MMM: Exploring Conditional Multi-Track Music Generation with the Transformer and the Johann Sebastian Bach Chorales Dataset.
Implementation of the paper "MMM: Exploring Conditional Multi-Track Music Generation with the Transformer" (paper). Uses OpenAI's GPT-2 to compose music.
Contact.
Find me on LinkedIn and say hello.
If you find and issue or have a feature request, report either here on GitHub.
Please be so kind and star the repository if you find it useful.
Acknowledgements.
This repository has been created in cooperation with Pyoneer. I am very grateful!
About.
This repository allows you to train GPT-2 on the Johann Sebastian Bach chorale dataset. You can train both MMMTrack and MMMBar from the paper.
How to run.
Requirements:
pip install transformers
pip install tokenizers
pip install torch
pip install music21
pip install note_seq
Training:
- Clone this repository
git clone https://github.com/AI-Guru/MMM-JSB.git. - Train MMMTrack with
python train_jsb_mmmtrack.py. - Train MMMBar with
python train_jsb_mmmbar.py.
Sampling: Run the jupyter notebook.
Training should take roughly one hour on a GPU per model for the JSB dataset.
Pretrained checkpoint.
A pretrained network can be found here: https://ai-guru.s3.eu-central-1.amazonaws.com/mmm-jsb/mmm_jsb_checkpoints.zip
What is missing?
- TensorFlow support is rudimentary.
- Data preprocessing and training on the Lakh dataset.
- Implementation as a tool or a DAW plugin.
License.
Released under the Apache-2.0 License.