +from torch.utils.data import Dataset +from tqdm import tqdm +from pathlib import Path + +from .data_augment import ( + augment_hsv, + letterbox, + mixup, + random_affine, + mosaic_augmentation, +) +from yolov6.utils.events import LOGGER + +# Parameters +IMG_FORMATS = ["bmp", "jpg", "jpeg", "png", "tif", "tiff", "dng", "webp", "mpo"] +# Get orientation exif tag +for k, v in ExifTags.TAGS.items(): + if v == "Orientation": + ORIENTATION = k + break + + +class TrainValDataset(Dataset): + # YOLOv6 train_loader/val_loader, loads images and labels for training and validation + def __init__( + self, + img_dir, + img_size=640, + batch_size=16, + augment=False, + hyp=None, + rect=False, + check_images=False, + check_labels=False, + stride=32, + pad=0.0, + rank=-1, + class_names=None, + task="train", + ): + assert task.lower() in ("train", "val", "speed"), f"Not supported task: {task}" + t1 = time.time() + self.__dict__.update(locals()) + self.main_process = self.rank in (-1, 0) + self.task = self.task.capitalize() + self.img_paths, self.labels = self.get_imgs_labels(self.img_dir) + if self.rect: + shapes = [self.img_info[p]["shape"] for p in self.img_paths] + self.shapes = np.array(shapes, dtype=np.float64) + self.batch_indices = np.floor( + np.arange(len(shapes)) / self.batch_size + ).astype( + np.int + ) # batch indices of each image + self.sort_files_shapes() + t2 = time.time() + if self.main_process: + LOGGER.info(f"%.1fs for dataset initialization." % (t2 - t1)) + + def __len__(self): + """Get the length of dataset""" + return len(self.img_paths) + + def __getitem__(self, index): + """Fetching a data sample for a given key. + This function applies mosaic and mixup augments during training. + During validation, letterbox augment is applied. + """ + # Mosaic Augmentation + if self.augment and random.random() < self.hyp["mosaic"]: + img, labels = self.get_mosaic(index) + shapes = None + + # MixUp augmentation + if random.random() < self.hyp["mixup"]: + img_other, labels_other = self.get_mosaic( + random.randint(0, len(self.img_paths) - 1) + ) + img, labels = mixup(img, labels, img_other, labels_other) + + else: + # Load image + img, (h0, w0), (h, w) = self.load_image(index) + + # Letterbox + shape = ( + self.batch_shapes[self.batch_indices[index]] + if self.rect + else self.img_size + ) # final letterboxed shape + img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment) + shapes = (h0, w0), ((h / h0, w / w0), pad) # for COCO mAP rescaling + + labels = self.labels[index].copy() + if labels.size: + w *= ratio + h *= ratio + # new boxes + boxes = np.copy(labels[:, 1:]) + boxes[:, 0] = ( + w * (labels[:, 1] - labels[:, 3] / 2) + pad[0] + ) # top left x + boxes[:, 1] = ( + h * (labels[:, 2] - labels[:, 4] / 2) + pad[1] + ) # top left y + boxes[:, 2] = ( + w * (labels[:, 1] + labels[:, 3] / 2) + pad[0] + ) # bottom right x + boxes[:, 3] = ( + h * (labels[:, 2] + labels[:, 4] / 2) + pad[1] + ) # bottom right y + labels[:, 1:] = boxes + + if self.augment: + img, labels = random_affine( + img, + labels, + degrees=self.hyp["degrees"], + translate=self.hyp["translate"], + scale=self.hyp["scale"], + shear=self.hyp["shear"], + new_shape=(self.img_size, self.img_size), + ) + + if len(labels): + h, w = img.shape[:2] + + labels[:, [1, 3]] = labels[:, [1, 3]].clip(0, w - 1e-3) # x1, x2 + labels[:, [2, 4]] = labels[:, [2, 4]].clip(0, h - 1e-3) # y1, y2 + + boxes = np.copy(labels[:, 1:]) + boxes[:, 0] = ((labels[:, 1] + labels[:, 3]) / 2) / w # x center + boxes[:, 1] = ((labels[:, 2] + labels[:, 4]) / 2) / h # y center + boxes[:, 2] = (labels[:, 3] - labels[:, 1]) / w # width + boxes[:, 3] = (labels[:, 4] - labels[:, 2]) / h # height + labels[:, 1:] = boxes + + if self.augment: + img, labels = self.general_augment(img, labels) + + labels_out = torch.zeros((len(labels), 6)) + if len(labels): + labels_out[:, 1:] = torch.from_numpy(labels) + + # Convert + img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB + img = np.ascontiguousarray(img) + + return torch.from_numpy(img), labels_out, self.img_paths[index], shapes + + def load_image(self, index): + """Load image. + This function loads image by cv2, resize original image to target shape(img_size) with keeping ratio. + + Returns: + Image, original shape of image, resized image shape + """ + path = self.img_paths[index] + im = cv2.imread(path) + assert im is not None, f"Image Not Found {path}, workdir: {os.getcwd()}" + + h0, w0 = im.shape[:2] # origin shape + r = self.img_size / max(h0, w0) + if r != 1: + im = cv2.resize( + im, + (int(w0 * r), int(h0 * r)), + interpolation=cv2.INTER_AREA + if r < 1 and not self.augment + else cv2.INTER_LINEAR, + ) + return im, (h0, w0), im.shape[:2] + + @staticmethod + def collate_fn(batch): + """Merges a list of samples to form a mini-batch of Tensor(s)""" + img, label, path, shapes = zip(*batch) + for i, l in enumerate(label): + l[:, 0] = i # add target image index for build_targets() + return torch.stack(img, 0), torch.cat(label, 0), path, shapes + + def get_imgs_labels(self, img_dir): + + assert osp.exists(img_dir), f"{img_dir} is an invalid directory path!" + valid_img_record = osp.join( + osp.dirname(img_dir), "." + osp.basename(img_dir) + ".json" + ) + img_info = {} + NUM_THREADS = min(8, os.cpu_count()) + # check images + if ( + self.check_images or not osp.exists(valid_img_record) + ) and self.main_process: + img_paths = glob.glob(osp.join(img_dir, "*"), recursive=True) + img_paths = sorted( + p for p in img_paths if p.split(".")[-1].lower() in IMG_FORMATS + ) + assert img_paths, f"No images found in {img_dir}." + + nc, msgs = 0, [] # number corrupt, messages + LOGGER.info( + f"{self.task}: Checking formats of images with {NUM_THREADS} process(es): " + ) + with Pool(NUM_THREADS) as pool: + pbar = tqdm( + pool.imap(TrainValDataset.check_image, img_paths), + total=len(img_paths), + ) + for img_path, shape_per_img, nc_per_img, msg in pbar: + if nc_per_img == 0: # not corrupted + img_info[img_path] = {"shape": shape_per_img} + nc += nc_per_img + if msg: + msgs.append(msg) + pbar.desc = f"{nc} image(s) corrupted" + pbar.close() + if msgs: + LOGGER.info("\n".join(msgs)) + + # save valid image paths. + with open(valid_img_record, "w") as f: + json.dump(img_info, f) + + # check and load anns + label_dir = osp.join( + osp.dirname(osp.dirname(img_dir)), "labels", osp.basename(img_dir) + ) + assert osp.exists(label_dir), f"{label_dir} is an invalid directory path!" + if not img_info: + with open(valid_img_record, "r") as f: + img_info = json.load(f) + assert ( + img_info + ), "No information in record files, please add option --check_images." + img_paths = list(img_info.keys()) + label_paths = [ + osp.join(label_dir, osp.basename(p).split(".")[0] + ".txt") + for p in img_paths + ] + if ( + self.check_labels or "labels" not in img_info[img_paths[0]] + ): # key 'labels' not saved in img_info + nm, nf, ne, nc, msgs = 0, 0, 0, 0, [] # number corrupt, messages + LOGGER.info( + f"{self.task}: Checking formats of labels with {NUM_THREADS} process(es): " + ) + with Pool(NUM_THREADS) as pool: + pbar = pool.imap( + TrainValDataset.check_label_files, zip(img_paths, label_paths) + ) + pbar = tqdm(pbar, total=len(label_paths)) if self.main_process else pbar + for ( + img_path, + labels_per_file, + nc_per_file, + nm_per_file, + nf_per_file, + ne_per_file, + msg, + ) in pbar: + if img_path: + img_info[img_path]["labels"] = labels_per_file + else: + img_info.pop(img_path) + nc += nc_per_file + nm += nm_per_file + nf += nf_per_file + ne += ne_per_file + if msg: + msgs.append(msg) + if self.main_process: + pbar.desc = f"{nf} label(s) found, {nm} label(s) missing, {ne} label(s) empty, {nc} invalid label files" + if self.main_process: + pbar.close() + with open(valid_img_record, "w") as f: + json.dump(img_info, f) + if msgs: + LOGGER.info("\n".join(msgs)) + if nf == 0: + LOGGER.warning( + f"WARNING: No labels found in {osp.dirname(self.img_paths[0])}. " + ) + else: + with open(valid_img_record) as f: + img_info = json.load(f) + if self.task.lower() == "val": + assert ( + self.class_names + ), "Class names is required when converting labels to coco format for evaluating." + save_dir = osp.join(osp.dirname(osp.dirname(img_dir)), "annotations") + if not osp.exists(save_dir): + os.mkdir(save_dir) + save_path = osp.join( + save_dir, "instances_" + osp.basename(img_dir) + ".json" + ) + if not osp.exists(save_path): + TrainValDataset.generate_coco_format_labels( + img_info, self.class_names, save_path + ) + + img_paths, labels = list( + zip( + *[ + ( + img_path, + np.array(info["labels"], dtype=np.float32) + if info["labels"] + else np.zeros((0, 5), dtype=np.float32), + ) + for img_path, info in img_info.items() + ] + ) + ) + self.img_info = img_info + LOGGER.info( + f"{self.task}: Final numbers of valid images: {len(img_paths)}/ labels: {len(labels)}. " + ) + return img_paths, labels + + def get_mosaic(self, index): + """Gets images and labels after mosaic augments""" + indices = [index] + random.choices( + range(0, len(self.img_paths)), k=3 + ) # 3 additional image indices + random.shuffle(indices) + imgs, hs, ws, labels = [], [], [], [] + for index in indices: + img, _, (h, w) = self.load_image(index) + labels_per_img = self.labels[index] + imgs.append(img) + hs.append(h) + ws.append(w) + labels.append(labels_per_img) + img, labels = mosaic_augmentation(self.img_size, imgs, hs, ws, labels, self.hyp) + return img, labels + + def general_augment(self, img, labels): + """Gets images and labels after general augment + This function applies hsv, random ud-flip and random lr-flips augments. + """ + nl = len(labels) + + # HSV color-space + augment_hsv( + img, + hgain=self.hyp["hsv_h"], + sgain=self.hyp["hsv_s"], + vgain=self.hyp["hsv_v"], + ) + + # Flip up-down + if random.random() < self.hyp["flipud"]: + img = np.flipud(img) + if nl: + labels[:, 2] = 1 - labels[:, 2] + + # Flip left-right + if random.random() < self.hyp["fliplr"]: + img = np.fliplr(img) + if nl: + labels[:, 1] = 1 - labels[:, 1] + + return img, labels + + def sort_files_shapes(self): + # Sort by aspect ratio + batch_num = self.batch_indices[-1] + 1 + s = self.shapes # wh + ar = s[:, 1] / s[:, 0] # aspect ratio + irect = ar.argsort() + self.img_paths = [self.img_paths[i] for i in irect] + self.labels = [self.labels[i] for i in irect] + self.shapes = s[irect] # wh + ar = ar[irect] + + # Set training image shapes + shapes = [[1, 1]] * batch_num + for i in range(batch_num): + ari = ar[self.batch_indices == i] + mini, maxi = ari.min(), ari.max() + if maxi < 1: + shapes[i] = [maxi, 1] + elif mini > 1: + shapes[i] = [1, 1 / mini] + self.batch_shapes = ( + np.ceil(np.array(shapes) * self.img_size / self.stride + self.pad).astype( + np.int + ) + * self.stride + ) + + @staticmethod + def check_image(im_file): + # verify an image. + nc, msg = 0, "" + try: + im = Image.open(im_file) + im.verify() # PIL verify + shape = im.size # (width, height) + im_exif = im._getexif() + if im_exif and ORIENTATION in im_exif: + rotation = im_exif[ORIENTATION] + if rotation in (6, 8): + shape = (shape[1], shape[0]) + + assert (shape[0] > 9) & (shape[1] > 9), f"image size {shape} <10 pixels" + assert im.format.lower() in IMG_FORMATS, f"invalid image format {im.format}" + if im.format.lower() in ("jpg", "jpeg"): + with open(im_file, "rb") as f: + f.seek(-2, 2) + if f.read() != b"\xff\xd9": # corrupt JPEG + ImageOps.exif_transpose(Image.open(im_file)).save( + im_file, "JPEG", subsampling=0, quality=100 + ) + msg += f"WARNING: {im_file}: corrupt JPEG restored and saved" + return im_file, shape, nc, msg + except Exception as e: + nc = 1 + msg = f"WARNING: {im_file}: ignoring corrupt image: {e}" + return im_file, None, nc, msg + + @staticmethod + def check_label_files(args): + img_path, lb_path = args + nm, nf, ne, nc, msg = 0, 0, 0, 0, "" # number (missing, found, empty, message + try: + if osp.exists(lb_path): + nf = 1 # label found + with open(lb_path, "r") as f: + labels = [ + x.split() for x in f.read().strip().splitlines() if len(x) + ] + labels = np.array(labels, dtype=np.float32) + if len(labels): + assert all( + len(l) == 5 for l in labels + ), f"{lb_path}: wrong label format." + assert ( + labels >= 0 + ).all(), f"{lb_path}: Label values error: all values in label file must > 0" + assert ( + labels[:, 1:] <= 1 + ).all(), f"{lb_path}: Label values error: all coordinates must be normalized" + + _, indices = np.unique(labels, axis=0, return_index=True) + if len(indices) < len(labels): # duplicate row check + labels = labels[indices] # remove duplicates + msg += f"WARNING: {lb_path}: {len(labels) - len(indices)} duplicate labels removed" + labels = labels.tolist() + else: + ne = 1 # label empty + labels = [] + else: + nm = 1 # label missing + labels = [] + + return img_path, labels, nc, nm, nf, ne, msg + except Exception as e: + nc = 1 + msg = f"WARNING: {lb_path}: ignoring invalid labels: {e}" + return None, None, nc, nm, nf, ne, msg + + @staticmethod + def generate_coco_format_labels(img_info, class_names, save_path): + # for evaluation with pycocotools + dataset = {"categories": [], "annotations": [], "images": []} + for i, class_name in enumerate(class_names): + dataset["categories"].append( + {"id": i, "name": class_name, "supercategory": ""} + ) + + ann_id = 0 + LOGGER.info(f"Convert to COCO format") + for i, (img_path, info) in enumerate(tqdm(img_info.items())): + labels = info["labels"] if info["labels"] else [] + path = Path(img_path) + img_id = int(path.stem) if path.stem.isnumeric() else path.stem + img_w, img_h = info["shape"] + dataset["images"].append( + { + "file_name": os.path.basename(img_path), + "id": img_id, + "width": img_w, + "height": img_h, + } + ) + if labels: + for label in labels: + c, x, y, w, h = label[:5] + # convert x,y,w,h to x1,y1,x2,y2 + x1 = (x - w / 2) * img_w + y1 = (y - h / 2) * img_h + x2 = (x + w / 2) * img_w + y2 = (y + h / 2) * img_h + # cls_id starts from 0 + cls_id = int(c) + w = max(0, x2 - x1) + h = max(0, y2 - y1) + dataset["annotations"].append( + { + "area": h * w, + "bbox": [x1, y1, w, h], + "category_id": cls_id, + "id": ann_id, + "image_id": img_id, + "iscrowd": 0, + # mask + "segmentation": [], + } + ) + ann_id += 1 + + with open(save_path, "w") as f: + json.dump(dataset, f) + LOGGER.info( + f"Convert to COCO format finished. Resutls saved in {save_path}" + ) diff --git a/yolov6/layers/common.py b/yolov6/layers/common.py new file mode 100644 index 0000000000000000000000000000000000000000..800659e8d9f709f188731038bed457d7b50ec8cc --- /dev/null +++ b/yolov6/layers/common.py @@ -0,0 +1,269 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- + +import warnings +from pathlib import Path + +import numpy as np +import torch +import torch.nn as nn + + +class SiLU(nn.Module): + '''Activation of SiLU''' + @staticmethod + def forward(x): + return x * torch.sigmoid(x) + + +class Conv(nn.Module): + '''Normal Conv with SiLU activation''' + def __init__(self, in_channels, out_channels, kernel_size, stride, groups=1, bias=False): + super().__init__() + padding = kernel_size // 2 + self.conv = nn.Conv2d( + in_channels, + out_channels, + kernel_size=kernel_size, + stride=stride, + padding=padding, + groups=groups, + bias=bias, + ) + self.bn = nn.BatchNorm2d(out_channels) + self.act = nn.SiLU() + + def forward(self, x): + return self.act(self.bn(self.conv(x))) + + def forward_fuse(self, x): + return self.act(self.conv(x)) + + +class SimConv(nn.Module): + '''Normal Conv with ReLU activation''' + def __init__(self, in_channels, out_channels, kernel_size, stride, groups=1, bias=False): + super().__init__() + padding = kernel_size // 2 + self.conv = nn.Conv2d( + in_channels, + out_channels, + kernel_size=kernel_size, + stride=stride, + padding=padding, + groups=groups, + bias=bias, + ) + self.bn = nn.BatchNorm2d(out_channels) + self.act = nn.ReLU() + + def forward(self, x): + return self.act(self.bn(self.conv(x))) + + def forward_fuse(self, x): + return self.act(self.conv(x)) + + +class SimSPPF(nn.Module): + '''Simplified SPPF with ReLU activation''' + def __init__(self, in_channels, out_channels, kernel_size=5): + super().__init__() + c_ = in_channels // 2 # hidden channels + self.cv1 = SimConv(in_channels, c_, 1, 1) + self.cv2 = SimConv(c_ * 4, out_channels, 1, 1) + self.m = nn.MaxPool2d(kernel_size=kernel_size, stride=1, padding=kernel_size // 2) + + def forward(self, x): + x = self.cv1(x) + with warnings.catch_warnings(): + warnings.simplefilter('ignore') + y1 = self.m(x) + y2 = self.m(y1) + return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1)) + + +class Transpose(nn.Module): + '''Normal Transpose, default for upsampling''' + def __init__(self, in_channels, out_channels, kernel_size=2, stride=2): + super().__init__() + self.upsample_transpose = torch.nn.ConvTranspose2d( + in_channels=in_channels, + out_channels=out_channels, + kernel_size=kernel_size, + stride=stride, + bias=True + ) + + def forward(self, x): + return self.upsample_transpose(x) + + +class Concat(nn.Module): + def __init__(self, dimension=1): + super().__init__() + self.d = dimension + + def forward(self, x): + return torch.cat(x, self.d) + + +def conv_bn(in_channels, out_channels, kernel_size, stride, padding, groups=1): + '''Basic cell for rep-style block, including conv and bn''' + result = nn.Sequential() + result.add_module('conv', nn.Conv2d(in_channels=in_channels, out_channels=out_channels, + kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, bias=False)) + result.add_module('bn', nn.BatchNorm2d(num_features=out_channels)) + return result + + +class RepBlock(nn.Module): + ''' + RepBlock is a stage block with rep-style basic block + ''' + def __init__(self, in_channels, out_channels, n=1): + super().__init__() + self.conv1 = RepVGGBlock(in_channels, out_channels) + self.block = nn.Sequential(*(RepVGGBlock(out_channels, out_channels) for _ in range(n - 1))) if n > 1 else None + + def forward(self, x): + x = self.conv1(x) + if self.block is not None: + x = self.block(x) + return x + + +class RepVGGBlock(nn.Module): + '''RepVGGBlock is a basic rep-style block, including training and deploy status + This code is based on https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py + ''' + def __init__(self, in_channels, out_channels, kernel_size=3, + stride=1, padding=1, dilation=1, groups=1, padding_mode='zeros', deploy=False, use_se=False): + super(RepVGGBlock, self).