FGD
CVPR 2022 Paper: Focal and Global Knowledge Distillation for Detectors
Install MMDetection and MS COCO2017
- Our codes are based on MMDetection. Please follow the installation of MMDetection and make sure you can run it successfully.
- This repo uses mmdet==2.11.0 and mmcv-full==1.2.4
- If you want to use higher mmdet version, you may have to change the optimizer in apis/train.py and build_detector in tools/train.py.
- For mmdet>=2.12.0, if you want to use inheriting strategy, you have to initalize the student with teacher's parameters after model.init_weights().
Add and Replace the codes
- Add the configs/. in our codes to the configs/ in mmdetectin's codes.
- Add the mmdet/distillation/. in our codes to the mmdet/ in mmdetectin's codes.
- Replace the mmdet/apis/train.py and tools/train.py in mmdetection's codes with mmdet/apis/train.py and tools/train.py in our codes.
- Add pth_transfer.py to mmdetection's codes.
- Unzip COCO dataset into data/coco/
Train
#single GPU
python tools/train.py configs/distillers/fgd/fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py
#multi GPU
bash tools/dist_train.sh configs/distillers/fgd/fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py 8
Transfer
# Tansfer the FGD model into mmdet model
python pth_transfer.py --fgd_path $fgd_ckpt --output_path $new_mmdet_ckpt
Test
#single GPU
python tools/test.py configs/retinanet/retinanet_r50_fpn_2x_coco.py $new_mmdet_ckpt --eval bbox
#multi GPU
bash tools/dist_test.sh configs/retinanet/retinanet_r50_fpn_2x_coco.py $new_mmdet_ckpt 8 --eval bbox
Results
| Model | Backbone | Baseline(mAP) | +FGD(mAP) | config | weight | code |
|---|---|---|---|---|---|---|
| RetinaNet | ResNet-50 | 37.4 | 40.7 | config | baidu | wsfw |
| RetinaNet | ResNet-101 | 38.9 | 41.7 | config | ||
| Faster RCNN | ResNet-50 | 38.4 | 42.0 | config | baidu | dgpf |
| Faster RCNN | ResNet-101 | 39.8 | 44.1 | config | ||
| RepPoints | ResNet-50 | 38.6 | 42.0 | config | baidu | qx5d |
| RepPoints | ResNet-101 | 40.5 | 43.8 | config | ||
| FCOS | ResNet-50 | 38.5 | 42.7 | config | baidu | sedt |
| MaskRCNN | ResNet-50 | 39.2 | 42.1 | config | baidu | sv8m |
| GFL | ResNet-50 | 40.2 | 43.5 | config |
| Model | Backbone | Baseline(Mask mAP) | +FGD(Mask mAP) | config | weight | code |
|---|---|---|---|---|---|---|
| SOLO | ResNet-50 | 33.1 | 36.0 | config | ||
| MaskRCNN | ResNet-50 | 35.4 | 37.8 | config | baidu | sv8m |
| Student | Teacher | Baseline(mAP) | +FGD(mAP) | config | weight | code |
|---|---|---|---|---|---|---|
| YOLOX-m | YOLOX-l | 45.9 | 46.6 | config | baidu | af9g |
- Please refer branch yolox
Citation
@article{yang2021focal,
title={Focal and Global Knowledge Distillation for Detectors},
author={Yang, Zhendong and Li, Zhe and Jiang, Xiaohu and Gong, Yuan and Yuan, Zehuan and Zhao, Danpei and Yuan, Chun},
journal={arXiv preprint arXiv:2111.11837},
year={2021}
}
Acknowledgement
Our code is based on the project MMDetection.
Thanks to the work GCNet and mmetection-distiller.