Many-to-Many Symbolic Multi-track Music Genre Transfer

This the pytorch implementation of StarGAN-based model to realize music style transfer between different musical domains with a single generator.

Dependencies

• Python 3.6+
• pytorch 1.0
• librosa
• tensorboardX
• scikit-learn
• numpy 1.14.2
• pretty_midi 0.2.8
• pypianoroll 0.1.3

Usage

Dataset

Link for Desert Camel MIDI Dataset adapted for the training and testing of the model: https://drive.google.com/open?id=1QZP1OCTZnAwasmsglbBxXpJs6C8kbT-A

Train

python main.py

Convert

python main.py --mode test --test_iters 200000 --src_style bossanova --trg_style "['rock','funk']"

Network Architecture

Generator

Discrminator

Domain Classifier

[Reference]

tensorflow StarGAN-VC code

pytorch-StarGAN-VC code

StarGAN-VC paper

Symbolic Music Genre Transfer with CycleGAN code

Symbolic Music Genre Transfer with CycleGAN paper

MGTGAN paper