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

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}
}