pyro-ppl/pyro

Multi-output multi-lengthscale GPs [feature request] [discussion]

Open

#2,889 opened on Jun 28, 2021

View on GitHub
 (6 comments) (0 reactions) (0 assignees)Python (8,211 stars) (981 forks)batch import
enhancementhelp wanted

Description

Hey everyone,

I have coded up a multi-output multi-lengthscale GP (i.e. separate kernel hyperparameter per input dimension and per output dimension). It is pretty efficient, and scalable combined with inducing points. I took the code from GP contrib and modified it for research purposes (neural data analysis for big data), but if there is enough interest here it should be pretty easy to put this into Pyro itself. The question I have is, how should I package it in? Replace the current GP, or leave the option to use simple GP versus multi-dimension GP, or provide a separate subpackage in addition to contrib.GP?

I am aware of the interfacing with GPyTorch. Initially I tried to use that, but it is not straightforward to perform all PPL features (the loss objective is limited in this interface, GPLVM is tricky etc.).

Contributor guide

Tech stack
pythonpytorch
Domain
machine learningai
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
Gaussian processesPyroPyTorch
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
Review the current GP implementation in Pyro's contrib module and discuss with maintainers how to integrate the multi output multi lengthscale GP, possibly as a separate subpackage. Examine GPyTorch for reference but note limitations with Pyro's PPL features. Consider feedback from the 6 comments on the issue for design direction.