__init__() + """ Intialization of the class. + Args: + in_channels (int): Number of channels in the input image + out_channels (int): Number of channels produced by the convolution + kernel_size (int or tuple): Size of the convolving kernel + stride (int or tuple, optional): Stride of the convolution. Default: 1 + padding (int or tuple, optional): Zero-padding added to both sides of + the input. Default: 1 + dilation (int or tuple, optional): Spacing between kernel elements. Default: 1 + groups (int, optional): Number of blocked connections from input + channels to output channels. Default: 1 + padding_mode (string, optional): Default: 'zeros' + deploy: Whether to be deploy status or training status. Default: False + use_se: Whether to use se. Default: False + """ + self.deploy = deploy + self.groups = groups + self.in_channels = in_channels + self.out_channels = out_channels + + assert kernel_size == 3 + assert padding == 1 + + padding_11 = padding - kernel_size // 2 + + self.nonlinearity = nn.ReLU() + + if use_se: + raise NotImplementedError("se block not supported yet") + else: + self.se = nn.Identity() + + if deploy: + self.rbr_reparam = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, + padding=padding, dilation=dilation, groups=groups, bias=True, padding_mode=padding_mode) + + else: + self.rbr_identity = nn.BatchNorm2d(num_features=in_channels) if out_channels == in_channels and stride == 1 else None + self.rbr_dense = conv_bn(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups) + self.rbr_1x1 = conv_bn(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=padding_11, groups=groups) + + def forward(self, inputs): + '''Forward process''' + if hasattr(self, 'rbr_reparam'): + return self.nonlinearity(self.se(self.rbr_reparam(inputs))) + + if self.rbr_identity is None: + id_out = 0 + else: + id_out = self.rbr_identity(inputs) + + return self.nonlinearity(self.se(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out)) + + def get_equivalent_kernel_bias(self): + kernel3x3, bias3x3 = self._fuse_bn_tensor(self.rbr_dense) + kernel1x1, bias1x1 = self._fuse_bn_tensor(self.rbr_1x1) + kernelid, biasid = self._fuse_bn_tensor(self.rbr_identity) + return kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid, bias3x3 + bias1x1 + biasid + + def _pad_1x1_to_3x3_tensor(self, kernel1x1): + if kernel1x1 is None: + return 0 + else: + return torch.nn.functional.pad(kernel1x1, [1, 1, 1, 1]) + + def _fuse_bn_tensor(self, branch): + if branch is None: + return 0, 0 + if isinstance(branch, nn.Sequential): + kernel = branch.conv.weight + running_mean = branch.bn.running_mean + running_var = branch.bn.running_var + gamma = branch.bn.weight + beta = branch.bn.bias + eps = branch.bn.eps + else: + assert isinstance(branch, nn.BatchNorm2d) + if not hasattr(self, 'id_tensor'): + input_dim = self.in_channels // self.groups + kernel_value = np.zeros((self.in_channels, input_dim, 3, 3), dtype=np.float32) + for i in range(self.in_channels): + kernel_value[i, i % input_dim, 1, 1] = 1 + self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device) + kernel = self.id_tensor + running_mean = branch.running_mean + running_var = branch.running_var + gamma = branch.weight + beta = branch.bias + eps = branch.eps + std = (running_var + eps).sqrt() + t = (gamma / std).reshape(-1, 1, 1, 1) + return kernel * t, beta - running_mean * gamma / std + + def switch_to_deploy(self): + if hasattr(self, 'rbr_reparam'): + return + kernel, bias = self.get_equivalent_kernel_bias() + self.rbr_reparam = nn.Conv2d(in_channels=self.rbr_dense.conv.in_channels, out_channels=self.rbr_dense.conv.out_channels, + kernel_size=self.rbr_dense.conv.kernel_size, stride=self.rbr_dense.conv.stride, + padding=self.rbr_dense.conv.padding, dilation=self.rbr_dense.conv.dilation, groups=self.rbr_dense.conv.groups, bias=True) + self.rbr_reparam.weight.data = kernel + self.rbr_reparam.bias.data = bias + for para in self.parameters(): + para.detach_() + self.__delattr__('rbr_dense') + self.__delattr__('rbr_1x1') + if hasattr(self, 'rbr_identity'): + self.__delattr__('rbr_identity') + if hasattr(self, 'id_tensor'): + self.__delattr__('id_tensor') + self.deploy = True + + +class DetectBackend(nn.Module): + def __init__(self, weights='yolov6s.pt', device=None, dnn=True): + + super().__init__() + assert isinstance(weights, str) and Path(weights).suffix == '.pt', f'{Path(weights).suffix} format is not supported.' + from yolov6.utils.checkpoint import load_checkpoint + model = load_checkpoint(weights, map_location=device) + stride = int(model.stride.max()) + self.__dict__.update(locals()) # assign all variables to self + + def forward(self, im, val=False): + y = self.model(im) + if isinstance(y, np.ndarray): + y = torch.tensor(y, device=self.device) + return y diff --git a/yolov6/models/efficientrep.py b/yolov6/models/efficientrep.py new file mode 100644 index 0000000000000000000000000000000000000000..4cb9f1ca431b466af460212becec3b0e7ea15d5c --- /dev/null +++ b/yolov6/models/efficientrep.py @@ -0,0 +1,102 @@ +from torch import nn +from yolov6.layers.common import RepVGGBlock, RepBlock, SimSPPF + + +class EfficientRep(nn.Module): + '''EfficientRep Backbone + EfficientRep is handcrafted by hardware-aware neural network design. + With rep-style struct, EfficientRep is friendly to high-computation hardware(e.g. GPU). + ''' + + def __init__( + self, + in_channels=3, + channels_list=None, + num_repeats=None, + ): + super().__init__() + + assert channels_list is not None + assert num_repeats is not None + + self.stem = RepVGGBlock( + in_channels=in_channels, + out_channels=channels_list[0], + kernel_size=3, + stride=2 + ) + + self.ERBlock_2 = nn.Sequential( + RepVGGBlock( + in_channels=channels_list[0], + out_channels=channels_list[1], + kernel_size=3, + stride=2 + ), + RepBlock( + in_channels=channels_list[1], + out_channels=channels_list[1], + n=num_repeats[1] + ) + ) + + self.ERBlock_3 = nn.Sequential( + RepVGGBlock( + in_channels=channels_list[1], + out_channels=channels_list[2], + kernel_size=3, + stride=2 + ), + RepBlock( + in_channels=channels_list[2], + out_channels=channels_list[2], + n=num_repeats[2] + ) + ) + + self.ERBlock_4 = nn.Sequential( + RepVGGBlock( + in_channels=channels_list[2], + out_channels=channels_list[3], + kernel_size=3, + stride=2 + ), + RepBlock( + in_channels=channels_list[3], + out_channels=channels_list[3], + n=num_repeats[3] + ) + ) + + self.ERBlock_5 = nn.Sequential( + RepVGGBlock( + in_channels=channels_list[3], + out_channels=channels_list[4], + kernel_size=3, + stride=2, + ), + RepBlock( + in_channels=channels_list[4], + out_channels=channels_list[4], + n=num_repeats[4] + ), + SimSPPF( + in_channels=channels_list[4], + out_channels=channels_list[4], + kernel_size=5 + ) + ) + + def forward(self, x): + + outputs = [] + x = self.stem(x) + x = self.ERBlock_2(x) + x = self.ERBlock_3(x) + outputs.append(x) + x = self.ERBlock_4(x) + outputs.append(x) + x = self.ERBlock_5(x) + outputs.append(x) + + return tuple(outputs) diff --git a/yolov6/models/effidehead.py b/yolov6/models/effidehead.py new file mode 100644 index 0000000000000000000000000000000000000000..4664f9373c745bf4324f3531e4afb282cc7393af --- /dev/null +++ b/yolov6/models/effidehead.py @@ -0,0 +1,211 @@ +import torch +import torch.nn as nn +import math +from yolov6.layers.common import * + + +class Detect(nn.Module): + '''Efficient Decoupled Head + With hardware-aware degisn, the decoupled head is optimized with + hybridchannels methods. + ''' + def __init__(self, num_classes=80, anchors=1, num_layers=3, inplace=True, head_layers=None): # detection layer + super().__init__() + assert head_layers is not None + self.nc = num_classes # number of classes + self.no = num_classes + 5 # number of outputs per anchor + self.nl = num_layers # number of detection layers + if isinstance(anchors, (list, tuple)): + self.na = len(anchors[0]) // 2 + else: + self.na = anchors + self.anchors = anchors + self.grid = [torch.zeros(1)] * num_layers + self.prior_prob = 1e-2 + self.inplace = inplace + stride = [8, 16, 32] # strides computed during build + self.stride = torch.tensor(stride) + + # Init decouple head + self.cls_convs = nn.ModuleList() + self.reg_convs = nn.ModuleList() + self.cls_preds = nn.ModuleList() + self.reg_preds = nn.ModuleList() + self.obj_preds = nn.ModuleList() + self.stems = nn.ModuleList() + + # Efficient decoupled head layers + for i in range(num_layers): + idx = i*6 + self.stems.append(head_layers[idx]) + self.cls_convs.append(head_layers[idx+1]) + self.reg_convs.append(head_layers[idx+2]) + self.cls_preds.append(head_layers[idx+3]) + self.reg_preds.append(head_layers[idx+4]) + self.obj_preds.append(head_layers[idx+5]) + + def initialize_biases(self): + for conv in self.cls_preds: + b = conv.bias.view(self.na, -1) + b.data.fill_(-math.log((1 - self.prior_prob) / self.prior_prob)) + conv.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) + for conv in self.obj_preds: + b = conv.bias.view(self.na, -1) + b.data.fill_(-math.log((1 - self.prior_prob) / self.prior_prob)) + conv.