Conditional Variational Autoencoder (CVAE) with learnable priors

Loss Function of a Variational Autoencoder

Loss Function of CVAE with learnable priors, whereby c denotes a one-hot label

Examples for conditionally sampled digits (default settings in main.py were used)

One example for mixed sampling from all digit-classes

Requirements

python 3.6.10
torch 1.2.0+cu92
tensorflow 1.14

References:

Discriminative Variational Autoencoder for Continual Learning with Generative Replay: https://openreview.net/pdf?id=SJxjPxSYDH (Blind-Submission at the ICLR 2020)