Notes on 22 Mar. 2024

Agenda

Open discussion

Minutes

Video link: No recording

Improvements to DataAugmentations.jl and MLDatasets.jl:

@adrhill has made https://github.com/adrhill/ImageNetDataset.jl
- Based on https://github.com/JuliaML/MLDatasets.jl/pull/146/files
Interest in porting these changes over to MLDatasets.jl and DataAugmentation.jl
Brian and Kyle need to review
Kyle needs to fix DataAugmentation docs so we can first cut a release
Changes to DataAugmentation to support "inverse" transforms
- Are these real inverses or viewing the output of a pipeline back in image space (so only reversing color transformations, etc.)?
- How much is already covered by https://github.com/FluxML/DataAugmentation.jl/blob/master/src/visualization.jl?

Lux has PR for distributed training support: https://github.com/LuxDL/Lux.jl/pull/500/files

Seems NCCL revival is complete for supporting DL use cases
DistributedUtils submodule already defines primitives for ComponentArrays and NamedTuples
Should be easy to make a Flux/Lux agnostic port depending only on Optimisers.jl