Skip to content

PyTorch implementation of MuseMorphose, a Transformer-based model for music style transfer.

License

Notifications You must be signed in to change notification settings

MaximeHoude/MuseMorphose

 
 

Repository files navigation

MuseMorphose

This repository contains the official implementation of the following paper:

  • Shih-Lun Wu, Yi-Hsuan Yang
    MuseMorphose: Full-Song and Fine-Grained Music Style Transfer with One Transformer VAE
    ArXiv preprint, May 2021 [arXiv] [demo website]

Prerequisites

  • Python >= 3.6
  • Install dependencies
pip3 install -r requirements.txt
  • GPU with >6GB RAM (optional, but recommended)

Preprocessing

# download REMI-pop-1.7K dataset
wget -O remi_dataset.tar.gz https://zenodo.org/record/4782721/files/remi_dataset.tar.gz?download=1
tar xzvf remi_dataset.tar.gz
rm remi_dataset.tar.gz

# compute attributes classes
python3 attributes.py

Training

python3 train.py [config file]
  • e.g.
python3 train.py config/default.yaml
  • Or, you may download the pretrained weights straight away
wget -O musemorphose_pretrained_weights.pt https://zenodo.org/record/5119525/files/musemorphose_pretrained_weights.pt?download=1

Generation

python3 generate.py [config file] [ckpt path] [output dir] [num pieces] [num samples per piece]
  • e.g.
python3 generate.py config/default.yaml musemorphose_pretrained_weights.pt generations/ 10 5

This script will randomly draw the specified # of pieces from the test set.
For each sample of a piece, the rhythmic intensity and polyphonicity will be shifted entirely and randomly by [-3, 3] classes for the model to generate style-transferred music.
You may modify random_shift_attr_cls() in generate.py or write your own function to set the attributes.

Customized Generation (To Be Added)

We welcome the community's suggestions and contributions for an interface on which users may

  • upload their own MIDIs, and
  • set their desired bar-level attributes easily

Citation BibTex

If you find this work helpful and use our code in your research, please kindly cite our paper:

@article{musemorphose21arxiv,
    title={{MuseMorphose}: Full-Song and Fine-Grained Music Style Transfer with One {Transformer VAE}},
    author={Shih-Lun Wu and Yi-Hsuan Yang},
    year={2021},
    journal={arXiv preprint arXiv:2105.04090},
}

About

PyTorch implementation of MuseMorphose, a Transformer-based model for music style transfer.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%