SFDA-DE

This is a third-party implementation for CVPR 2022 paper Source-Free Domain Adaptation via Distribution Estimation.

Requirements

• Python 3.7
• scipy
• PyYAML
• easydict
• torch==1.9.1
• torchvision==0.10.1

Please use a conda env and run pip install -r requirements.txt

Dataset

The structure of the dataset should be like

visda2017
|_ category.txt
|_ train
|  |_ aeroplane
|     |_ <im-1-name>.jpg
|     |_ ...
|     |_ <im-N-name>.jpg
|  |_ bicycle
|     |_ <im-1-name>.jpg
|     |_ ...
|     |_ <im-N-name>.jpg
|  |_ ...
|_ validation
|  |_ aeroplane
|     |_ <im-1-name>.jpg
|     |_ ...
|     |_ <im-N-name>.jpg
|  |_ bicycle
|     |_ <im-1-name>.jpg
|     |_ ...
|     |_ <im-N-name>.jpg
|  |_ ...

The "category.txt" contains the names of all the categories, which is like

aeroplane
bicycle
bus
car
...

"category.txt" can be found in ./experiments/dataset/VisDA-2017/category.txt

Remember to change the DATAROOT item in the 3rd line of the cfg file

./experiments/config/VisDA-2017/visda17_train2val_cfg.yaml

Training

On VisDA-2017 dataset, first download the pretrained weight, password: sfda

https://pan.baidu.com/s/1RbLqvBtqJWNtBJN6Eh3Xzg

place the file 'visda_pretrained_r101.pth' in root folder ./, then run:

python train.py --cfg ./experiments/config/VisDA-2017/visda17_train2val_cfg.yaml --exp_name SFDA-DE

The experiment log and checkpoints will be stored at ./experiments/ckpt/*{exp_name}**

Citing

Please cite the paper if you find this code helpful:

@inproceedings{ding2022source,
  title={Source-Free Domain Adaptation via Distribution Estimation},
  author={Ding, Ning and Xu, Yixing and Tang, Yehui and Xu, Chao and Wang, Yunhe and Tao, Dacheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={7212--7222},
  year={2022}
}

Thanks to third party

The code is based on https://github.com/kgl-prml/Contrastive-Adaptation-Network-for-Unsupervised-Domain-Adaptation.