Pay attention to your loss: understanding misconceptions about 1-Lipschitz Neural networks

Béthune Louis, Boissin Thibaut, Serrurier Mathieu, Mamalet Franck, Friedrich Corentin, González-Sanz Alberto

Official repository of the paper "Pay attention to your loss: understanding misconceptions about 1-Lipschitz Neural networks" accepted at Neurips 2022. Arxiv version: https://arxiv.org/abs/2104.05097

To cite us:

@inproceedings{bethune2022pay,
  year = {2022},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  title = {Pay attention to your loss: understanding misconceptions about 1-Lipschitz neural networks},
  author = {Béthune, Louis and Boissin, Thibaut and Serrurier, Mathieu and Mamalet, Franck and Friedrich, Corentin and González-Sanz, Alberto},
}

The repository contains:

• The slides slides.pdf
• The poster poster.pdf
• and the code !

Code

Our code for large scale experiments can be found in .py files. Smaller scale experiments are in notebooks .ipynb: those can be uploaded on Google Colab and have been thought to work out of the box.

Dependencies

The Deel-Lip library is among the dependencies of the the library -- we embed its wheel wheels/deel_lip-1.2.0-py2.py3-none-any.whl. However notice that this library can be found online here.

The code uses custom data loaders and data augmentation pipelines in Deep Learning Toolbox (DLT) -- we embed its wheel wheels/DLT-0.1.0-py2.py3-none-any.whl.

Take note of PEP427: "A wheel is a ZIP-format archive with a specially formatted file name and the .whl extension.".

Experiments in Appendix

Other Pareto front in Appendix can be generated using the pareto-front-experiment.py with appropriate arguments.