Variational-Auto-Encoder

PyTorch implementation of Variational Auto-Encoder as described in Auto-Encoding Variational Bayes from ICLR 2014.

Randomly Sampled Images for 2D Latent Space

Latent Space

This model was trained to encode 784 dimensional MNIST images to just 2 dimensions and to then reconstruct it. The image below is a grid of outputs generated by walking through the 2D latent space Z.

Implementation Details

• The encoder and decoder are symmetrical MLPs with 256 neurons in each’s hidden layer.
• This implementation is inspired by Building Autoencoders in Keras.
• All the code was written and ran on Google Colab.

Requirements

torch
torchvision
numpy
matplotlib

References

1. Diederik P. Kingma, et al. Auto-Encoding Variational Bayes ICLR 2014[arxiv]