This repository contains code demonstrating the Co-learn++ method in our IJCV 2024 paper Source-Free Domain Adaptation Guided by Vision and Vision-Language Pre-Training. This is an extension of the Co-learn method in our ICCV 2023 paper Rethinking the Role of Pre-Trained Networks in Source-Free Domain Adaptation.

Prerequisites:

We used NVIDIA container image for PyTorch, release 22.01, to run experiments.

Install additional libraries with pip install -r requirements.txt.

Dataset:

• Please manually download the datasets Office, Office-Home, VisDA-C, DomainNet from the official websites, and modify the path of images in each '.txt' under the folder ./code/data/. Scripts to generate the txt files are in the respective data folders.

Training:

• Training scripts in ./code/uda/scripts. Run eval_target_zeroshot.sh for zero-shot CLIP and train_target_two_branch.sh for co-learning with CLIP encoder.
• Results consolidation scripts in ./code/uda/consolidation_scripts.

Citation

@article{zhang2024colearnplus,
    author = {Zhang, Wenyu and Shen, Li and Foo, Chuan-Sheng},
    year = {2024},
    month = {08},
    pages = {1-23},
    title = {Source-Free Domain Adaptation Guided by Vision and Vision-Language Pre-training},
    journal = {International Journal of Computer Vision},
    doi = {10.1007/s11263-024-02215-3}
}

@inproceedings{zhang2023colearn,
    author = {Zhang, Wenyu and Shen, Li and Foo, Chuan-Sheng},
    booktitle = {2023 IEEE/CVF International Conference on Computer Vision (ICCV)},
    title = {Rethinking the Role of Pre-Trained Networks in Source-Free Domain Adaptation},
    year = {2023},
    volume = {},
    issn = {},
    pages = {18795-18805},
    doi = {10.1109/ICCV51070.2023.01727},
    url = {https://doi.ieeecomputersociety.org/10.1109/ICCV51070.2023.01727},
    publisher = {IEEE Computer Society},
    address = {Los Alamitos, CA, USA},
    month = {oct}
}

Acknowledgements

Our implementation is based off SHOT++. Thanks to the SHOT++ implementation.