Satrat / LooperGP

Masters thesis repo, transformer model for generating loopable music phrases

LooperGP

Generating loopable symbolic music using TransformerXL and the DadaGP dataset. Sound and Music Computing Masters Thesis at Queen Mary University of London, published at EvoMUSART 2023

ArXiv Link: https://arxiv.org/abs/2303.01665

How to Use

Install dependencies

python -m pip install -r requirements.txt

Modify paths and other configuration details for training and inference in full-data-config_5_lat1024.yml

Run training

python train.py

Generate outputs (without extracting loops)

python inference.py
cd data_parse
python convert_folder ./inference_attempts/yyyymmdd-hhmmss OUTPUT FOLDER

Generate ouputs and extract loops: run data_parse/extract_ex.ipynb notebook

Repo Layout

full-data-config_5_lat1024.yml

main config file with all the parameters for generation and training

model_ead.py

script for model backbone, adapted from https://github.com/YatingMusic/compound-word-transformer

modules.py

script for model backbone, adapted from https://github.com/YatingMusic/compound-word-transformer

model weights/

folder where model weights and model config file should go

data/

folder for storing mappings between token strings and integer IDs, as well as dataset in npz format

data_parse/

Scripts loop extraction, file format conversion, and generation examples

inference.py

Script for generation. Depends on the main config file, vocab pickle files, and npz dataset

train.py

Script for training. Depends on the main config file, vocab pickle files, and npz dataset

Files that need to be downloaded externally

• data/fulldataset-song-artist-train_data_XL.npz (for training)
• model_weights/ep_40.pth.tar (for inference)