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) + + def forward(self, x): + z = [] + for i in range(self.nl): + x[i] = self.stems[i](x[i]) + cls_x = x[i] + reg_x = x[i] + cls_feat = self.cls_convs[i](cls_x) + cls_output = self.cls_preds[i](cls_feat) + reg_feat = self.reg_convs[i](reg_x) + reg_output = self.reg_preds[i](reg_feat) + obj_output = self.obj_preds[i](reg_feat) + if self.training: + x[i] = torch.cat([reg_output, obj_output, cls_output], 1) + bs, _, ny, nx = x[i].shape + x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() + else: + y = torch.cat([reg_output, obj_output.sigmoid(), cls_output.sigmoid()], 1) + bs, _, ny, nx = y.shape + y = y.view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() + if self.grid[i].shape[2:4] != y.shape[2:4]: + d = self.stride.device + yv, xv = torch.meshgrid([torch.arange(ny).to(d), torch.arange(nx).to(d)]) + self.grid[i] = torch.stack((xv, yv), 2).view(1, self.na, ny, nx, 2).float() + if self.inplace: + y[..., 0:2] = (y[..., 0:2] + self.grid[i]) * self.stride[i] # xy + y[..., 2:4] = torch.exp(y[..., 2:4]) * self.stride[i] # wh + else: + xy = (y[..., 0:2] + self.grid[i]) * self.stride[i] # xy + wh = torch.exp(y[..., 2:4]) * self.stride[i] # wh + y = torch.cat((xy, wh, y[..., 4:]), -1) + z.append(y.view(bs, -1, self.no)) + return x if self.training else torch.cat(z, 1) + + +def build_effidehead_layer(channels_list, num_anchors, num_classes): + head_layers = nn.Sequential( + # stem0 + Conv( + in_channels=channels_list[6], + out_channels=channels_list[6], + kernel_size=1, + stride=1 + ), + # cls_conv0 + Conv( + in_channels=channels_list[6], + out_channels=channels_list[6], + kernel_size=3, + stride=1 + ), + # reg_conv0 + Conv( + in_channels=channels_list[6], + out_channels=channels_list[6], + kernel_size=3, + stride=1 + ), + # cls_pred0 + nn.Conv2d( + in_channels=channels_list[6], + out_channels=num_classes * num_anchors, + kernel_size=1 + ), + # reg_pred0 + nn.Conv2d( + in_channels=channels_list[6], + out_channels=4 * num_anchors, + kernel_size=1 + ), + # obj_pred0 + nn.Conv2d( + in_channels=channels_list[6], + out_channels=1 * num_anchors, + kernel_size=1 + ), + # stem1 + Conv( + in_channels=channels_list[8], + out_channels=channels_list[8], + kernel_size=1, + stride=1 + ), + # cls_conv1 + Conv( + in_channels=channels_list[8], + out_channels=channels_list[8], + kernel_size=3, + stride=1 + ), + # reg_conv1 + Conv( + in_channels=channels_list[8], + out_channels=channels_list[8], + kernel_size=3, + stride=1 + ), + # cls_pred1 + nn.Conv2d( + in_channels=channels_list[8], + out_channels=num_classes * num_anchors, + kernel_size=1 + ), + # reg_pred1 + nn.Conv2d( + in_channels=channels_list[8], + out_channels=4 * num_anchors, + kernel_size=1 + ), + # obj_pred1 + nn.Conv2d( + in_channels=channels_list[8], + out_channels=1 * num_anchors, + kernel_size=1 + ), + # stem2 + Conv( + in_channels=channels_list[10], + out_channels=channels_list[10], + kernel_size=1, + stride=1 + ), + # cls_conv2 + Conv( + in_channels=channels_list[10], + out_channels=channels_list[10], + kernel_size=3, + stride=1 + ), + # reg_conv2 + Conv( + in_channels=channels_list[10], + out_channels=channels_list[10], + kernel_size=3, + stride=1 + ), + # cls_pred2 + nn.Conv2d( + in_channels=channels_list[10], + out_channels=num_classes * num_anchors, + kernel_size=1 + ), + # reg_pred2 + nn.Conv2d( + in_channels=channels_list[10], + out_channels=4 * num_anchors, + kernel_size=1 + ), + # obj_pred2 + nn.Conv2d( + in_channels=channels_list[10], + out_channels=1 * num_anchors, + kernel_size=1 + ) + ) + return head_layers diff --git a/yolov6/models/loss.py b/yolov6/models/loss.py new file mode 100644 index 0000000000000000000000000000000000000000..b86e4830826d94b7927c173e7805889d2dcb2217 --- /dev/null +++ b/yolov6/models/loss.py @@ -0,0 +1,411 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- + +# The code is based on +# https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/models/yolo_head.py +# Copyright (c) Megvii, Inc. and its affiliates. + +import torch +import torch.nn as nn +import numpy as np +import torch.nn.functional as F +from yolov6.utils.figure_iou import IOUloss, pairwise_bbox_iou + + +class ComputeLoss: + '''Loss computation func. + This func contains SimOTA and siou loss. + ''' + def __init__(self, + reg_weight=5.0, + iou_weight=3.0, + cls_weight=1.0, + center_radius=2.5, + eps=1e-7, + in_channels=[256, 512, 1024], + strides=[8, 16, 32], + n_anchors=1, + iou_type='ciou' + ): + + self.reg_weight = reg_weight + self.iou_weight = iou_weight + self.cls_weight = cls_weight + + self.center_radius = center_radius + self.eps = eps + self.n_anchors = n_anchors + self.strides = strides + self.grids = [torch.zeros(1)] * len(in_channels) + + # Define criteria + self.l1_loss = nn.L1Loss(reduction="none") + self.bcewithlog_loss = nn.BCEWithLogitsLoss(reduction="none") + self.iou_loss = IOUloss(iou_type=iou_type, reduction="none") + + def __call__( + self, + outputs, + targets + ): + dtype = outputs[0].type() + device = targets.device + loss_cls, loss_obj, loss_iou, loss_l1 = torch.zeros(1, device=device), torch.zeros(1, device=device), \ + torch.zeros(1, device=device), torch.zeros(1, device=device) + num_classes = outputs[0].shape[-1] - 5 + + outputs, outputs_origin, gt_bboxes_scale, xy_shifts, expanded_strides = self.get_outputs_and_grids( + outputs, self.strides, dtype, device) + + total_num_anchors = outputs.shape[1] + bbox_preds = outputs[:, :, :4] # [batch, n_anchors_all, 4] + bbox_preds_org = outputs_origin[:, :, :4] # [batch, n_anchors_all, 4] + obj_preds = outputs[:, :, 4].unsqueeze(-1) # [batch, n_anchors_all, 1] + cls_preds = outputs[:, :, 5:] # [batch, n_anchors_all, n_cls] + + # targets + batch_size = bbox_preds.shape[0] + targets_list = np.zeros((batch_size, 1, 5)).tolist() + for i, item in enumerate(targets.cpu().numpy().tolist()): + targets_list[int(item[0])].append(item[1:]) + max_len = max((len(l) for l in targets_list)) + + targets = torch.from_numpy(np.array(list(map(lambda l:l + [[-1,0,0,0,0]]*(max_len - len(l)), targets_list)))[:,1:,:]).to(targets.device) + num_targets_list = (targets.sum(dim=2) > 0).sum(dim=1) # number of objects + + num_fg, num_gts = 0, 0 + cls_targets, reg_targets, l1_targets, obj_targets, fg_masks = [], [], [], [], [] + + for batch_idx in range(batch_size): + num_gt = int(num_targets_list[batch_idx]) + num_gts += num_gt + if num_gt == 0: + cls_target = outputs.new_zeros((0, num_classes)) + reg_target = outputs.new_zeros((0, 4)) + l1_target = outputs.new_zeros((0, 4)) + obj_target = outputs.new_zeros((total_num_anchors, 1)) + fg_mask = outputs.new_zeros(total_num_anchors).bool() + else: + + gt_bboxes_per_image = targets[batch_idx, :num_gt, 1:5].mul_(gt_bboxes_scale) + gt_classes = targets[batch_idx, :num_gt, 0] + bboxes_preds_per_image = bbox_preds[batch_idx] + cls_preds_per_image = cls_preds[batch_idx] + obj_preds_per_image = obj_preds[batch_idx] + + try: + ( + gt_matched_classes, + fg_mask, + pred_ious_this_matching, + matched_gt_inds, + num_fg_img, + ) = self.get_assignments( + batch_idx, + num_gt, + total_num_anchors, + gt_bboxes_per_image, + gt_classes, + bboxes_preds_per_image, + cls_preds_per_image, + obj_preds_per_image, + expanded_strides, + xy_shifts, + num_classes + ) + + except RuntimeError: + print( + "OOM RuntimeError is raised due to the huge memory cost during label assignment. \ + CPU mode is applied in this batch. If you want to avoid this issue, \ + try to reduce the batch size or image size." + ) + torch.cuda.empty_cache() + print("------------CPU Mode for This Batch-------------") + + _gt_bboxes_per_image = gt_bboxes_per_image.cpu().float() + _gt_classes = gt_classes.cpu().float() + _bboxes_preds_per_image = bboxes_preds_per_image.cpu().float() + _cls_preds_per_image = cls_preds_per_image.cpu().float() + _obj_preds_per_image = obj_preds_per_image.cpu().float() + + _expanded_strides = expanded_strides.cpu().float() + _xy_shifts = xy_shifts.cpu() + + ( + gt_matched_classes, + fg_mask, + pred_ious_this_matching, + matched_gt_inds, + num_fg_img, + ) = self.get_assignments( + batch_idx, + num_gt, + total_num_anchors, + _gt_bboxes_per_image, + _gt_classes, + _bboxes_preds_per_image, + _cls_preds_per_image, + _obj_preds_per_image, + _expanded_strides, + _xy_shifts, + num_classes + ) + + gt_matched_classes = gt_matched_classes.cuda() + fg_mask = fg_mask.cuda() + pred_ious_this_matching = pred_ious_this_matching.cuda() + matched_gt_inds = matched_gt_inds.cuda() + + torch.cuda.empty_cache() + num_fg += num_fg_img + if num_fg_img > 0: + cls_target = F.one_hot( + gt_matched_classes.to(torch.int64), num_classes + ) * pred_ious_this_matching.unsqueeze(-1) + obj_target = fg_mask.unsqueeze(-1) + reg_target = gt_bboxes_per_image[matched_gt_inds] + + l1_target = self.get_l1_target( + outputs.new_zeros((num_fg_img, 4)), + gt_bboxes_per_image[matched_gt_inds], + expanded_strides[0][fg_mask], + xy_shifts=xy_shifts[0][fg_mask], + ) + + cls_targets.append(cls_target) + reg_targets.append(reg_target) + obj_targets.append(obj_target) + l1_targets.append(l1_target) + fg_masks.append(fg_mask) + + cls_targets = torch.cat(cls_targets, 0) + reg_targets = torch.cat(reg_targets, 0) + obj_targets = torch.cat(obj_targets, 0) + l1_targets = torch.cat(l1_targets, 0) + fg_masks = torch.cat(fg_masks, 0) + + num_fg = max(num_fg, 1) + # loss + loss_iou += (self.iou_loss(bbox_preds.view(-1, 4)[fg_masks].T, reg_targets)).sum() / num_fg + loss_l1 += (self.l1_loss(bbox_preds_org.view(-1, 4)[fg_masks], l1_targets)).sum() / num_fg + + loss_obj += (self.bcewithlog_loss(obj_preds.view(-1, 1), obj_targets*1.0)).sum() / num_fg + loss_cls += (self.bcewithlog_loss(cls_preds.view(-1, num_classes)[fg_masks], cls_targets)).sum() / num_fg + + total_losses = self.reg_weight * loss_iou + loss_l1 + loss_obj + loss_cls + return total_losses, torch.cat((self.reg_weight * loss_iou, loss_l1, loss_obj, loss_cls)).detach() + + def