This repository contains the implementation of our paper: "Sliced-Wasserstein Autoencoder: An Embarrassingly Simple Generative Model" using Keras and Tensorflow. The proposed method ameliorates the need for adversarial networks in training generative models, and it provides a stable optimization while having a very simple implementation.

A PyTorch implementation of the SWAE algorithm was kindly provided by Emmanuel Fuentes.

This notebook trains the SWAE on the MNIST dataset with a uniform distribution in the embedding space. The figure below visualizes the embedded data and the embedding space for the MNIST dataset:

Similarly, this notebook trains the SWAE on the MNIST dataset with a disk distribution in the embedding space. The figure below visualizes the embedded data and the embedding space for the MNIST dataset:

Similarly, this notebook trains the SWAE on the MNIST dataset with a ring distribution in the embedding space. The figure below visualizes the embedded data and the embedding space for the MNIST dataset:

The pretrained SWAE modules are also uploaded:

• LearnedModels/MNIST_uniform(circle)(ring)_autoencoder.h5
• LearnedModels/MNIST_uniform(circle)(ring)_encoder.h5
• LearnedModels/MNIST_uniform(circle)(ring)_decoder.h5