Chord Types and Ornamentation

This repository contains the code and data used for the paper "Chord Types and Ornamentation".

Authors: Christoph Finkensiep, Petter Ericson, Sebastian Klassmann, Martin Rohrmeier.

Data

ABC+ (DCML)

Based on the DCML harmonic annotation standard. Includes

• Beethoven, String Quartets
• Mozart, Piano Sonatas
• Corelli, Trio Sonatas
• the Romantic Piano Corpus
  • Beethoven, Piano Sonatas
  • Chopin, Mazurkas
  • Debussy, Suite Bergamasque
  • Dvorak, Silhouettes
  • Grieg, Lyric Pieces
  • Liszt, Années de Pelerinage
  • Medtner, Tales
  • Schumann, Kinderszenen
  • Tchaikovsky, The Seasons

The exact version of the data is included in this repository as git submodules under data/dcml_corpora and data/romantic_piano_corpus. To check out the submodules, either clone the repository with --recures-submodules:

$ git clone --recurse-submodules https://github.com/DCMLab/probabilistic_harmony_model

or run the following command after cloning:

git submodule update --init --recursive

The DCML dataset can be preprocessed using preprocess_dcml.py. The processed chord data can be found in data/dcml.tsv. A list of used pieces has been recorded in data/preprocess_dmcl.txt.

EWLD

The Enhanced Wikifonia Leadsheet Dataset (Simonetta et al. 2018). This corpus is not included in this repository for licensing reasons, but can be requested from the authors for scientific purposes. The preprocessing script expects the unpacked corpus under data/ewld.

The EWLD can be processed using preprocess_ewld.py and the resulting chord data can be found in data/ewld.tsv. A list of used pieces is recorded in data/preprocess_ewld.txt.

Code and Results

The dependencies for the preprocessing and model code are listed in requirements.txt (or requirements-amd.txt for AMD GPUs). An example for how to set up a virtual environment and installying a corresponding Jupyter kernel is given in init.sh. The default version will work with CPUs and Nvidia GPUs. An alternative pip install command for AMD GPUs is included in init.sh. The minimal library requirements without specified version are listed in requirements-any.txt.

The main model is implemented as a Jupyter notebook in model.ipynb. Its output is stored in results/dcml_params.{json,pt} and results/ewld_params.{json,pt}. The model can also be run as a normal Python script (without Jupyter) as model.py. This will show all plots in separate windows at the end of the computation, so running the notebook version is generally preferred. The model can be run on CPU or GPU.

The cluster model is implemented in model_clustering.ipynb / model_clustering.py. Its output is stored in results/clustering/. This notebook is intended to be run with GPU support and therefore contains instructions for how to run it on Google Colab. Running it on CPU is possible but can take much longer.

The plots shown in the paper are generated using plots.ipynb / plots.py and can be found in plots/ directory. Because of mathematical notation in some of the plots, running this notebook requires a LaTeX installation. Alternatively, LaTeX calls can be disabled at the top of the notebook.

Plots of the remaining posterior distributions not shown in the paper can be found in additional-results.pdf.

Licences

All code in this repository (i.e., all *.py and *.ipynb files, init.sh, and requirements*.txt) is licensed under the BSD 3-clause license (see LICENSE-BSD). The licenses of the DCML datasets are provided in the respective subdirectories under data/. All remaining documents, images, and data files are licensed under the CC-BY 4.0 license (see LICENSE-CCBY).