Generative-Adversarial-Learning

Abstract:

The goal of this assignment is to generate Pokemon type images from a GAN by training through a dataset of 819 images of real pokemons. The type of GAN implemented is DCGAN with two different approaches of loading the images. First part of the assignment uses DCGAN with greyscale images and the second part uses RGB-Red Blue Green colored images. It is observed that the quality of images generated with color pixels looks more appealing and with perfect amount of epochs a new pokemon can be visualised.

Part 1: Deep Convolutional Generative Adversarial Network (DCGAN) using greyscale pokemon images¶

We use ImageDataGenerator() to load images and encode them to an array of numbers. The arguments to be needed are directory of the images, shuffle=True, target_size of the image (28,28)

Conclusion for Part 1:

Initially, the GAN model is trained with the help of discriminator and with the help of losses, the optimiser tries to minimise the loss using 'Adam' optimiser and loss estimimated using cross entropy is minimised after a significantly large number of epochs.

Part 2: DCGAN with colored RGB images

Conclusion Part 2:

Finally a nearly shapes are generated by Generator with an accuracy of 97% for fake generated images. The process is enhanced by using the more accurate activation function and bias variable.