Skip to content

Latest commit

 

History

History
85 lines (59 loc) · 2.07 KB

README.md

File metadata and controls

85 lines (59 loc) · 2.07 KB
Raw

SGMCMCJax

Quickstart | Samplers | Documentation

SGMCMCJax is a lightweight library of stochastic gradient Markov chain Monte Carlo (SGMCMC) algorithms. The aim is to include both standard samplers (SGLD, SGHMC) as well as state of the art samplers.

The target audience for this library is researchers and practitioners: simply plug in your JAX model and easily obtain samples.

Note that this library is still in its early stages so expect the API to change a bit.

Example usage

We show the basic usage with the following example of estimating the mean of a D-dimensional Gaussian from data using a Gaussian prior.

import jax.numpy as jnp
from jax import random
from sgmcmcjax.samplers import build_sgld_sampler


# define model in JAX
def loglikelihood(theta, x):
    return -0.5*jnp.dot(x-theta, x-theta)

def logprior(theta):
    return -0.5*jnp.dot(theta, theta)*0.01

# generate dataset
N, D = 10_000, 100
key = random.PRNGKey(0)
X_data = random.normal(key, shape=(N, D))

# build sampler
batch_size = int(0.1*N)
dt = 1e-5
my_sampler = build_sgld_sampler(dt, loglikelihood, logprior, (X_data,), batch_size)

# run sampler
Nsamples = 10_000
samples = my_sampler(key, Nsamples, jnp.zeros(D))

Samplers

The library includes several SGMCMC algorithms with their pros and cons briefly discussed in the documentation.

The current list of samplers is:

  • SGLD
  • SGLD-CV
  • SVRG-Langevin
  • SGHMC
  • SGHMC-CV
  • SVRG-SGHMC
  • pSGLD
  • SGLDAdam
  • BAOAB
  • SGNHT
  • SGNHT-CV
  • BADODAB
  • BADODAB-CV

Installation

Create a virtual environment and either install a stable version using pip or install the development version.

Stable version

To install the latest stable version run:

pip install sgmcmcjax

Development version

To install the development version run:

git clone https://github.com/jeremiecoullon/SGMCMCJax.git
cd SGMCMCJax
python setup.py develop

Then run the tests with pip install pytest; make