PyTorch版YOLOv4更新了，不僅適用於自定義資料集，還整合了注意力和MobileNet

距離 YOLO v4 的推出，已經過去 5 個多月。YOLO 框架採用 C 語言作為底層程式碼，這對於慣用 Python 的研究者來說，實在是有點不友好。因此網上出現了很多基於各種深度學習框架的 YOLO 復現版本。近日，就有研究者在 GitHub 上更新了基於 PyTorch 的 YOLOv4。

• Nvida GeForce RTX 2080TI

• CUDA10.0

• CUDNN7.0

• windows 或 linux 系統

• python 3.6

• DO-Conv (https://arxiv.org/abs/2006.12030) (torch>=1.2)

• Attention

• fp_16 training

• Mish

• Custom data

• Data Augment (RandomHorizontalFlip, RandomCrop, RandomAffine, Resize)

• Multi-scale Training (320 to 640)

• focal loss

• CIOU

• Label smooth

• Mixup

• cosine lr

pip3 install -r requirements.txt --user

git clone github.com/argusswift/YOLOv4-PyTorch.git

# Download the data.
cd $HOME/data
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
# Extract the data.
tar -xvf VOCtrainval_11-May-2012.tar
tar -xvf VOCtrainval_06-Nov-2007.tar
tar -xvf VOCtest_06-Nov-2007.tar

#step1: download the following data and annotation   
2017 Train images [118K/18GB]   
2017 Val images [5K/1GB]   2017 Test images [41K/6GB]   
2017 Train/Val annotations [241MB]   
#step2: arrange the data to the following structure   
COCO   
---train   
---test   
---val   
---annotations

• 將資料集放入目錄，更新 config/yolov4_config.py 中的 DATA_PATH 引數。

• （對於 COCO 資料集）使用 coco_to_voc.py 將 COCO 資料型別轉換為 VOC 資料型別。

• 轉換資料格式：使用 utils/voc.py 或 utils/coco.py 將 pascal voc *.xml 格式（或 COCO *.json 格式）轉換為 *.txt 格式（Image_path xmin0,ymin0,xmax0,ymax0,class0 xmin1,ymin1,xmax1,ymax1,class1 ...）。

CUDA_VISIBLE_DEVICES=0 nohup python -u train.py  --weight_path weight/yolov4.weights --gpu_id 0 > nohup.log 2>&1 &

CUDA_VISIBLE_DEVICES=0 nohup python -u train.py  --weight_path weight/last.pt --gpu_id 0 > nohup.log 2>&1 &

for VOC dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_voc.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval --mode det
for COCO dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_coco.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval --mode det

for VOC dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_voc.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval -

CUDA_VISIBLE_DEVICES=0 python3 eval_coco.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval --mode val

type=bbox
Running per image evaluation...      DONE (t=0.34s).
Accumulating evaluation results...   DONE (t=0.08s). 
Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.438  
Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.607  
Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.469  
Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.253  
Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.486  
Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.567  
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.342  
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.571  
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.632  
Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.458  
Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.691  
Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.790

for VOC dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_voc.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval
for COCO dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_coco.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval