This repository contains code for training a variational autoencoder with categorical latents on the MNIST dataset. It's meant to accompany this blog post: https://jxmo.io/posts/variational-autoencoders
The training can also be visualized through Weights and Biases, like in this run here: https://wandb.ai/jack-morris/categorical-vae/runs/36nfcetj. Weights & Biases is a really nice tool that lets you visualize loss curves, gradients, and auto-encoded images and see how they change across training.
Model code (including code for the Gumbel-softmax trick) is in models.py. Training code (including the KL divergence computation) is in train.py. To run the thing, you can just type:
python train.py
(You'll need to install numpy, torchvision, torch, wandb, and pillow to get things running.)