How to pass function that returns gradient to Nonconvex.jl

[jonasmac16]

Hi,
I am looking into linking Nonconvex.jl with GalacticOptim.jl. I was running into the issue that GalacticOptim takes care of the AD stuff for the optimisation problem but I couldn't see an easy way to pass this info into Nonconvex.jl. Is it possible or does the AD of the optimisation problem has to take place in Nonconvex.jl?

I hope my question is somewhat clear.

[mohamed82008]

Define a new callable struct GOFunction that wraps a function and GO's AD-backend. Define a chain rule for GOFunction to make use of GO's AD mechanism. Nonconvex will be calling Zygote on the top level but that's a very shallow layer over GO's AD which can be using anything.

[mohamed82008]

I am closing this now for lack of activity.

[jonasmac16]

Thanks for the pointer, I will have a try implementing this in the coming week. Do you mind if I tag you in the PR and get back to you if I have anymore questions?

[mohamed82008]

Great, I don't mind at all!