Rodan or Mothra-Rodan for iterative neume-shape classifying

[kyrieb-ekat]

Placing this here to have a log of decisions and discussion. Downstream (Summer 2026) we will be needing to iteratively train a model to recognize neume symbols and classify them. The Interactive Classifier (IC) in Rodan currently does this, but has several drawbacks:

• memory currently becomes an issue, and it becomes quite slow if you have too many neume shapes and components
• it has some quirks (see https://github.com/DDMAL/Rodan/issues?q=is%3Aissue%20state%3Aopen%20IC )

Where I am curious as to the allotment of time and resources to fixing the IC as it stands in its current state inside Rodan, or do we want to take the bones of what we have and build something up using our current resources and approaches.

[fujinaga]

Can you clarify what you mean by "take the bones of what we have and build something up using our current resources and approaches."?

[kyrieb-ekat]

By "bones of what we have" I mean our current script(s) and infrastructure that can either be improved or replaced with more recent libraries, or approaches (is our previous annotation approach still the best way to go? are we able to get away with different memory/processing power now with improved machines/instances/etc?). Not moving away from kNN, but that with claude code etc we can potentially revamp faster than we could have five years ago, if we did want to make a new version.

[fujinaga]

Thank you for the explanation. We should try to fix one component at a time, so that the entire workflow wouldn't break.