Pianist AI

A transformer-based model that can generate sequences of piano notes and their respective durations and velocities.

Example Music Generation

Watch the video

This is an example piano music generation from a model trained on 10000 training iterations with a learning rate of 3e-4 with 256 emedding dimensions, 6 transformer blocks, and 6 heads in each multi-head self attention layer. The script to convert the tokens, durations, and velocities into an audio file has its limitations. Namely, that each note's reverb will be cut off when the next notes begins; however, this example still exhibits the models understanding of melody, chord progression, etc., despite the lossiness in the audio file creation scheme.

Prerequisites

Before you begin, ensure you have the following installed on your system:

• Docker: Install Docker from here.
• NVIDIA Container Toolkit (for GPU support): Follow the installation guide here.

Getting Started

Clone the Repository

First, clone this repository to your local machine:

git clone https://github.com/SamuelReeder/piano.git
cd piano

Using Docker

This project is fully Dockerized, meaning all dependencies and environment setup is handled via Docker.

Building the Docker Image

To build the Docker image, use the Makefile:

make build

Running the Docker Container

To run the Docker container with GPU support, use the following command:

make run

This command will:

• Start the Docker container interactively.
• Mount the current directory to /workspace inside the container, so any changes made are reflected on your host system.
• Enable GPU support (assuming you have the NVIDIA Container Toolkit installed).

Running the Scripts

Once inside the Docker container, you can run the following scripts:

Training

To train a model, run the following command:

python train.py <model_name>

This will train a model with the specified name and save it to the models directory. Please update the parameters in the train.py script to customize the training.

Generation

To generate a musical piece and a .wav file consisting of the simulated audio, run the following command:

python generate.py <model_name> <max_tokens> <output_file_name>