Update: slowly migrating to PyTorch...
This project was developed and presented for the Computer Music course as a semester project. The goal is to generate classical pieces that follow the musical patterns of the training data.
The results were not disappointing nor rewarding as the generated pieces didn't really have those musical patterns in them but they have a lot of potential.
That alone sets the challenge.
Install required modules using
pip install -r requirements.txt
First download the project using git clone
.
To train the model and generate a new note sequence into a midi file, run:
python src/main.py --data dataset/classical_music_midi --composers <composer names>
Composer names should be space separated. For the available composers check the folders in dataset/classical_music_midi directory.
Some ideas worth trying:
- better preprocessing, tweaks in input format
- encode more musical information into the input data
- remodeling of the network architecture, maybe move to Transformers as well
This project is licensed under the MIT License.
PS: trained model files exceeded GitHub's file size limit of 100.00 MB, so they weren't uploaded