Michael Abbott

Relatedly, it would be nice if the MLJFlux models listed here https://github.com/FluxML/model-zoo#examples-elsewhere could be updated to use latest Flux, and avoid implicit gradients. Examples of similar upgrades: https://github.com/FluxML/model-zoo/issues?q=is%3Aclosed+label%3Aupdate+explicit In the...

`parameters = Flux.params(chain)` outside the `gradient` context will only work in the implicit style -- changing the explicit local `m` will not change `pen`. (Edit -- as ToucheSir says, while...

Can you sketch what this does, for people who don't know anything about Arrow? At the moment calling `destructure` on the tree of states doesn't work, as Leaf is a...

I have not followed Optimisers.jl closely, but in trying to solve https://github.com/FluxML/Flux.jl/issues/1826 re-invented `fmap(f, xs...)`... and then started to wonder if it has the right behaviour. See examples here: https://github.com/FluxML/Flux.jl/issues/1826#issuecomment-1013783952...

No. The most relevant line is: ```julia fmap(println, (x=[1,2], y=(a=[3,4], b=[5,6])), (x=[0,2], y=nothing)) ``` This will happen all the time with gradients, since Zygote collapses nothings. The second most relevant...

Yes, maybe `fmap` is off-topic. I have not tried hard to digest what this PR does. But I don't see any attempt in this package to test tricky cases, and...

I had a branch somewhere to change the printing to `ClipNorm(1) => Descent(0.1)`, and also digest such strings of `=>`. 2 characters and no brackets. Pushed as #139 now

This package still likes to think of itself as the new little thing moved outside Flux... it could be splintered further but it's a bit of a pain to manage...

Ok. I guess my vote is that such a thing should just depend on this package, until it's clear that the burden of loading more code than it needs outweighs...

> what are the increased maintenance effort More juggling of which packages are `dev`'d locally. More different version numbers, more ways for conflicts to occur. More waiting for registration of...