Agustinus Kristiadi

Can you use this instead? https://github.com/aleximmer/Laplace/issues/254 See also docs: https://aleximmer.github.io/Laplace/huggingface_example/#laplace-on-a-subset-of-an-llms-weights

Yes, I think this would be worthwhile to add! The starting point is here: https://github.com/aleximmer/Laplace/blob/6f170995582d07405462a3a89a252a12d412410b/laplace/baselaplace.py#L567-L591 Then, take a look at: https://github.com/aleximmer/Laplace/blob/6f170995582d07405462a3a89a252a12d412410b/laplace/baselaplace.py#L214-L241 Let me know if you have further questions!

The `CurvlinopsGGN/EF` with the diagonal structure should be able to handle non-module layers (as long as its params are included in `model.parameters()`) since it's just a pure `torch.func`. For custom...

If it's a BackPACK issue, then maybe switching backend will help. Can you try the following? ``` python from laplace import Laplace from laplace.curvature import AsdlGGN la = Laplace(model, ...,...

For sequential decision making (BO & contextual bandits), there doesn't seem to be any difference between post-hoc and online marglik, in my experience. See * https://arxiv.org/abs/2310.00137 * https://arxiv.org/abs/2304.08309 I also...

Still WIP test cases

Ready to review! Discussion points: 1. Do BackPACK, ASDL, Asdfghjkl even support multiple output dims? That is, if we flatten `logits = logits.view(-1, logits.size(logit_class_dim))`, do they even compute the correct...

Merged with `main` and ready for review!

It seems more complicated than anticipated. This PR is useful for models with image outputs like diffusion models. Considering v0.2 is all about LLMs, let's defer this to v0.3!

So what do you have in mind regarding the wording? Something like this in README.md? ```markdown ## Caveats - Currently, this library always assumes that the model has an output...