discrete-embeddings

Link to repository

(Ongoing) I explore a class of Variational Autoencoders (VAEs) that bottleneck into discrete representations, which are called Vector Quantized Variational Autoencoders (VQ-VAE), introduced in van den Oord et al. (2017). I implement the models in TensorFlow 2 and experiment with a few different datasets. I also implement state-of-the-art generative models (such as PixelCNN) to learn priors over the discrete representations, which enables high-quality data generation.

I will continue working on this project to implement and test VQ-VAE 2, Jukebox and PixelSNAIL.

I also plan to conduct a little research experiment into few-shot image generation in the discretized embedding space, instead of the full image.