Repository navigation
Feature/baseline updates - #68
Closed
laurenceyoon wants to merge 9 commits into
Closed
laurenceyoon wants to merge 9 commits into
laurenceyoon wants to merge 9 commits into
Conversation
CQTProcessor, CQTSpectralFluxProcessor (OPHMM) and RawSpectrumProcessor (SKF) passed dtype=np.float32 to librosa.cqt / librosa.stft. librosa 0.11 then returns a real array, dropping the imaginary part, so np.abs gave |Re X| instead of |X| (about 36% mean magnitude error). Compute the complex transform and cast its modulus to float32. OPHMM's score templates already used the complex modulus.
The audio OPHMM used the symmetric +-3 transition table of the MIDI model
(Nakamura et al. 2014, Table 3, with the +1 entry mistyped as 0.94531 for
0.84531), filled the whole matrix with 1e-6 outside the band that the forward
step uses, boosted the exit probability by spectral flux, and defaulted to
s_j = 1e-5. Follow the audio paper instead (Sec. III-B, IV-B2): nbh(i) =
{j; 0 <= i - j <= 2} with a_{i,i+2} = 1e-50 and no backward moves, no mass
outside the band, no flux boost, and s_j = 1e-100 (Sec. IV-C).
build_chord_sequence measured note lengths in score beats divided by 4, which is wrong whenever the beat is not a quarter note; use onset_quarter. The tempo read from a score marking now honours the marking's unit (e.g. half = 100 is 200 quarter-note BPM).
arzt_tempo: frame-level Arzt OLTW with the simple tempo model of Arzt & Widmer (SMC 2010), computed from a backward path (dtw_loop records backpointers). arzt_multi_fold: for folded scores, one Arzt matcher per unfolded version of the score, outputting the matcher with the lowest recent cost (adapted from Arzt, Widmer & Dixon, ECAI 2008, and Widmer & Arzt, Vienna Talk 2010).
gt_from_match took performance onsets from note_array()'s onset_sec, which
partitura stores as float32. A MIDI follower stamps its updates with the
float64 onset it receives, so the GT time of the same onset could fall a few
tens of nanoseconds earlier, and the perf -> score lookup ("last position at
or before t") then missed the update made at that onset and scored the
previous event. In the eval fold about a third of MIDI onsets were scored one
event late this way; 6/8 scores, where one event is up to a beat, failed the
tracking criterion on paths that are correct. Use the PerformedPart notes'
note_on (float64) instead: GT and follower times now coincide exactly.
The bottom-level self-transitions took the chord durations from score beats times 60 / tempo. The tempo is quarter-note BPM, but score beats follow the time signature, so in 6/8 (eighth-note beats) every chord was expected to last twice as long, and in 2/2 half as long. Use onset_quarter; the reported score positions stay in beats.
…notes A .match file and the MIDI file a follower streams can store the same notes at different tick resolutions ((n)ASAP: 480 vs 384 ppq), so their onsets differ by up to half a tick. A GT time a fraction of a millisecond before the follower's update was scored against the position before that update. resolve_gt/gt_from_match take the streamed performance and use its onsets when every matched note has the same id and pitch in it.
ParticleFilterMidiProcessor returned the middle of the polling window, which precedes the onsets in the window's second half and follows those in its first half. Use the first onset of the frame, as the other MIDI processors do.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Baseline fixes and two new audio baselines for the TISMIR 2026 benchmark reruns.
dtypedropped the imaginary part).D1=0,D2=2),a_{i,i+2}=1e-50, no flux boost, class defaults_j=1e-100.gt_from_match: GT onsets from float64note_oninstead of float32onset_sec.arzt_tempo(Arzt & Widmer 2010) andarzt_multi_fold(needsunfold_score=False).run_examples.py: options for piece, files, unfolding (now off by default) and method kwargs.dtw_loop.pyx: optionalbackpointersrecord the backward path used forarzt_tempo's tempo estimate; rebuild the Cython extension.methods.yamlkeeps audio OPHMMs_j: 0.0on purpose (benchmark setting for performances without repeats).matchmaker_benchmarkfeature/baseline-updates, which needsarzt_tempoindevelop.