Normalizing flows to generate MNIST Digits

This code base implements Normalizing Flows as proposed in Rezende et al. to generate MNIST digits using Tensorflow.

Usage:

python main.py [plot_or_not] [number of flows]

No need to download MNIST, tensorflow does it for you!

Outputs

• 10 files with latent states of each number
• 1 file with combined latent states
• Graph plotting latent states as in the diagram if plot_or_not=1
• Folder names Out is generated containing samples of generations after interval of 100 iterations

2D Latent Vector Representation, Left is vanilla VAE, Right is with Normalizing flows. As can be seen, the one with normalizing flows has a flexible multi-modal distribution, opposed to unimodal gaussian for vanilla vae

If you use the code base, please cite us at

@article{saxena2017variational,
  title={Variational Inference via Transformations on Distributions},
  author={Saxena, Siddhartha and Dohare, Shibhansh and Kapoor, Jaivardhan},
  journal={arXiv preprint arXiv:1707.02510},
  year={2017}
}