Generating Polyphonic Music with Adversarial Networks

Final Project for CS 236G: General Adversarial Networks. Using DCGAN's and approaches from StyleGAN to generate raw audio of EDM music. Link to final paper: Final Paper

Environment

Create new conda environment from environment.yml

$ conda env create -f environment.yml

Data

For the project, I trained on the Beatport EDM Dataset, which can be found here: Beatport EDM Key Dataset | Zenodo The models can be trained on any dataset however, as long the as the files are in a common sound format, like mp3 or wav.

Preparing Data

To configure your dataset, split the data into train, test, and dev folders in a data/ directory. To prepare and save the data into tensor format (for faster startup at training), run the following command for one of train, dev, or test:

$ python dataset.py -d [[tr]ain, [d]ev, [t]est]

To verify that the data is saved, you can run

$ python dataset.py -d [[tr]ain, [d]ev, [t]est] --load

Training

Usage: train.py [-h] [--model WaveGAN] [--epochs 20]
                [--epochs_per_save 10] [--batch_size 4]
                [--n_critic 5] [--phase_shuffle] [--spectral_norm]
                [--warmup] [--style_gan]

optional arguments:
   -h, --help           show this help message and exit
  --model MODEL         Model architecture [WaveGAN, TransGAN]. Default: WaveGAN
  --epochs EPOCHS       Training epochs. Default: 20
  --epochs_per_save     Save model every n epochs. Default: 10
  --batch_size          Batch size. Default: 4
  --n_critic            n disc updates for 1 gen update. Default: 5
  --phase_shuffle       Use phase shuffle. Default: False
  --spectral_norm       Use spectral norm. Default: False
  --warmup              Use warmup. Default: False
  --style_gan           Use AdaIN from StyleGAN. Default: False

Evaluation

$ python test.py

Running test.py will calculate and print the evaluation metrics. Open notebook.ipynb to listen to generated audio samples from the generator.