Exclude thresholds of 0.0 and 1.0 in precision recall curve

[jacanchaplais]

🚀 Feature

Instead of using threshold values in the inclusive range [0, 1], it would be good to have (at least the option for) thresholds in the exclusive range (0, 1). Custom threshold values would be even better, in my opinion.

Motivation

I am tracking metrics for memory intensive models, and don't want to keep track of values which are not useful. Also, it clutters up my TensorBoard, which is mildly annoying.

Pitch

I guess you could just add in an optional boolean argument like

def BinnedPrecisionRecallCurve(..., exclusive: bool = False):
    ...

to exclude the edge values.

Alternatives

If you wanted to allow user specified thresholds, which I would prefer but might be confusing with "binned" being part of the name, perhaps you could allow:

pr = BinnedPrecisionRecallCurve(num_thresholds=None, thresholds=[0.1, 0.5, 0.9], ...)

[SkafteNicki]

@maximsch2 any opinion here?

[maximsch2]

I'd rather directly take thresholds instead of making people figure out what exclusive means. Given people's knowledge of the model they might also want to have unevenly spaced thresholds to save even more memory.