decode_output(self, output, k, stride, dtype, device): + grid = self.grids[k].to(device) + batch_size = output.shape[0] + hsize, wsize = output.shape[2:4] + if grid.shape[2:4] != output.shape[2:4]: + yv, xv = torch.meshgrid([torch.arange(hsize), torch.arange(wsize)]) + grid = torch.stack((xv, yv), 2).view(1, 1, hsize, wsize, 2).type(dtype).to(device) + self.grids[k] = grid + + output = output.reshape(batch_size, self.n_anchors * hsize * wsize, -1) + output_origin = output.clone() + grid = grid.view(1, -1, 2) + + output[..., :2] = (output[..., :2] + grid) * stride + output[..., 2:4] = torch.exp(output[..., 2:4]) * stride + + return output, output_origin, grid, hsize, wsize + + def get_outputs_and_grids(self, outputs, strides, dtype, device): + xy_shifts = [] + expanded_strides = [] + outputs_new = [] + outputs_origin = [] + + for k, output in enumerate(outputs): + output, output_origin, grid, feat_h, feat_w = self.decode_output( + output, k, strides[k], dtype, device) + + xy_shift = grid + expanded_stride = torch.full((1, grid.shape[1], 1), strides[k], dtype=grid.dtype, device=grid.device) + + xy_shifts.append(xy_shift) + expanded_strides.append(expanded_stride) + outputs_new.append(output) + outputs_origin.append(output_origin) + + xy_shifts = torch.cat(xy_shifts, 1) # [1, n_anchors_all, 2] + expanded_strides = torch.cat(expanded_strides, 1) # [1, n_anchors_all, 1] + outputs_origin = torch.cat(outputs_origin, 1) + outputs = torch.cat(outputs_new, 1) + + feat_h *= strides[-1] + feat_w *= strides[-1] + gt_bboxes_scale = torch.Tensor([[feat_w, feat_h, feat_w, feat_h]]).type_as(outputs) + + return outputs, outputs_origin, gt_bboxes_scale, xy_shifts, expanded_strides + + def get_l1_target(self, l1_target, gt, stride, xy_shifts, eps=1e-8): + + l1_target[:, 0:2] = gt[:, 0:2] / stride - xy_shifts + l1_target[:, 2:4] = torch.log(gt[:, 2:4] / stride + eps) + return l1_target + + @torch.no_grad() + def get_assignments( + self, + batch_idx, + num_gt, + total_num_anchors, + gt_bboxes_per_image, + gt_classes, + bboxes_preds_per_image, + cls_preds_per_image, + obj_preds_per_image, + expanded_strides, + xy_shifts, + num_classes + ): + + fg_mask, is_in_boxes_and_center = self.get_in_boxes_info( + gt_bboxes_per_image, + expanded_strides, + xy_shifts, + total_num_anchors, + num_gt, + ) + + bboxes_preds_per_image = bboxes_preds_per_image[fg_mask] + cls_preds_ = cls_preds_per_image[fg_mask] + obj_preds_ = obj_preds_per_image[fg_mask] + num_in_boxes_anchor = bboxes_preds_per_image.shape[0] + + # cost + pair_wise_ious = pairwise_bbox_iou(gt_bboxes_per_image, bboxes_preds_per_image, box_format='xywh') + pair_wise_ious_loss = -torch.log(pair_wise_ious + 1e-8) + + gt_cls_per_image = ( + F.one_hot(gt_classes.to(torch.int64), num_classes) + .float() + .unsqueeze(1) + .repeat(1, num_in_boxes_anchor, 1) + ) + + with torch.cuda.amp.autocast(enabled=False): + cls_preds_ = ( + cls_preds_.float().sigmoid_().unsqueeze(0).repeat(num_gt, 1, 1) + * obj_preds_.float().sigmoid_().unsqueeze(0).repeat(num_gt, 1, 1) + ) + pair_wise_cls_loss = F.binary_cross_entropy( + cls_preds_.sqrt_(), gt_cls_per_image, reduction="none" + ).sum(-1) + del cls_preds_, obj_preds_ + + cost = ( + self.cls_weight * pair_wise_cls_loss + + self.iou_weight * pair_wise_ious_loss + + 100000.0 * (~is_in_boxes_and_center) + ) + + ( + num_fg, + gt_matched_classes, + pred_ious_this_matching, + matched_gt_inds, + ) = self.dynamic_k_matching(cost, pair_wise_ious, gt_classes, num_gt, fg_mask) + + del pair_wise_cls_loss, cost, pair_wise_ious, pair_wise_ious_loss + + return ( + gt_matched_classes, + fg_mask, + pred_ious_this_matching, + matched_gt_inds, + num_fg, + ) + + def get_in_boxes_info( + self, + gt_bboxes_per_image, + expanded_strides, + xy_shifts, + total_num_anchors, + num_gt, + ): + expanded_strides_per_image = expanded_strides[0] + xy_shifts_per_image = xy_shifts[0] * expanded_strides_per_image + xy_centers_per_image = ( + (xy_shifts_per_image + 0.5 * expanded_strides_per_image) + .unsqueeze(0) + .repeat(num_gt, 1, 1) + ) # [n_anchor, 2] -> [n_gt, n_anchor, 2] + + gt_bboxes_per_image_lt = ( + (gt_bboxes_per_image[:, 0:2] - 0.5 * gt_bboxes_per_image[:, 2:4]) + .unsqueeze(1) + .repeat(1, total_num_anchors, 1) + ) + gt_bboxes_per_image_rb = ( + (gt_bboxes_per_image[:, 0:2] + 0.5 * gt_bboxes_per_image[:, 2:4]) + .unsqueeze(1) + .repeat(1, total_num_anchors, 1) + ) # [n_gt, 2] -> [n_gt, n_anchor, 2] + + b_lt = xy_centers_per_image - gt_bboxes_per_image_lt + b_rb = gt_bboxes_per_image_rb - xy_centers_per_image + bbox_deltas = torch.cat([b_lt, b_rb], 2) + + is_in_boxes = bbox_deltas.min(dim=-1).values > 0.0 + is_in_boxes_all = is_in_boxes.sum(dim=0) > 0 + + # in fixed center + gt_bboxes_per_image_lt = (gt_bboxes_per_image[:, 0:2]).unsqueeze(1).repeat( + 1, total_num_anchors, 1 + ) - self.center_radius * expanded_strides_per_image.unsqueeze(0) + gt_bboxes_per_image_rb = (gt_bboxes_per_image[:, 0:2]).unsqueeze(1).repeat( + 1, total_num_anchors, 1 + ) + self.center_radius * expanded_strides_per_image.unsqueeze(0) + + c_lt = xy_centers_per_image - gt_bboxes_per_image_lt + c_rb = gt_bboxes_per_image_rb - xy_centers_per_image + center_deltas = torch.cat([c_lt, c_rb], 2) + is_in_centers = center_deltas.min(dim=-1).values > 0.0 + is_in_centers_all = is_in_centers.sum(dim=0) > 0 + + # in boxes and in centers + is_in_boxes_anchor = is_in_boxes_all | is_in_centers_all + + is_in_boxes_and_center = ( + is_in_boxes[:, is_in_boxes_anchor] & is_in_centers[:, is_in_boxes_anchor] + ) + return is_in_boxes_anchor, is_in_boxes_and_center + + def dynamic_k_matching(self, cost, pair_wise_ious, gt_classes, num_gt, fg_mask): + matching_matrix = torch.zeros_like(cost, dtype=torch.uint8) + ious_in_boxes_matrix = pair_wise_ious + n_candidate_k = min(10, ious_in_boxes_matrix.size(1)) + topk_ious, _ = torch.topk(ious_in_boxes_matrix, n_candidate_k, dim=1) + dynamic_ks = torch.clamp(topk_ious.sum(1).int(), min=1) + dynamic_ks = dynamic_ks.tolist() + + for gt_idx in range(num_gt): + _, pos_idx = torch.topk( + cost[gt_idx], k=dynamic_ks[gt_idx], largest=False + ) + matching_matrix[gt_idx][pos_idx] = 1 + del topk_ious, dynamic_ks, pos_idx + + anchor_matching_gt = matching_matrix.sum(0) + if (anchor_matching_gt > 1).sum() > 0: + _, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0) + matching_matrix[:, anchor_matching_gt > 1] *= 0 + matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1 + fg_mask_inboxes = matching_matrix.sum(0) > 0 + num_fg = fg_mask_inboxes.sum().item() + fg_mask[fg_mask.clone()] = fg_mask_inboxes + matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0) + gt_matched_classes = gt_classes[matched_gt_inds] + + pred_ious_this_matching = (matching_matrix * pair_wise_ious).sum(0)[ + fg_mask_inboxes + ] + + return num_fg, gt_matched_classes, pred_ious_this_matching, matched_gt_inds diff --git a/yolov6/models/reppan.py b/yolov6/models/reppan.py new file mode 100644 index 0000000000000000000000000000000000000000..2571d687155fb754dbe013a0f0f9468af487f8ca --- /dev/null +++ b/yolov6/models/reppan.py @@ -0,0 +1,108 @@ +import torch +from torch import nn +from yolov6.layers.common import RepBlock, SimConv, Transpose + + +class RepPANNeck(nn.Module): + """RepPANNeck Module + EfficientRep is the default backbone of this model. + RepPANNeck has the balance of feature fusion ability and hardware efficiency. + """ + + def __init__( + self, + channels_list=None, + num_repeats=None + ): + super().__init__() + + assert channels_list is not None + assert num_repeats is not None + + self.Rep_p4 = RepBlock( + in_channels=channels_list[3] + channels_list[5], + out_channels=channels_list[5], + n=num_repeats[5], + ) + + self.Rep_p3 = RepBlock( + in_channels=channels_list[2] + channels_list[6], + out_channels=channels_list[6], + n=num_repeats[6] + ) + + self.Rep_n3 = RepBlock( + in_channels=channels_list[6] + channels_list[7], + out_channels=channels_list[8], + n=num_repeats[7], + ) + + self.Rep_n4 = RepBlock( + in_channels=channels_list[5] + channels_list[9], + out_channels=channels_list[10], + n=num_repeats[8] + ) + + self.reduce_layer0 = SimConv( + in_channels=channels_list[4], + out_channels=channels_list[5], + kernel_size=1, + stride=1 + ) + + self.upsample0 = Transpose( + in_channels=channels_list[5], + out_channels=channels_list[5], + ) + + self.reduce_layer1 = SimConv( + in_channels=channels_list[5], + out_channels=channels_list[6], + kernel_size=1, + stride=1 + ) + + self.upsample1 = Transpose( + in_channels=channels_list[6], + out_channels=channels_list[6] + ) + + self.downsample2 = SimConv( + in_channels=channels_list[6], + out_channels=channels_list[7], + kernel_size=3, + stride=2 + ) + + self.downsample1 = SimConv( + in_channels=channels_list[8], + out_channels=channels_list[9], + kernel_size=3, + stride=2 + ) + + def forward(self, input): + + (x2, x1, x0) = input + + fpn_out0 = self.reduce_layer0(x0) + upsample_feat0 = self.upsample0(fpn_out0) + f_concat_layer0 = torch.cat([upsample_feat0, x1], 1) + f_out0 = self.Rep_p4(f_concat_layer0) + + fpn_out1 = self.reduce_layer1(f_out0) + upsample_feat1 = self.upsample1(fpn_out1) + f_concat_layer1 = torch.cat([upsample_feat1, x2], 1) + pan_out2 = self.Rep_p3(f_concat_layer1) + + down_feat1 = self.downsample2(pan_out2) + p_concat_layer1 = torch.cat([down_feat1, fpn_out1], 1) + pan_out1 = self.Rep_n3(p_concat_layer1) + + down_feat0 = self.downsample1(pan_out1) + p_concat_layer2 = torch.cat([down_feat0, fpn_out0], 1) + pan_out0 = self.Rep_n4(p_concat_layer2) + + outputs = [pan_out2, pan_out1, pan_out0] + + return outputs diff --git a/yolov6/models/yolo.py b/yolov6/models/yolo.py new file mode 100644 index 0000000000000000000000000000000000000000..5d3d86be4fa6e9ceab089bbf1c655f5bf86163bf --- /dev/null +++ b/yolov6/models/yolo.py @@ -0,0 +1,83 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import math +import torch.nn as nn +from yolov6.layers.common import * +from yolov6.utils.torch_utils import initialize_weights +from yolov6.models.efficientrep import EfficientRep +from yolov6.models.reppan import RepPANNeck +from yolov6.models.effidehead import Detect, build_effidehead_layer + + +class Model(nn.Module): + '''YOLOv6 model with backbone, neck and head. + The default parts are EfficientRep Backbone, Rep-PAN and + Efficient Decoupled Head. + ''' + def __init__(self, config, channels=3, num_classes=None, anchors=None): # model, input channels, number of classes + super().