Spectral Normalization for Keras

The simple Keras implementation of ICLR 2018 paper, Spectral Normalization for Generative Adversarial Networks. [openreview][arixiv][original code(chainer)]

[Hackmd][github]

Result

CIFAR10

DCGAN architecture

10epoch	With SN	Without SN
With GP
Without GP

100epoch	With SN	Without SN
With GP
Without GP

200epoch	With SN	Without SN
With GP
Without GP

300epoch	With SN	Without SN
With GP
Without GP

400epoch	With SN	Without SN
With GP
Without GP

500epoch	with SN	without SN
With GP
Without GP

Loss	with SN	without SN
With GP
Without GP

ResNet architecture

10epoch	With SN	Without SN
With GP
Without GP

100epoch	With SN	Without SN
With GP
Without GP

200epoch	With SN	Without SN
With GP
Without GP

300epoch	With SN	Without SN
With GP
Without GP

400epoch	With SN	Without SN
With GP
Without GP

500epoch	with SN	without SN
With GP
Without GP

Loss	with SN	without SN
With GP
Without GP

How to use?

1. Move SpectralNormalizationKeras.py in your dir
2. Import these layer class

from SpectralNormalizationKeras import DenseSN, ConvSN1D, ConvSN2D, ConvSN3D

3. Use these layers in your discriminator as usual

Example notebook

CIFAR10 with DCGAN architecture

CIFAR10 with ResNet architecture

Model Detail

Architecture

DCGAN

Generator

Discriminator

ResNet GAN

Generator

Generator UpSampling ResBlock

Dicriminator

Discriminator DownSampling ResBlock

Discriminator ResBlock

Issue

• [x] Compare with WGAN-GP
• [ ] Projection Discriminator

Acknowledgment

• Thank @anshkapil pointed out and @IFeelBloated corrected this implementation.