An implementation of Adversarial Variational Bayes Autoencoder (as per [1]) using Keras with TensorFlow as backend. The code reproduces the generative experiments from the paper (synthetic dataset and MNIST) and was written as part of a Deep Learning lab course at TUM.
Check out also the original AVB repository: https://github.com/LMescheder/AdversarialVariationalBayes
The code for training and testing the models is compatible with Python 2, however, for metric evaluation only
Python 3 will interpret the third-party ITE package. Hence it is strongly recommended that you configure a
Python 3 interpreter. Please refer to the requirements.txt for a working package versions.
Note: for running the code on GPU, you will have to change the tensorflow requirement to tensorflow-gpu.
add details
These are some examples of the reproduced experiments on the latent space of the 4-points synthetic dataset and random generated samples trained on the binarized MNIST dataset. Both examples use AVB without AC. More images are coming soon.
Many thanks to Maximilian Karl for the fruitful discussions and the invaluable advice. This implementation makes use of the ITE package by Zoltán Szabó.
[1]: Mescheder, Lars, Sebastian Nowozin, and Andreas Geiger. "Adversarial variational bayes: Unifying variational autoencoders and generative adversarial networks." arXiv preprint arXiv:1701.04722 (2017).