CSCI 374 Final Project: Machine and Music (M&M)

Yuanzhe Liu, Kumo (Yiyun) Shao, Han Shao

This code implements multi-layer Recurrent Neural Network for training music generation models. In other words the model takes MIDI files as input and trains a Recurrent Neural Network that learns to predict the next message following the previous message sequence.

Requirements

This code is written in Python and requires Tensorflow 1.15 and Mido libraries. You can install Tensorflow 1.15 and Mido through

$ pip3 install tensorflow==1.15
$ pip3 install mido

Usage

Data

All input MIDI files should be stored in an input directory. You'll notice that there is an example dataset included in the repo (dataset) which consisted of 15 songs from performer Jean-Selim Abdelmoula.

Training

Start training the model and generating music using mm.py.

$ python3 mm.py --data_dir=DATA_DIR

Music generated after each 10 epochs will be stored in results/DATA_DIR.

Checkpoints. While the model is training, it will periodically save models to folder resuls/DATA_DIR/model for each 10 epochs.