Just Play

Live Demo

A CharRNN to create some music.The inspiration comes from MIT Lectures.

Dataset

I used subset of Nottingham Music Database.To be more precise I used the following collections:

• Jigs
• Hornpipes
• Morris
• Playford
• Reels A-C
• Reels D-G
• Reels H-L

Each collection contains 15 to 340 different tunes. In total I used 709 different tunes. I randomly selected 20% of tunes for the validation set.

Network Structure

The neural network have 4 layers.

1. Embedding Layer
   • Input Size:92
   • Output Size: 256
2. Two LSTM Layers
   • Input Size: 256
   • Output Size: 1024
3. Fully Connected Linear Layer
   • Input Size: 1024
   • Output Size: 92

I used Google Colab to train the neural network.

Deployment

I used Flask framework to create the web application. For deployment platform, my choice was Heroku. To run locally, first install necessary packages:

apt-get install abcmidi timidity
pip install flask

Set environment variables:

export FLASK_APP=server.py
export FLASK_ENV=development

Then run:

flask run

I only tested on macOS High Sierra.

TODO

• Add download option for created musics.
• Try different networks.
• Add different networks to web application.
• Use different data sources.(different music genres?)
• Add tests 😄