shuowang666 / detrdistill

[ICCV2023] DETRDistill: A Universal Knowledge Distillation Framework for DETR-families

DETRDistill:A Universal Knowledge Distillation Framework for DETR-families

A Universal Knowledge Distillation Framework for DETR-families

Jiahao Chang*, Shuo Wang*, Hai-Ming Xu*, Zehui Chen, Chenhongyi Yang, Feng Zhao.

[Paper]

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.

pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install mmcv-full==1.4.0
cd our project
pip install -r requirements/build.txt
pip install -v -e .

• Unzip COCO dataset into data/coco/
• Modify pretraining R-101/R-50 and our teacher model weight path

Train

# Deformable DETR
cd deformdetr_project
bash tools/dist_train.sh  cfg_distill/deformdetr_r101_2x_distill_r50_LayerbyLayer_CL_teachergroup.py  8

Transfer

# Tansfer the saved distilled model into mmdet model
python pth_transfer.py --ckpt_path $ckpt --output_path $new_mmdet_ckpt

Performance

MS COCO Val set

Model	Backbone	mAP	config	weight	log	mAP & weight & log (reprod.)
Deformable DETR	ResNet-101	44.8	config	baidu	baidu
Deformable DETR	ResNet-50	44.1	config	baidu	baidu
Deformable DETR-Distill	ResNet-50	46.6(+2.5)	config		baidu	46.5(+2.4) baidu baidu
Deformable DETR	ResNet-18	40.0	config	baidu	baidu
Deformable DETR-Distill	ResNet-18	43.3(+3.3)	config	baidu	baidu

NOTE

This repository is an initial draft, we will release more code (AdaMixer, Conditional DETR) soon.

Acknowledgement

Our code is based on the project MMDetection. Thanks to the work FGD.