OpenDG: Open Domain Generalization

Code release for Open Domain Generalization with Domain-Augmented Meta-Learning (CVPR2021)

Datasets

• Here are some datasets you may need. PACS and Office-Home for the open domain generalization experiments on them. Office-31, STL-10, Visda2017 and DomainNet for the Multi-Datasets experiment.

Dataset	Link
PACS	https://dali-dl.github.io/project_iccv2017.html
Office-Home	https://www.hemanthdv.org/officeHomeDataset.html
Office-31	http://www.eecs.berkeley.edu/~mfritz/domainadaptation/
STL-10	https://cs.stanford.edu/~acoates/stl10/
Visda2017	http://ai.bu.edu/visda-2017/
DomainNet	http://ai.bu.edu/M3SDA/

• We provide the labels and train-val-test splits for these datasets in data/ folder.

Requirements

• Python 3.8
• PyTorch 1.5.0

Quick Start

• Download the DATASET you need. Move the image_list folder of the DATASET (which we provide in data/DATASET/) to the directory of the DATASET.
• We provide scripts in src/scripts/. Complete the configuration of experiments, such as the path to the DATASET, then bash run_train.sh for training on source domains and testing on target domain data from known classes.
• After training and saving the model checkpoints, bash run_validate.sh for testing on the whole target domain, including both known and unknown classes.

Citation

If you find this code or our paper useful, please consider citing:

@inproceedings{shu2021open,
  title={Open Domain Generalization with Domain-Augmented Meta-Learning},
  author={Shu, Yang and Cao, Zhangjie and Wang, Chenyu and Wang, Jianmin and Long, Mingsheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9624--9633},
  year={2021}
}

Contact

If you have any problems about our code, feel free to contact

• shu-y18@mails.tsinghua.edu.cn