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