music_transformer

Music Generation

• Model learns notes and durations from the midi files
• Difference from the original algorithm in Ref [1] is that instead of keeping only the last note from each chord, we keep the highest 4 dropping the same lower notes that are in different octaves.
• The drawback is that, if notes in a chord have different durations, longer notes won't be sustained, but replayed. This will result in a characteristic hammering pattern. Also, durations are not learned as good as notes, resulting in a jazz like pattern.
• The vocabulary consists not only of single notes (singlets) but of duplets, triplets and quadruplets of notes. However, since not all combinations of notes are likely the total size of the dictionary is significantly smaller than the combinatorial value.
• This repo comes with a pre-trained model
• The midi output is generated inside the data/generated folder.

Installation

Clone this repo via

https://github.com/iohanngrig/music_transformer.git

Create a conda environment from .yml file

conda env create -f environment.yml

or

Create a new conda environment (versions matter)

conda create -n env python=3.9

Install Tensorflow specific packages (select your version of cuda)

conda install -c conda-forge cudatoolkit=11.8 cudnn=8.1.0
pip install "tensorflow<2.11"

Install MuseScore desktop software, and after music21 library:

pip install music21
python -m music21.configure 

wait until the configuration finds the path to MuseScore4.exe file.

Install the rest of the packages (repetitions are present):

pip install -r requirements.txt

Generation

You can generate via generate.py script

python generate.py

or with optional parameters, e.g. python generate.py --number [number of generated samples: int>0] --length [multiple of generating length: float>0] --temperature [variability: float>0] --seed [random seed: int>0] --draw [draw/show musical score: 0 or 1]

Dataset

The dataset consists of works by J.S. Bach. Midi files for training are placed inside the data/midi folder.

Training

You have to train transformer by running

python train_transformer.py

monitor training by running tensorboard --logdir runs/logs/train.

If training from scratch or on a new dataset, modify config.json, and set PARSE_MIDI_FILES: true and LOAD_MODEL: false.

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

References

[1] This work is based on the original code developed by David Foster inside the following repository Generative_Deep_Learning_2nd_Edition that contains codes for the eponymous textbook.

License

This project is licensed under the MIT license