https://github.com/Information-Fusion-Lab-Umass/NuX

Normalizing Flows using JAX
https://github.com/Information-Fusion-Lab-Umass/NuX

Last synced: about 2 months ago
JSON representation

Normalizing Flows using JAX

• Host: GitHub
• URL: https://github.com/Information-Fusion-Lab-Umass/NuX
• Owner: Information-Fusion-Lab-Umass
• Created: 2020-03-09T00:38:49.000Z (over 4 years ago)
• Default Branch: main
• Last Pushed: 2023-11-30T23:00:01.000Z (10 months ago)
• Last Synced: 2024-07-18T23:27:41.039Z (2 months ago)
• Language: Python
• Homepage:
• Size: 86.5 MB
• Stars: 82
• Watchers: 9
• Forks: 4
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-normalizing-flows - NuX

README

        
### Check out an improved version of NuX called [generax](https://github.com/EddieCunningham/generax)!

# NuX - Normalizing Flows using JAX

[![.github/workflows/CI.yml](https://github.com/MilesCranmer/NuX/actions/workflows/CI.yml/badge.svg)](https://github.com/MilesCranmer/NuX/actions/workflows/CI.yml)

## What is NuX?

NuX is a library for building [normalizing flows](https://arxiv.org/pdf/1912.02762.pdf) using [JAX](https://github.com/google/jax).

## What are normalizing flows?

Normalizing flows learn a parametric model over an unknown probability density function using data.  We assume that data points are sampled i.i.d. from an unknown distribution p(x).  Normalizing flows learn a parametric approximation of the true data distribution using maximum likelihood learning.  The learned distribution can be efficiently sampled from and has a log likelihood that can be evaluated exactly.

## Why use NuX?

It is easy to build, train and evaluate normalizing flows with NuX

```python

import jax

import jax.numpy as jnp

import nux

# Generate some fake data

rng_key = jax.random.PRNGKey(0)

batch_size, dim = 8, 3

x = jax.random.normal(rng_key, (batch_size, dim))

# Create your normalizing flow

flow = nux.Sequential([nux.DenseMVP(),

                       nux.SneakyReLU(),

                       nux.UnitGaussianPrior()])

# Initialize the flow with data-dependent initialization

z, log_px = flow(x, rng_key=rng_key)

# Retrieve the initialized parameters from the flow

params = flow.get_params()

# Generate samples.  The prior layer will use the input's shape for sampling.

x_samples, log_px_samples = flow(jnp.zeros_like(x), params=params, inverse=True, rng_key=rng_key)

# Pass the samples to the latent space

z, log_px = flow(x_samples, params=params, inverse=False)

# Reconstruct the samples

x_reconstr, log_px_reconstr = flow(z, params=params, inverse=True, reconstruction=True)

assert jnp.allclose(x_samples, x_reconstr)

assert jnp.allclose(log_px_samples, log_px)

assert jnp.allclose(log_px, log_px_reconstr)

```

## Get started

The easiest way to use NuX is to clone this repo and install the prerequisites with the "pip install ." command.  JAX should be manually installed (see [this](https://github.com/google/jax#installation)) because GPU support is system dependent.  The NuX package available on pip is outdated and does not have much of the functionality of the current code.

NuX is in active development, so expect the API to change over time.  Contributions are welcome!