Skip to content

IAMIQBAL/Spectrally-Normalized-GAN-Flowers102

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Generating Flowers - Using Spectrally Normalized GAN

This repository contains the Spectrally Normalized (SN) GAN architecture, model weights and training resuls. SN-GAN is based on the paper Spectral Normalization for Generative Adversarial Networks.

Architecture Details

SN-GAN uses spectral normalization, which is a weight normalization technique, to stabilize the training of Discriminator. The spectral norm of a weight matrix W can be obtained by Singular Value Decomposition (SVD), which helps finding the matrix's largest singular value.

Once the largest singular value is obtained, it is divided by every value in the weight matrix.

  • In pytorch, we use torch.nn.utils.spectral_norm which is wrapped around each nn.Conv2d layer

Unlike batch norm, which is used to normalize the activations of each layer, spectral norm normalizes the weights of each layer.

Dataset

I have used Flowers102 (Oxford 102 Flower) dataset. It is available by calling torchvision.datasets.Flowers102. It consists of 102 flower categories. The flowers were chosen to be flowers commonly occurring in the United Kingdom. I have resized the images to 64 x 64 for faster training.

Results

Below are some generated flowers after 20, 120 and 200 epochs of training respectively:


20 epochs



120 epochs



200 epochs



Feel free to load the model weights and generate some flowers ;)

About

Spectrally Normalized GAN trained on Flowers102 dataset

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published