midi2hands 0.0.3

Training and inference for midi hand assignment.

midi2hands

This project is about training and evaluating deep learning models for the task of predicting what hand is supposed to play what note in a piano piece. There are two main architectures used in this project:

1. LSTM
2. Transformer

The models can also be trained in two different ways,

1. Generative: Auto-regressive models that predict the next note given the previously predicted notes.
2. Discriminative: Models that predict the hand of the note given the note.

This is discussed in more detail in the report.

This project is integrated in midi2vid, a tool that can generate videos of piano performances from midi files.

Installation

For inference, you can install the package with pip:

pip install midi2hands

For training, you can clone the repository and install the package with pip:

git clone ...
cd midi2hands
pip install -e .[train]

Usage

Here is an example for inference:

midi2hands -i src/midi2hands/data/test/000-Faure_ClaireDeLune_Op46No2.mid

You can also train new models by running train.py. The models can be configured from the script. Here is an example:

python src/midi2hands/train.py

Data

The data is included in the repository. The data comes from the following project, hannds.

License

This project is licensed under the MIT License. See the LICENSE file for details.