This repo contains the scripts our team used for this project. It contains the training scripts of two models: VAE and GAN.
Dataset
To run the models, you need to download the Piano dataset. Make sure the individual .wav files are put in a directory named "piano". After the data is downloaded, run $ python3 GAN_utils.py to sample the audios to chunks(you need to change the path of the data). The default chunk length is 4s. You can adjust it by changing the paramter in create_chunks(chunk_len=4)
GAN
These instructions assume you have already downloaded and chunked the data.
To train a model with either WaveGAN or CMD WaveGAN discriminator/generator, run the command
$ python3 train.py
The possible command-line arguments are as follows:
--epochs", type=int, default=500: number of epochs
--lr", type=float, default=0.0001: learning rate
--batch_size", type=int, default=64: batch_size
--z_size", type=int, default=400: random vector size in generator
--G_d", type=int, default=64: generator dimensionality parameter
--D_d", type=int, default=64: discriminator dimensionality parameter
--G_type, type=str, default="old": generator type, old (vanilla WaveGAN) vs new (CMD WaveGAN)
--D_type, type=str, default="old": discriminator type, old (vanilla WaveGAN) vs new (CMD WaveGAN)
--k, type=int, default=5: number of discriminator sub-epochs per generator epoch
--l, type=float, default=10: lambda parameter, weighting on gradient penalty
--subsample_ratio, type=float, default=1: subsampling ratio of data
--inject_noise_var, type=float, default=0: variance of injected gaussian noise to data
--save_results", type=bool, default=True: if we wish to save results to a folder. If this is true, a directory named 'GAN_results_piano' needs to exist.
VAE
The VAE model was implemented in jupyter notebook and one can run the VAE by following the scripts in the notebook.
Robustness Experiment
To run the robustness analysis, follow the example in the script robustness_eval_fns.py