This repository contains our team project for Data 515 in the University of Washington Masters in Data Science program.

The requirements for the project are described here [access required].

Installation

NOTE: Installation of some of the packages required for EasyMusicGenerator can take significant time. The install process described here can take 30+ minutes.

Clone this repository to your local computer.

git clone https://github.com/PrestonStringham/DATA515-MusicGeneration

Change directories to the local copy of the repository.

cd DATA515-MusicGeneration

The pip installer has difficulty installing some Python packages used for data science. We recommend that you create a custom conda environment using the provided env-music-gen.yml file.

conda env create -f env-music-gen.yml

Activate the newly created conda environment, music-gen-515.

conda activate music-gen-515 

Install the package--and remaining package requirements--using the following PIP command line.

pip install .

If after using EasyMusicGenerator, you'd like to remove the recommended conda environment, use the following commands.

conda activate
conda remove --name music-gen-515 --all

Verify removal.

conda env list

Usage

Make sure that you are inside the DATA515-MusicGeneration directory before running the following code.

from easy_music_generator import easy_music_generator as emg
emg_obj = emg.EasyMusicGenerator()
input_file_path = 'easy_music_generator/music/'
emg_obj.analyze(input_file_path)
output_file_path = 'output/'
number_of_bars = 10
emg_obj.generate(number_of_bars, output_file_path)

This code is also available in examples/example.py, but your must be in the examples directory in order to run that file.

Project Structure

|- DATA515-MusicGeneration/
      |.vscode
      |- doc/
        |-...
      |- easy_music_generator/
         |- music/
            |-...
         |- preprocessor/
            |- tests
                |- __init__.py
                |- test_chord_distribution.py
                |- test_note_distribution.py
                |- test_preprocessor.py
            |- __init__.py
            |- chord_distribution.py
            |- note_distribution.py
            |- preprocessor.py
         |- tests/
            |-__init__.py
            |-test_easy_music_generator.py
         |- __init__.py
         |- __version.py
         |- easy_music_generator.py
         |- pregenerator.py
         |- lakh2_polyphony_rnn.mag  
      |- examples
        |- music
            |- ...
        |- example.py
        |- example_notebook.ipynb
      |- tests
        |- emg-import-test.py
      |- .gitignore
      |- .travis.yml
      |- LICENSE
      |- README.md
      |- requirements.txt
      |- setup.py

DATA

We trained our model on a segment of the Lakh MIDI Dataset.

The Lakh MIDI Dataset is distributed with a CC-BY 4.0 license.

Colin Raffel. "Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-MIDI Alignment and Matching". PhD Thesis, 201.

MODEL

We used Magenta's Polyphony RNN.

--- END ---