Project 5 - Generative Adversarial Networks (Due 11/09)
Objectives:
The objective of this project is to learn how GAN works through a
simple generative task. You also learn how to train a conditional GAN
such that you have some level of control over the type of images you
generated.
Data set used:
MNIST
Requirements:
- Task 1: Generate MNIST digit images based on DCGAN. The sample
code can be found here.
- Task 2: Train a conditional GAN model. Based on the DCGAN code
in Task 1,
concatenate labels (e.g., class labels) to the input of the generator and discriminator to
control the generated images. You need to go through [ConditionalGAN:2014].
Report
- Read [Arjovsky:2017] and [Goodfellow:2014]. Discuss the reasons
why the vanilla GAN is unstable. Explain what unstable means in this
context. Read [DCGAN:2016] and [ConditionalGAN:2014] and explain how
they use different approaches to stablize GAN. Write
the report in less than two pages.