Gaul

Quickstart

To sample from a 10 dimensional Gaussian with HMC, ADVI, or Laplace/quadratic approximation:

import jax.numpy as jnp
import jax.scipy.stats as stats

from gaul import hmc, advi, quap

def ln_posterior(params):
    return stats.norm.logpdf(params).sum()

params = jnp.zeros(10)

samples_hmc = hmc.sample(
  ln_posterior,
  params,
  step_size=0.2
  n_chains=1000,
  n_warmup=1000,
  n_samples=100
)

samples_advi = advi.sample(
  ln_posterior,
  params,
  lr=0.2,
  n_steps=1000,
)

samples_quap = quap.sample(ln_posterior, params)

Documentation

User Guide

• Installation
  • Package manager
  • From source
  • Tests

Examples and Tutorials

• Linear regression with Laplace approximation
• Eight schools - hierarchical modelling with HMC
• AR(2) model with variational inference
• Gaussian process regression

Notes

• API Reference
  • gaul.hmc module
  • gaul.quap module
    • sample()
  • gaul.advi module
    • batch_elbo()
    • diag_gaussian_logpdf()
    • diag_gaussian_sample()
    • elbo()
    • sample()
  • gaul.utils.pbar module
    • progress_bar_scan()
  • gaul.utils.tree_utils module
    • dense_hessian()
    • make_tree_hessian()
    • tree_ones_like()
    • tree_random_normal_like()
    • tree_split_keys_like()
    • tree_stack()
    • tree_zeros_like()
• Changelog
  • 0.4.2 (2022-04-07)
  • 0.4.1 (2022-04-06)
  • 0.4.0 (2022-04-05)
  • 0.3.0 (2022-04-05)
  • 0.2.2 (2022-04-05)
  • 0.2.1 (2022-02-27)
  • 0.2.0 (2022-02-26)
  • 0.1.0 (2022-02-25)

License

Copyright 2022, Jeff Shen.

The source code is free software dual licensed under MIT and Apache 2.0. For more details, see the LICENSE-MIT and/or LICENSE-APACHE files.