SymRep : A Systematic Evaluation of Symbolic Music Representations

This repo contains the code for the paper: [Symbolic Music Representations for Classification Tasks: A Systematic Evaluation]()

It contains the models and experiments for evaluating three symbolic music representations:

• Matrix-like (Pianoroll)
• Graph-like
• Sequence-like (Using Tokenized representations mainly from MidiTok)

We use two datasets for evaluation:

• ASAP
• ATEPP

To use please download the corresponding datasets and refer to the path in conf/config.yaml

Requirements

• Python 3.6 or newer

To install all python dependencies, run:

pip install -r requirements.txt

Usage

To run the experiments, run the following command and change the parameters:

./experiments/crossval_run.sh

Configuration

The configurations regarding experiment, task, dataset, as well as the three representations, can be modified in conf/config.yaml. The parameters management are supported by Hydra.

Experiments for this repository are logged using Weights and Biases.

Structure

The code is organized into the following folders:

• conf: configuration files
• model: contains the code for the models used in the paper
• experiments: contains the code for the experiments in the paper
• converters: contains the code for converting each symbolic representation
  • miditok: contains the custom tokenizers for musicxml