This is a Variational Autoencoder (VAE) implementation using Tensorflow on Python. It uses of convolutional layers and fully connected layers in encoder and decoder. The loading functions are designed to work with CIFAR-10 dataset. New loading functions need to be written to handle other datasets. In deep learning literature, it is reported that VAEs produce blurry images when trained on CIFAR-10 dataset. We have written this code to experiment on ways to generate realistic-looking images.
Thanks: While writing this code, we got inspired by J. H. Metzen's implementation, and got help from the data loading scripts distributed in UC Berkeley's CS 294-129 and Stanford's CS231n courses.
Authors: Orhan Ocal and Raaz Dwivedi.