Gumbel-Softmax Variational Autoencoder with Keras.

This Keras based implementation of Categorical Reparameterization with Gumbel-Softmax. The code is based on Eric Jang's tutorial

You can run the Binder to get results like the following input-latent code-output triplet:

Also, you can visualize the latent code in 2D using Hamming distance based t-SNE:

I think this is a nice way to look into image hashing for image retrieval/search/etc. I'm excited to check how it scales up to real images...