/GANotebooks

wgan, wgan2(improved, gp), infogan, and dcgan implementation in lasagne, keras, pytorch

Primary LanguageJupyter NotebookMIT LicenseMIT

Generative Adversarial Notebooks

Collection of my Generative Adversarial Network implementations

Most codes are for python3, most notebooks works on

CycleGAN

  • CycleGAN-lasagne
  • CycleGAN-keras

CycleGAN results

Result after 3 hours and 58 epochs on a GTX 1080. From top to bottom: Input, Fake, Recreate of the input.

Face-off result. From top to bottom: Input, Fake, Recreate of the input. [youtube video](https://www.youtube.com/watch?v=Fea4kZq0oFQ)

pix2pix

  • pix2pix-keras: pix2pix GAN Keras implementation
  • pix2pix-lasagne: pix2pix GAN Lasagne implementation
  • pix2pix-torch: pix2pix GAN pytorch implementation

pix2pix sample results

Validation result of edges-to-shoes after 12 epochs. From top to bottom: Input, Ground truth, the result.

Validation result of facades dataset after 150 epochs using resnet. From top to bottom: Input, Ground truth, the result.

WGAN on CIFAR10

WGAN2 (improved WGAN/WGAN-gp)

  • wgan2-lasagne: improved WGAN Lasagne implementation (on CIFAR10)
  • wgan2-keras: improved WGAN Keras implementation (on CIFAR10)
  • wgan2-lasagne-anime: WGAN on anime face images, lasagne
  • wgan2-AC-lasagne: improved WGAN Lasagne implementation with Auxillary classfier

WGAN2 sample results

  • cifar10 dataset

  • cifar10 dataset with Auxillary classfier

  • anime face dataset

InfoGAN

  • mnist-infogan: InfoGAN Lasagne on MNIST dataset
  • mnist-infogan-paper-uniform: InfoGAN Lasagne on MNIST dataset (fllowing the paper implementation)

InfoGAN sample results

DCGAN

  • dcgan-lasagne: DCGAN in Lasagne

DCGAN sample results