open-mmlab / mmyolo

OpenMMLab YOLO series toolbox and benchmark. Implemented RTMDet, RTMDet-Rotated,YOLOv5, YOLOv6, YOLOv7, YOLOv8,YOLOX, PPYOLOE, etc.
https://mmyolo.readthedocs.io/zh_CN/dev/
GNU General Public License v3.0
3k stars 540 forks source link

The RTMDet distillation algorithm cannot reproduce the accuracy on coco dataset #752

Open wintercat1994 opened 1 year ago

wintercat1994 commented 1 year ago

Prerequisite

🐞 Describe the bug

I use RTMDet-s as teacher and RTMDet-tiny as student to reproduce the giving distillation algorithm. But the RTMDet distillation algorithm cannot reproduce the accuracy of coco/bbox_mAP: 0.41 on coco dataset , I can only achieve coco/bbox_mAP: 0.40 , which is 1 percentage points different from the official result. The only modification I made was to change the batch from 32 to 20, what is the reason?

Environment

NVCC: Cuda compilation tools, release 11.0, V11.0.221 GCC: gcc (GCC) 4.8.5 20150623 (Red Hat 4.8.5-44) PyTorch: 1.12.0 PyTorch compiling details: PyTorch built with:

TorchVision: 0.13.0 OpenCV: 4.7.0 MMEngine: 0.7.2 MMCV: 2.0.0 MMDetection: 3.0.0 MMYOLO: 0.5.0+dc85144

Additional information

No response

hhaAndroid commented 1 year ago

@wintercat1994 If you have modified the batch and card numbers, please scale the other hyperparameters accordingly