Merge branch 'meituan:main' into patch-1

.pre-commit-config.yaml

@@ -0,0 +1,7 @@
+repos:
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.3.0
+    hooks:
+      - id: check-yaml
+      - id: end-of-file-fixer
+      - id: trailing-whitespace

README.md

@@ -18,6 +18,7 @@ YOLOv6 is composed of the following methods:
 
 - [ ] YOLOv6 m/l/x model.
 - [ ] Deployment for MNN/TNN/NCNN/CoreML...
+- [ ] Quantization tools
 
 
 ## Quick Start
@@ -27,23 +28,23 @@ YOLOv6 is composed of the following methods:
 ```shell
 git clone https://github.com/meituan/YOLOv6
 cd YOLOv6
-pip install -r requirements.txt  
+pip install -r requirements.txt
 ```
 
-### Inference 
+### Inference
 
 First, download a pretrained model from the YOLOv6 release
 
 Second, run inference with `tools/infer.py`
 
 ```shell
 python tools/infer.py --weights yolov6s.pt --source [img.jpg / imgdir]
-                                yolov6n.pt  
+                                yolov6n.pt
 ```
 
 ### Training
 
-Single GPU 
+Single GPU
 
 ```shell
 python tools/train.py --batch 256 --conf configs/yolov6s.py --data data/coco.yaml --device 0
@@ -96,4 +97,4 @@ python tools/eval.py --data data/coco.yaml  --batch 32 --weights yolov6s.pt --ta
 - 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.
+- Speed results of other methods are tested in our environment using official codebase and model if not found from the corresponding official release.

configs/yolov6_tiny.py

@@ -2,7 +2,7 @@
 model = dict(
     type='YOLOv6t',
     pretrained=None,
-    depth_multiple=0.25,  
+    depth_multiple=0.25,
     width_multiple=0.50,
     backbone=dict(
         type='EfficientRep',
@@ -35,12 +35,12 @@
     weight_decay=0.0005,
     warmup_epochs=3.0,
     warmup_momentum=0.8,
-    warmup_bias_lr=0.1       
+    warmup_bias_lr=0.1
 )
- 
+
 data_aug = dict(
-    hsv_h=0.015,  
-    hsv_s=0.7, 
+    hsv_h=0.015,
+    hsv_s=0.7,
     hsv_v=0.4,
     degrees=0.0,
     translate=0.1,

configs/yolov6_tiny_finetune.py

@@ -2,7 +2,7 @@
 model = dict(
     type='YOLOv6t',
     pretrained='./weights/yolov6t.pt',
-    depth_multiple=0.25,  
+    depth_multiple=0.25,
     width_multiple=0.50,
     backbone=dict(
         type='EfficientRep',
@@ -24,7 +24,7 @@
         strides=[8, 16, 32],
         iou_type='ciou'
     )
-)   
+)
 
 solver = dict(
     optim='SGD',
@@ -35,9 +35,9 @@
     weight_decay=0.00036,
     warmup_epochs=2.0,
     warmup_momentum=0.5,
-    warmup_bias_lr=0.05 
+    warmup_bias_lr=0.05
 )
- 
+
 data_aug = dict(
     hsv_h=0.0138,
     hsv_s=0.664,

configs/yolov6n.py

@@ -2,7 +2,7 @@
 model = dict(
     type='YOLOv6n',
     pretrained=None,
-    depth_multiple=0.33,  
+    depth_multiple=0.33,
     width_multiple=0.25,
     backbone=dict(
         type='EfficientRep',
@@ -35,12 +35,12 @@
     weight_decay=0.0005,
     warmup_epochs=3.0,
     warmup_momentum=0.8,
-    warmup_bias_lr=0.1       
+    warmup_bias_lr=0.1
 )
 
 data_aug = dict(
-    hsv_h=0.015,  
-    hsv_s=0.7, 
+    hsv_h=0.015,
+    hsv_s=0.7,
     hsv_v=0.4,
     degrees=0.0,
     translate=0.1,
@@ -50,4 +50,4 @@
     fliplr=0.5,
     mosaic=1.0,
     mixup=0.0,
-)
+)

configs/yolov6n_finetune.py

@@ -24,7 +24,7 @@
         strides=[8, 16, 32],
         iou_type='ciou'
     )
-)   
+)
 
 solver = dict(
     optim='SGD',
@@ -35,9 +35,9 @@
     weight_decay=0.00036,
     warmup_epochs=2.0,
     warmup_momentum=0.5,
-    warmup_bias_lr=0.05 
+    warmup_bias_lr=0.05
 )
- 
+
 data_aug = dict(
     hsv_h=0.0138,
     hsv_s=0.664,

configs/yolov6s.py

@@ -35,12 +35,12 @@
     weight_decay=0.0005,
     warmup_epochs=3.0,
     warmup_momentum=0.8,
-    warmup_bias_lr=0.1       
+    warmup_bias_lr=0.1
 )
 
 data_aug = dict(
-    hsv_h=0.015,  
-    hsv_s=0.7, 
+    hsv_h=0.015,
+    hsv_s=0.7,
     hsv_v=0.4,
     degrees=0.0,
     translate=0.1,

configs/yolov6s_finetune.py

@@ -2,7 +2,7 @@
 model = dict(
     type='YOLOv6s',
