Localized Adversarial Domain Generalization

Installation

Our code is developed based on the wilds benchmark. Please follow the instruction to install wilds.

Requirements: Pytorch, torchvision, tqdm, wandb, wilds=1.2.2.

You need a GPU to run the code, and results will be logged with wandb.

Data Preparation

Please follow wilds benchmark to download the required datasets to ../../data/wild/, e.g. ../../data/wild/camelyon17_v1.0.

Experiments

We provide the script for Camelyon17 and Povertymap.

1. Conduct experiments on Camelyon17 by running

 python examples/run_expt_camelyon17.py

Please change the random_seed to reproduce our results.

2. Conduct experiments on Poverty by running

python examples/run_exp_poverty.py  

Please change fold to reproduce our results.

Thanks for your interests.

Citation

@inproceedings{zhu2022localized,
  title={Localized Adversarial Domain Generalization},
  author={Zhu, Wei and Lu, Le and Xiao, Jing and Han, Mei and Luo, Jiebo and Harrison, Adam P},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={7108--7118},
  year={2022}
}