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.
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_normwhich is wrapped around eachnn.Conv2dlayer
Unlike batch norm, which is used to normalize the activations of each layer, spectral norm normalizes the weights of each layer.
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.
Below are some generated flowers after 20, 120 and 200 epochs of training respectively:
