- 代码：https://github.com/ikuokuo/start-scaled-yolov4
Scaled-YOLOv4
- 代码: https://github.com/WongKinYiu/ScaledYOLOv4
- 论文: https://arxiv.org/abs/2011.08036
- 文章: https://alexeyab84.medium.com/scaled-yolo-v4-is-the-best-neural-network-for-object-detection-on-ms-coco-dataset-39dfa22fa982
环境准备
基础环境
- Nvidia 显卡的主机
- Ubuntu 18.04
- 系统安装，可见 制作 USB 启动盘，及系统安装
- Nvidia Driver
- 驱动安装，可见 Ubuntu 初始配置 - Nvidia 驱动
开发环境
下载并安装 Anaconda ,之后于 Terminal 执行：
- Anaconda: https://www.anaconda.com/products/individual#Downloads
# 创建 Python 虚拟环境
conda create -n scaled-yolov4 python=3.8 -y
conda activate scaled-yolov4
# 安装 PyTorch with CUDA
conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.2 -c pytorch -y
注意：
pytorch等版本请对照表 torch, torchvision, python- https://github.com/pytorch/vision#installation
cudatoolkit版本请对照表 CUDA Toolkit and Compatible Driver Versions- https://docs.nvidia.com/deploy/cuda-compatibility/index.html#binary-compatibility__table-toolkit-driver
- cuda toolkit: https://developer.nvidia.com/cuda-toolkit-archive
conda下载过慢，可尝试配置国内镜像源
下载 CUDA Toolkit ，其版本也注意对应 Nvidia 驱动版本。下一步需要。命令参考：
代码语言：javascript复制wget https://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run
sudo sh cuda_10.2.89_440.33.01_linux.run
注意：安装时，请手动取消驱动安装选项。
下载 mish-cuda 并安装：
- mish-cuda: https://github.com/JunnYu/mish-cuda
# install mish-cuda, if you use different pytorch version, you could try https://github.com/thomasbrandon/mish-cuda
git clone https://github.com/JunnYu/mish-cuda
cd mish-cuda
python setup.py build install
下载 ScaledYOLOv4-large: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-large
代码语言：javascript复制git clone -b yolov4-large https://github.com/WongKinYiu/ScaledYOLOv4
脚本依赖
代码语言：javascript复制conda activate scaled-yolov4
cd start-scaled-yolov4/
pip install -r scripts/requirements.txt
模型准备
下载官方的 yolov4-p5.pt, yolov4-p6.pt, yolov4-p7.pt 权重文件到 ScaledYOLOv4/weights/ 目录。
- yolov4-p5.pt: https://drive.google.com/file/d/1aXZZE999sHMP1gev60XhNChtHPRMH3Fz/view?usp=sharing
- yolov4-p6.pt: https://drive.google.com/file/d/1aB7May8oPYzBqbgwYSZHuATPXyxh9xnf/view?usp=sharing
- yolov4-p7.pt: https://drive.google.com/file/d/18fGlzgEJTkUEiBG4hW00pyedJKNnYLP3/view?usp=sharing
现有模型测试
准备 COCO 数据集
下载 COCO 数据集 http://cocodataset.org/，
代码语言：javascript复制coco2017
├── annotations
│ ├── instances_train2017.json
│ └── instances_val2017.json
├── test2017
├── train2017
└── val2017
转成 YOLOv5 数据集结构 https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data#3-organize-directories，
代码语言：javascript复制export COCO_DIR=~/datasets/coco2017
export OUTPUT_DIR=~/datasets/coco2017_yolov5
# train2017 训练集
# - 图片：目录软链到 images/
# - 标注：转换存储进 labels/*.txt
# - 物体类型：全部记录进 *.names
# - 图片列表：有物体标注的记录进 *.txt, 无的进 *.txt.ignored
python scripts/coco2yolov5.py
--coco_img_dir $COCO_DIR/train2017/
--coco_ann_file $COCO_DIR/annotations/instances_train2017.json
--output_dir $OUTPUT_DIR
# val2017 验证集
# - 物体类型：依照训练集的记录，保证顺序
python scripts/coco2yolov5.py
--coco_img_dir $COCO_DIR/val2017/
--coco_ann_file $COCO_DIR/annotations/instances_val2017.json
--output_dir $OUTPUT_DIR
--obj_names_file $OUTPUT_DIR/train2017.names
如下：
代码语言：javascript复制coco2017_yolov5/
├── images
│ ├── train2017 -> /home/john/datasets/coco2017/train2017
│ └── val2017 -> /home/john/datasets/coco2017/val2017
├── labels
│ ├── train2017
│ └── val2017
├── train2017.names
├── train2017.txt
├── train2017.txt.ignored
├── val2017.txt
└── val2017.txt.ignored
coco2017_yolov5 软链到 ScaledYOLOv4/ 目录，并添加 ScaledYOLOv4/data/coco2017_yolov5.yaml 文件，描述数据集：
# train and val datasets (image directory or *.txt file with image paths)
train: ./coco2017_yolov5/images/train2017
val: ./coco2017_yolov5/images/val2017
test: ./coco2017_yolov5/images/val2017
