vae

VAE

A basic implementation of Variational Auto-Encoders based on the tensorflow tutorial with some small changes. This directory contains the following files:

• model.py: defines the model architecture;
• train_mnist.py: trains the model on the MNIST Handwriten Digits Dataset;
• make_gif.py: transforms saved generated images for a training run into a GIF file such as the one above;
• make_2d_plots.py: plots part of the latent space by decoding points into images. The result is an image as the one below;

Additionally, the generated_samples/ and plots/ directories contain examples from a training run.

Experiment

As a proof of concept, this model was trained for 100 epochs on the MNIST dataset. The results are shown below:

Animation of generated images

The image below depicts the original test set images:

Below is the animation for the training run. Each frame corresponds to a batch of generated images after an epoch.

Latent space visualization

Below is a visualization of the latent space for the trained model.