Brian Chen

Another interesting bit: adding _any_ operation around `o` in the problematic line resolves the issue: ```julia julia> function f(x) y = [[x]', [x]] r = 0.0 o = 1.0 for...

The assertion in question: https://github.com/FluxML/Zygote.jl/blob/v0.6.40/src/lib/base.jl#L27. I don't see any reason to not support this form of mutation as well, but perhaps there are some edge cases I've not thought of.

This looks to be strictly a Zygote thing and could probably moved there (or closed, if we think it's an inherent design limitation) instead of Flux.

Perhaps we could kill two birds with one stone if https://github.com/JuliaDiff/ChainRules.jl/issues/314 gets implemented. Moving to ChainRules would also get us ProjectTo, which in theory could handle more array types (modulo...

Following up on this, could the `accumulate!` behaviour be worked around by adding an explicit `@opt_out` for that function?

Zygote tries to diff through the kwsorter definition (i.e. https://docs.julialang.org/en/v1/devdocs/functions/#Keyword-arguments), which includes control flow. It's very difficult to make this type stable because it requires saving a different set of...

After looking into Diffractor, I think whatever it does happens outside the actual AD transform (perhaps leaving control flow intact is enough), but the ability to have unused branches/blocks in...

> So the only question I have where are the array versions of the `+`, `-` in ChainRules, or am I missing something here? See https://github.com/JuliaDiff/ChainRules.jl/blob/2e6491c9dd20608b725b7ecdca2dbe872cf7833b/src/rulesets/Base/arraymath.jl#L306-L329. I'm not sure if...

One could always go maximally explicit with something like `Tables.subset_along_row_dim`, but I imagine folks would not be terribly enthused about that name :P

Should the vocab CSV files be checked in? I would've assumed they would be artifacts or DataDeps as well.