__init__() + # Build network + num_layers = config.model.head.num_layers + self.backbone, self.neck, self.detect = build_network(config, channels, num_classes, anchors, num_layers) + + # Init Detect head + begin_indices = config.model.head.begin_indices + out_indices_head = config.model.head.out_indices + self.stride = self.detect.stride + self.detect.i = begin_indices + self.detect.f = out_indices_head + self.detect.initialize_biases() + + # Init weights + initialize_weights(self) + + def forward(self, x): + x = self.backbone(x) + x = self.neck(x) + x = self.detect(x) + return x + + def _apply(self, fn): + self = super()._apply(fn) + self.detect.stride = fn(self.detect.stride) + self.detect.grid = list(map(fn, self.detect.grid)) + return self + + +def make_divisible(x, divisor): + # Upward revision the value x to make it evenly divisible by the divisor. + return math.ceil(x / divisor) * divisor + + +def build_network(config, channels, num_classes, anchors, num_layers): + depth_mul = config.model.depth_multiple + width_mul = config.model.width_multiple + num_repeat_backbone = config.model.backbone.num_repeats + channels_list_backbone = config.model.backbone.out_channels + num_repeat_neck = config.model.neck.num_repeats + channels_list_neck = config.model.neck.out_channels + num_anchors = config.model.head.anchors + num_repeat = [(max(round(i * depth_mul), 1) if i > 1 else i) for i in (num_repeat_backbone + num_repeat_neck)] + channels_list = [make_divisible(i * width_mul, 8) for i in (channels_list_backbone + channels_list_neck)] + + backbone = EfficientRep( + in_channels=channels, + channels_list=channels_list, + num_repeats=num_repeat + ) + + neck = RepPANNeck( + channels_list=channels_list, + num_repeats=num_repeat + ) + + head_layers = build_effidehead_layer(channels_list, num_anchors, num_classes) + + head = Detect(num_classes, anchors, num_layers, head_layers=head_layers) + + return backbone, neck, head + + +def build_model(cfg, num_classes, device): + model = Model(cfg, channels=3, num_classes=num_classes, anchors=cfg.model.head.anchors).to(device) + return model diff --git a/yolov6/solver/build.py b/yolov6/solver/build.py new file mode 100644 index 0000000000000000000000000000000000000000..c18c97bb759913bbf4e8139e550bf08bf1aacc77 --- /dev/null +++ b/yolov6/solver/build.py @@ -0,0 +1,41 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import os +import math + +import torch +import torch.nn as nn + +def build_optimizer(cfg, model): + """ Build optimizer from cfg file.""" + g_bnw, g_w, g_b = [], [], [] + for v in model.modules(): + if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): + g_b.append(v.bias) + if isinstance(v, nn.BatchNorm2d): + g_bnw.append(v.weight) + elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): + g_w.append(v.weight) + + assert cfg.solver.optim == 'SGD' or 'Adam', 'ERROR: unknown optimizer, use SGD defaulted' + if cfg.solver.optim == 'SGD': + optimizer = torch.optim.SGD(g_bnw, lr=cfg.solver.lr0, momentum=cfg.solver.momentum, nesterov=True) + elif cfg.solver.optim == 'Adam': + optimizer = torch.optim.Adam(g_bnw, lr=cfg.solver.lr0, betas=(cfg.solver.momentum, 0.999)) + + optimizer.add_param_group({'params': g_w, 'weight_decay': cfg.solver.weight_decay}) + optimizer.add_param_group({'params': g_b}) + + del g_bnw, g_w, g_b + return optimizer + + +def build_lr_scheduler(cfg, optimizer, epochs): + """Build learning rate scheduler from cfg file.""" + if cfg.solver.lr_scheduler == 'Cosine': + lf = lambda x: ((1 - math.cos(x * math.pi / epochs)) / 2) * (cfg.solver.lrf - 1) + 1 + else: + LOGGER.error('unknown lr scheduler, use Cosine defaulted') + + scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) + return scheduler, lf diff --git a/yolov6/utils/Arial.ttf b/yolov6/utils/Arial.ttf new file mode 100644 index 0000000000000000000000000000000000000000..ab68fb197d4479b3b6dec6e85bd5cbaf433a87c5 Binary files /dev/null and b/yolov6/utils/Arial.ttf differ diff --git a/yolov6/utils/checkpoint.py b/yolov6/utils/checkpoint.py new file mode 100644 index 0000000000000000000000000000000000000000..3ce2ede485e62131554acc374c990331bcbfbac5 --- /dev/null +++ b/yolov6/utils/checkpoint.py @@ -0,0 +1,60 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import os +import shutil +import torch +import os.path as osp +from yolov6.utils.events import LOGGER +from yolov6.utils.torch_utils import fuse_model + + +def load_state_dict(weights, model, map_location=None): + """Load weights from checkpoint file, only assign weights those layers' name and shape are match.""" + ckpt = torch.load(weights, map_location=map_location) + state_dict = ckpt['model'].float().state_dict() + model_state_dict = model.state_dict() + state_dict = {k: v for k, v in state_dict.items() if k in model_state_dict and v.shape == model_state_dict[k].shape} + model.load_state_dict(state_dict, strict=False) + del ckpt, state_dict, model_state_dict + return model + + +def load_checkpoint(weights, map_location=None, inplace=True, fuse=True): + """Load model from checkpoint file.""" + LOGGER.info("Loading checkpoint from {}".format(weights)) + ckpt = torch.load(weights, map_location=map_location) # load + model = ckpt['ema' if ckpt.get('ema') else 'model'].float() + if fuse: + LOGGER.info("\nFusing model...") + model = fuse_model(model).eval() + else: + model = model.eval() + return model + + +def save_checkpoint(ckpt, is_best, save_dir, model_name=""): + """ Save checkpoint to the disk.""" + if not osp.exists(save_dir): + os.makedirs(save_dir) + filename = osp.join(save_dir, model_name + '.pt') + torch.save(ckpt, filename) + if is_best: + best_filename = osp.join(save_dir, 'best_ckpt.pt') + shutil.copyfile(filename, best_filename) + + +def strip_optimizer(ckpt_dir): + for s in ['best', 'last']: + ckpt_path = osp.join(ckpt_dir, '{}_ckpt.pt'.format(s)) + if not osp.exists(ckpt_path): + continue + ckpt = torch.load(ckpt_path, map_location=torch.device('cpu')) + if ckpt.get('ema'): + ckpt['model'] = ckpt['ema'] # replace model with ema + for k in ['optimizer', 'ema', 'updates']: # keys + ckpt[k] = None + ckpt['epoch'] = -1 + ckpt['model'].half() # to FP16 + for p in ckpt['model'].parameters(): + p.requires_grad = False + torch.save(ckpt, ckpt_path) diff --git a/yolov6/utils/config.py b/yolov6/utils/config.py new file mode 100644 index 0000000000000000000000000000000000000000..7f9c13a3085e0738a3547fc35c5106defed4c489 --- /dev/null +++ b/yolov6/utils/config.py @@ -0,0 +1,101 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- +# The code is based on +# https://github.com/open-mmlab/mmcv/blob/master/mmcv/utils/config.py +# Copyright (c) OpenMMLab. + +import os.path as osp +import shutil +import sys +import tempfile +from importlib import import_module +from addict import Dict + + +class ConfigDict(Dict): + + def __missing__(self, name): + raise KeyError(name) + + def __getattr__(self, name): + try: + value = super(ConfigDict, self).__getattr__(name) + except KeyError: + ex = AttributeError("'{}' object has no attribute '{}'".format( + self.__class__.__name__, name)) + except Exception as e: + ex = e + else: + return value + raise ex + + +class Config(object): + + @staticmethod + def _file2dict(filename): + filename = str(filename) + if filename.endswith('.py'): + with tempfile.TemporaryDirectory() as temp_config_dir: + shutil.copyfile(filename, + osp.join(temp_config_dir, '_tempconfig.py')) + sys.path.insert(0, temp_config_dir) + mod = import_module('_tempconfig') + sys.path.pop(0) + cfg_dict = { + name: value + for name, value in mod.__dict__.items() + if not name.startswith('__') + } + # delete imported module + del sys.modules['_tempconfig'] + else: + raise IOError('Only .py type are supported now!') + cfg_text = filename + '\n' + with open(filename, 'r') as f: + cfg_text += f.read() + + return cfg_dict, cfg_text + + @staticmethod + def fromfile(filename): + cfg_dict, cfg_text = Config._file2dict(filename) + return Config(cfg_dict, cfg_text=cfg_text, filename=filename) + + def __init__(self, cfg_dict=None, cfg_text=None, filename=None): + if cfg_dict is None: + cfg_dict = dict() + elif not isinstance(cfg_dict, dict): + raise TypeError('cfg_dict must be a dict, but got {}'.format( + type(cfg_dict))) + + super(Config, self).