Michael Abbott

See #54 for a start on CategoricalArrays support... and having package extensions.

Timing like this mostly measures compilation time. There's some chance this is much slower on your server because it's the first run with these package versions, or something like that....

Aside, I highly recommend changing this to use "explicit" style, as in current docs https://fluxml.ai/Flux.jl/stable/guide/training/training/ . This "implicit" style with `Flux.params` will stop working soon. ```julia loss(m, x, y) =...

Xref https://discourse.julialang.org/t/flux-slows-down-by-10x-when-moving-from-local-system-to-high-performance-cluster/119753 No time to look closely now. But there is some chance your mac is just fast! M-processor memory is very quick. Some chance matmul isn't taking advantage of...

Fine to ask in a few places, I just like cross-linking so that anyone can check whether what they're about to say has already been typed up nicely elsewhere. I'm...

Not so clear what the input is on that Enzyme issue, but the output is a Matrix. But xref https://github.com/JuliaLang/julia/issues/52590, where the output is an MMatrix: ```julia julia> stack([SA[1,2], SA[3,4],...

I'm not so sure what you mean. Can you specify what the input is? All legal values of `dims` do already work, e.g. with the input above: ``` julia> stack(SA[SA[1,2],...

Note that `dims` is optional, and cases without it can be type-stable (when the input has enough static axes). At present these give `MMatrix`, as that's what falls out of...

Note that unlike the original issue, this one is `ProjectTo{AbstractArray}(:: Tuple{Float64, Float64, Float64})` i.e. it's about NTuple-vs-Array, not 1-Tuple-vs-Number. I think this is part of it: ```julia julia> gradient(x ->...

Maybe a Zygote issue? The gradient of `Tuple(::Vector)` seems to be giving the wrong type.