PyTorch Implementation of Auto-Encoding Variational Bayes

Prerequisites

• PyTorch
• torchvision

Training

• Train on MNIST

  python vae.py --cuda --hidden_size 2
  

  During training, samples generated from random normal Gaussian distribution will be saved under folder samples. hidden_size is the dimension of z, if hidden_size=2, an interpolation image will be generated at the end of the last epoch, shown as below:

  

Image Quality w.r.t Hidden Size

I observed that larger hidden_size gives shaper but harder to recognize images, while smaller hidden_size gives more blur but easier to recognize images, shown as below(From left to right is 2D, 5D, 10D hidden_size). My guess is that a 2-dimensional Gaussian is much easier to approximate than a 20-dimensional Gaussian, while smaller hidden_size blocks lots of information going through the auto-encoder.

Reference

1. https://github.com/yunjey/pytorch-tutorial
2. https://github.com/pytorch/examples