Classical Music Generator

For this project, we will use Python3.7, managing the necessary Python packages using Pipenv: Python Dev Workflow for Humans. You can install it following the steps shown at their installation section.

To install the requirements listed in the Pipfile, run:

$ pipenv install

To add a new package to the project:

$ pipenv install <package>

To activate the python shell, run the following command, which will spawn a new shell subprocess, which can be deactivated by using exit.

$ pipenv shell

First of all, to work with this project, we are going to need a Environment variable pointing to our root. E.g. if our project is at /home/projects/bachpropagation, just add to your`favorite bash profile:

export BACHPROPAGATION_ROOT_PATH=/home/projects/bachpropagation

To download the raw midi files, run the script at src/dataset called crawler.py, which will download all MIDI files into res/dataset/raw. Then, to make them easier to process, we will convert those into CSV files by running src/dataset/parser.py that will write the processed files into res/dataset/processed.

To train the model, run main.py, which will automatically read and instantiate the Datasets and start the training process. For hyperparameter tuning, change some default folders configuration... visit constants.py.

⚠️ Warning: If not using PyCharm or any other powerful IDE, you'll have to specify the Python sources path manually:

# Standard running command
$ python main.py

# Extend Python path
$ PYTHONPATH="${PYTHONPATH}:src" python main.py

For each iteration it's going to print an output like the following:

Epoch 1: 100%|███████████| 10/10 [00:15<00:00,  1.88s/it]
Generator loss: -0.684159 | Discriminator loss: -1.386941
INFO:root:Creating Sample 1
INFO:root:Converting to CSV...
INFO:root:Parsing note data...
INFO:root:Writing MIDI file at ~/bachpropagation/res/results/

There is an option to visualize the training evolution with Visdom, already installed in the pipenv dev-dependencies (running pipenv install --dev). To run the Visdom server just write the following comand in the console (you can also specify custom host, port and other configurations):

$ visdom

Be sure to have the Visdom server running before training the model, then run the model with the visualization flag:

$ python model.py --viz