Report nan in PR curve above max threshold

[ytang137]

🚀 Feature

Currently the PR curve metrics report (0,0) at threshold greater than the maximum pred confidence. For example:

from torchmetrics.classification import BinaryPrecisionRecallCurve
import torch

preds = torch.tensor([0.3,0.4,0.5,0.6,0.7,0.8])
targets = torch.tensor([0,1,1,0,1,0])
pr_curve = BinaryPrecisionRecallCurve(thresholds=10)
pr_curve(preds, targets)

This code returns the following result:

(tensor([0.5000, 0.5000, 0.5000, 0.6000, 0.5000, 0.3333, 0.5000, 0.0000, 0.0000, 0.0000, 1.0000]),
  tensor([1.0000, 1.0000, 1.0000, 1.0000, 0.6667, 0.3333, 0.3333, 0.0000, 0.0000, 0.0000, 0.0000]),
  tensor([0.0000, 0.1111, 0.2222, 0.3333, 0.4444, 0.5556, 0.6667, 0.7778, 0.8889, 1.0000]))

Since there are no pred confidence>=0.889, I think it makes more sense to return nans for precision and recall. This makes it possible to filter out (0,0) points artificially introduced by the fix threshold grid.

Motivation

Make it possible to filter out (0,0) points artificially introduced by the fix threshold grid in PR curve metrics.

Pitch

Return nans for precision and recall for thresholds greater than the max pred confidence.

Alternatives

Additional context