Add MultiMNIST to MLDatasets?

[manuelbb-upb]

To do some experiments along the lines of [1], I reimplemented their data generation routine in Julia. For MultiMNIST, we want to have two digits in one 28x28 image and 2D labels. That's why the MNIST type needs to be changed to

struct MultiMNIST <: SupervisedDataset
    metadata::Dict{String, Any}
    split::Symbol
    features::Array{<:Any, 3}
    targets::Matrix{Int}  # `MNIST` has Vector{Int}
end

Here is a constructor to create a MultiMNIST data set from an MNIST object. For that, I use ImageTransformations and Interpolations.

Does it seem like a good idea to try integrating something like that into MLDatasets? It would add those two dependencies...

[1] O. Sener and V. Koltun, “Multi-Task Learning as Multi-Objective Optimization,” arXiv:1810.04650 [cs, stat], Jan. 2019, Accessed: Jan. 24, 2022. [Online]. Available: http://arxiv.org/abs/1810.04650

[manuelbb-upb]

This will likely better be its own package