Multi-label ROCs

[FlorianMF]

🚀 Feature

Similary to issue #100 it would be nice to be able to make roc work with multi-label inputs.

Motivation & Pitch

auc and hence _auroc_compute do work with multi-label inputs and return a AUROC value for each label/class by iterating over range(num_classes) when passing average=None.
_roc_compute and hence roc differentiate only between binary and multi-class (by checking if num_classes == 1)

I would expect _roc_update to similarly return a mode using _input_format_classification(preds, target) and return a list of [fpr, tpr, threshold] of length=num_classes.
The easiest would be the format [[fpr, tpr, thres]]*5

[github-actions]

Hi! thanks for your contribution!, great first issue!

[FlorianMF]

a possible implementation would be :

if mode == 'binary':
   ...
elif mode == 'multi-label':
    output= [roc(preds=preds[:, i],
             target=target[:, i],
             num_classes=1,
             pos_label=1,
             sample_weights=sample_weights)
                      for i in range(num_classes]
     fpr = [o[0] for o in output]
     tpr = [o[1] for o in output]
     thresholds= [o[2] for o in output]
else:
...