SLL Lipschitz Initialization

This repository is the code for the paper 1-Lipschitz Network Initialization for Certifiably Robust Classification Applications: A Decay Problem.

To cite please use,

@misc{juston20251lipschitznetworkinitializationcertifiably,
      title={1-Lipschitz Network Initialization for Certifiably Robust Classification Applications: A Decay Problem}, 
      author={Marius F. R. Juston and William R. Norris and Dustin Nottage and Ahmet Soylemezoglu},
      year={2025},
      eprint={2503.00240},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.00240}, 
}

Paper figures

To generate the figures in the paper, run for generating the distribution of the intialization properties

python src/distribution/distribution.py

to run the training of the networks run

python src/train.py

and to generate the plots you would do,

python src/plot.py --runs_dir ../runs_lipschitz --outdir figs/tb_plots_lipschitz --compare_runs_dir runs --compare_outdir figs/tb_plots