This is the repository associated to the paper :
Sample compression unleashed : New generalization bounds for real valued losses
The PBB and regression forests experiments were run using Python 3.12.2 and a virtual environment as defined in the following file : requirements_pbb.txt. All the experiments on MNIST were run using Python 3.12.3 and a virtual environment as defined in the following file : requirements_p2l.txt.
To run the grid search, use the file main.py. For the baselines, use the files baseline.py and tree_baseline.py. The results were parsed using the jupyter notebooks. In both cases, the configs for the datasets can be found in configs/parameter_configs and the configs for the type of models can be found in configs/sweep_configs.
If you use our code, please cite our paper :
Bazinet, M., Zantedeschi, V., & Germain, P. (2024). Sample compression unleashed: New generalization bounds for real valued losses. arXiv preprint arXiv:2409.17932.
@article{bazinet2024sample,
title={Sample compression unleashed: New generalization bounds for real valued losses},
author={Bazinet, Mathieu and Zantedeschi, Valentina and Germain, Pascal},
journal={arXiv preprint arXiv:2409.17932},
year={2024}
}