onecold does not work on CuMatrix

[AzamatB]

julia> xs = rand(2,3)
2×3 Array{Float64,2}:
 0.0758553  0.779633    0.218497
 0.840066   0.00234058  0.402085

julia> Flux.onecold(xs, ["a","b"])
3-element Array{String,1}:
 "b"
 "a"
 "b"

julia> xs = cu(xs)
2×3 CuArray{Float32,2}:
 0.0758553  0.779633    0.218497
 0.840066   0.00234058  0.402085

julia> Flux.onecold(xs, ["a","b"])
ERROR: Type String does not have a definite size.
Stacktrace:
 [1] sizeof at ./essentials.jl:452 [inlined]
 [2] CuArray{String,1}(::UndefInitializer, ::Tuple{Int64}) at /home/azamat/.julia/packages/CuArrays/wXQp8/src/array.jl:38
 [3] similar at /home/azamat/.julia/packages/CuArrays/wXQp8/src/array.jl:69 [inlined]
 [4] similar(::CuArray{Float32,1}, ::Type{String}, ::Tuple{Base.OneTo{Int64}}) at ./abstractarray.jl:635
 [5] #mapslices#111(::Int64, ::typeof(mapslices), ::getfield(Flux, Symbol("##25#26")){Tuple{Array{String,1}}}, ::CuArray{Float32,2}) at ./abstractarray.jl:2009
 [6] #mapslices at ./none:0 [inlined]
 [7] onecold(::CuArray{Float32,2}, ::Array{String,1}) at /home/azamat/.julia/packages/Flux/dkJUV/src/onehot.jl:121
 [8] top-level scope at none:0

[CarloLucibello]

It does work if you replace ["a","b"] with numbers. Also it may be not too slow in some cases to perform the computation on the cpu (e.g. large network, i.e. slow forward pass, and move only last output to the cpu. This is what Knet does)

[dnabanita7]

Since onecold only accepts numbers, Can I try fixing this?

[darsnack]

I don't see how this works for strings:

onecold(y::OneHotMatrix, labels...) =
  mapreduce(x -> Flux.onecold(x, labels...), |, y.data, dims = 2, init = 0)

What does | do for strings? Why do we need to reduce at all? Isn't the following sufficient?

onecold(y::OneHotMatrix, labels...) = map(x -> Flux.onecold(x, labels...), y.data)