Variational Autoencoder

Implementation of a Variational Autoencoder model applied to several datasets.

The model is implemented in TensorFlow 2 using its Keras API tf.keras. The primary goal of this project was not to find the most performant ANN architecture for a general VAE or even to find very well-tuned hyperparameters. The primary goal wast to have a small, maintanable, reliable & well-tested code which is easy to extend, and present some simple results to demonstrate that the model is quite able to create new images instead of just reconstructing training instances.

Below you find results for some of the most well-known datasets.

Results

MNIST dataset

Latent dimension 25

The following animation shows how the decoder of the VAE model creates images for a set of 16 points randomly sampled in the latent space held fixed over all epochs.

Generated	Training set

Fashion-MNIST dataset

Latent dimension 100

The following animation shows how the decoder of the VAE model creates images for a set of 16 points randomly sampled in the latent space held fixed over all epochs.

Generated	Training set

Visualization of 10k points of the train set in the latent space

The points from the 100 dimensional latent space have been projected into the plane using PCA.

Visualization of 10k points of the test set in the latent space

The points from the 100 dimensional latent space have been projected into the plane using PCA.

CIFAR-10 dataset

Latent dimension 100

The following animation shows how the decoder of the VAE model creates images for a set of 16 points randomly sampled in the latent space held fixed over all epochs.

Generated	Training set

Visualization of 10k points of the train set in the latent space

The points from the 100 dimensional latent space have been projected into the plane using PCA.

Visualization of 10k points of the test set in the latent space

The points from the 100 dimensional latent space have been projected into the plane using PCA.