W2D

This repo is the official implementation of our CVPR 2022 paper ["The Two Dimensions of Worst-case Training and the Integrated Effect for Out-of-domain Generalization"]().

## Getting Started ### Data Preparation * Downloader datasets (except NICO and CelebA datasets) ``` python3 -m domainbed.scripts.download \ --data_dir=./domainbed/data ``` * Download CelebA dataset from [here](https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) * Download clean NICO dataset (provided by ours) from [here]() * The directory structures are discribed at [OoD-Bench](https://github.com/ynysjtu/ood_bench) ### Install * Pytorch ### Launch a sweep ``` cd /ood_bench/DomainBed bash sweep/"dataset_name"/run.sh launch ../datasets 0 ``` * To change the training setting, modify the scripts under /ood_bench/DomainBed/sweep. * If you have any questions about the scripts, more details are discribed at [OoD-Bench](https://github.com/ynysjtu/ood_bench) and [DomainBed](https://github.com/facebookresearch/DomainBed). * Note: Since ResNet is not used in Colored_MNIST dataset, when you train on Colored_MNIST, uncomment line 992-1020 at algorithms.py. ### View the results ``` python -m domainbed.scripts.collect_results\ --input_dir="sweep_output_path" ``` ## Acknowledgments The codebase is built upon [OoD-Bench](https://github.com/ynysjtu/ood_bench) and [DomainBed](https://github.com/facebookresearch/DomainBed).