use MLUtils

[CarloLucibello]

This is replacing some utility functions and the DataLoader with their equivalent versions in MLUtils.jl.

On the DataLoader inferrability

After this, the DataLoader will accept generic datasets exposing the MLUtils.numobs and MLUtils.getobs interface (something that we should probably document somewhere). This has been discussed in #1282.

Unfortunately we lose inferrability, e. g. @inferred first(loader) will fail. This is because the previous implementation was just wrong for generic datasets. It was

Base.eltype(loader::DataLoader{D}) where D = D

where D is the underlying data type. So this PR will reopen #1159. This is an example of the current problematic behavior on master:

julia> using Flux.Data: Dataloader

julia> d = DataLoader(1:5);

julia> [x for x in d]  # iteration is fine
5-element Vector{Vector{Int64}}:
 [1]
 [2]
 [3]
 [4]
 [5]

julia> collect(d)
ERROR: MethodError: Cannot `convert` an object of type Vector{Int64} to an object of type UnitRange{Int64}
Closest candidates are:
  convert(::Type{T}, ::LinearAlgebra.Factorization) where T<:AbstractArray at ~/julia/julia-1.7.1/share/julia/stdlib/v1.7/LinearAlgebra/src/factorization.jl:58
  convert(::Type{T}, ::AbstractRange) where T<:AbstractRange at ~/julia/julia-1.7.1/share/julia/base/range.jl:215
  convert(::Type{T}, ::T) where T<:AbstractArray at ~/julia/julia-1.7.1/share/julia/base/abstractarray.jl:16
  ...
Stacktrace:
 [1] setindex!(A::Vector{UnitRange{Int64}}, x::Vector{Int64}, i1::Int64)
   @ Base ./array.jl:903
 [2] copyto!(dest::Vector{UnitRange{Int64}}, src::DataLoader{UnitRange{Int64}, Random._GLOBAL_RNG})
   @ Base ./abstractarray.jl:893
 [3] _collect(cont::UnitRange{Int64}, itr::DataLoader{UnitRange{Int64}, Random._GLOBAL_RNG}, #unused#::Base.HasEltype, isz::Base.HasLength)
   @ Base ./array.jl:655
 [4] collect(itr::DataLoader{UnitRange{Int64}, Random._GLOBAL_RNG})
   @ Base ./array.jl:649
 [5] top-level scope
   @ REPL[9]:1
 [6] top-level scope
   @ ~/.julia/packages/CUDA/bki2w/src/initialization.jl:52

TODO:

• Add news
• check if it is closing Cannot do double iteration of DataLoader #1227. It does!

Close #1227

[CarloLucibello]

Merge?

[CarloLucibello]

mergity merge?

[darsnack]

I guess the main issue here is that the new DataLoader does not work with something that isn't an array but implemented getindex and not getobs. Can we just add the getindex fallback to MLUtils.jl?

Otherwise this is a drop-in replacement, so good enough for me.

[darsnack]

Other question is what happens when MLUtils -> MLDataUtils?

[darsnack]

No more concerns on my end

[mcabbott]

There are many warnings like this in the tests:

┌ Warning: `unsqueeze(x::AbstractArray, dims::Int)` is deprecated, use `unsqueeze(x; dims = dims)` instead.
│   caller = macro expansion at Test.jl:463 [inlined]
└ @ Core ~/.julia/dev/julia/usr/share/julia/stdlib/v1.8/Test/src/Test.jl:463
┌ Warning: `stack(x, dims)` is deprecated, use `stack(x; dims = dims)` instead.
│   caller = macro expansion at utils.jl:249 [inlined]
└ @ Core ~/.julia/dev/Flux/test/utils.jl:249

No rush but it would be nice to update all of them.

[cossio]

Can we have a release with the nicer DataLoader?

[darsnack]

These are the blockers: https://github.com/FluxML/Flux.jl/milestone/3