/Generative-Dog-Images-GAN

This project comes from a Kaggle Competiton named Generative-Dog-Images. Deep Convolutional GAN (DCGAN) and Conditional GAN (cGAN) are applied to generate dog images. Created a model to randomly generate dog images which are not existed in the original dataset.

Primary LanguagePython

Generative-Dog-Images-GAN

This project comes from a Kaggle Competiton named Generative-Dog-Images. Data is based on the Stanford Dogs Dataset.

Team

Authors: Gaofeng Huang, Jun Ying, Xi Zhang. Final project of the course Machine Learning II (by Prof. Amir Jafari).

GAN Method

This project applied a technique called Generative Adversarial Network (GAN). GAN is a the-state-of-art deep learning algorithm proposed by Ian Goodfellow in 2014. GAN is a combination of generator network and discriminator network. These two networks compete with each other and form a zero-sum game. A zero-sum game means if one gets better, another will must be worse. As two networks optimize their perfomance when their opponents get better. There is a Nash Equilibrium when both generator and discriminator reach to their best performance.

Objective

This project is going to create a model, which can generate new dog images beyond the training dataset.