This repository contains the code to reproduce the experiments of Statistical and Topological Properties of Sliced Probability Divergences, accepted as a spotlight presentation at NeurIPS 2020. Please cite our paper if you use any of our code.
Joblib, Matplotlib, Multiprocessing, Numpy, POT, PyTorch
compare_dist.py: Illustration of the topological result in Theorem 2. Run it to reproduce Figure 1.study_complexity.py: Illustration of the sample and projection complexity results for Sliced-Wasserstein and Sliced-Sinkhorn, on the synthetical setting. Run it to reproduce Figures 2 and 3 (main doc) and Figures S1 and S2 (supplementary doc).real_data_exp.py: Two-sample testing problem (for data integration) on MNIST and CIFAR-10. Run it to reproduce Figure 4.sinkhorn_pointcloud.py: Contains the function that computes the optimal transport regularized cost and Sinkhorn divergences.utils.py: Contains the functions that compute different divergences and their sliced versions (Wasserstein, Sinkhorn, MMD).