consistency between metrics (specifically F-measure)

[andyr0id]

Hi, thanks for this very useful library.

I have noticed a small inconsistency between some metric function outputs that are shared between some of the submodules. F-measure is one of them.

When evaluating based of F-measure it is often useful to know the precision and recall values. Calculating F-measure in the "onset" submodule returns this information, whereas the "beat" submodule does not.

[craffel]

mir_eval.beat.f_measure only returns F-measure because only F-measure is used by the beat evaluation toolbox and MIREX; mir_eval.onset.f_measure returns all three because MIREX uses all three. I'd potentially support changing mir_eval.beat.f_measure to return all three, but FWIW mir_eval.beat.f_measure and mir_eval.onset.f_measure are functionally identical (except that the threshold/window is .05 and .07 in onset and beat respectively) so you can safely use mir_eval.onset.f_measure in place of mir_eval.beat.f_measure to get all three if you set the window kwarg to .07. At any rate, to implement this change I'd like to get some sort of community consensus first.

[andyr0id]

Yes that's what I've been doing as a work around.

I see your point, but the counter argument is that by open sourcing this library it is now not necessarily tied to the one MIREX implementation, but a community resource.

Either way it's a small issue that will come down to the ethos of the maintainers.