AutoEncoderToolkit.jl: A Julia package for training (Variational) Autoencoders

With the advent of generative models, the field of unsupervised learning has exploded in the last decade.One of the most popular generative models is the variational autoencoder (VAE) (Kingma & Welling, 2014).VAEs assume the existence of a joint probability distribution between a high-dimensional data space and a lower-dimensional latent space.Using a variational inference approach, the VAE parametrizes this joint distribution with two neural networks-an encoder and a decoder.This approach allows the model to approximate the underlying lowdimensional structure that generated the observed data and generate new samples by sampling from the learned latent space.Several variations of the original VAE have been proposed to extend its capabilities and tackle different problems.Here, we present AutoEncoderToolkit.jl,a Julia package for training VAEs and its extensions.The package is built on top of the Flux.jldeep learning library (Innes, 2018) and provides a simple and flexible interface for training different flavors of VAEs.Furthermore, the package offers a set of utilities for the geometric analysis of the learned latent space.