Michael Abbott

No great suggestions, but if https://github.com/JuliaLang/Compat.jl/pull/663 is ever merged then Compat will provide the new Slices type.

The eager `Align` is `stack`, and my understanding is that people were not so keen on adding a lazy version. While a view from `eachcol` is almost always as good...

Somehow the gradient of trace needs to be a `Diagonal{..., CuArray{...`, made by either by calling `similar`, or by having a special rule for CuArray. Rule in CR makes always...

I haven't investigated but those do look a bit non-standard, the usual form is `getindex(x::AbstractArray{

Maybe the first question should be: What comes after this layer? In this package, I think the efficient methods are that `*(::Matrix, ::OneHotMatrix)` is just indexing, used by `Flux.Embedding`, and...

Can you explain more the motivation? As in: - what is this used for? - why is any modification to OneHotArray better than wrapping it in existing Transpose / PermutedDimsArray...

> expect this layer to be followed by `*(...)` primarily. Since `*` is only for matrices & vectors, perhaps you just want `transpose(onehotbatch([1,1,2], 1:4))`? That doesn't specialise but it could:...

> I also tried going in the Transpose / PermutedDimsArray route (https://github.com/FluxML/OneHotArrays.jl/pull/36) , but I haven't been able to get type inference/stability right. Ah sorry, I mis-read, or forgot that...

Attempting to construct the minimal object: ```julia julia> using CategoricalArrays, OneHotArrays julia> cv = CategoricalArrays.CategoricalValue('b', CategoricalArray('a':'z')) CategoricalValue{Char, UInt32} 'b' julia> dump(cv) CategoricalValue{Char, UInt32} pool: CategoricalPool{Char, UInt32, CategoricalValue{Char, UInt32}} levels: Array{Char}((26,))...

See #54 for a start. Probably need someone to come up with a list of CategoricalArrays examples worth testing.