MIDI Generator

A typical MIDI file can be viewed as a sequence of notes and chords with specified offsets that show their place in a melody. From this point of view, a chord is a group of notes that are played simultaneously. For melody generation, the next note or chord must be predicted based on an already-existing part. It's an example of a seq2seq problem and the GRU model could be used to solve it. For simplicity’s sake, note offsets in the code do not show the specific place of a note or chord in a piece, they show the time delay that required before the predicted element is played.

This project shows a simple example of .mid files generation. A small model based on GRU architecture is used in order to learn patterns (orders of notes and chords) from existing melodies (see data directory).

The project is created from the Neuro Platform Project Template and designed to run on Neuro Platform, so you can jump into problem-solving right away using the instruction from Quick Start section.

Quick Start

First, to run this project on Neuro Platform install the neuro client and clone the project repository:

pip install -U neuromation
neuro login
git clone git@github.com:neuromation/ml-recipe-midi-generator.git
cd ml-recipe-midi-generator

This repository already contains pre-trained models, so that you can run Jupyter Notebook with inference code and play with these models. To do so, just copy the following command:

make setup && make upload && make jupyter

It will ask you to log in on Neuro Platform when the command is run for the first time. A Github account can be used for this purpose. You also can log in with the neuro login command. For a deeper understanding of what this command does and of the project’s structure read the explanation below.

Development Environment

Directory structure

Local directory	Description	Storage URI	Environment mounting point
data/	Data	storage:midi-generator/data/	/midi-generator/data/
midi-generator/	Directory with code	storage:midi-generator/midi-generator/	/midi-generator/cmidi-generator/
notebooks/	Jupyter notebooks	storage:midi-generator/notebooks/	/midi-generator/notebooks/
results/	Logs and results	storage:midi-generator/results/	/midi-generator/results/

How to train your own model or run an existing model

Follow the instructions below to set up the environment and start a development session.

Log into Neuro Platform

Before environment setup, platform login is required:

neuro login

It can be done with a Github account.

Setup the development environment

The first thing you need to do is to specify an environment on the neuro platform. To do so, run this command in the project’s root directory:

make setup

What happens during this command:

• Several files from the local project are uploaded to the platform storage (namely, requirements.txt, apt.txt, setup.cfg).
• A new job is started in our base environment.
• Pip requirements from requirements.txt and apt applications from apt.txt are installed in this environment.
• The updated environment is saved under a new project-dependent name and is used further on.

After that, the image with the name neuromation-midi-generator will be created. To see a list of images the followed command may be used:

neuro image ls

Uploading / Downloading Resources

After the development environment is set up, the next step is to upload repository resources on the platform with:

make upload

This command will upload all repository directories on the platform. To specify exactly what you want to upload:

Command	Uploaded directory
make upload-code	midi-generator
make upload-data	data
make upload-notebooks	notebooks
make upload-results	results

WARNING: After each local code modification, you must update the code on the platform by uploading it. Resources also can be downloaded with the make download* command.

Generate a new MIDI file

To generate a new midi file you can run a job on the platform:

make generate

After the command is executed, download the resulting file from the platform: make download-results. The result can be opened with timidity. To setup generation parameters, modify the generate_config.cfg file, but do not forget to upload it onto the platform after modification.

Model training

You can also train your own model. To do so, use the following command:

make training CONFIG_NAME=base_train_config.cfg

To specify model and training parameters, modify the base_train_config.cfg file. If you want to change the training config, just specify the CONFIG_NAME variable in the command (see midi-generator/configs for the full list of available predefined train configs). The training process can be followed with TensorBoard:

make tensorboard

Development mode

Remote terminal can be run to execute code manually:

make developing && connect-developing

All modified code can be downloaded on a local machine with the make download command.

Run Jupyter with GPU

make jupyter

• The content of code and notebooks directories is uploaded to the platform storage.
• A job with Jupyter is started, and its web interface is opened in the local web browser window.

Kill Jupyter

make kill-jupyter

• The job with Jupyter Notebooks is terminated. The notebooks are saved on the platform storage. You may run make download-notebooks to download them to the local notebooks/ directory.

Customization

To modify listed above commands or clarify what they do, see Makefile.