Camelyon17 Loss Function

[niki-amini-naieni]

Hello,

Can someone please confirm that this is the loss function (with the same arguments) that was used for the experiments on the Camelyon17 dataset in the original WILDS paper?

It seems as though to transform the output to a scalar, you eventually compute the mean:

I am wondering why then you set reduction='none' instead of leaving reduction=mean, which is the default for PyTorch.

Thanks for all the help!

[teetone]

Hi @niki-amini-naieni, yes, that's correct.

We don't initialize the loss with reduction=mean because we use the unreduced loss to compute group-wise metrics for GroupDRO (see def _compute_group_wise under def _compute).

[niki-amini-naieni]

Thank you very much!