This project implements an unsupervised generative modeling technique called Wasserstein Auto-Encoders (WAE), proposed by Tolstikhin, Bousquet, Gelly, Schoelkopf (2017).
wae.py - everything specific to WAE, including encoder-decoder losses, various forms of a distribution matching penalties, and training pipelines
run.py - master script to train a specific model on a selected dataset with specified hyperparameters
Example of output pictures
The following picture shows various characteristics of the WAE-MMD model trained on CelebA after 50 epochs: