GraphMuse Documentation
GraphMuse is a research toolkit for graph-based deep learning on symbolic music. It provides heterogeneous graph representations for scores, sampling utilities, and graph neural network architectures that are tailored to music analysis tasks.
Quick Links
Features
Score graph creation pipelines with accelerated C extensions for performant preprocessing.
Sampling strategies and dataloaders optimised for symbolic music graphs.
A collection of graph neural network layers and models designed for metrical, cadence, and hybrid tasks.
Utilities to generate synthetic graphs for experimentation and prototyping.
Repository Structure
graphmuse– Library source code. -io– Input/output utilities for interacting with score files. -loader– Sampling dataloaders for heterogeneous music graphs. -nn– Graph neural network modules and reference models. -samplers– Interfaces to high-performance C sampling routines. -utils– Helpers for graph creation and experimentation.include– C headers that back the sampling extension.docs– Documentation sources prepared for Read the Docs.tests– PyTest-based regression tests.
Processing Pipeline
The typical workflow when using GraphMuse is:
Create a score graph from symbolic music (e.g., MusicXML) via
graphmuse.create_score_graph().Sample sub-graphs with
graphmuse.loader.MuseNeighborLoaderor other strategies.Train music-specific graph neural networks from
graphmuse.nn.Evaluate the resulting model on held-out data.
The high-level GraphMuse pipeline.