GAN-Tutorial

GANs Tutorial

Very simple implementation of GANs, DCGANs, CGANs, WGANs, and etc. with PyTorch for various dataset (MNIST, CARS, CelebA).

You can run the code at Jupyter Notebook. And actually you can also run these codes by using Google Colab immediately (needed downloading some dataset)!

Sometimes ipynb files do not work in Github, please clone and run it in your server.

Requirements

• python 3.6 (Anaconda)
• pytorch 1.0.0 (updated from 0.4.0. If you want to use the previous version, then find previous commit.)

Implementation List

MNIST

• Vanilla GAN
• Conditional GANs
• DCGAN
• Wasserstain GAN
• WGAN-gp
• infoGAN

CARS (Stanford dataset)

• you can download the dataset from https://ai.stanford.edu/~jkrause/cars/car_dataset.html
• DCGAN
• But I think the dataset is small and needs preprocessing a lot.

CelebA (aligned dataset)

• you can download the dataset from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
• BEGAN
• R1GAN

Experimental Results

• You can also see the samples at ipynbs.
• After DCGAN, DCGAN with condition is a base model.
• Trained 30 epochs respectively.

Vanilla GAN

Conditional GAN

DC GAN

WGAN-gp

infoGAN w/ walking code 1

infoGAN w/ walking code 2

BEGAN random samples (20 epochs)

BEGAN interpolation

GAN with R1 regularization random samples (20 epochs)

GAN with R1 regularization interpolation

Colab

• https://colab.research.google.com
• You can run it with GPU(K80) Runtime mode
• Training Vanilla GAN takes only 10 minutes.