FolkDuet: When Counterpoint Meets Chinese Folk Melodies

Introduction

This is the official implementation of When Counterpoint Meets Chinese Folk Melodies (NeurIPS'2020) paper.

In this work, we propose a system named FolkDuet to automatically generate countermelodies for Chinese folk melodies, modelling the counterpoint concept in Western music theory while maintaining the Chinese folk style. FolkDuet is designed to support real-time human-machine collaborative duet improvisation, hence the algorithm is causal.

Dependencies

the following python packages are required

torch==0.4.1 (we do not know why, but torch 1.x will give different (and of course worse) results, so please use torch 0.4.1)
numpy
music21
glog

Usage

How to run

1.train Bach-HM and Bach-M model

python3 main_note.py --arch BachHM --batch_size 256 --lr 0.05 --nfc_left 256 --nhid 128 --exp_dir results/bachHM
python3 main_note.py --arch BachM --batch_size 256 --lr 0.05 --nfc_left 512 --nhid 256 --exp_dir results/bachM

2.train initialization models for Generator and StyleRewarder

python3 main_note.py --arch Generator --batch_size 512 --folk --lr 0.1 --nfc_left 512 --nhid 256 --exp_dir results/generator_init --raw
python3 main_note.py --arch StyleRewarder --batch_size 512 --folk --lr 0.05 --nfc_left 512 --nhid 256 --exp_dir results/style_init

3.use RL and IRL to train the Generator and StyleRewarder

python3 irl.py --bach_both results/bachHM --bach_self results/bachM --check_dir results/generator_init --reward_dir results/style_init --exp_dir results/irl --raw

How to sample music

python3 sample.py --check_dir results/pretrained

Citations

Please consider citing our paper in your publications, if the project helps your research. BibTeX reference is as follows.

@article{jiang2020counterpoint,
  title={When Counterpoint Meets Chinese Folk Melodies},
  author={Jiang, Nan and Jin, Sheng and Duan, Zhiyao and Zhang, Changshui},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

License

This code is freely available for free non-commercial use, and may be redistributed under these conditions. Please, see the LICENSE for further details. Third-party datasets and softwares are subject to their respective licenses.

jin-s13/FolkDuet