     pretrained='./weights/yolov6s.pt',
-    depth_multiple=0.33,  
+    depth_multiple=0.33,
     width_multiple=0.50,
     backbone=dict(
         type='EfficientRep',
@@ -24,7 +24,7 @@
         strides=[8, 16, 32],
         iou_type='siou'
     )
-)   
+)
 
 solver = dict(
     optim='SGD',
@@ -35,9 +35,9 @@
     weight_decay=0.00036,
     warmup_epochs=2.0,
     warmup_momentum=0.5,
-    warmup_bias_lr=0.05 
+    warmup_bias_lr=0.05
 )
- 
+
 data_aug = dict(
     hsv_h=0.0138,
     hsv_s=0.664,

data/coco.yaml

@@ -16,4 +16,3 @@ names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', '
          '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' ]
-

deploy/ONNX/README.md

@@ -2,13 +2,13 @@
 
 ### Check requirements
 ```shell
-pip install onnx>=1.10.0                             
+pip install onnx>=1.10.0
 ```
 
 ### Export script
 ```shell
 python deploy/ONNX/export_onnx.py --weights yolov6s.pt --img 640 --batch 1
-                                  
+
 ```
 
 ### Download

deploy/ONNX/export_onnx.py

@@ -3,12 +3,12 @@
 import argparse
 import time
 import sys
-import os  
+import os
 import torch
 import torch.nn as nn
 import onnx
 
-ROOT = os.getcwd()                                
+ROOT = os.getcwd()
 if str(ROOT) not in sys.path:
     sys.path.append(str(ROOT))
 
@@ -22,9 +22,9 @@
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--weights', type=str, default='./yolov6s.pt', help='weights path')  
+    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('--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('--simplify', action='store_true', help='simplify onnx model')
@@ -36,9 +36,9 @@
 
     # Check device
     cuda = args.device != 'cpu' and torch.cuda.is_available()
-    device = torch.device('cuda:0' if cuda else 'cpu') 
+    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    
+    # 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):
@@ -57,7 +57,7 @@
                 m.act = SiLU()
         elif isinstance(m, EffiDeHead):
             m.inplace = args.inplace
-       
+
     y = model(img)  # dry run
 
     # ONNX export

deploy/OpenVINO/README.md

@@ -3,13 +3,13 @@
 ### Check requirements
 ```shell
 pip install --upgrade pip
-pip install openvino-dev                             
+pip install openvino-dev
 ```
 
 ### Export script
 ```shell
 python deploy/OpenVINO/export_openvino.py --weights yolov6s.pt --img 640 --batch 1
-                                  
+
 ```
 
 ### Download

deploy/OpenVINO/export_openvino.py

@@ -3,13 +3,13 @@
 import argparse
 import time
 import sys
-import os  
+import os
 import torch
 import torch.nn as nn
 import onnx
 import subprocess
 
-ROOT = os.getcwd()                                
+ROOT = os.getcwd()
 if str(ROOT) not in sys.path:
     sys.path.append(str(ROOT))
 
@@ -22,9 +22,9 @@
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--weights', type=str, default='./yolov6s.pt', help='weights path')  
+    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('--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')
@@ -35,9 +35,9 @@
 
     # Check device
     cuda = args.device != 'cpu' and torch.cuda.is_available()
-    device = torch.device('cuda:0' if cuda else 'cpu') 
+    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    
+    # 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):
@@ -56,7 +56,7 @@
                 m.act = SiLU()
         elif isinstance(m, EffiDeHead):
             m.inplace = args.inplace
-       
+
     y = model(img)  # dry run
 
     # ONNX export

docs/Test_speed.md

@@ -8,7 +8,7 @@ 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. 
+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.
 
@@ -31,12 +31,11 @@ To  get inference speed with TensorRT in FP16 mode on T4, you can follow the ste
 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] 
+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
 ```
-