Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

Adjustment of initial transcription offsets, also affects pitches #1

Closed
00sapo opened this issue Aug 24, 2021 · 1 comment
Closed

Comments

@00sapo
Copy link
Collaborator

00sapo commented Aug 24, 2021

self.perfm -= np.min(self.perfm[:, 1:3])

self._ground_truth -= np.min(self._ground_truth[:, 1:3])

self.score -= np.min(self.score[:, 1:3])

should be like:

 self.perfm[:, 1:3] -= np.min(self.perfm[:, 1:3]) 

Such an error could affect TAFE, EIFE, SEBA, and transcription evaluation. Specifically:

  • Since the transcription first onsets are likely the same as the ground-truth, the transcription evaluation is only partially affected, likely in a negligible way
  • Since in ASMD the first onset of the score is 0, the SEBA method is not affected at all, because it only uses the pitches of the score
  • TAFE and EIFE could may be affected by such error in that they compare the score pitches to the transcription pitches, and this error makes them not comparable; we can expect an improvement in both the two methods

As consequence, the main conclusions of the paper, should not change

  • EIFE is far better than TAFE on piano music
  • SEBA is better than TAFE and EIFE on non-piano music
  • Bytedance model works better than Omnizart model
@00sapo
Copy link
Collaborator Author

00sapo commented Dec 18, 2021

Fixed in 09041c9

New results are being computed.

@00sapo 00sapo closed this as completed Dec 18, 2021
# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant