Agustinus Kristiadi

Hmm interesting. The major change would be `AsdlGGN` -> `CurvlinopsGGN` as the default backend. Could you try this and let me know the runtime? ```diff +from laplace.curvature import AsdlGGN ......

Ok great to narrow down the issue. Could you please paste here your full Laplace arguments? The one in dots in `la = Laplace(...)`.

As for grid search, please open a separate issue---let's discuss it there for clarity sake! Thanks!

I talked to @f-dangel, the author of [f-dangel/curvlionps-for-pytorch](https://github.com/f-dangel/curvlinops/tree/main). The issue is because of this line(s): https://github.com/aleximmer/Laplace/blob/83c5de70bb976a4266fe2c66f06f8726eba669fe/laplace/curvature/curvlinops.py#L117-L120 It is implemented as for-loop over num. of parameters => slow. I will put...

The above would indeed be the limitations of `curvlinops`. I would suggest to simply use `hessian_factorization=kron`---it should work best for all problems I could imagine. If you must use the...

Thanks a lot for the contribution! Looks great at a glance! I will review it in-depth this weekend. But for now, could you please make sure it conforms to both...

No worries! We should've been explicit about contributing in the README in the first place. The linting and autoformatting I requested are just for the scope of this PR. The...

Thanks for the changes! The code looks good now and passes all checks. Only small bits of changes needed this time.

@aleximmer please review if you have chance.

Hi Thomas, yes, this is currently not supported. At the current moment, this repo does not have an active maintainer. Not sure about the others, but I do not have...