GKD

WACV 2024 Paper: Gradient-Guided Knowledge Distillation for Object Detectors

Install MMDetection and grad-cam

• Our codes are based on MMDetection-2.x. Please follow the installation of MMDetection and make sure you can run it successfully.
• This repo uses mmdet==2.25.3 and mmcv-full==1.7.0
• install grad-cam by: pip install grad-cam

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/, use filter_traffic.py to get coco-traffic dataset.
• Unzip KITTI dataset into data/kitti/ (We group classes as 3 and split orignal training set into 8:2 as training and validation sets)

Train

#single GPU
python tools/train.py configs/distillers/gkd/gkd_faster_rcnn_r50_r101_fpn_1x_coco.py

#multi GPU
bash tools/dist_train.sh configs/distillers/gkd/gkd_faster_rcnn_r50_r101_fpn_1x_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/distillers/gkd/gkd_faster_rcnn_r50_r101_fpn_1x_coco.py $new_mmdet_ckpt --eval bbox

#multi GPU
bash tools/dist_test.sh configs/distillers/gkd/gkd_faster_rcnn_r50_r101_fpn_1x_coco.py $new_mmdet_ckpt 8 --eval bbox

Citation

@inproceedings{lan2024gradient,
  title={Gradient-guided knowledge distillation for object detectors},
  author={Lan, Qizhen and Tian, Qing},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={424--433},
  year={2024}
}

Acknowledgement

Our code is based on the project MMDetection.

Thanks to the work FGD and pytorch-grad-cam.