NumPyro is a lightweight probabilistic programming library utilizing JAX for autograd and JIT compilation, supporting GPU/TPU/CPU. It’s under active development, so expect changes. NumPyro focuses on Pyro primitives, providing a flexible platform for building models with regular Python and NumPy code. Inference algorithms include Hamiltonian Monte Carlo and No U-Turn Sampler. NumPyro aims to speed up Hamiltonian Monte Carlo through JIT compilation, and offers a variety of other algorithms like Variational Inference. The NumPyro.distributions module simplifies working with distribution classes, constraints, and bijective transforms. Unique features include effect handlers and Predictive utility for generating predictions. The library also offers various examples for Bayesian regression, time series forecasting, hidden Markov models, and more.
https://github.com/pyro-ppl/numpyro