Michael Abbott

xref #599, although not sure it's identical.

Does Zygote's forward mode use ChainRules? If so, then removing SpecialFunctions should be easy. Which others might be worth trying to do quickly, and holding up 0.6? If it's just...

OK. It sounds like this `ChainRules_Distances_Glue.jl` idea would work, if someone is happy to set it up. (Maybe ChainRules is a good candidate for that mult-package-one-repository story?) The transition to...

Maybe the short-term work-around is to define `f(c) = Zygote.forwarddiff(c -> itp[c], c)`.

I haven't looked in ages but yes, sure. Maybe the "Debugging" section should start with ways to get things running at all, not with profiling?

I don't think there's a way to do this with Zygote's implicit parameter dictionary, `params(model)`. Or at least, nobody has written one. But if I'm reading correctly, your Python code...

Happens with `sum` too. (Edit -- with `abs`!) I think the problem is complex broadcasting, for CuArrays broadcasting uses ForwardDiff. Note that it needs 3 arguments, to hit generic broadcasting:...

Oh right, sorry, my example should give that error! Fixed above. (I guess we could make `sensitivity` work when the imaginary part is zero, that would be OK, but it...

Maybe. There's something to be said for keeping legacy tests as they were, so that you can be quite sure what's changed if they fail. And, if spending time re-writing...

Functors.trainable and ChainRulesCore.ProjectTo have quite a bit in common, it's possible they should get to know each other better. I'm not precisely sure why this doesn't work today, but with...