python version 3.8, torch 1.8.1, torchvision 0.9.1:
pip install torch==1.8.1 torchvision==0.9.1
数据文件夹结构如下:
datasets/
JPEGImages/ # VOC2007 + VOC2012 all images
img1.jpg
img2.jpg
.
.
.
train.txt
val.txt
python train.py --input_size 448 448 --batch_size 32 --epochs 80 --nb_classes 20 --finetune ./weights/resnet50_ram-a26f946b.pth --data_path ./datasets/ --output_dir ./output_dir
python eval.py --input_size 448 448 --weights ./output_dir/best.pth --data_path ./datasets/ --nb_classes 20
python predict.py --input_size 448 448 --weights ./output_dir/best.pth --image_path ./person.jpg --nb_classes 20
python export_onnx.py --input_size 448 448 --weights ./output_dir/best.pth --nb_classes 20
python -m onnxsim best.onnx best_sim.onnx
model | backbone | mAP (VOC2007 test) |
---|---|---|
YOLOv1 | ResNet50 | 0.65 |