Beyond Invariance: Test-Time Label-Shift Adaptation for Distributions with "Spurious" Correlations

Available: https://arxiv.org/abs/2211.15646

To cite our paper

@misc{sun2022invariance,
      title={Beyond Invariance: Test-Time Label-Shift Adaptation for Distributions with "Spurious" Correlations}, 
      author={Qingyao Sun and Kevin Murphy and Sayna Ebrahimi and Alexander D'Amour},
      year={2022},
      eprint={2211.15646},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
}

PyTorch Implementation (WIP)

We have a work-in-progress PyTorch implementation available at https://github.com/nalzok/BalancingGroups .

How to Reproduce

1. Install Pipenv and Pyenv.
2. Install dependence with PIP_FIND_LINKS=https://storage.googleapis.com/jax-releases/libtpu_releases.html pipenv install --deploy.
3. Run experiments with make paper-mnist, make paper-chexpert-embedding, make paper-chexpert-pixel, and make tree.
4. Aggregate experimental results and generate figures with make merge.