1-Lipschitz Signed Distance Function

Learn to approximate the Sign Distance Function (SDF) to the boundary of a distribution (e.g a dataset or a point cloud).

Use a combination of Lipschitz networks, adversarial training and Hinge Kantorovich Rubinstein loss (HKR).

Structure of the repository

The repository is organized as follow:

• run_*.ipynb notebooks launchs predefined configurations, datasets, create model, train it, and log the results.
  • run_toy2d.ipynb for toy examples in 2D (see figure above).
  • run_tabular.ipynb for tabular data (Thyroid, Mammography, etc) in anomaly detection.
  • run_mnist.ipynb for simple images from Mnist.
  • run_fashion_mnist.ipynb for harder task on Fashion Mnist.
  • run_sdf.ipynb for implicit surface parametrization in 3D.
  • run_cat_dogs.ipynb for challenging experiments on the high dimensional cats versus dogs dataset.
• ocml/: contains all source files.
• experiments/: contains the scripts to launch several experiments sequentially and upload the results to a wandb account. You should login on Wandb before running the scripts.
• legacy_notebooks/: old notebooks for early experiments and prototypes. Saved for reproducibility and archiving. Should be avoided for new experiments.

Remarks

Wandb is used experiment tracking, plotly and seaborn are used for plotting. Latest version of deel-lip is recommanded.

Run the code

The following directories will be populated:

• images/: record images produced for uploading to wandb.
• weights/: contain weights of the network architecture in .h5 format.
• wandb/: if wandb is used - to store local variables.