wind222 / SN-GAN

pyTorch implementation of Spectral Normalization for Generative Adversarial Networks

SN-GAN

PyTorch implementation of Spectral Normalization for Generative Adversarial Networks

Prerequistes

• Python 2.7 or Python 3.4+
• PyTorch
• Numpy

Usage

• Clone this repo:

git clone https://github.com/godisboy/SN-GAN.git
cd SN-GAN

• Train the standard model for 32*32 image size

python train.py --cuda(optional) 

• Train the Res-Generator and Res-Discriminator for 64*64 image size

python train-res.py --cuda --dataPath /Path/to/yourdataset/

---

1. Result of SN-GAN on CIFAR10 dataset

• Generated

• Generated

• Generated samples with Standard Generator and ResDiscriminator

Note: The ResBlock of Res-Generator is different from what implemented in original paper. This repo use UpsamplingBilinear instead of Uppooling for Upsampling operation.

To Do

• Conditional version of SNGAN with Conditional BatchNormalization
• ImageNet 1000 classes
• Latent Inference
• ...

Acknowledgments

Based on the implementation DCGAN and official implementation with Chainer sngan_projection