Agustinus Kristiadi

### Discussion result with @f-dangel Let `J(x)SJ(x)^T: \R^k \to R^k`, i.e. `(J(x)SJ(x)^T)(v)` for `v \in \R^k`. #### Computation with vmap (for small `K` or want differentiability) Then `vmap(J(x)SJ(x)^T)(I)` gives us...

@keyvan-amiri could you try this with our new backend? Install the `main` branch of the `laplace-torch` ``` pip uninstall laplace-torch pip install git+https://[email protected]/aleximmer/laplace ``` Then use the diagonal GGN/EF from...

Thanks for the input, Andreas! I wonder if something like this works for your case: ```diff def _glm_predictive_samples( self, f_mu: torch.Tensor, f_var: torch.Tensor, + link_function: Optional[Callable[[torch.Tensor], torch.Tensor]] n_samples: int, diagonal_output:...

Thanks @joemrt! I confirm via this test https://github.com/aleximmer/Laplace/commit/c182504070e3ae47e3ba374a009846f9b5c12f46 that this is indeed an unintended behavior.

> I guess one small potential downside is that in this case you can't use __call__ with another link function. Maybe this is even a good thing since using functional_samples...

@aleximmer could you please double-check and merge?

Thanks for the progress! * **Documentation:** Since we are going to integrate it into `laplace`'s core, we should follow `laplace`'s code standard, in particular, tests, typehinting and docstrings. I think...

Thanks for opening this issue. This shouldn't be the case since both `FullLLLaplace` and `KronLLLaplace` inherit `LLLaplace` which has `state_dict` defined. Moreover, this functionality is also well-tested https://github.com/aleximmer/Laplace/blob/main/tests/test_serialization.py. Could you...

I believe this commit fixes the issue: https://github.com/danielfalk/smart-open.nvim/commit/99ea9c3cfc407d25204147dae93520ab6acc39a5 It's still only in the `main` and `0.1` branches, though.

WIP adding an article based on `lm_example.py` to the docs.