/music-transformer

Primary LanguageJupyter NotebookMIT LicenseMIT

Music Transformer with 🤗 Transformers

This project implements the Music Transformer, a powerful attention-based neural network for generating long pieces of music with improved long-term coherence. The Music Transformer is built using the Hugging Face 🤗 Transformers library, specifically the LLaMA model, which allows for efficient training and generation of high-quality music.

The model is trained using PyTorch Lightning for efficient and scalable training, and Hydra is employed for configuration management.

Dataset

The model is trained on a combination of the Lakh MIDI Dataset and/or the MAESTRO dataset, providing a diverse collection of musical pieces for learning the structure and patterns in music.

Features

  • Utilizes the powerful LLAMA model from the Hugging Face Transformers library
  • Trains the model using PyTorch Lightning for efficient and scalable training
  • Employs Hydra for configuration management, making it easy to experiment with different settings
  • Supports training on the Lakh MIDI Dataset and the MAESTRO dataset

Installation

  1. Clone the repository:

    git clone https://github.com/georgi/music-transformer.git

  2. Install the required dependencies:

    pip install -r requirements.txt

Usage

  1. Prepare your training data:

    python preprocess.py lmd
python chunker.py

  2. Train the Music Transformer model:

    python run.py

  3. Generate new music pieces: TBD

  4. Enjoy the generated music!

Contributing

Contributions are welcome! If you find any issues or have suggestions for improvements, please open an issue or submit a pull request.

License

This project is licensed under the MIT License.

Acknowledgements