Awesome

IRMCon-IPW

This repository is the PyTorch implementation for ECCV 2022 Paper "Class Is Invariant to Context and Vice Versa: On Learning Invariance for Out-Of-Distribution Generalization", For the detailed theories, please refer to our paper. If you have any questions or suggestions, please email me, (I do not usually browse my Github, so the reply to issues may be not on time).

Note that this repository is based on the LfF and DomainBed. (Note, the data processing in Lff should be checked when you use their code, the input range seems abnormal.)

If you find this work is useful in your research, please kindly consider citing:

@inproceedings{qi2022class,
  title={Class Is Invariant to Context and Vice Versa: On Learning Invariance for Out-Of-Distribution Generalization},
  author={Qi, Jiaxin and Tang, Kaihua and Sun, Qianru and Hua, Xian-Sheng and Zhang, Hanwang},
  booktitle={ECCV},
  year={2022}
}

Dependencies

python 3.9.4, pytorch 1.7.1, torchvision 0.8.2

Preparing

Download biased data from here and unzip it under the path ./Biased_dataset/data (Note the result should be ""./Biased_dataset/data/cmnist/...")

Training (examples)

1.Biased dataset, Check your download data path and the set data path in the code.

1.1.Training for Colored MNIST

ERM baseline: python train_cmnist_erm.py --dir_name bias0.05

Ours: python train_cmnist_ours.py --dir_name bias0.05

1.2.Training for Corrupted Cifar-10

ERM baseline: python train_ccifar10_erm.py --dir_name bias0.05

Ours: python train_ccifar10_ours.py --dir_name bias0.05

1.3.Training for BAR

ERM baseline: python train_bar_erm.py --ratio 0.05

Ours: python train_bar_ours.py --ratio 0.05

2.Training for PACS (codebase is from DomainBed, find the full version from here) (The differences are in dataset (we need augmented images), dataloader, and settings (like no pretraining), we need 16G card)

2.1.download data(PACS) from DomainBed and put into ./data

2.2.run baseline: python train_origin.py --test_envs 0

2.3.run ours: python train_ours.py --test_envs 0

Acknowledgements

Thanks for the source code from LfF and DomainBed.