Classification of Themes and Motifs in Musical Composition with MIDI

A multi-label classifier algorithm to predict motifs/themes in musical composition.

• About the Project
  • Built With
• Getting Started
  • Prerequisites
  • Installation
• Usage
• Roadmap
• License
• Contact

About the Project

This project is for my individual dissertation of my Bachelor's.

Built With

• Python v3.9.7

Getting Started

Prerequisites

Python related dependencies can be installed using:

pip install -r requirements.txt

Installation

Clone the repo:

git clone https://github.com/chuangcaleb/music-theme-recognition

Usage

The project is subdivided into four modules, to decouple the project workflow.

1. collecting_data

There are two stages to collecting the required dataset.

1a. scraping_midi

Different scripts download MIDI files from various sources into a bin directory. Manually downloaded MIDI files can be manually added in here as well.

1b. building_dataset

1. create_db.py goes through the bin directory and builds a database containing the ids of the samples.
2. From here, I have manually added the theme labels as columns, as well as metadata columns such as duplicate.
3. Then, the samples are slowly labelled, marking songs that I've looked through with a 1 in the recognizable column if I have labelled them, and 0 if not (This means that if that field is empty/null, it has not been identified yet).
4. process_db.py converts all 'p's in the database into '1's.¹
5. db_stats.py is a convenience script that returns some statistics about the label dataset so far.

2. calculating_dataset

1. generate_jsymbolic_config.py builds a configuration script based on the MIDI files found in the bin directory.
2. Run jSymbolic with theme_jsymb_config.txt as the configuration script.
3. Finally, run clean_db.py to clean up the database for use. 'definition_dump/py` dumps the definition data from xml to csv.

These three steps can (and should) be automatically executed.

Here is a script file that I've used — modify it to point to your jSybolic2.jar.

python3 calculating_dataset/generate_jsymbolic_config.py

java -Xmx3072m -jar [PATH_TO_YOUR_JSYMBOLIC]/jSymbolic2.jar -configrun data/features/theme_jsymb_config.txt

python3 calculating_dataset/clean_db.py
python3 calculating_dataset/definition_dump.py

3. building_model

From here, it is mostly automated.

model.py is the main script to run. You should never need to fiddle with it because the parameters can all be configured with config.py.

4. evaluating_results

Similarly to the building_model module, a config.py file handles configurations for this module.

4a. Process Results

results_stats.py calculates relevant statistics about the result set.

flatten_json.py is a temporary utility script that converts the json results into a flat table in csv.

4b. Analyze Results

plot_trees.py draws the best Decision Trees from the specified run.

feature_importances.py shows a plot of the feature value distributions for each theme label for the top 10 most-important features as calculated from the best Random Forest, from the specified run.

graph.py plots results from the specified run.

Roadmap

See the kanban for active tasks.

License

Distributed under the MIT License. See LICENSE for more information.

Contact

20204134 Chuang Caleb hcycc2

Footnotes

1. This is because I have sped up the hand-labelling process by marking fields with '0' or 'p', since they are closeby on the keyboard. The script later turns the 'p's into '1's. ↩