# 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']
测试 YOLOv4-P5
代码语言：javascript复制cd ScaledYOLOv4
conda activate scaled-yolov4
pip install opencv-python pyyaml scipy tqdm
python test.py
--img 896
--conf 0.001
--batch 8
--device 0
--data data/coco2017_yolov5.yaml
--weights weights/yolov4-p5.pt
结果如下：
代码语言：javascript复制Fusing layers... Model Summary: 331 layers, 7.07943e 07 parameters, 6.81919e 07 gradients
Scanning labels coco2017_yolov5/labels/val2017.cache (4952 found, 0 missing, 48 empty, 0 duplicate, for 5000 images): 100%|█| 5000/5000 [00:00<00:
Class Images Targets P R mAP@.5 mAP@.5:.95: 100%|█████████████| 625/625 [01:34<00:00, 6.64it/s]
all 5e 03 3.68e 04 0.456 0.76 0.687 0.494
Speed: 14.3/1.5/15.8 ms inference/NMS/total per 896x896 image at batch-size 8
进行推断，
代码语言：javascript复制python detect.py
--img 896
--conf 0.5
--device 0
--weights weights/yolov4-p5.pt
--source demo.jpg
结果如下，
代码语言：javascript复制Fusing layers... Model Summary: 331 layers, 7.07943e 07 parameters, 6.81919e 07 gradients
image 1/1 /home/john/Codes/ScaledYOLOv4/demo.jpg: 768x896 1 cats, 1 dogs, Done. (0.029s)
Results saved to inference/output
Done. (0.133s)
测试 YOLOv4-P7
代码语言：javascript复制python test.py
--img 1536
--conf 0.001
--batch 6
--device 0
--data data/coco2017_yolov5.yaml
--weights weights/yolov4-p7.pt
结果如下：
代码语言：javascript复制Fusing layers... Model Summary: 503 layers, 2.87475e 08 parameters, 2.7862e 08 gradients
Scanning labels coco2017_yolov5/labels/val2017.cache (4952 found, 0 missing, 48 empty, 0 duplicate, for 5000 images): 100%|█| 5000/5000 [00:00<00:
Class Images Targets P R mAP@.5 mAP@.5:.95: 100%|█████████████| 834/834 [06:57<00:00, 2.00it/s]
all 5e 03 3.68e 04 0.435 0.804 0.719 0.531
Speed: 78.2/1.6/79.8 ms inference/NMS/total per 1536x1536 image at batch-size 6
进行推断，
代码语言：javascript复制python detect.py
--img 1536
--conf 0.5
--device 0
--weights weights/yolov4-p7.pt
--source demo.jpg
结果如下，
代码语言：javascript复制Fusing layers... Model Summary: 503 layers, 2.87475e 08 parameters, 2.7862e 08 gradients
image 1/1 /home/john/Codes/ScaledYOLOv4/demo.jpg: 1152x1536 1 cats, 1 dogs, 1 chairs, 1 couchs, 1 potted plants, Done. (0.079s)
Results saved to inference/output
Done. (0.282s)
自定义数据集训练
准备数据集
这里从 COCO 数据集拿出一个子集，作为自定义数据集的演示：
代码语言：javascript复制cat <<EOF > subset.names
cat
dog
EOF
export COCO_DIR=~/datasets/coco2017
export OUTPUT_DIR=~/datasets/coco2017_yolov5_subset
python scripts/coco2yolov5.py
--coco_img_dir $COCO_DIR/train2017/
--coco_ann_file $COCO_DIR/annotations/instances_train2017.json
--output_dir $OUTPUT_DIR
--obj_names_file subset.names
python scripts/coco2yolov5.py
--coco_img_dir $COCO_DIR/val2017/
--coco_ann_file $COCO_DIR/annotations/instances_val2017.json
--output_dir $OUTPUT_DIR
--obj_names_file subset.names
coco2017_yolov5_subset 软链到 ScaledYOLOv4/ 目录，并添加 ScaledYOLOv4/data/coco2017_yolov5_subset.yaml 文件，描述数据集：
# train and val datasets (image directory or *.txt file with image paths)
train: ./coco2017_yolov5_subset/train2017.txt
val: ./coco2017_yolov5_subset/val2017.txt
test: ./coco2017_yolov5_subset/val2017.txt
# number of classes
nc: 2
# class names
names: ['cat', 'dog']
准备参数文件
这里以
YOLOv4-P6为例，P5,P7一样。
复制 ScaledYOLOv4/models/yolov4-p6.yaml 到 ScaledYOLOv4/models/coco2017_yolov5_subset/yolov4-p6.yaml 文件，修改 nc 参数：
nc: 2 # number of classes
训练模型
代码语言：javascript复制conda activate scaled-yolov4
pip install tensorboard
python train.py -h
参数，
代码语言：javascript复制optional arguments:
-h, --help show this help message and exit
--weights WEIGHTS initial weights path
--cfg CFG model.yaml path
--data DATA data.yaml path
--hyp HYP hyperparameters path, i.e. data/hyp.scratch.yaml
--epochs EPOCHS
--batch-size BATCH_SIZE
total batch size for all GPUs
--img-size IMG_SIZE [IMG_SIZE ...]
