Michael Abbott

This turns out to be tricker than expected. See discussion in https://github.com/FluxML/NNlib.jl/pull/434, but the tl;dr is that `istraining` above relies on AD not being smart enough to notice that the...

That would be nice to have, but seems tricky to ensure. It does this: ``` a = wrapdims(rand(2), UniqueVector, x=1:2) perm = sortperm(a.x) wrapdims(parent(a)[perm], x=a.x[perm]) ``` and `a.x[perm]` doesn't know...

I get an error when I try `sort!`: ``` u = UniqueVector([41, 46, 19, 47, 21, 27, 16, 25, 45]) findfirst(isequal(u[5]), u) # fast method su = sort!(u) # ArgumentError:...

Simple views behave well, e.g. `@which findfirst(isequal(47), view(u, 2:7))`. But for `u[2:7]` the cost of re-generating the lookup dictionary was felt to be too much, in https://github.com/garrison/UniqueVectors.jl/pull/9. Maybe you could...

I agree it would be nice to make this all work. Trying to recover what's actually going on first, this works: ```julia a = AcceleratedArrays.Interval(sd, ed) A(in(a), :) # in(...)...

Not blocking, but should we make some effort to start organising a larger set of tests for this package? Zygote's tests contain a lot of knowledge about potentially awkward edge...

Are there any relevant weird cases where these may differ? An irrelevant one is this (because adjoint vectors have their own path): ```julia julia> map(sqrt, [1,2,3]') 1×3 adjoint(::Vector{Float64}) with eltype...

OK, that's interesting. Nothing about ProjectTo was closely looked at for latency, sadly.

> Currently, structured zero tangents are allowed to pass through during projection for (I believe) all number types. This was certainly the intention. The strong zero should mean no further...

> However, this breaks https://github.com/FluxML/Zygote.jl/blob/v0.6.43/test/features.jl#L528 because [gradient].y becomes nothing instead of zero. That test was added fairly recently, in https://github.com/FluxML/Zygote.jl/pull/1059 . I'd probably expect a hard Zero there (i.e. `nothing`)...