pymc-devs/pymc

Compute only dependent logp in step samplers

Open

#7,591 创建于 2024年11月26日

在 GitHub 查看
 (0 评论) (0 反应) (0 负责人)Python (7,926 star) (1,902 fork)batch import
enhancementshelp wantedsamplers

描述

Description

As discussed in: https://discourse.pymc.io/t/semantik-description-of-kruschke-diagrams-model-to-graphviz/16142/8 we could be making more use of the conditional partition of the logp graph for partial step samplers.

When the proposal depends on a ratio between probabilities (BinaryMetropolois, BinaryGibbsMetropolis, CategoricalGibbsMetropolis, Slice, NUTS), we don't need to compute terms that don't depend on the updated variables.

Step samples that explicitly request delta_logp (Metropolis, DEMetropolis) achieve this implicitly via PyTensor rewrites.

I suggest adding a model.dependent_logp(vars) that helps us get the logp we need (the variables of interest + conditional dependent variables)

贡献者指南

技术栈
python
领域
machine learningperformance
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。
4
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。
1-2 days
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。
fresh
清晰度议题是否清楚说明期望改动、验收标准和下一步。
clear
前置要求
Understanding of PyTensor graphFamiliarity with PyMC step methodsKnowledge of conditional logp
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。
20
研究方向
Examine the current logp computation in PyMC's step samplers, particularly how delta logp is used in Metropolis samplers. Implement a helper function `model.dependent logp(vars)` that computes only the logp of variables that depend on `vars` using the conditional partition of the logp graph. Test with BinaryMetropolis and Slice samplers to ensure correctness. Refer to the discourse discussion at https://discourse.pymc.io/t/semantik description of kruschke diagrams model to graphviz/16142/8 for context.