NF

Normalizing flows for density estimation with built-in support for sampling.

This repo provides an interface for creating normalizing flows. It hosts two modules, NF and BAM.

The NF module provides a basic interface for normalizing flows, and defines some transformations like rotation, permutation, etc. NF transformations can be stacked to create more complex NF transformations. A NF transformation can be converted to a generative model that can be trained to model an unknown target distribution and sample from it. See RealNVP_test.ipynb for an example of how the NF module can be used to implement the RealNVP architecture from the paper "Density estimation using Real NVP" (https://arxiv.org/pdf/1605.08803.pdf).

The BAM module provides a basic interface for blockwise autoregressive monotonic transformations, and defines some of those (most notably the BAM.Linear, which is the basic building block for the B-NAF architecture from the paper "Block Neural Autoregressive Flow" (https://arxiv.org/pdf/1904.04676.pdf)). Individual BAM transformations can be stacked/concatenated/summed to create more complex transformations. Similar to the NF module, the BAM module can easily be extended by custom transformations. A BAM transformation can be converted to a NF transformation. B-NAF_test.py provides an example of how the BAM module can be used to create a custom normalizing flow.

Python notebooks sometimes do not render on github. In this case, you can use nbviewer:

• https://nbviewer.jupyter.org/github/sshish/NF/blob/master/RealNVP_test.ipynb
• https://nbviewer.jupyter.org/github/sshish/NF/blob/master/B-NAF_test.ipynb

This repo was inspired by

• https://github.com/senya-ashukha/real-nvp-pytorch, which is a PyTorch implementation of RealNVP
• https://github.com/nicola-decao/BNAF, which is the authors' original implementation of B-NAF

Suggestions for improvement are welcome.

Requirements

To use NF and BAM modules:

• PyTorch version 1.3.0 or above

To run the RealNVP_test.ipynb and B-NAF_test.py examples:

• CUDA support
• scikit-learn
• matplotlib