js fakes 4bars

Introduction

The JS-FAKES-4BARS model utilizes the GPT-2 language model architecture for music generation. By treating musical pieces as text, the model generates Bach-like chorales, producing 4 bars at a time with four voices: soprano, alto, tenor, and bass. This project demonstrates the potential of using language models for music composition.

Architecture

The model is based on GPT-2, configured with 6 decoders and 8 attention heads per decoder. It features a context length of 512 and embedding dimensions of 512. The vocabulary size is 119, enabling it to generate structured musical outputs.

Training

The model is trained on the TristanBehrens/js-fakes-4bars dataset, focusing on a small corpus of music data. This limited dataset results in significant overfitting, but it serves as a proof of concept for music generation using language models.

Guide: Running Locally

1. Clone the repository and navigate to the project directory.
2. Access the provided notebook, which can be executed in Google Colab for ease of use.
3. Ensure you have the necessary dependencies installed, including PyTorch and the Transformers library.

For optimal performance, especially during model inference, it is recommended to use cloud GPUs such as those available on Google Colab or AWS.

License

Details regarding the license for this model and its usage can be found in the repository's main documentation or LICENSE file. Users should review these terms to ensure compliance with the model's usage guidelines.