train,test sizes
--rect rectangular training
--resume [RESUME] resume from given path/last.pt, or most recent run if blank
--nosave only save final checkpoint
--notest only test final epoch
--noautoanchor disable autoanchor check
--evolve evolve hyperparameters
--bucket BUCKET gsutil bucket
--cache-images cache images for faster training
--name NAME renames results.txt to results_name.txt if supplied
--device DEVICE cuda device, i.e. 0 or 0,1,2,3 or cpu
--multi-scale vary img-size /- 50%
--single-cls train as single-class dataset
--adam use torch.optim.Adam() optimizer
--sync-bn use SyncBatchNorm, only available in DDP mode
--local_rank LOCAL_RANK
DDP parameter, do not modify
--logdir LOGDIR logging directory
训练，
代码语言：javascript复制python train.py
--batch-size 2
--img 1280 1280
--data data/coco2017_yolov5_subset.yaml
--cfg models/coco2017_yolov5_subset/yolov4-p6.yaml
--weights ''
--sync-bn
--device 0,1
--name yolov4-p6
--epochs 100
信息如下：
如要恢复训练：
代码语言：javascript复制python train.py
--batch-size 2
--img 1280 1280
--data data/coco2017_yolov5_subset.yaml
--cfg models/coco2017_yolov5_subset/yolov4-p6.yaml
--weights 'runs/exp0_yolov4-p6/weights/last.pt'
--sync-bn
--device 0,1
--name yolov4-p6
--resume
错误 `RuntimeError: main thread is not in main loop`
代码语言：javascript复制Exception ignored in: <function Image.__del__ at 0x7f609bf9bd30>
Traceback (most recent call last):
File "/home/john/anaconda3/envs/scaled-yolov4/lib/python3.8/tkinter/__init__.py", line 4014, in __del__
self.tk.call('image', 'delete', self.name)
RuntimeError: main thread is not in main loop
Tcl_AsyncDelete: async handler deleted by the wrong thread
Aborted (core dumped)
如果发生此错误，可于 train.py __main__ 修改 GUI 的 backend：
if __name__ == '__main__':
import matplotlib.pyplot as plt
plt.switch_backend("agg")
训练指标
训练完成后，内容如下：
代码语言：javascript复制runs/exp0_yolov4-p6/
├── events.out.tfevents.1610070159.john-ubuntu18.17638.0
├── hyp.yaml
├── labels.png
├── opt.yaml
├── results.png
├── results.txt
├── test_batch0_gt.jpg
├── test_batch0_pred.jpg
├── train_batch0.jpg
├── train_batch1.jpg
├── train_batch2.jpg
└── weights
├── best_yolov4-p6.pt
├── best_yolov4-p6_strip.pt
├── last_000.pt
├── last_yolov4-p6.pt
└── last_yolov4-p6_strip.pt
labels.png: 标注分布图results.png: 训练过程图results.txt: 训练过程日志
results.png 要训练完成后才有，如果训练过程中要查看，可用 tensorboard：
$ tensorboard --logdir runs
TensorFlow installation not found - running with reduced feature set.
Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
TensorBoard 2.4.0 at http://localhost:6006/ (Press CTRL C to quit)
打开 http://localhost:6006/ 可见：
测试模型
代码语言：javascript复制python test.py
--img 1280
--conf 0.001
--batch 8
--device 0
--data data/coco2017_yolov5_subset.yaml
--weights runs/exp0_yolov4-p6/weights/best_yolov4-p6_strip.pt
进行推断，
代码语言：javascript复制python detect.py
--img 1280
--conf 0.5
--device 0
--weights runs/exp0_yolov4-p6/weights/best_yolov4-p6_strip.pt
--source demo.jpg
