pyro-ppl/pyro

Implement Nonparametric Hamiltonian Monte Carlo

Open

#2,903 创建于 2021年7月15日

在 GitHub 查看
 (0 评论) (0 反应) (0 负责人)Python (8,211 star) (981 fork)batch import
discussionenhancementhelp wanted

描述

This issue proposes adding an implementation of nonparametric Hamiltonian Monte Carlo, a modified version of HMC that provably correctly handles models with stochastic control flow, unlike vanilla HMC.

@fzaiser (one of the paper's authors) has open-sourced a PyTorch reference implementation, and turning this into a Pyro inference algorithm would be a great project for contributors with some preexisting interest in and knowledge of the line of research (on extensions of gradient-based approximate inference algorithms to discontinuous models) embodied in the paper.

Some questions for discussion:

  1. What extra information about the model is necessary to support the algorithms used in the paper's experiments?
  2. Can this information be extracted automatically, or would it require additional user-supplied annotations?
  3. Would it make more sense to implement this functionality in a new MCMC kernel or to modify our existing HMC kernels?
  4. Would NumPyro be a better substrate for this algorithm since Jax does not compile away control flow?

贡献者指南

技术栈
pythonpytorch
领域
machine learningai
议题类型
feature
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
5
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
over 1 week
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
stale
清晰度议题是否清楚说明期望改动、验收标准和下一步。
unclear
前置要求
Knowledge of Hamiltonian Monte CarloFamiliarity with PyTorchUnderstanding of probabilistic programming
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
10
研究方向
Research the paper referenced (arxiv.org/abs/2106.10238) and the PyTorch reference implementation at github.com/fzaiser/nonparametric hmc. Investigate how to adapt the algorithm to Pyro's existing MCMC kernel structure. The issue raises several open questions: determining necessary model information, automation vs. user annotations, whether to implement as new kernel or modify existing HMC kernels, and whether NumPyro (with JAX) would be a better substrate due to JAX's handling of control flow. Consider exploring these questions through prototyping and consulting Pyro's source code for examples of existing inference algorithms.