VAE-tutorial

A simple tutorial of Variational AutoEncoder(VAE) models. This repository contains the implementations of following VAE families.

• Variational AutoEncoder (VAE, D.P. Kingma et. al., 2013)
• Vector Quantized Variational AutoEncoder (VQ-VAE, A. Oord et. al., 2017)

Requirements

• Anaconda
• python=3.7
• pytorch=1.7
• tqdm
• numpy

How-to-use

simply run the <file_name>.ipynb files using jupyter notebook.

Experimental Results

Variational AutoEncoder (VAE)

• trained on MNIST dataset for 20 epochs
• groundtruth(left) vs. generated(reconstructed, right)

• generated random samples from noise vector

Vector Quantized Variational AutoEncoder (VQ-VAE)

• trained on CIFAR-10 dataset for 50 epochs
• groundtruth(top) vs. reconstruction(bottom)

• randomly sampled codes from codebook