Brian Chen

Results 857 comments of Brian Chen

Having followed the previous multi-month, multi-person effort towards this (compiler plugins and more stable user-facing compiler APIs) quite closely before it fizzled out, I guess my reaction is "once bitten,...

Hmm, perhaps we're actually talking about different parts of the compiler? I was thinking of how Mixtape tries to provide an equivalent to https://github.com/JuliaLang/julia/pull/44950 and https://github.com/JuliaLang/julia/pull/41632, which is pretty orthogonal...

This won't help FastAI, but I think Optimisers.jl might be able to handle this directly. You might be able to workaround the lack of a `Flux.Optimise.update!(::AbstractOptimiser, ::TorchModuleWrapper, _)` by manually...

Edit: ignore everything below, `params` works as mentioned above! `Flux.trainable` defaults to `functor`, so unless you want it to return something other than `TorchModuleWrapper#params` then no additional action should be...

I fear the answer is no :frowning_face:. I'm not sure if there's anything we can do in Zygote to track implicit gradients for nested parameter arrays when the entire module...

I agree WRT not going the fully static route. Pragmatically, I don't think it even works right now because custom AbstractInterpreters are lost during dynamic dispatch. The tricky part of...

I don't quite understand either. Another benefit would be avoiding some of the gnarlier internal functions such as `_apply(_iterate)` which are usually compiled away in normal code. > In addition,...

Thanks for the explanation. I presume we're all mostly thinking about use cases like AD, which is how this came up in the first place. One of the big takeaways...

My first test was to search for `mean`. Not only was https://docs.julialang.org/en/v1/stdlib/Statistics/#Statistics.mean not near the top of the results like in the current docs search, it didn't show up at...

I would also add that the type of data in a batch can be quite heterogeneous. Think nested dicts, strings and whatnot. If MLJ(Flux) wants to handle that level of...