Redefining default output representations and metrics

The paper originally described the six conventional tasks, and the alternative method based on the 75-most common Roman numerals.

A Roman numeral can also be reconstructed by predicting the notes conforming the chord. This has become the main approach in this version.

Some other additions include:

• An updated README with instructions for inference and training
• A notebook to run an inference demo
• An updated version of the dataset with the new output representations

Minor improvements

This is generally the same version of the AugmentedNet network presented in the 2021 paper, with some minor improvements.

• The HarmonicRhythm representation was improved to account for imbalanced classes. This provides better chord segmentation.
• A new representation based on Closed-Position chord outputs is added to the outputs. The number went from 11 tonal tasks to 14. These new representations are used to predict the final Roman numeral label, instead of all previous methods
• Some minor bugfixes

Full Changelog: v1.0.0...v1.1.0

Camera-ready submission

This is the version of the code used for all the revised experiments in the camera-ready paper.

Accompanying this release will be the experiment logs, pre-processed data, and data splits of the paper.

One (and maybe the only) substantial difference between the architecture submitted for the initial submission and this one, is the use of rmsprop as the optimizer, instead of adam.

I found that rmsprop was generally better and decided to use it as the default.

Every revised experiment in the camera-ready paper uses rmsprop. As observed before, the results went slightly above just because of that.

Submitted version (May 2021)

This is the commit version accompanying the paper submission.

The code is not "production ready", but it is the code used to produce the final results in the paper submitted to ISMIR 2021.

The plan is to clean this code (without making changes to the architecture/output of the model) before releasing to the public.

Initial version

This version is the result of experiments that have been going for a while.

It is not stable yet, but I am planning on removing several of the experimental scripts I wrote, mostly when I was validating the data from different collections such as ABC, TAVERN, Haydn Op. 20, and When-in-Rome.

I have a pretty standard workflow for pairing scores and annotations nowadays, so most of these scripts (e.g., measure_alignment.py) just clutter the repository.

Nothing major should have been omitted from this release.