FluxML/Flux.jl

Allow BatchNorm training on CUDA with `track_stats=false`

Open

#1,606 opened on May 31, 2021

View on GitHub
 (4 comments) (0 reactions) (0 assignees)Julia (4,725 stars) (619 forks)batch import
good first issuehelp wanted

Description

Gathered from https://discourse.julialang.org/t/batchnorm-only-track-stats-true-supported-on-gpu/62091.

This would most likely require changes in NNlibCUDA as well. I'm not sure how interchangeable the various cudnnBatchNormalizationForward* functions are, so putting a pin in this until someone more knowledgeable can comment.

Contributor guide

Tech stack
None
Domain
machine learning
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.
mostly clear
Prerequisites
JuliaFlux.jlCUDACuDNN
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
Review the discourse thread linked for context on the request. Investigate the CuDNN API to determine if there is a way to perform batch normalization training without updating running statistics (i.e., track stats=false). Then identify the necessary changes in NNlibCUDA and Flux.jl to expose this option. Check for any existing work or discussions in the linked repository forks or related issues.