- 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
# Deformable DETR
cd deformdetr_project
bash tools/dist_train.sh cfg_distill/deformdetr_r101_2x_distill_r50_LayerbyLayer_CL_teachergroup.py 8
# Tansfer the saved distilled model into mmdet model
python pth_transfer.py --ckpt_path $ckpt --output_path $new_mmdet_ckpt
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 |
This repository is an initial draft, we will release more code (AdaMixer, Conditional DETR) soon.
Our code is based on the project MMDetection. Thanks to the work FGD.