Agustinus Kristiadi comments

Results 90 comments of


                                            Agustinus Kristiadi

Bringing laplace-torch to foundation-model era

Might as well fixes #156 while we're at it.

Bringing laplace-torch to foundation-model era

All tasks finished!

Bringing laplace-torch to foundation-model era

Double-checked and everything looks fine! Merging this.

[Discussion] Last-layer Laplace for img2img problem

Beyond CIFAR images, though, the GGN computation will also be an issue. E.g. in ImageNet the output dim is `224*224*3=150528`, much larger than `3024` of CIFAR. I talked Felix about...

[Discussion] Last-layer Laplace for img2img problem

For predictions/reconstructions, my proposal is to use https://github.com/f-dangel/unfoldNd. Using this, then `conv_transpose2d` is just a matrix multiplication under the original weights/filters, implying that we can easily obtain $p(f(x))$. ```python import...

[Discussion] Last-layer Laplace for img2img problem

Full, self-contained prototype here: https://gist.github.com/wiseodd/b8d57fa029f876e00b336b7b3b5052bd

[Discussion] Last-layer Laplace for img2img problem

Unfortunately, there's no update on this. Partly because the loss function usually used in image problems (BCELoss) is not supported by the Hessian backends, and partly because my research agenda...

[Discussion] Last-layer Laplace for img2img problem

This issue should be easier to solve once #145 is merged. Will work on this after the release of milestone 0.2.

Serialization: `torch.load(..., map_location="cpu")` doesn't map all tensors in `Laplace` instance

Thanks, Steve. It seems like a neat addition to #148 and I don't see any downside to this. Feel free to open a pull request! Anyway, I'm just curious about...

Last Layer Laplace predictions could be computed much faster and becomes problematic for large label classification.

Hi Carles, they are indeed good suggestions! I wrote the formula for KronLLLaplace in App. B.1 of [my paper](https://arxiv.org/pdf/2002.10118.pdf); not sure why I haven't managed to implement that yet here!...