The precision recall curve (and maybe others) should returned a thresholds object with the same size as Xaxis and Yaxis

[Yann-CV]

🐛 Bug

The thresholds returned by the precision recall curve classes/functions (and maybe others) is missing the last value (threshold > max_prediction)

To Reproduce

pr_curve = PrecisionRecallCurve(task="binary")
precision, recall, thresholds = pr_curve(pred, target)
assert precision.size() == recall.size()  == thresholds.size()  ---------> fails

Expected behavior

pr_curve = PrecisionRecallCurve(task="binary")
precision, recall, thresholds = pr_curve(pred, target)
assert precision.size() == recall.size()  == thresholds.size()  ---------> does not fail

Environment

TorchMetrics 1.2

Additional context

This is an issue when you want to display the threshold value for all points in the curve (even though selecting this threshold makes no sense). It is also an issue when you want to create pandas.Dataframe from the returned tuple.

Quick fix on our side: we manually added a max_prediction + epsilon at the end of thresholds

[Borda]

seems as the same issue discovered by @matsumotosan in #2154 (comment)

[SkafteNicki]

This is intended behavior as it is consistent with sklearn:

https://scikit-learn.org/stable/modules/generated/sklearn.metrics.precision_recall_curve.html

[Yann-CV]

@SkafteNicki sorry I did not noticed your answer... Then thanks.

If the point is to be sklearn compatible, then I guess it is a good point to follow their choice.

On my side, because I see Torchmetrics as a substitute to sklearn regarding metric computation, I still personally think this is a bad design choice.