Table Notes

• Results of the mAP and speed are evaluated on COCO val2017 dataset with the input resolution of 640×640.
• Speed is tested with TensorRT 7.2 and TensorRT 8.5 on T4 GPU.

• python requirements

  pip install -r requirements.txt

• data:

  prepare COCO dataset, YOLO format coco labels and specify dataset paths in data.yaml

• Step 1: Training a base model

  Be sure to open use_dfl mode in config file (use_dfl=True, reg_max=16)

  python -m torch.distributed.launch --nproc_per_node 8 tools/train.py \
  									--batch 128 \
  									--conf configs/gold_yolo-n.py \
  									--data data/coco.yaml \
  									--epoch 300 \
  									--fuse_ab \
  									--use_syncbn \
  									--device 0,1,2,3,4,5,6,7 \
  									--name gold_yolo-n

• Step 2: Self-distillation training

  Be sure to open use_dfl mode in config file (use_dfl=True, reg_max=16)

  python -m torch.distributed.launch --nproc_per_node 8 tools/train.py \
  									--batch 128 \
  									--conf configs/gold_yolo-n.py \
  									--data data/coco.yaml \
  									--epoch 300 \
  									--device 0,1,2,3,4,5,6,7 \
  									--use_syncbn \
  									--distill \
  									--teacher_model_path runs/train/gold_yolo_n/weights/best_ckpt.pt \
  									--name gold_yolo-n

• Step 1: Training a base model

  Be sure to open use_dfl mode in config file (use_dfl=True, reg_max=16)

  python -m torch.distributed.launch --nproc_per_node 8 tools/train.py \
  									--batch 256 \
  									--conf configs/gold_yolo-s.py \ # gold_yolo-m/gold_yolo-l
  									--data data/coco.yaml \
  									--epoch 300 \
  									--fuse_ab \
  									--use_syncbn \
  									--device 0,1,2,3,4,5,6,7 \
  									--name gold_yolo-s # gold_yolo-m/gold_yolo-l

• Step 2: Self-distillation training

  Be sure to open use_dfl mode in config file (use_dfl=True, reg_max=16)

  python -m torch.distributed.launch --nproc_per_node 8 tools/train.py \
  									--batch 256 \ # 128 for distillation of gold_yolo-l
  									--conf configs/gold_yolo-s.py \ # gold_yolo-m/gold_yolo-l
  									--data data/coco.yaml \
  									--epoch 300 \
  									--device 0,1,2,3,4,5,6,7 \
  									--use_syncbn \
  									--distill \
  									--teacher_model_path runs/train/gold_yolo-s/weights/best_ckpt.pt \
  									--name gold_yolo-s # gold_yolo-m/gold_yolo-l

python tools/eval.py --data data/coco.yaml --batch 32 --weights weights/Gold_s_pre_dist.pt --task val --reproduce_640_eval

Please refer to Test speed

The implementation is based on YOLOv6, and some implementations borrowed from Topformer. Thanks for their open source code.