flowMC: Normalizing-flow enhanced sampling package for probabilistic inference in Jax

TL;DR

flowMC is a Python library that combines local gradient-based MCMC methods with normalizing flow models to efficiently sample complex posterior distributions, leveraging JAX for high performance and flexibility.

Contribution

It introduces a novel sampling framework that integrates deep generative models with traditional MCMC, supporting complex distributions and accelerators.

Findings

Handles multimodal and correlated distributions effectively

Supports gradient-based samplers like HMC and MALA via JAX

Achieves high performance with GPU/TPU acceleration

Abstract

flowMC is a Python library for accelerated Markov Chain Monte Carlo (MCMC) leveraging deep generative modeling. It is built on top of the machine learning libraries JAX and Flax. At its core, flowMC uses a local sampler and a learnable global sampler in tandem to efficiently sample posterior distributions. While multiple chains of the local sampler generate samples over the region of interest in the target parameter space, the package uses these samples to train a normalizing flow model, then uses it to propose global jumps across the parameter space. The flowMC sampler can handle non-trivial geometry, such as multimodal distributions and distributions with local correlations. The key features of flowMC are summarized in the following list: * Since flowMC is built on top of JAX, it supports gradient-based samplers through automatic differentiation such as MALA and Hamiltonian Monte Carlo (HMC). * flowMC uses state-of-the-art normalizing flow models such as Rational-Quadratic Splines to power its global sampler. These models are very efficient in capturing important features within a relatively short training time. * Use of accelerators such as GPUs and TPUs are natively supported. The code also supports the use of multiple accelerators with SIMD parallelism. * By default, Just-in-time (JIT) compilations are used to further speed up the sampling process. * We provide a simple black box interface for the users who want to use flowMC by its default parameters, yet provide at the same time an extensive guide explaining trade-offs while tuning the sampler parameters. The tight integration of all the above features makes flowMC a highly performant yet simple- to-use package for statistical inference.