Closed kukas closed 5 years ago
Thanks for catching this. Happy to merge after @rabitt or @justinsalamon confirm this is correct behavior.
Gave it a quick look (it's late in Europe), in principle fix looks correct, but will try to give a second look tomorrow to confirm.
Didn't have time today, but note that the regression tests fail (due to the change in the chroma metric). I imagine that, assuming the fix is confirmed to be correct, the regression data will have to be updated as part of this PR before it can be merged.
See my comment above, but in summary:
(1) the fix is almost correct, but doesn't work correctly when estimates are reported as negative frequencies.
(2) the same bug exists for raw_pitch_accuracy
, but in practice, we probably haven't hit any cases where the value of the metric was affected.
Good catch, @kukas do you mind implementing @rabitt's version and then we can merge?
Good catch, @kukas do you mind implementing @rabitt's version and then we can merge?
Sure, I will fix it on Thursday, if it can wait. Thanks @rabitt for the review!
Thanks @rabitt and @juanjobosch for reviewing this and good catch on the neg case (sorry, totally swamped over here with on-boarding).
@kukas note @rabitt renamed correct_voicing
to matching_voicing
, which I think is a semantically more helpful name :)
Now it should be fixed, I used @rabitt `s patch and applied it also to raw pitch accuracy.
@kukas thanks, you'll need to fix the PEP-8 errors before Travis lets you pass though. https://travis-ci.org/craffel/mir_eval/jobs/490483158#L9358
@rabitt mind giving this a look and LGTM?
Thanks!
Fix #311