0.2.3

Patches 0.2.2 with the following changes:

• restore compatibility with python 3.7 for mcmc.
• impose cache size limit in MCMC utilities.

0.2.2

Breaking changes

• Minor interface changes to MCMC utility functions. All experimental interfaces are marked as such in the documentation.

New Features

• A numpyro.factor primitive that adds an arbitrary log probability factor to a probabilistic model.

Enhancements and Bug Fixes

• Addressed a bug where multiple invocations of MCMC.run would wrongly use the previously cached arguments.
• MCMC reuses compiled model code whenever possible. e.g. when re-running with different but same sized model arguments.
• Ability to reuse the same warmup state for subsequent MCMC runs using MCMC.warmup.

0.2.1

Breaking changes

• Code reorganization - numpyro.mcmc is moved to numpyro.infer.mcmc but all major classes are exposed in the numpyro.infer module.
• rng argument to many classes and the seed handler has been more accurately renamed to rng_key.
• Deprecated functions that formed the old interface like mcmc and svi have been removed.

New Features

• Improved turning condition for NUTS that results in much higher effective sample size for many models.
• A numpyro.plate context manager, which records conditional independence information in the trace and does automatic broadcasting, like in Pyro.
• Inclusion of AutoMultivariateNormal, AutoLaplaceApproximation to the autoguide module.
• More distributions like LowRankMultivariateNormal, LKJ, BetaBinomial, GammaPoisson, ZeroInflatedPoisson, and OrderedLogistic.
• More transforms: MultivariateAffineTransform, InvCholeskyTransform, OrderedTransform.
• A numpyro.compat module that supports the pyro generic API for modeling and inference that can dispatch to multiple Pyro backends.
• Inclusion of Independent distribution and Distribution.to_event method to convert independent batch dimensions to dependent event dimensions. See the Pyro tutorial on tensor shapes for more details.
• A Predictive utility for generating samples from prior models, predictions from models using SVI's guide, or posterior samples from MCMC.
• A log_likelihood utility function that can compute the log likelihood for observed data by conditioning latent sites to values from the posterior distribution.
• New ClippedAdam optimizer to prevent exploding gradients.
• New RenyiELBO loss for Renyi divergence variational inference and importance weighted variational inference.

Enhancements and Bug Fixes

• MCMC does not throw an error on models with no latent sites.
• numpyro.seed handler can be used as a context manager like:

   with numpyro.seed(rng_seed=1):
       ...

• Utilities to enable validation checks for distributions, set host device count, and platform.
• More efficient sampling from Binomial / Multinomial distributions.
• The evidence lower bound loss for SVI is now a class called ELBO.
• Add energy field to HMCState, which is used to compute Bayesian Fraction of Missing Information for diagnostics.
• Add init_strategy arg to HMC/NUTS classes, which allows users select various initialization strategies.

0.2.0

Highlights

• Interface Changes to MCMC and SVI: The interface for inference algorithms have been simplified, and is much closer to Pyro. See MCMC and SVI.
• Multi-chain Sampling for MCMC: There are three options provided: parallel (default), sequential, and vectorized. Currently, parallel method is the fastest among the three.

Breaking changes

• The primitives param, sample are moved to primitives module. All primities are exposed in numpyro namespace.

New Features

MCMC

• In MCMC, we have the option to collect fields other than just the samples such as number of steps or step size, using collect_fields arg in MCMC.run. This can be useful when gathering diagnostic information during debugging.
• diverging field is added to HMCState. This field is useful to detect divergent transitions.
• Support improper prior through param primitives. e.g.

def model(data):
    loc = numpyro.param('loc', 0.)
    scale = numpyro.param('scale', 0.5, constraint=constraints.positive)
    return numpyro.sample('obs', dist.Normal(loc, scale), obs=data)

Primitives / Effect Handlers

• module primitive to support JAX style neural network. See VAE example.
• condition handler for conditioning sample sites to observed data.
• scale handler for rescaling the log probability score.

Optimizers

JAX optimizers are wrapped in the numpyro.optim module, so that the optimizers can be passed in directly to SVI.

Distributions

• New distributions: Delta, GaussianRandomWalk, InverseGamma, LKJCholesky (with both cvine and onion methods for sampling), MultivariateNormal.
• New transforms: CorrCholeskyTransform (which is vectorized), InverseAutoregressiveTransform, LowerCholeskyTransform, PermuteTransform, PowerTransform.

Utilities

• predictive utility for vectorized predictions from the posterior predictive distribution.

Autoguides

An experimental autoguide module, with more autoguides to come.

New Examples

• Sparse Linear Regression - fast Bayesian discovery of pairwise interactions in high dimensional data.
• Gaussian Process - sample from the posterior over the hyperparameters of a gaussian process.
• HMC on Neal's Funnel - automatic reparameterization through transform distributions.

Enhancements and Bug Fixes

• Improve compiling time in MCMC.
• Better PRNG splitting mechanism in SVI (to avoid reusing PRNG keys).
• Correctly handle models with dynamically changing distribution constraints. e.g.

def model():
    x = numpyro.sample('x', dist.Uniform(0., 2.))
    y = numpyro.sample('y', dist.Uniform(0., x))  # y's support is not static.

• Fixes step_size getting NaN in MCMC when it becomes extremely small.

First Public Release

Refer to the README for details.