sleep-yearning/magenta

Magenta: Music and Art Generation with Machine Intelligence

Magenta is a research project exploring the role of machine learning in the process of creating art and music. This Fork was made to build on top of the coconet music generation model and expand it.

Notably, the ability to train a model from a folder of midi files has been developed and easy management through a GUI.

Getting Started

• Requirements & Installation
• Using the coconet GUI training

Installation

We recommend to create an anaconda environment (python 3.5-3.7).

conda create -n coconet python=3.7

Install the package from github per:

conda activate coconet
pip install https://github.com/sleep-yearning/magenta/archive/master.zip

If you want to enable GPU support and can run it, you can create a version which depends on tensorflow-gpu instead of tensorflow (change in setup.py) and install it.

NOTE: In order to install the rtmidi package that magenta depends on, you may need to install headers for some sound libraries. On Ubuntu, these are the necessary packages:

• build-essential
• libasound2-dev
• libjack-dev

On Arch based systems, those packages are:

• base-devel
• jack
• alsa-lib

If you run into problems with this package, you can try and install a copy of our package without the dependency (found in setup.py). This might work since coconet doesn't really use it, but will probably break if you want to use the same magenta installation for something else. You specifically won't be able to use magentas midi-interface for playback.

Using CocoNet GUI

The easiest way to interface with CocoNet is to download a small selection of files from the usage-api branch. Once downloaded, you can start those files in the coconet environment where magenta is installed.

python /coconet/GUI

From the GUI you can either select your folder of midi files to train on, or generate new midi files from already trained models. We put some selected MIDI folders into /coconet/train_data/, which is also part of the usage-api download.

If you don't want to use the GUI, just call the python files with their needed/optional arguments. For example:

python /coconet/prepare_and_train.py path-to-midi-folder --grouped
python /coconet/coconet_sample.py path-to-trained-model-folder igibbs midi-output-folder

If you want to use the package more flexible and specify more hyperparameters without the GUI, you can use the package in your python code.

from magenta.models.coconet.prepare_and_train import prepare, train
from magenta.models.coconet.coconet_sample import main

And then call those functions in your code.

Generated files

Training database & model checkpoints are saved into the MIDI folder by default. This way you can find all the relevant files for one interpret in that folder. You can specify custom output directories for the generated samples in the GUI or as parameter when used on the command line.

The original magenta implementation can be found here. Some further background information can be found in the readme of the coconet folder.

This version was forked from here because necessary changes for CocoNet to work with python 3 where added there.