| Model |
Backbone |
Batch |
Scale |
APval
0.5:0.95
| APval
0.5
| FLOPs
(G)
| Params
(M)
| Weight |
| YOLOv2 |
ResNet-18 |
1xb16 |
640 |
|
|
38.0 |
21.5 |
ckpt |
- For training, we train redesigned YOLOv2 with 150 epochs on COCO. We also gradient accumulate.
- For data augmentation, we only use the large scale jitter (LSJ), no Mosaic or Mixup augmentation.
- For optimizer, we use SGD with momentum 0.937, weight decay 0.0005 and base lr 0.01.
- For learning rate scheduler, we use linear decay scheduler.
Train YOLOv2
Single GPU
Taking training YOLOv2 on COCO as the example,
python train.py --cuda -d coco --root path/to/coco -m yolov2 -bs 16 -size 640 --wp_epoch 3 --max_epoch 150 --eval_epoch 10 --no_aug_epoch 10 --ema --fp16 --multi_scale
Multi GPU
Taking training YOLOv2 on COCO as the example,
python -m torch.distributed.run --nproc_per_node=8 train.py --cuda -dist -d coco --root /data/datasets/ -m yolov2 -bs 128 -size 640 --wp_epoch 3 --max_epoch 150 --eval_epoch 10 --no_aug_epoch 20 --ema --fp16 --sybn --multi_scale --save_folder weights/
Test YOLOv2
Taking testing YOLOv2 on COCO-val as the example,
python test.py --cuda -d coco --root path/to/coco -m yolov2 --weight path/to/yolov2.pth -size 640 -vt 0.3 --show
Evaluate YOLOv2
Taking evaluating YOLOv2 on COCO-val as the example,
python eval.py --cuda -d coco-val --root path/to/coco -m yolov2 --weight path/to/yolov2.pth
Demo
Detect with Image
python demo.py --mode image --path_to_img path/to/image_dirs/ --cuda -m yolov2 --weight path/to/weight -size 640 -vt 0.3 --show
Detect with Video
python demo.py --mode video --path_to_vid path/to/video --cuda -m yolov2 --weight path/to/weight -size 640 -vt 0.3 --show --gif
Detect with Camera
python demo.py --mode camera --cuda -m yolov2 --weight path/to/weight -size 640 -vt 0.3 --show --gif
yjh0410
8ed35eaf32
release