Implementation in Pytorch of Neural Discrete Representation Learning, Van den Oord et al (VQVAE). Similar to regular VAE the code optimizes the ELBO, however, instead of using a continuous latent variable z, the paper proposes to use a discrete vector which is used to index embeddings from a codebook. Also different, the framework learns the prior instead of using a fixed distribution; this is done with a PixelCNN. The code for the pixelCNN used in this repository is my own implementation which is available here.
Some results on FashionMnist:
