Benchamrking symbolic music features

This is the code used for benchmarking different feature sets, including musif. Please, cite us:

Simonetta F., Llorens A., Serrano M., García-Portugués E., Torrente A., "Optimizing Feature Extraction for Symbolic Music", ISMIR 2023.

Dependencies

• Python 3.10 (e.g. via conda or pyenv)
• pdm, you barely have three options:
  • pipx install pdm (need pipx, recommended)
  • pip install pdm (environment specific)
  • see https://pdm.fming.dev/latest/ for other alternatives
• pdm sync to create the environment and install python packages
• Alternatively to pdm, see cluster.md for bare venv approach
• MuseScore: download AppImage (4.0.1 has a bug, use 3.6.2, instead)
• Java: install using you OS package manager and check that the java command is available in the PATH
• jSymbolic 2.2: download and unzip
• GCC and make: install using your OS package manager
• humdrum:
  1. git submodule update
  2. cd humdrum-tools
  3. make update
  4. make

In symbolic_features/settings.py set the paths to MuseScore and jSymbolic executables.

Datasets

Download the following datasets and set the paths to the root of each one in symbolic_features/settings.py

1. Josquin - La Rue
2. ASAP
3. Didone
4. EWLD
5. String quartets:

• Haydn
• Mozart
• Beethoven
• unzip the above three zips into one directory, e.g.: quartets/haydn, quartets/mozart, quartets/beethoven

Preprocessing

Fix invalid file names: pdm fix_names. This will fix names containing , and ; that cause errors in csv files.

Convert any file to MIDI: pdm convert2midi. You will need to run Xvfb :99 & export DISPLAY=:99 if you are running without display (e.g. in a remote ssh session)

Feature extraction

Reproduce experiments: ./extract_all.sh

Detailed commands:

1. jSymbolic: pdm extract --jsymbolic --extension .mid
2. musif:

• pdm extract --musif --extension .mid
• pdm extract --musif --extension .xml
• pdm extract --musif --extension .krn

3. music21:

• pdm extract --music21 --extension .mid
• pdm extract --music21 --extension .xml
• pdm extract --music21 --extension .krn

Classification accuracy

Reproduce experiments: pdm validation

Detailed commands

• pdm classification: run all experiments with original features
• pdm classification --use_first_10_pc: run all experiments with first 10 Principal Components from each task (where a task is a combination of dataset, feature set, and extension)
• pdm plot: plot the AutoML optimization score across time
• pdm classification --featureset='music21' --dataset='EWLD' --extension='mid' --automl_time=60: run an experiment on a single task for 60 seconds