YuheD / ProxyMix

NN 2023

Official Implementation for ProxyMix

ProxyMix: Proxy-based Mixup Training with Label Refinery for Source-Free Domain Adaptation

This implementation is based on ATDOC.

Framework:

1. train on the source domain;
2. Construct the proxy source domain and train on target dataset.

Prerequisites:

• python == 3.6.8
• pytorch ==1.1.0
• torchvision == 0.3.0
• numpy, scipy, sklearn, PIL, argparse, tqdm

Dataset:

• Please manually download the datasets Office, Office-Home, VisDA-C from the official websites, and modify the path of images in each '.txt' under the folder './data/'. [**How to generate such txt files could be found in https://github.com/tim-learn/Generate_list **]

Training:

1. Office-31 dataset

   # train source model
   python train_source.py --dset office --s 0 --max_epoch 50
   # train target model
   python train_target.py --dset office --easynum 5 --output test --gpu_id 7 --s 0 --t 1
   python train_target.py --dset office --easynum 5 --output test --gpu_id 7 --s 0 --t 2

2. Office-Home dataset

   # train source model
   python train_source.py --dset office-home --s 0 --max_epoch 50
   # train target model
   python train_target.py --dset office-home --easynum 10 --output test --gpu_id 7 --s 0 --t 1 --max_epoch 50

3. VISDA-C dataset

   # train source model
   python train_source.py --dset VISDA-C --s 0 --max_epoch 5
   # train target model
   python train_target.py --dset VISDA-C --easynum 50 --output test --gpu_id 7 --s 0 --t 1 --max_epoch 1

Citation

If you find this code useful for your research, please cite our papers

@article{ding2022proxymix,
  title={ProxyMix: Proxy-based Mixup Training with Label Refinery for Source-Free Domain Adaptation},
  author={Ding, Yuhe and Sheng, Lijun and Liang, Jian and Zheng, Aihua and He, Ran},
  journal={arXiv preprint arXiv:2205.14566},
  year={2022}
}