write datasets in a JLD2 or Arrow format for faster read

[CarloLucibello]

We could have a "processed" folder in each dataset folder where we write the dataset object the first time we create it. In the following creations, e.g. d = MNIST() we just load the JLD2 file.

Example:

function MNIST(...)
    dataset_dir = ...
    processed_file = joinpath(dataset_dir, "processed", "dataset.jld2") 
    if isfile(processed_file) 
        return FileIO.load(processed_file, "dataset")
    end 

    mnist = ...
    if isfile(processed_file) 
        FileIO.save(processed_file, Dict("dataset" => mnist))
    end 
    return mnist
end

[lorenzoh]

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 the JLD2 files, though of course it shouldn't be a problem for MNIST. Arrow.jl can also be a good format with built-in compression when the data has samples made up of primitive types and arrays.

[CarloLucibello]

What's to be expected from the JLD2 sizes? hopefully not larger than the size of the original data, right?

[lorenzoh]

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

[zsz00]

I agree too Arrow.jl is a good format:

1. built-in compression
2. Cross-language processing dataset

[CarloLucibello]

HuggingFace's datasets library also uses Arrow: https://huggingface.co/docs/datasets/about_arrow

[CarloLucibello]

some code showing how to read/write color arrays from/to arrow tables
https://gist.github.com/CarloLucibello/51d713ec4a1612b46e6c90e53c0f88e8