thomassutter / mvhg

Python library for the differentiable hypergeometric distribution

Learning Group Importance using the Differentiable Hypergeometric Distribution

Python library for the differentiable hypergeometric distribution

This is the official code for the ICLR 2023 Paper (Spotlight) "Learning Group Importance using the Differentiable Hypergeometric Distribution".

Link to Openreview

Link to Arxiv

We are still working on the code and the repository. Feedback and requests are very welcome.

How to get started

We provide an environment file env_mvhg.yml that helps you with setting up a conda environment. For help on how to install conda, please follow the guidelines on the offical webiste (link to the offical website)

To create the conda environment needed, please run the following command

conda env create -f env_mvhg.yml
conda activate mvhg
pip install "[.pt]"

The conda environment runs on python 3.8.

Minimal Example

We provide a minimal example, which learn the class weights from samples. The minimal example uses pytorch lightning, weights & biases, and hydra config.

In the root directory, run the following command

python main_minimal_app.py

Citation

If you use our model in your own, please cite us using the following citation

@inproceedings{sutter2023,
  title={Learning Group Importance using the Differentiable Hypergeometric Distribution},
  author={Sutter, Thomas M and Manduchi, Laura and Ryser, Alain and Vogt, Julia E},
  year = {2023},
  booktitle = {International Conference on Learning Representations},
}

Questions, Requests, Feedback

For any questions or requests, please reach out to: Thomas Sutter ([e-mail])