Could you try with #764 ?

@dhairyagandhi96, #764 makes no difference, it makes some changes to onecold which is not called in this example.

julia> ohb = float.(onehotbatch(rand(1:10, 100), 1:10)) |> gpu
10×100 CuArray{Float32,2,Nothing}:
....
or
julia> ohb = cu.(onehotbatch(rand(1:10, 100), 1:10)) |> gpu
10×100 CuArray{Float32,2,Nothing}:
....

julia> dl(ohb)
5×100 CuArray{Float32,2,Nothing}:
....

Right, I misread that, apologies for that. Minimizing shouldn't be hard, it's probably in the matmul, but would need to cross check.

As far as I understand, the point of having OneHotMatrix as a separate type is to replace the expensive matmul with inexpensive indexing operation. This optimisation is lost in the workaround proposed by @mrchaos .

It also does not currently work properly with CuArrays, as demonstrated in the OP. I suppose this code from onehot.jl is supposed to make it work, but evidently it does not.

adapt_structure(T, xs::OneHotMatrix) = OneHotMatrix(xs.height, adapt(T, xs.data))

import .CuArrays: CuArray, cudaconvert
import Base.Broadcast: BroadcastStyle, ArrayStyle
BroadcastStyle(::Type{<:OneHotMatrix{<:CuArray}}) = ArrayStyle{CuArray}()
cudaconvert(x::OneHotMatrix{<:CuArray}) = OneHotMatrix(x.height, cudaconvert(x.data))

I am not 100% sure of the correct way Adapt.jl functions should be used to achieve the desired behaviour (docs are a bit scarce). One workaround I can think of is that instead of holding a Vector of OneHotVectors inside OneHotMatrix, it should hold an array of indices (Vector{Int}).

In any case, we don't have this problem when computing crossentropies (which is the main reason why OneHotMatrix is there).

using Flux, CuArrays
CuArrays.allowscalar(false)
using Flux: onehotbatch

ohb = onehotbatch(rand(1:10, 100), 1:10) |> gpu;
ŷ = CuArrays.rand(size(ohb)...)
Flux.crossentropy(ŷ, ohb)

I'm not sure wether OneHotMatrix was ever meant to be a fully-fledged AbstractMatrix to use as an input to a model

I'm not sure wether OneHotMatrix was ever meant to be a fully-fledged AbstractMatrix to use as an input to a model

One-hot encoding is a standard technique used in language modelling and other applications of deep learning. When training these models on GPU with Flux, one currently has to revert to dense arrays with zeros and ones for that purpose. This is not ideal; either OneHotMatrix needs to be fixed to support one-hot encoding (seems like a natural fit), or a separate type should be provided.

#958

Fixed by JuliaGPU/CUDA.jl#90. Make sure Julia 1.5 is used, as well as CUDA instead of CuArrays.