Generating traditional celtic tunes with LSTM neural nets.
This project is mostly cloned from Hannah Shaw's wonderful tutorial on generative poetry - https://vivshaw.github.io/blog/electric-pentameter/
The orginal dataset comes from Jeremy's great work at The Session - https://github.com/adactio/TheSession-data
Training took about 45 minutes/epoch on an AWS ec2 p2.xlarge GPU Ubuntu deep learning instance , using a TensorFlow/CUDA backend.
The files:
slicetunes.py extracts just the abc-formatted notes (and in this case, just from Reels to avoid potential Cronenberged odd-meter monstrosities).
build_models.py prepares the data and trains the neural net
generate.py seeds the model and produces several bars of raw abcjs formatted notes.