__setattr__('_cfg_dict', ConfigDict(cfg_dict)) + super(Config, self).__setattr__('_filename', filename) + if cfg_text: + text = cfg_text + elif filename: + with open(filename, 'r') as f: + text = f.read() + else: + text = '' + super(Config, self).__setattr__('_text', text) + + @property + def filename(self): + return self._filename + + @property + def text(self): + return self._text + + def __repr__(self): + return 'Config (path: {}): {}'.format(self.filename, + self._cfg_dict.__repr__()) + + def __getattr__(self, name): + return getattr(self._cfg_dict, name) + + def __setattr__(self, name, value): + if isinstance(value, dict): + value = ConfigDict(value) + self._cfg_dict.__setattr__(name, value) diff --git a/yolov6/utils/ema.py b/yolov6/utils/ema.py new file mode 100644 index 0000000000000000000000000000000000000000..104d97b36cb357d1f0c408751ed4aa9529511007 --- /dev/null +++ b/yolov6/utils/ema.py @@ -0,0 +1,59 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +# The code is based on +# https://github.com/ultralytics/yolov5/blob/master/utils/torch_utils.py +import math +from copy import deepcopy +import torch +import torch.nn as nn + + +class ModelEMA: + """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models + Keep a moving average of everything in the model state_dict (parameters and buffers). + This is intended to allow functionality like + https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage + A smoothed version of the weights is necessary for some training schemes to perform well. + This class is sensitive where it is initialized in the sequence of model init, + GPU assignment and distributed training wrappers. + """ + + def __init__(self, model, decay=0.9999, updates=0): + self.ema = deepcopy(model.module if is_parallel(model) else model).eval() # FP32 EMA + self.updates = updates + self.decay = lambda x: decay * (1 - math.exp(-x / 2000)) + for param in self.ema.parameters(): + param.requires_grad_(False) + + def update(self, model): + with torch.no_grad(): + self.updates += 1 + decay = self.decay(self.updates) + + state_dict = model.module.state_dict() if is_parallel(model) else model.state_dict() # model state_dict + for k, item in self.ema.state_dict().items(): + if item.dtype.is_floating_point: + item *= decay + item += (1 - decay) * state_dict[k].detach() + + def update_attr(self, model, include=(), exclude=('process_group', 'reducer')): + copy_attr(self.ema, model, include, exclude) + + +def copy_attr(a, b, include=(), exclude=()): + """Copy attributes from one instance and set them to another instance.""" + for k, item in b.__dict__.items(): + if (len(include) and k not in include) or k.startswith('_') or k in exclude: + continue + else: + setattr(a, k, item) + + +def is_parallel(model): + # Return True if model's type is DP or DDP, else False. + return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) + + +def de_parallel(model): + # De-parallelize a model. Return single-GPU model if model's type is DP or DDP. + return model.module if is_parallel(model) else model diff --git a/yolov6/utils/envs.py b/yolov6/utils/envs.py new file mode 100644 index 0000000000000000000000000000000000000000..10159a9484ed525ad5ef3826ec3db4bf70b4c9cc --- /dev/null +++ b/yolov6/utils/envs.py @@ -0,0 +1,54 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- +import os +import random +import numpy as np + +import torch +import torch.backends.cudnn as cudnn +from yolov6.utils.events import LOGGER + + +def get_envs(): + """Get PyTorch needed environments from system envirionments.""" + local_rank = int(os.getenv('LOCAL_RANK', -1)) + rank = int(os.getenv('RANK', -1)) + world_size = int(os.getenv('WORLD_SIZE', 1)) + return local_rank, rank, world_size + + +def select_device(device): + """Set devices' information to the program. + Args: + device: a string, like 'cpu' or '1,2,3,4' + Returns: + torch.device + """ + if device == 'cpu': + os.environ['CUDA_VISIBLE_DEVICES'] = '-1' + LOGGER.info('Using CPU for training... ') + elif device: + os.environ['CUDA_VISIBLE_DEVICES'] = device + assert torch.cuda.is_available() + nd = len(device.strip().split(',')) + LOGGER.info(f'Using {nd} GPU for training... ') + cuda = device != 'cpu' and torch.cuda.is_available() + device = torch.device('cuda:0' if cuda else 'cpu') + return device + + +def set_random_seed(seed, deterministic=False): + """ Set random state to random libray, numpy, torch and cudnn. + Args: + seed: int value. + deterministic: bool value. + """ + random.seed(seed) + np.random.seed(seed) + torch.manual_seed(seed) + if deterministic: + cudnn.deterministic = True + cudnn.benchmark = False + else: + cudnn.deterministic = False + cudnn.benchmark = True diff --git a/yolov6/utils/events.py b/yolov6/utils/events.py new file mode 100644 index 0000000000000000000000000000000000000000..6a3dd509423b00182ba80e87dd20adc2971a029b --- /dev/null +++ b/yolov6/utils/events.py @@ -0,0 +1,41 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- +import os +import yaml +import logging +import shutil + + +def set_logging(name=None): + rank = int(os.getenv('RANK', -1)) + logging.basicConfig(format="%(message)s", level=logging.INFO if (rank in (-1, 0)) else logging.WARNING) + return logging.getLogger(name) + + +LOGGER = set_logging(__name__) +NCOLS = shutil.get_terminal_size().columns + + +def load_yaml(file_path): + """Load data from yaml file.""" + if isinstance(file_path, str): + with open(file_path, errors='ignore') as f: + data_dict = yaml.safe_load(f) + return data_dict + + +def save_yaml(data_dict, save_path): + """Save data to yaml file""" + with open(save_path, 'w') as f: + yaml.safe_dump(data_dict, f, sort_keys=False) + + +def write_tblog(tblogger, epoch, results, losses): + """Display mAP and loss information to log.""" + tblogger.add_scalar("val/mAP@0.5", results[0], epoch + 1) + tblogger.add_scalar("val/mAP@0.50:0.95", results[1], epoch + 1) + + tblogger.add_scalar("train/iou_loss", losses[0], epoch + 1) + tblogger.add_scalar("train/l1_loss", losses[1], epoch + 1) + tblogger.add_scalar("train/obj_loss", losses[2], epoch + 1) + tblogger.add_scalar("train/cls_loss", losses[3], epoch + 1) diff --git a/yolov6/utils/figure_iou.py b/yolov6/utils/figure_iou.py new file mode 100644 index 0000000000000000000000000000000000000000..f3a0f3f7b7c477d38840ddbcddba1de405aaf164 --- /dev/null +++ b/yolov6/utils/figure_iou.py @@ -0,0 +1,114 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +import math +import torch + + +class IOUloss: + """ Calculate IoU loss. + """ + def __init__(self, box_format='xywh', iou_type='ciou', reduction='none', eps=1e-7): + """ Setting of the class. + Args: + box_format: (string), must be one of 'xywh' or 'xyxy'. + iou_type: (string), can be one of 'ciou', 'diou', 'giou' or 'siou' + reduction: (string), specifies the reduction to apply to the output, must be one of 'none', 'mean','sum'. + eps: (float), a value to avoid devide by zero error. + """ + self.box_format = box_format + self.iou_type = iou_type.lower() + self.reduction = reduction + self.eps = eps + + def __call__(self, box1, box2): + """ calculate iou. box1 and box2 are torch tensor with shape [M, 4] and [Nm 4]. + """ + box2 = box2.T + if self.box_format == 'xyxy': + b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3] + b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3] + elif self.box_format == 'xywh': + b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2 + b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2 + b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2 + b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2 + + # Intersection area + inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \ + (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0) + + # Union Area + w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + self.eps + w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + self.eps + union = w1 * h1 + w2 * h2 - inter + self.eps + iou = inter / union + + cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex width + ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height + if self.iou_type == 'giou': + c_area = cw * ch + self.eps # convex area + iou = iou - (c_area - union) / c_area + elif self.iou_type in ['diou', 'ciou']: + c2 = cw ** 2 + ch ** 2 + self.eps # convex diagonal squared + rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + + (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center distance squared + if self.iou_type == 'diou': + iou = iou - rho2 / c2 + elif self.iou_type == 'ciou': + v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2) + with torch.no_grad(): + alpha = v / (v - iou + (1 + self.eps)) + iou = iou - (rho2 / c2 + v * alpha) + elif self.iou_type == 'siou': + # SIoU Loss https://arxiv.org/pdf/2205.12740.pdf + s_cw = (b2_x1 + b2_x2 - b1_x1 - b1_x2) * 0.5 + s_ch = (b2_y1 + b2_y2 - b1_y1 - b1_y2) * 0.5 + sigma = torch.pow(s_cw ** 2 + s_ch ** 2, 0.5) + sin_alpha_1 = torch.abs(s_cw) / sigma + sin_alpha_2 = torch.abs(s_ch) / sigma + threshold = pow(2, 0.5) / 2 + sin_alpha = torch.where(sin_alpha_1 > threshold, sin_alpha_2, sin_alpha_1) + angle_cost = torch.cos(torch.arcsin(sin_alpha) * 2 - math.pi / 2) + rho_x = (s_cw / cw) ** 2 + rho_y = (s_ch / ch) ** 2 + gamma = angle_cost - 2 + distance_cost = 2 - torch.exp(gamma * rho_x) - torch.exp(gamma * rho_y) + omiga_w = torch.abs(w1 - w2) / torch.max(w1, w2) + omiga_h = torch.abs(h1 - h2) / torch.max(h1, h2) + shape_cost = torch.pow(1 - torch.exp(-1 * omiga_w), 4) + torch.pow(1 - torch.exp(-1 * omiga_h), 4) + iou = iou - 0.5 * (distance_cost + shape_cost) + loss = 1.0 - iou + + if self.reduction == 'sum': + loss = loss.sum() + elif self.reduction == 'mean': + loss = loss.mean() + + return loss + + +def pairwise_bbox_iou(box1, box2, box_format='xywh'): + """Calculate iou. + This code is based on https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/utils/boxes.py + """ + if box_format == 'xyxy': + lt = torch.max(box1[:, None, :2], box2[:, :2]) + rb = torch.min(box1[:, None, 2:], box2[:, 2:]) + area_1 = torch.prod(box1[:, 2:] - box1[:, :2], 1) + area_2 = torch.prod(box2[:, 2:] - box2[:, :2], 1) + + elif box_format == 'xywh': + lt = torch.max( + (box1[:, None, :2] - box1[:, None, 2:] / 2), + (box2[:, :2] - box2[:, 2:] / 2), + ) + rb = torch.min( + (box1[:, None, :2] + box1[:, None, 2:] / 2), + (box2[:, :2] + box2[:, 2:] / 2), + ) + + area_1 = torch.prod(box1[:, 2:], 1) + area_2 = torch.prod(box2[:, 2:], 1) + valid = (lt < rb).type(lt.type()).prod(dim=2) + inter = torch.prod(rb - lt, 2) * valid + return inter / (area_1[:, None] + area_2 - inter) diff --git a/yolov6/utils/nms.py b/yolov6/utils/nms.py new file mode 100644 index 0000000000000000000000000000000000000000..9c61b7cc4567b03cd2977b505b89c76e0e1d6769 --- /dev/null +++ b/yolov6/utils/nms.py @@ -0,0 +1,106 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- +# The code is based on +# https://github.com/ultralytics/yolov5/blob/master/utils/general.py + +import os +import time +import numpy as np +import cv2 +import torch +import torchvision + + +# Settings +torch.set_printoptions(linewidth=320, precision=5, profile='long') +np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format}) # format short g, %precision=5 +cv2.setNumThreads(0) # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader) +os.environ['NUMEXPR_MAX_THREADS'] = str(min(os.cpu_count(), 8)) # NumExpr max threads + + +def xywh2xyxy(x): + # Convert boxes with shape [n, 4] from [x, y, w, h] to [x1, y1, x2, y2] where x1y1 is top-left, x2y2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x + y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y + y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x + y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y + return y + + +def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False, max_det=300): + """Runs Non-Maximum Suppression (NMS) on inference results. + This code is borrowed from: https://github.com/ultralytics/yolov5/blob/47233e1698b89fc437a4fb9463c815e9171be955/utils/general.py#L775 + Args: + prediction: (tensor), with shape [N, 5 + num_classes], N is the number of bboxes. + conf_thres: (float) confidence threshold. + iou_thres: (float) iou threshold. + classes: (None or list[int]), if a list is provided, nms only keep the classes you provide. + agnostic: (bool), when it is set to True, we do class-independent nms, otherwise, different class would do nms respectively. + multi_label: (bool), when it is set to True, one box can have multi labels, otherwise, one box only huave one label. + max_det:(int), max number of output bboxes. + + Returns: + list of detections, echo item is one tensor with shape (num_boxes, 6), 6 is for [xyxy, conf, cls]. + """ + + num_classes = prediction.shape[2] - 5 # number of classes + pred_candidates = prediction[..., 4] > conf_thres # candidates + + # Check the parameters. + assert 0 <= conf_thres <= 1, f'conf_thresh must be in 0.0 to 1.0, however {conf_thres} is provided.' + assert 0 <= iou_thres <= 1, f'iou_thres must be in 0.0 to 1.0, however {iou_thres} is provided.' + + # Function settings. + max_wh = 4096 # maximum box width and height + max_nms = 30000 # maximum number of boxes put into torchvision.ops.nms() + time_limit = 10.0 # quit the function when nms cost time exceed the limit time. + multi_label &= num_classes > 1 # multiple labels per box + + tik = time.time() + output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0] + for img_idx, x in enumerate(prediction): # image index, image inference + x = x[pred_candidates[img_idx]] # confidence + + # If no box remains, skip the next process. + if not x.shape[0]: + continue + + # confidence multiply the objectness + x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf + + # (center x, center y, width, height) to (x1, y1, x2, y2) + box = xywh2xyxy(x[:, :4]) + + # Detections matrix's shape is (n,6), each row represents (xyxy, conf, cls) + if multi_label: + box_idx, class_idx = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T + x = torch.cat((box[box_idx], x[box_idx, class_idx + 5, None], class_idx[:, None].float()), 1) + else: # Only keep the class with highest scores. + conf, class_idx = x[:, 5:].max(1, keepdim=True) + x = torch.cat((box, conf, class_idx.float()), 1)[conf.view(-1) > conf_thres] + + # Filter by class, only keep boxes whose category is in classes. + if classes is not None: + x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)] + + # Check shape + num_box = x.shape[0] # number of boxes + if not num_box: # no boxes kept. + continue + elif num_box > max_nms: # excess max boxes' number. + x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence + + # Batched NMS + class_offset = x[:, 5:6] * (0 if agnostic else max_wh) # classes + boxes, scores = x[:, :4] + class_offset, x[:, 4] # boxes (offset by class), scores + keep_box_idx = torchvision.ops.nms(boxes, scores, iou_thres) # NMS + if keep_box_idx.shape[0] > max_det: # limit detections + keep_box_idx = keep_box_idx[:max_det] + + output[img_idx] = x[keep_box_idx] + if (time.time() - tik) > time_limit: + print(f'WARNING: NMS cost time exceed the limited {time_limit}s.') + break # time limit exceeded + + return output diff --git a/yolov6/utils/torch_utils.py b/yolov6/utils/torch_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..9e67c2ab18e048f0d05ddd61750843ebd73669de --- /dev/null +++ b/yolov6/utils/torch_utils.py @@ -0,0 +1,109 @@ +#!/usr/bin/env python3 +# -*- coding:utf-8 -*- + +import time +from contextlib import contextmanager +from copy import deepcopy +import torch +import torch.distributed as dist +import torch.nn as nn +import torch.nn.functional as F +from yolov6.utils.events import LOGGER + +try: + import thop # for FLOPs computation +except ImportError: + thop = None + + +@contextmanager +def torch_distributed_zero_first(local_rank: int): + """ + Decorator to make all processes in distributed training wait for each local_master to do something. + """ + if local_rank not in [-1, 0]: + dist.barrier(device_ids=[local_rank]) + yield + if local_rank == 0: + dist.barrier(device_ids=[0]) + + +def time_sync(): + # Waits for all kernels in all streams on a CUDA device to complete if cuda is available. + if torch.cuda.is_available(): + torch.cuda.synchronize() + return time.time() + + +def initialize_weights(model): + for m in model.modules(): + t = type(m) + if t is nn.Conv2d: + pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') + elif t is nn.BatchNorm2d: + m.eps = 1e-3 + m.momentum = 0.03 + elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]: + m.inplace = True + + +def fuse_conv_and_bn(conv, bn): + # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/ + fusedconv = ( + nn.Conv2d( + conv.in_channels, + conv.out_channels, + kernel_size=conv.kernel_size, + stride=conv.stride, + padding=conv.padding, + groups=conv.groups, + bias=True, + ) + .requires_grad_(False) + .to(conv.weight.device) + ) + + # prepare filters + w_conv = conv.weight.clone().view(conv.out_channels, -1) + w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) + fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape)) + + # prepare spatial bias + b_conv = ( + torch.zeros(conv.weight.size(0), device=conv.weight.device) + if conv.bias is None + else conv.bias + ) + b_bn = bn.bias - bn.weight.mul(bn.running_mean).div( + torch.sqrt(bn.running_var + bn.eps) + ) + fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) + + return fusedconv + + +def fuse_model(model): + from yolov6.layers.common import Conv + + for m in model.modules(): + if type(m) is Conv and hasattr(m, "bn"): + m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv + delattr(m, "bn") # remove batchnorm + m.forward = m.forward_fuse # update forward + return model + + +def get_model_info(model, img_size=640): + """Get model Params and GFlops. + Code base on https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/utils/model_utils.py + """ + from thop import profile + stride = 32 + img = torch.zeros((1, 3, stride, stride), device=next(model.parameters()).device) + flops, params = profile(deepcopy(model), inputs=(img,), verbose=False) + params /= 1e6 + flops /= 1e9 + img_size = img_size if isinstance(img_size, list) else [img_size, img_size] + flops *= img_size[0] * img_size[1] / stride / stride * 2 # Gflops + info = "Params: {:.2f}M, Gflops: {:.2f}".format(params, flops) + return info diff --git a/yolov6s.pt b/yolov6s.pt new file mode 100644 index 0000000000000000000000000000000000000000..e2fdde47b2b765dab48d48dd7b6955d57ca3a714 --- /dev/null +++ b/yolov6s.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7b05fef6f9bdf1fc9f8c201bf50f6ac3ce77b7a91ece9ab32d2e2b6b8518723 +size 38102844