DDSP VQ-VAE

This is a 6 credits team project done by me and two of my fellow students.

Getting started

Install the project's requirements using pip install -r requirements.txt
- We have used ddsp==0.5.1 but I just saw that version 0.10.0 has been published. Our code might work on that newer version but it is also possible that it will not be compatible.
- Make sure to use the correct numba version because a too new version caused trouble for us. Might have been fixed in 0.10.0.

#Latest update:

Use the virtual environment - workdir(/project/oktoberfest/) for the environment setup using the latest librares for the project.
command: source workdir/bin/activate
To configure the virtual environment with jupyter notebooks: python3 -m ipykernel install --user --name==workdir
Prepare your Dataset as TFRecords: https://github.com/magenta/ddsp/tree/master/ddsp/training/data_preparation
Add in all you local paths in the train_vq_vae.sh script
Execute the train_vq_vae.sh script

A dataset must be created using the latent variabls and the loudness and the fundamental frequency values. Dataset must be saved in a folder named saved_latent_spaces/.
The following command will create the dataset using a trained DDSP model and save it in the necessary location.
$python3 create_latent_space_dataset.py
Then running the gru.py file using the command below will run the autoregresive model on the above dataset.
$python3 gru.py --folder_name=tensorboard_folder
The tensorboard information will be stored in a folder named tensorboard_folder
Finally, the functions in the utils folder can be used to generate audio using the trained model.