/GAN-wGAN

3 experiments of GAN/wGAN on simple gaussian distribution, MNIST dataset and exploration of music generation by MuseGAN.

Primary LanguageJupyter NotebookMIT LicenseMIT

GAN-wGAN

Source Code

GAN-Gaussian

The source code is GAN_Gaussian.ipynb.

WGAN-Gaussian

The source code is wGAN_Gaussian.ipynb.

Application on the mnist data set

We use the file MNIST- GAN&WGAN.ipynb to generate number images based on the mnist data set by GAN and WGAN.

Run all the cells in order and then two number images will be generated by GAN and WGAN. You can change the code EPOCHS = 500 into the number of epochs you want to run.

Application on generating music: simple model

The source code is in file MUSIC_SIMPLE.ipynb.

  1. The data set need to be download from https://salu133445.github.io/lakh-pianoroll-dataset/dataset .

On this website:

Data of music sample is under entry LPD-5 and called "lpd-5-cleansed". Then, we need to put it in directory data/music_simple/lpd_5.

Data of file IDs and the matched MSD IDs is under entry LPC-cleansed and called "cleansed_ids.txt". Then, we need to put it in directory data/music_simple/amg.

  1. A result of generated music is saved at results/music_simple/generated_music.mp3.

MuseGAN

The source code is at https://github.com/salu133445/musegan.

A result of generated music is saved at Music/MuseGAN/best_samples.mp3.

Result

All our results are in the results folder.

Other Folders

  • imgs and models are used to stored intermediate result. They are by default empty.
  • data is the folder to store data input of MNIST experiment.