Michael Abbott

Something like this might be useful, thanks for digging into the somewhat messy show code... At present `_nan_show(io, trainables(layer))` (a few lines down) aims to warn you about Inf/NaN, and...

I think I agree with your diagnosis. `ProjectTo{AbstractArray` should sort out the trailing dimensions, but doesn't get a chance because `map` switches from its N-dim form, to its just-iterate form:...

I suspect that `j` is a bug, also encountered here: https://github.com/FluxML/Zygote.jl/issues/1573

That does look wrong, but probably `Tuple{Vararg{Any,N}}` will be better for type stability than `Tuple`, while still allowing `nothing`: ```julia julia> NTuple{3}([1, 2, 3, 4]) (1, 2, 3) julia> NTuple{3}([1,...

These things are annoying. We tried to build a general solution, once upon a time... This could be solved by adding a rule, ideally to ChainRules.jl. One work-around might to...

+1, the code-vs-comment is a much more important distinction than function-vs-bracket within a comment. (Julia 1.12's display of docstrings at the REPL has the same problem, although I can't find...

I have no memory of this, but making lazy things fast is tricky. Some chance this code is missing a `Broadcast.instantiate`? Some chance the loop should just be replaced with...

Well, the only problem is that I don't know what `can_setindex` does, either. I just thought this gripe-issue was long enough with just one function. So perhaps this should move...

Second commit makes overwriting a 1 with a 0 increment the label. Then `copyto!` still works, and `x .= 0` gives an error rather than illegal answer. Although `x .=...

Mostly we need to be convinced this is doing the right thing. There were many other behaviours proposed earlier. Are we OK with `x .= 1` not giving an error?...