lorenzoh

Will be good to have ImageNet support! I'm wondering if there may be a simpler implementation for this, though. It seems the dataset has the same format as the (derived)...

Have done this for large vision datasets like COCO that have annotations in JSON which can be slow to parse. One thing to keep in mind is the size of...

Depends. If you have a large dataset of .jpg images and store them as arrays (hence losslessly), size can be multiples.

For reference, a [`ManualDataDep`](https://www.oxinabox.net/DataDeps.jl/dev/z40-apiref/#DataDeps.ManualDataDep) may be useful for when a dataset requires the user to perform some manual steps.

I feel it may be cleaner to tell users to use `mapobs` and give some examples. AFAICT there would be no functional difference, right?

They're supposed to work on any folder containing the dataset in the right format, not just the included datasets. Also some datasets can be used for multiple different tasks, so...

Have you seen https://github.com/lorenzoh/FeatureRegistries.jl? In FastAI.jl, it's used to create such a list, make it easier to search for datasets and handle the downloading: https://fluxml.ai/FastAI.jl/dev/references/FastAI.Registries.datasets

FastAI.jl will allow loading datasets from MLDatasets.jl. If MLDatasets.jl wants to include a feature registry for datasets, I'd also be happy to merge additional datasets currently in FastAI.jl into MLDatasets.jl...

How about using FeatureRegistries.jl for this?

The epoch number should definitely be rounded, but unfortunately combining both tables is not possible, since they are created individually at the end of each phase. There would be a...