Agustinus Kristiadi

Unfortunately, it's not supported yet because the Hessian-computation libraries we use (BackPACK and ASDL) do not support BCE yet. Once they've added support to BCE loss, it'll be easy to...

* Linewidth 100 sounds about right to me (I usually use 120) * About string with `'` or `"`, I'm used to the former. But I don't have any preference---PEP8...

Hi Artur, it seems that there is indeed an issue with the `glm` predictive. When I change this line in your `main.py`: ```python pred = la_model.la(X) ``` into ```python pred...

Unfortunately no. I used VGG instead.

Hi @SouLeo, multi-output models are indeed still in our backlog. For now, I think this paper along with the code can be very useful for you .

This can be done easily with `torch.func` and Curvlinops. Useful too, for img2img problems. Should be worked on once milestone 0.2 is done.

This PR is currently blocked by https://github.com/f-dangel/curvlinops/issues/86. Need to make `curvlinops` supports Huggingface datasets first.

Regarding the addition of `logit_class_idx` argument: I did a "smoke test" in implementing this, and realized that the changes would be substantial---even without the present PR, `laplace-torch` always assumes that...

So, the remaining TODOs for this PR are: 1. Decide the conversations about `matrix.py`. @runame & @aleximmer could you please handle this? 2. Wait until https://github.com/f-dangel/curvlinops/pull/100 is merged Then we...

I replaced `get_nll` in crossval with `RunningNLLMetrics()`. So this PR will close #160 . (I thought before that I `RunningNLLMetics` hadn't been implemented, so would have involve major work.)