There might be a small bug in optimize_prior_precision

[ruili-pml]

Hi,

There might be a small bug in

Laplace/laplace/utils/metrics.py

Line 35 in 6b0618a

targets = targets.view(-1)

as for classification, it requires the target to be the label, instead of one hot vector.

Since in the train_loader it requires one hot vector (or maybe the label also work, in my case I only feed into the one hot vector), it might be consistent to let the val_loader require one hot vector or make both format work?

Best,
Rui

[ruili-pml]

Btw, at

Laplace/laplace/baselaplace.py

Line 468 in 6b0618a

interval = torch.logspace(log_prior_prec_min, log_prior_prec_max, grid_size)

as the default base for torch.logspace is 10 (https://pytorch.org/docs/stable/generated/torch.logspace.html), the grid search is in a super large range. I was wondering is it on purpose or? Feels like search in the base of e is more natural.

[wiseodd]

Sorry I missed this issue!

I don't quite get the first issue: In the doc, we specified that the targets tensor follows PyTorch's CrossEntropyLoss convention, i.e. it is an integer tensor of (...) or (..., 1) where ... indicates any leading dimensions (see e.g. this).

As for your second question, I think it's a good point to be more flexible when defining the interval. E.g., we could pass the base parameter of torch.logspace as a method parameter of optimize_prior_precision.

Feel free to open a pull request! (I will eventually work on this, but I have so much in my queue.)