TONNETZ-CAD

A Data-Set for GDL containing heterogeneous graphs produced by Tonnetz trajectories with cadence labels

Overview

Cadence and Voice Leading Detection in Symbolic Classical Music is a challenging task. This Repository provides Datasets with different representations for applying Graph learning.

A typical Heterogenous graph modelling of the musical score

The Tonnetz Representation with Bass and Soprano Connections

Motivation

The original dataset contains the fully annotated scores but it presents some challenges for the task.

• Boudaries of Cadences The context boundaries of the cadences are ambiguous we solve this buy providing preassinged labels.
• Complicated Representation Essencially the format is enriched XML scores and tsv tables. We needed a setting that can be put to use directly for cadence segmentation in a Machine learning context.

We developed TONNETZ-CAD to address these issues.

Getting the dataset

Here some minimal examples of how to hands on the dataset:

Node Classification

from sripts import MPGD_cad

# Load Dataset from csv urls
dataset = MPGD_cad()
# Get the heterogenous graph (a big graph like a social net)
hetero_graph = dataset[0]

# The node type to predict
category = dataset.predict_category
# The number of classes
num_classes = dataset.num_classes
# A predifined train_mask
train_mask = hetero_graph.nodes[category].data.pop('train_mask')
# A predifined test_mask
test_mask = hetero_graph.nodes[category].data.pop('test_mask')
# Labels of the predict type nodes
labels = hetero_graph.nodes[category].data.pop('labels')
# Print the edge types of the graph
print(hetero_graph.e_types)
>>>["follows", "during", "onset"]

Graph Classification with Tonnetz Representation

import requests, pickle

url = 'https://github.com/melkisedeath/tonnetzcad/raw/master/graph_classification/t345.pkl'
r = requests.get(url, allow_redirects=True)
open('./t345.pkl', 'wb').write(r.content)

with open('./t345.pkl', 'rb') as handle:
    data = pickle.load(handle)
    
data.keys()

>>> dict_keys(['x', 'x_test', 'y', 'y_test', 't', 'templates'])  # these are NumPy arrays

A slightly better way to do things is to clone this repo and then use the get_dataset method in data.py to do essentially the same thing.

Constructing the dataset

This is a synthetically-generated dataset which, by default, consists of 32 fully annotated Sonatas. The test sample contains 10 Sonatas but you can change that if you wish. Each example contains either the entire Tonnetz trajectory of a Sonata as graph in the Node Classification folder or a 2 bar segment Tonentz trajectory of a Sonata in the Graph Classification folder. You can of course choose in which Tonnetz you prefer the representation. These patterns are analogous to the cadences in the original Mozart Piano Sonatas Data-Set.

Dependencies

• NumPy
• PyTorch
• DGL
• pandas