Theme Transformer

<p> <a href=""><img alt="LICENSE" src="https://img.shields.io/github/license/atosystem/ThemeTransformer"/> </a> <a href=""><img alt="STAR" src="https://img.shields.io/github/stars/atosystem/ThemeTransformer"/> </a> <a href="https://github.com/atosystem/ThemeTransformer/issues"><img alt="ISSUE" src="https://img.shields.io/github/issues/atosystem/ThemeTransformer" /></a> <a href="https://github.com/atosystem/ThemeTransformer/pulls"><img alt="PR" src="https://img.shields.io/github/issues-pr/atosystem/ThemeTransformer" /></a> </p>

This is the official implementation of Theme Transformer.

Checkout our demo and paper : Demo | arXiv

Environment:

• Clone this Repo

  git clone https://github.com/atosystem/ThemeTransformer.git -b main --single-branch
  

• using python version 3.6.8

• install python dependencies:

  pip install -r requirements.txt

To train the model with GPU:

python train.py --cuda

To generate music from theme

python inference.py --cuda --theme <theme midi file> --out_midi <output midi file>

Details of the files in this repo

.
├── ckpts                   For saving checkpoints while training
├── data_pkl                Stores train and val data
│   ├── train_seg2_512.pkl
│   └── val_seg2_512.pkl
├── inference.py            For generating music. (Detailed usage are written in the file)
├── logger.py               For logging
├── mymodel.py              The overal Theme Transformer Architecture
├── myTransformer.py        Our transformer revision code 
├── parse_arg.py            Some arguments for training
├── preprocess              For data preprocessing  
│   ├── music_data.py       Theme Transformer pytorch dataset definition
│   └── vocab.py            Our vocabulary for transformer
├── randomness.py           For fixing random seed
├── readme.txt              Readme
├── tempo_dict.json         The original tempo information from POP909 (used in inference time)
├── theme_files/            The themes from our testing set.
├── trained_model           The model we trained.
│   └── model_ep2311.pt
└── train.py                Code for training Theme Transformer

Citation

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

@article{shih2022theme,
  title={Theme Transformer: Symbolic Music Generation with Theme-Conditioned Transformer},
  author={Yi-Jen Shih and Shih-Lun Wu and Frank Zalkow and Meinard Müller and Yi-Hsuan Yang},
  journal={IEEE Transactions on Multimedia},
  year={2022},
  publisher={IEEE}
}