Heuristic pitch finding fails sometimes (can be reproduced)

[JoyfulGen]

Occasionally, when running an e2e OMR with an MS73 folio, the Heuristic_Pitch_Finding job with a message like this:

RuntimeError: Image view dimensions out of range for data nrows 42 offset_y 511 data nrows 41 
data offset_y 512 ncols 56 offset_x 2560 data ncols 252 data offset_x 2560 

The numbers change, but the general message is always the same. We don't know why this is happening, but we do know this:

• We can't predict when the job will fail, but if the job does fail it will fail consistently with the same resources.
• Whether the job fails might be related to how well the layer separation happens. We have two sets of layer separation models for MS73. If we run a folio with its "correct" models, the Heuristic job will probably work. If we run the same folio with the other set of models, the Heuristic job will probably fail.

Folio 133 is an example of this. If you put these resources into the Heuristic_Pitch_Finding job, the job will run correctly:

Miyao Staff Finding - 133 finished.json
Non-Interactive Classifier - 133 finished.xml.zip

And if you put these resources in instead, the job will fail:

Miyao Staff Finding - 133 failed.json
Non-Interactive Classifier - 133 failed.xml.zip

I was bold and put a high priority, because we can't run e2e OMR workflows if this job is failing.

[JoyfulGen]

Confirmed that this issue is only on production!

[kyrieb-ekat]

I have reproduced the above items on production using the items Gen provided, and using the 133 folio. I also then experimented a bit and used the consistently failing model in a run, but with a halved image; it then ran successfully!

[JoyfulGen]

Unfortunate update: the issue is on staging as well.

Here are resources that will make Heuristic fail on staging:
Miyao Staff Finding - 261 heuristic fail.json.zip
Non-Interactive Classifier - 261 heuristic fail.xml.zip

The folio is 129v-261