vanilla_vae

Implementation of variational autoencoder (AEVB) algorithm, using Lasagne framework, as in: [1] arXiv:1312.6114 [stat.ML] (Diederik P Kingma, Max Welling 2013)

A lot of codes is borrowed from https://github.com/Lasagne/Recipes/blob/master/examples/variational_autoencoder/variational_autoencoder.py

Two experiments are performed on MNIST (Binary) and Frey Faces (continuous) datasets.

To run the training:

1. you could change the network architecture in the config.json file
2. run the command python vae_train.py mnist or python vae_train.py faces

After training, model parameters are stored and you could visualize the learned manifold by running python vae_visualize.py mnist or python vae_visualize.py faces

Visualisations of learned data manifold with two-dimensional latent space learned with VAE

Learned MNIST manifold

Learned Frey Faces manifold