SynthMIDI, a synthesized dataset for a single-note classification training. The dataset consists of a customizable data generation with all possible MIDI programs and notes. The goal of this dataset is to be used for the academic purpose.

1. Commit History
2. Requirements
3. Usage: Generating Dataset
   • Running on Docker
   • Running on local
   • Dataset config
4. Usage: Training Baseline
   • Training on Docker
   • Training on local
   • Training config
5. Baseline
   • Stratify Splitting
   • Leave Instrument Out
6. Contributing
7. Author
8. Acknowledgement

(26/06/2022) Refactor code structures 🪡

• Refactor code structure to isolate training and data generation pipeline
• Fix bugs where note generation mapped the keys wrongly
• Add customizable baseline model training pipeline

(25/06/2022) initial commit 🍢

• Push the initial code.
• The only preprocessing function includes a simple noise injection and minor scaling

(26/06/2022) midisynth_dataset-v1 (download here)

• The very first version of MIDI generated from mido library
• The only augmentation applied was gaussian noise injection

We assume the reader use docker to run this script, or else some modification will be needed. If you didn't have one, you can download it here.

To begin with, download or clone this repository. This can be done by running the following command:

git clone https://github.com/vistec-AI/SynthMIDI.git
cd SynthMIDI

We recommend you to run this script using docker, as we test only on a docker environment.

To build the docker, run:

docker build -t synthmidi/data-generator -f dataset.Dockerfile .

Once finished, you should have a built image on your docker. Run docker images to see if the image was built properly.

Once you've built the docker, run the following command on your terminal:

docker run -it -v $PWD/conf:/workspace/conf -v /path/to/saved/dataset:/root/dataset --name your-docker-container-name synthmidi/data-generator

If you're still insist of running on local, that's fine. But note that we use python3.8, so make sure you use the same python version to prevent any bug.

It is wise to work on a separated environment for each project. If you're not skipping this section even though you saw that is is optional, you're a man of culture.

python3 -m venv dataset.venv  # create virtual environment
source dataset.venv/bin/activate  # activate virtual environment

There're several dependencies you needed to run this script. To install those libraries, run:

pip install -r dataset.requirements.txt

We also recommended you to install soundfont by the following command:

wget https://musical-artifacts.com/artifacts/538/Roland_SC-88.sf2 -O ~/.fluidsynth/default_soundfont.sf2

This step is optional as the script will handle this anyway.

This step is also optional. If you skip this step, the dataset will be saved on the default directory at ~/dataset. However, if you feel like this is not a good place to store the dataset, please feels free to override save path on config.yaml.

To create the dataset, run:

python generate_datset.py

The config.yaml receives the following keys:

Note that the total number of samples generated is equal to len(available_notes) * len(available_program) * samples_per_program_note.

To begin with, download or clone this repository. This can be done by running the following command:

git clone https://github.com/vistec-AI/SynthMIDI.git
cd SynthMIDI

docker build -t synthmidi/trainer -f train.Dockerfile .

Once finished, you should have a built image on your docker. Run docker images to see if the image was built properly.

Once you have built your docker, run the following commands:

docker run -it -v $PWD/conf:/workspace/conf -v /path/to/saved/weights:/workspace/weights --shm-size=16G --name synthmidi-trainer synthmidi/trainer

Make sure you use python 3.8.x!

python3 -m venv dataset.venv  # create virtual environment
source dataset.venv/bin/activate  # activate virtual environment

pip install -r train.requirements.txt

Make sure your config file was setup according to your environment. Once you finished setting up config file, run:

python train.py

Then, you're done!

See an example at baseline.yaml

We construct two baseline for this experiment:

1. Stratify splitting - Simply apply stratify splitting the dataset.
2. Leave instrument out - Leaving specific instruments out as a testing data. This is way more challenging.

The split file and the whole dataset was available on dataset section

*For Leave instrument out setting, it's better to conduct it in a k-fold cross validation setting, but I'll leave this as an open contribution.

*Precision, Recall, and F1 are computed using macro average

There're several improvements that can be done and I'll be add several feature needed in the issue pannel. If you wish to contribute, fork this repository, and open the pull request.

Chompakorn Chaksangchaichot

This works is made possible with several helps and guidance from Noppawat Tantisiriwat. Also, Kudos to the AI Builder Program, P' Charin Polpanumas and Aj. Ekapol Chuangsuwanich for making this dataset possible.