yifei-he / GOAT

Generative Gradual Domain Adaptation with Optimal Transport (GOAT)

This is the official implementation for the algorithm Gradual DOmain Adaptation with Optimal Transport (GOAT) in the paper "Gradual Domain Adaptation: Theory and Algorithms.". The algorithm design is motivated by our previous work, "Understanding gradual domain adaptation: Improved analysis, optimal path and beyond", published in ICML 2022.

Install the repo

git clone https://github.com/yifei-he/GOAT.git
cd GOAT
pip install -r requirements.txt

Prepare Data

The covertype dataset can be downloaded from: https://archive.ics.uci.edu/dataset/31/covertype.

The portraits dataset can be downloaded from here. We follow the same data preprocessing procedure from https://github.com/p-lambda/gradual_domain_adaptation. Namely after downloading, extract the tar file, and copy the "M" and "F" folders inside a folder called dataset_32x32 inside the current folder. Then run "python create_dataset.py".

Run Experiment

To run experiments, follow the following syntax.

python experiments.py --dataset color_mnist --gt-domains 1 --generated-domains 2

Here, dataset can be selected from [mnist, portraits, covtype, color_mnist]; gt-domains and generated-domains are the number of given ground-truth intermediate domains (only available for the two MNSIT datasets) and domains generated by GOAT respectively, both default to be 0.

Citation

@misc{he2023gradual,
      title={Gradual Domain Adaptation: Theory and Algorithms}, 
      author={Yifei He and Haoxiang Wang and Bo Li and Han Zhao},
      year={2023},
      eprint={2310.13852},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

@inproceedings{wang2022understanding,
  title={Understanding gradual domain adaptation: Improved analysis, optimal path and beyond},
  author={Wang, Haoxiang and Li, Bo and Zhao, Han},
  booktitle={International Conference on Machine Learning},
  pages={22784--22801},
  year={2022},
  organization={PMLR}
}