Agustinus Kristiadi
Agustinus Kristiadi
@aleximmer what's the best way to go about this? For `KronLLLaplace` seems like this should be implemented in `KronDecomposed` in `matrix.py`. This change would be very useful in relation to...
Hi again, we use early stopping, so we usually don't look at the test results of the last epoch. This is mentioned in our paper. So instead, we look at...
Yes, that hotfix should work without any side effects. Let us know if you found any bugs, though!
Yes, should be good. Both are already covered with passing tests in #144. Closing this.
Revisiting this issue using the attached quick script. This problem happens in last-layer Laplace (all-layer is fine), for any Hessian structures, with the following backends: * `AsdlGGN` * `BackPackGGN` *...
This is actually a limitation of Bayesian linear regression with an isotropic Gaussian prior in general. Let $f(x) = W \phi(x)$ be the model where $f(x) \in \mathbb{R}^c$, $W \in...
Sorry @ArturPrzybysz for taking 2 years to answer this issue!
This is part of #144. The `mc-subset2` of that PR was a branch of `integrate-asdl-latest`. Once it's merged, this should be closed.
For Bayesian optimization, the standard way to do this is by first doing the linearized Laplace to obtain $p(f(x) \mid \mathcal{D})$ and then doing Thompson sampling on this functional posterior....
This should be fixed by #144. That PR has been successfully used for https://github.com/wiseodd/lapeft-bayesopt; i.e. Laplace on Huggingface LLMs like LLAMA-2-7B and T5.