Reduce hcat creates dense matrix

[simonmandlik]

See here: FluxML/Flux.jl#1596

[mcabbott]

Existing methods are here, a specialisation for reduce(hcat, ... could be added:

OneHotArrays.jl/src/array.jl

Lines 112 to 130 in 9c43cb7

	function Base.cat(x::OneHotLike{<:Any, <:Any, N}, xs::OneHotLike...; dims::Int) where N
	if isone(dims) || _notall_onehot(x, xs...)
	return cat(map(x -> convert(_onehot_bool_type(x), x), (x, xs...))...; dims = dims)
	else
	L = _nlabels(x, xs...)
	return OneHotArray(cat(_indices(x), _indices.(xs)...; dims = dims - 1), L)
	end
	end
	Base.hcat(x::OneHotLike, xs::OneHotLike...) = cat(x, xs...; dims = 2)
	Base.vcat(x::OneHotLike, xs::OneHotLike...) =
	vcat(map(x -> convert(_onehot_bool_type(x), x), (x, xs...))...)
	# optimized concatenation for matrices and vectors of same parameters
	Base.hcat(x::OneHotMatrix, xs::OneHotMatrix...) =
	OneHotMatrix(reduce(vcat, _indices.(xs); init = _indices(x)), _nlabels(x, xs...))
	Base.hcat(x::OneHotVector, xs::OneHotVector...) =
	OneHotMatrix(reduce(vcat, _indices.(xs); init = _indices(x)), _nlabels(x, xs...))

I suspect these methods could use some stress-testing & tidying, e.g. cat(ohm, ohm; dims=Val(2)) looks like it will be an error (instead of falling back).