Agustinus Kristiadi

An issue with Olympus ’s run (and thus benchmark) method: ```python oly = Olympus() for dataset in datasets: oly.run(dataset, ...) values = oly.campaign.observations.get_values().flatten() ``` When looking at the values of...

Efficient, universal, standalone Jacobian backend

2

- Functorch = memory blowup due to `vmap` - Asdl/asdfghjkl = can't backprop through the Jacobians => can't be used for continuous BO - BackPACK = requires inflexible extension We...

Closes #147 Closes #193 @runame please review and feel free to recommend the wording about the class-dim caveat.

Replace hardcoded class index with `logit_class_dim` argument

7

Closes #163 Please wait until #144 is merged.

Conform to PyTorch convention in the loss

1

Currently we always assume that the class index in the outputs of the model to be `-1`. While this works for standard models and Huggingface LLMs, there are some important...

Follow-up of #145

`torch.func` jacobian OOM

1

Minimum example to get this error: https://gist.github.com/wiseodd/b29973cd96f96f3af620ca131571eaa4 Fix: use `AsdlGGN` backend and then in `_glm_predictive_distribution` (in `baselaplace.py`), use: ``` python Js, f_mu = self.backend.jacobians(X) ``` instead of: ``` python Js,...

Currently, we still assume that `logits.shape == (batch_size, n_classes)`. E.g.: https://github.com/aleximmer/Laplace/blob/a4d3ed631c87ddb55a1bc7aa053da249180480f7/laplace/curvature/curvature.py#L332-L334

Enhancement on top of https://github.com/aleximmer/Laplace/pull/145. Need more investigation if this can be done efficiently.

Support for structured Kronecker factors

1

- [ ] Add support for [SINGD](https://github.com/f-dangel/singd) baseline. This enables structured Kronecker factors, e.g. diag & full, sparse, etc. - [ ] Update `KronLaplace` and `matrix.py` to accommodate the above.