chathasphere / pno-ai

Music Transformer Sequence Generation in Pytorch
MIT License
101 stars 29 forks source link

pno-ai

Pianistic Neuralnetworks Output Aleatoric Improvisations

About:

An implementation of Google Magenta's Music Transformer in Python/Pytorch. This library is designed to train a neural network on Piano MIDI data to generate musical samples. MIDIs are encoded into "Event Sequences", a dense array of musical instructions (note on, note off, dynamic change, time shift) encoded as numerical tokens. A custom transformer model learns to predict instructions on training sequences, and in generate.py a trained model can randomly sample from its learned distribution. (It is recommended to 'prime' the model's internal state with a MIDI input.)

Training Data:

The initial dataset comes from several years of recordings from the International Piano-e-Competition: over 1,000 performances played by professional pianists on a Yamaha Disklavier. Obtainable here. A sufficiently large dataset (order of 50 MB) of piano MIDIs should be sufficient to train a model.

Deploying Model:

Google Colab is a good free option for experimentation. But for running a model for multiple hours, I would recommend using a Google Cloud compute instance with GPU driver(s). This tutorial from fast.ai shows you how to get up and running on a GPU instance. (I would recommend using a non-preemptible instance for sanity's sake.) Once you're ssh'ed in...

Bibliography: