Scaled-YOLOv4 快速開始，訓練自定義資料集
阿新 • • 發佈：2021-01-09
Scaled-YOLOv4
- 程式碼: https://github.com/WongKinYiu/ScaledYOLOv4
- 論文: 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
環境準備
基礎環境
- Nvidia 顯示卡的主機
- Ubuntu 18.04
- 系統安裝，可見 製作 USB 啟動盤，及系統安裝
- 系統安裝，可見
- Nvidia Driver
- 驅動安裝，可見 Ubuntu 初始配置 - Nvidia 驅動
開發環境
下載並安裝 Anaconda ,之後於 Terminal 執行：
# 建立 Python 虛擬環境
conda create -n scaled-yolov4 python=3.8 -y
conda activate scaled-yolov4
# 安裝 PyTorch with CUDA
conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.2 -c pytorch -y
注意：
pytorch等版本請對照表 torch, torchvision, pythoncudatoolkit版本請對照表 CUDA Toolkit and Compatible Driver Versionsconda下載過慢，可嘗試配置國內映象源
下載 CUDA Toolkit ，其版本也注意對應 Nvidia 驅動版本。下一步需要。命令參考：
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：安裝時，請手動取消驅動安裝選項。
下載 mish-cuda 並安裝：
# 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
git clone -b yolov4-large https://github.com/WongKinYiu/ScaledYOLOv4
指令碼依賴
conda activate scaled-yolov4
cd start-scaled-yolov4/
pip install -r scripts/requirements.txt
模型準備
下載官方的 yolov4-p5.pt, yolov4-p6.pt, yolov4-p7.pt 權重檔案到 ScaledYOLOv4/weights/ 目錄。
現有模型測試
準備 COCO 資料集
下載 COCO 資料集，
coco2017
├── annotations
│ ├── instances_train2017.json
│ └── instances_val2017.json
├── test2017
├── train2017
└── val2017
轉成 YOLOv5 資料集結構，
export COCO_DIR=~/datasets/coco2017
export OUTPUT_DIR=~/datasets/coco2017_yolov5
# train2017 訓練集
# - 圖片：目錄軟鏈到 images/
# - 標註：轉換儲存進 labels/*.txt
# - 物體型別：全部記錄進 *.names
# - 圖片列表：有物體標註的記錄進 *.txt, 無的進 *.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 驗證集
# - 物體型別：依照訓練集的記錄，保證順序
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：
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 軟鏈到 ScaledYOLOv4/ 目錄，並新增 ScaledYOLOv4/data/coco2017_yolov5.yaml 檔案，描述資料集：
# 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']
測試 YOLOv4-P5
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
結果如下：
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 [email protected] [email protected]:.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
進行推斷，
python detect.py \
--img 896 \
--conf 0.5 \
--device 0 \
--weights weights/yolov4-p5.pt \
--source demo.jpg
結果如下，
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)
測試 YOLOv4-P7
python test.py \
--img 1536 \
--conf 0.001 \
--batch 6 \
--device 0 \
--data data/coco2017_yolov5.yaml \
--weights weights/yolov4-p7.pt
結果如下：
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 [email protected] [email protected]:.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
進行推斷，
python detect.py \
--img 1536 \
--conf 0.5 \
--device 0 \
--weights weights/yolov4-p7.pt \
--source demo.jpg
結果如下，
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)
自定義資料集訓練
準備資料集
這裡從 COCO 資料集拿出一個子集，作為自定義資料集的演示：
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 軟鏈到 ScaledYOLOv4/ 目錄，並新增 ScaledYOLOv4/data/coco2017_yolov5_subset.yaml 檔案，描述資料集：
# 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']
準備引數檔案
這裡以
YOLOv4-P6為例，P5,P7一樣。
複製 ScaledYOLOv4/models/yolov4-p6.yaml 到 ScaledYOLOv4/models/coco2017_yolov5_subset/yolov4-p6.yaml 檔案，修改 nc 引數：
nc: 2 # number of classes
訓練模型
conda activate scaled-yolov4
pip install tensorboard
python train.py -h
引數，
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
訓練，
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
資訊如下：
如要恢復訓練：
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
錯誤 RuntimeError: main thread is not in main loop
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)
如果發生此錯誤，可於 train.py __main__ 修改 GUI 的 backend：
if __name__ == '__main__':
import matplotlib.pyplot as plt
plt.switch_backend("agg")
訓練指標
訓練完成後，內容如下：
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: 標註分佈圖results.png: 訓練過程圖results.txt: 訓練過程日誌
results.png 要訓練完成後才有，如果訓練過程中要檢視，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)
開啟 http://localhost:6006/ 可見：
測試模型
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
進行推斷，
python detect.py \
--img 1280 \
--conf 0.5 \
--device 0 \
--weights runs/exp0_yolov4-p6/weights/best_yolov4-p6_strip.pt
--source demo.jpg
GoCoding 個人實踐的經驗分享，可關注公眾號！