OvO support on AUROC

[chris-tkinter]

🚀 Feature

Hi in the sklearn API there is a multi_class parameters which can be set to ‘ovr or ‘ovo'. But it seems that this is not yet implemented in the torchmetrics.

Motivation

In sklearn we can compute ovo AUC, but currently this is not supported

tm = AUROC(num_classes=7, average='macro')
x = torch.randn(100, 7).softmax(1)
y = torch.randint(7, (100, ))
print(tm(x, y), roc_auc_score(y, x, average="macro", multi_class="ovo"))

Additional context

Using torchmetrics 0.6.0

[SkafteNicki]

Issue will be closed by classification refactor: see this issue #1001 and this PR #1195 for all changes

Small recap: This issue ask for support that one-vs-one approach for calculating auroc in multiclass settings. During the refactor it was determined that we do not want to pursue this approach and still only support one-vs-the-rest for calculating roc, auroc etc.

Issue will be closed when #1195 is merged.