MusicDiff crashes on some input

[rigaux]

Hi there

We start running musicdiff on our OMR results. We compare two MEI: a first one, produced by our fellow musicologists, and another one, produced by a workflow OMR -> music21 -> MEI using your converter21 package. We would like to use (maybe with some adaptation) the metric OMR-NED published this year at ISMIR.

We obtain the following error on one of our inputs:

Traceback (most recent call last):
File "/Users/philippe/Documents/workspace/utilities/diff/diff_utils.py", line 101, in
main()
File "/Users/philippe/Documents/workspace/utilities/diff/diff_utils.py", line 75, in main
Visualization.mark_diffs(score1, score2, diff_list)
File "/Users/philippe/.pyenv/versions/neuma_env/lib/python3.12/site-packages/musicdiff/visualization.py", line 679, in mark_diffs
note1 = noteOrChord1.notes[op[4]]
^^^^^^^^^^^^^^^^^^
AttributeError: 'Note' object has no attribute 'notes'. Did you mean: 'sites'?

I send you both MEI files attached

predicted.mei.xml

ground_truth.mei.xml

Thanks for your help

Philippe

[gregchapman-dev]

Thank you, Philippe. I'll look at this soon. My guess (without looking at the code) is that there is a missing check for isinstance(noteOrChord1, m21.chord.Chord) before referencing .notes. Or that a non-Chord should never have gotten to this point in the code.

[gregchapman-dev]

Ah! Can you try updating the main branch again? I have a proposed fix there. (A non-Chord should never have taken that code branch.)

[rigaux]

Yes it works ! Thanks a lot.

[gregchapman-dev]

Oh dear. I am going to need to revert that change. I had a "how did this ever work?" moment, and looked deeper. The issue is caused by xml:id problems in predicted.mei, where I see:

                     <staff n="3">
                        <layer xml:id="2" n="1" label="2">
                           <note xml:id="nh_853_2720" dur="2" pname="g" oct="3" stem.dir="up" color="#a637b0" />
                           <note xml:id="nh_1072_2716" dur="2" pname="g" oct="3" stem.dir="up" color="#a637b0" />
                        </layer>
                        <layer xml:id="3" n="2" label="3">
                           <note xml:id="nh_853_2751" dur="2" pname="e" oct="3" stem.dir="down" color="#5034D4" accid="f" />
                           <chord xml:id="nh_1072_2716" dur="2">
                              <note xml:id="nh_1073_2761" pname="d" oct="3" stem.dir="down" color="#5034D4" />
                              <note xml:id="nh_1072_2716" pname="g" oct="3" stem.dir="down" color="#5034D4" />
                           </chord>
                        </layer>
                     </staff>

Note that "nh_1072_2716" is the xml:id of a standalone note, and a chord, and a note in that chord. musicdiff (and XML in general) requires that xml:id be unique within the document. The crash occurs when musicdiff is trying to color the note in the chord red. musicdiff looks it up by xml:id in the music21 score, and the first GeneralNote found with that id is the standalone note. Then musicdiff (expecting to find a chord) tries to color a particular note within that chord, and crashes because what should be "that chord" is actually the standalone note, which has no notes within.

So you need to fix your predicted MEI to have unique xml:ids throughout.

I will apply a different fix to musicdiff, so it will not crash, but it will not ever be able to color the correct notes when there are non-unique xml:ids. It will either color the wrong note, or it will not color any note at all.

[rigaux]

Oups sorry about that, definitely our mistake. Will be fixed tomorrow;

[gregchapman-dev]

Thanks!