Michael Abbott

This warning should be clearer that it's only about the gradient, but it seems like a tricky problem. If different parts of the range of `x` are done in parallel,...

It would be nice to be more organised about this, ideally both as "advertising" (i.e. to document what it is and isn't good at) and for development / regressions. Locally...

This `_normInf_back` function, and others like it which write into an array of zeros, should probably eventually call the gradient of `getindex`. This is because mutation won't play nicely with...

Ideally it would probably share code, have a function which for `mean` gets `scale=1/size(...)` or something. Xref #529 which is trying to re-work that rule.

What about the following pattern: ``` a, b, c = if cond(x) alpha, beta, gamma else nothing, f(beta), g(gamma) end ``` IIRC there are cases where one assignment outside an...

Have never thought about that. Fundamentally this package wants to make one big array from the RHS, and then there's another set of procedures to reshape/slice as desired by the...

FWIW, behaviour on version 0.4 is that these work without tweaks, and should probably be added to the tests: ```julia julia> x = rand(Int8, 3, 5) 3×5 Matrix{Int8}: 76 -128...

On this topic, StructArrays allows for efficient broadcasting out to several arrays, as for instance here: https://github.com/JuliaDiff/ChainRules.jl/blob/main/src/unzipped.jl#L6 It might not be hard to allow the pattern `@cast v[k], i[k] :=...

This should probably be an error. Right now it does not notice the indexing on the left, and treats it like `v[j]`: ```julia julia> v ≈ @reduce v2[j] := mean(i)...

I agree this could be given meanings. However I'm hoping not to add too many more features to this package! For now at least it shouldn't silently do the wrong...