Nested Dropout

Implementation of nested dropout (https://arxiv.org/abs/1402.0915), but modernized. Uses Gumbel-Bernoulli distribution for the latent vector.

Generates an ordered code for the latent space of the autoencoder, this has many uses e.g. adaptive compression.

Requirements

• tensorflow
• tensorflow_probability
• numpy
• matplotlib

Usage

First pretrain a model with the pretrain_model.py script (should only take a few minutes). Then feed this pretrain model to the train_model.py script (this will take awhile).

TODO

• Add KL term so it's similar to a VAE
• code clean up