keras-io/GauGAN-Image-generation

Model description

In this, GauGAN architecture has been implemented for conditional image generation which was proposed in Semantic Image Synthesis with Spatially-Adaptive Normalization.

GauGAN uses a Generative Adversarial Network (GAN) to generate realistic images that are conditioned on cue images and segmentation maps.

This repo contains the model for the notebook GauGAN for conditional image generation

Full credits go to Soumik Rakshit & Sayak Paul

Training and evaluation data

Here, the Facades dataset is used for training GauGAN model. Some custom layers that were added into the model are - SPADE (SPatially-Adaptive (DE) normalization), Residual block including SPADE & Gaussian sampler. Also, the GauGAN encoder consists of a few downsampling blocks. It outputs the mean and variance of a distribution as shown in this image.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

name	learning_rate	decay	rho	momentum	epsilon	centered	training_precision
RMSprop	0.0010000000474974513	0.0	0.8999999761581421	0.0	1e-07	False	float32

Model Plot

View Model Plot

Model Reproduced By Kavya Bisht