Unsupervised Multi-source Domain Adaptation Without Access to Source Data (CVPR '21 Oral)

This repository is a PyTorch implementation of the paper Unsupervised Multi-source Domain Adaptation Without Access to Source Data published at CVPR 2021. This code is based on the SHOT repository.

Create a conda environment with environment.yml.

• Manually download the datasets Office, Office-Home, Office-Caltech from the official websites.
• Move gen_list.py inside data directory.
• Generate '.txt' file for each dataset using gen_list.py (change dataset argument in the file accordingly).

• Train source models (shown here for Office with source A)

python train_source.py --dset office --s 0 --max_epoch 100 --trte val --gpu_id 0 --output ckps/source/

• Adapt to target (shown here for Office with target D)

python adapt_multi.py --dset office --t 1 --max_epoch 15 --gpu_id 0 --output_src ckps/source/ --output ckps/adapt

• Distill to single target model (shown here for Office with target D)

python distill.py --dset office --t 1 --max_epoch 15 --gpu_id 0 --output_src ckps/adapt --output ckps/dist

If you use this code in your research please consider citing

@article{ahmed2021unsupervised,
  title={Unsupervised Multi-source Domain Adaptation Without Access to Source Data},
  author={Ahmed, Sk Miraj and Raychaudhuri, Dripta S and Paul, Sujoy and Oymak, Samet and Roy-Chowdhury, Amit K},
  journal={arXiv preprint arXiv:2104.01845},
  year={2021}
}