Some Fundamental Aspects about Lipschitz Continuity of Neural Networks

Code for the paper "Some Fundamental Aspects about Lipschitz Continuity of Neural Networks", accepted for ICLR 2024.

Paper link: OpenReview, arXiv.

Requirements

• Mandatory:
  • For package versioning pipenv is required (regardless of the installation).
• Optional:
  • Python version specified in .python-version is controlled by pyenv. Installing other python versions could be done using other methods.

To install the required packages, run pipenv install. You can also manually inspect the Pipfile and decide what to install.

Most important files

• code/lipschitz.py - contains all Lipschitz constant estimates;
• code/visual_example.ipynb - contains the code for the intuition example for the fidelity of the Lipschitz Lower bound;
• code/train.py - contains training code.