pymc-devs/pymc

Implement DiffTransform for RandomWalk distributions

Open

#6,098 opened on Sep 4, 2022

View on GitHub
 (1 comment) (0 reactions) (0 assignees)Python (1,902 forks)batch import
enhancementshelp wantedpytensor

Repository metrics

Stars
 (7,926 stars)
PR merge metrics
 (Avg merge 11d) (7 merged PRs in 30d)

Description

My guess is that unobserved [Gaussian/MvNormal/MvStudentT]RandomWalk would sample much better if NUTS could propose innovation values instead of the "absolute" timeseries values. This is what is done when one manually creates a random walk model by

x_raw = pm.Normal("x_raw")
x = pm.Deterministic("x", x_raw.cumsum())

If we just added the transform, there would some redundancies in the logp, because the proposed diffed values would be cumsumed (in the back-transform) and then differentiated again in the randomwalk logprob, but we can probably teach Aesara to optimize those away.

Once we have that, the user won't be worse off when creating unobserved random walks via the specialized PyMC distribution classes.

Contributor guide

Research direction
Understand the current RandomWalk logp implementation in PyMC, then design and implement a DiffTransform that reparameterizes the random walk to sample innovations. Study how Aesara optimizes redundant operations and ensure the transform integrates with the existing distribution classes.
Tech stack
python
Domain
backend
Issue type
Feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
Half day
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
PythonPyMCprobabilistic programming concepts
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
55