DuetHLSTM: A Hierarchical LSTM Model for Two-Track Music Generation
10-707 Final Project

Jocelyn Huang, Menglan Ji, Daniel Martin

DuetHLSTM is a model that generates two-track, polyphonic music using a
hierarchical LSTM. You can find the training data here:

https://salu133445.github.io/lakh-pianoroll-dataset/

Save the train_x_lpd_5_phr.npz in the same directory as the .py files. To train, run:

python process_data.py

on the training data, which will load the data into .npy format. Then, to train, run:

python music_gen.py

This will output a set of samples at every epoch in .npy format in a
results/ subdirectory, which can be converted to MIDI files using

python midi.py in.npy out.mid

To evaluate using objective metrics, run metrics.py on the .npy samples, which will print metrics and save generated images in the working directory.

python metrics.py in.npy name