p2333 / rectified-rejection Goto Github PK

Hi,
I read your paper about Rectified Confidence recently.
It's very interesting and compelling, especially the notion of T-Con oracle and the novel mild condition.
In utils.py, I found the implementation of calculation of TPR-95 accuracy, i.e., the 'calculate_FPR_TPR' function.
I feel confused about the last three lines:

    index_c = (torch.where(correct > T)[0]).size(0)
    index_w = (torch.where(wrong > T)[0]).size(0)
    acc = index_c / (index_c + index_w)

It seems that, TPR-95 accuracy is TP / (TP + FP).
Could I assume that low-confidence decisions, correct or wrong, are all excluded in this case?

In addition, before I read the codes, I thought a fixed threshold 1/2 would be applied to R-Con according to Theorem 1.
But the threshold $t$ in paper is exactly the 'T' in line 49 ?

Thanks.

Hi! Thank you for your intresting work!
Is there lack of the implementation code for ATRO_loss which is required in train_cifar.py ?
It would be appreciated that adding the code for it.