wenqinglim/graphmusicgen

A graph-based music generator

polyphemus

A graph-based deep MIDI music generator. Official repository of the paper "Graph-based Polyphonic Multitrack Music Generation".

You can interact with the model in the dedicated Hugging Face space and listen to the samples from the experimental phase on the paper's site.

The source code is written in Python 3.7, using PyTorch and PyTorch Geometric to implement and train Deep Graph Networks.

Prerequisites

To run Polyphemus, follow these steps:

1. Create an Environment: Use conda to create a Python 3.7 environment and activate it:

   conda create -n polyphemus-env python=3.7
   conda activate polyphemus-env

   Alternatively, you can install Python 3.7 and create a new virtual environment using venv.

2. Clone the Repository:

   git clone https://github.com/EmanueleCosenza/polyphemus

3. Install the Required Python Packages: Navigate to the directory where the repository is cloned and install the required packages from requirements.txt:

   pip3 install -r requirements.txt

4. Prepare Your Environment for Audio Generation (Recommended): If you intend to generate audio from MIDI files directly with the provided scripts, you will need a sound synthesizer like fluidsynth and a SoundFont .sf2 file that tells fluidsynth which audio samples should be played for each MIDI instrument.

   You can get fluidsynth with the following command:

   sudo apt-get install fluidsynth

   or, if you prefer, through conda as follows:

   conda install -c conda-forge fluidsynth

   To quickly get a working SoundFont, you can run the following:

   sudo apt-get install fluid-soundfont-gm

   This installs the Fluid (R3) General MIDI SoundFont, a SoundFont that is known to work well with fluidsynth. By default, after the installation, the SoundFont will be stored in /usr/share/sounds/sf2/FluidR3_GM.sf2.

   If you don't have root access to your machine, you can do the same by running the shell script provided in the repo:

   ./download_soundfont.sh soundfonts/

   This automatically finds the FluidR3_GM.sf2 file in the fluid-soundfont-gm package and downloads it in the soundfonts/ directory.

   You can get many other SoundFonts by following the links provided in the FluidSynth project Wiki.

   ⚠️ After downloading the SoundFont in one of the ways shown above, make sure to properly set the SOUNDFONT_PATH variable in the generation_config.yaml file.

5. Download the Trained Models: Run the download_models.py script to download the models trained on the LMD dataset from Polyphemus' Hugging Face model repo:

   python3 download_models.py models/

   The script will download LMD2, a model that generates short 2-bar sequences of music, and LMD16, which generates longer 16-bar sequences, in the models/ directory. Both models have been trained on 4/4 music and are designed to process sequences comprising 4 tracks (=instruments). These are, in order, drums, bass, guitar and strings. The guitar track typically plays accompaniment chords and it can also be intepreted as a piano/accompaniment track. The strings track typically contains a lead melody. Finally, for both models, each bar in the sequence is composed of 32 timesteps, which means that each timestep has a rhythmic value of 1/32.

Generation

If you have already installed Polyphemus and downloaded the pretrained models, in order to generate music, you can simply run the generate.py script as follows:

python3 generate.py models/LMD2/ music/ --n 10 --n_loops 4

By executing this command, the LMD2 model located in models/LMD2/ will generate 10 sequences and the results will be saved in music/. The --n_loops argument tells the script to output, for each sequence, an additional extended sequence obtained by looping the generated one 4 times (this is advisable for 2-bar sequences, since they are quite short).

By default, generate.py outputs .wav audio files in addition to MIDI files. You can change the way audio is generated from MIDI data by editing the generation_config.yaml file. Here, you can specify the SoundFont file that has to be used while generating audio and the MIDI programs that have to be associated to each track.

Run generate.py with the --help flag to get a complete list of all the arguments you can pass to the script.

Structure Conditioning

With Polyphemus, you can condition generation by specifying which tracks (i.e. instruments) should be activated at specific timesteps. You can do so by editing the structure.json file and by passing the --s_file argument to generate.py as follows:

python3 generate.py models/LMD2/ music/ --n 10 --n_loops 4 --s_file structure.json

The structure.json file represents the rhythmic structure of the music to be generated and can be interpreted as a sort of pianoroll for instrument activations. The file contains a JSON representation of a [n_bars, n_tracks, n_timesteps] binary tensor. To give a concrete example, consider the following JSON data:

[
    # Bar 1
    [
        [0, 0, 0, 0,  0, 0, 0, 0], # Track 1
        [1, 0, 0, 0,  0, 0, 0, 0], # Track 2
        [0, 0, 0, 0,  0, 0, 0, 0], # Track 3
        [0, 0, 0, 0,  1, 0, 0, 0]  # Track 4
        
    ],
    # Bar 2
    [
        [0, 0, 0, 0,  0, 0, 0, 0], # Track 1
        [0, 0, 0, 0,  1, 0, 0, 0], # Track 2
        [0, 0, 0, 0,  0, 0, 0, 0], # Track 3
        [0, 0, 0, 0,  0, 0, 0, 0]  # Track 4
        
    ]
]

Here, for simplicity, n_timesteps=8. In the first bar, Track 2 and Track 4 will be activated, respectively, in the first and fifth timestep. In the second bar, just the second track will be activated in the fifth timestep. All other tracks will be silenced throughout the sequence.

Notice that the binary activations above refer to actual note activations and that sustain is not considered. This means that a track could end up playing a sustained note even if the corresponding value has been set to 0 in the file. For instance, Track 2, which has been activated in the first timestep of the sequence, could actually sustain its activated notes for more than a timestep.

The number of tracks for each bar in the JSON file must be strictly equal to the number of tracks used to train the model. The number of bars, instead, may be smaller than the number of bars that can be handled by the model. If this is the case, the script will just replicate the partial binary structure to reach n_bars. In this way, it is possible e.g. to compile the binary activations for just one bar and let the model use this same structure for each bar.

When editing this file for the LMD2 and LMD16 models, remember that n_timesteps=32, which means that each timestep has a rhythmic value of 1/32, and that the tracks used to train these models are, in order, drums, bass, guitar and strings.

Training a New Model from Scratch

Preprocessing

Before you can train Polyphemus from scratch, you have to preprocess a MIDI dataset. This can be done by running the preprocess.py script as follows:

python3 preprocess.py midi_dataset_dir preprocessed_dir

where midi_dataset_dir is the directory of the MIDI dataset and preprocessed_dir is the directory to save the preprocessed dataset. For the script to work, the midi_dataset_dir directory must only contain .mid files in a flat or hierarchical fashion (i.e. in a tree of subdirectories).

If you want to preprocess the Lahk MIDI Dataset (LMD-matched), you can first download it from here, or just execute the following:

wget http://hog.ee.columbia.edu/craffel/lmd/lmd_matched.tar.gz

Then, you can decompress the archive by using tar.

Training

You can train a new model from scratch by using the dedicated script:

python3 train.py dataset_dir model_dir config_file --model_name test --use_gpu

where dataset_dir is the directory of the preprocessed dataset to be used for training, model_dir is the directory to save the trained model, and config_file is the path to a JSON training configuration file. An example of this file is provided in the repo as training.json.

Make sure to run the command with the --help flag to find out how you can customize the training procedure.

No integration with TensorBoard or W&B is provided, but you can still check the progress of training with the training_stats.ipynb Jupyter Notebook.

License

This project is licensed under the MIT License.