Download as pdf or txt
Download as pdf or txt
You are on page 1of 26

A deep dive into GAN’s

Styled on Styled on
The Starry Night Woman with a hat.
What is a GAN?
GENERATIVE ADVERSARIAL NETWORK
The Architecture
The loss function
Generator Image generation
Upsampling and Convolution layers used
OVERVIEW OF DISCUSSION

GAN
Add
Conditionality
conditionality
inferred implicitly by
Explicitly
model

Conditional GAN InfoGAN

Use Paired Use unpaired


images for input input images

Pix2Pix GAN Cycle GAN


Conditional GAN

GAN’s can be extended to


conditional GAN if both the
generator and discriminator
are conditioned on some
extra information y(typically
the class label)
Pix2Pix GAN

Generator takes input random noise as well as a conditional image


InfoGAN
In InfoGAN, the generator not only takes as input random noise but also a set of "latent
codes that encode specific factors of variation in the generated output.

These latent codes can be learned in a disentangled manner


A disentangled representation
separates the different factors of
variation in the data, making it possible
to manipulate one factor while keeping
the others fixed.
The infoGAN splits the Generator input into two parts: 1 . the noise vector 2. a new “latent
code” vector.

These latent codes are learnt by the network and used for further generation.
Mutual Info of two random
variables is a measure of
the mutual dependence
between the two variables.

We maximize mutual
information between
generator’s output and
latent variable c
a separate neural network Q

Train Q using output of


generator and c as labels

Train D,G,Q
CycleGAN’s?
Image to image translation.
Loss Functions of CycleGAN
Cycle Consistency
Problems of GANs
Mode Collapse

The generator gets stuck


at a time producing
limited variety of
samples

Generator doesn’t learn


all necessary properties
to model the distribution
Reaching Nash Equilibrium
Applications of GAN

1. Generate Datasets for Model Training

MNIST dataset
generated by Deep
Convolutional CGAN
2. Style Transfer
3. Photo Inpainting
4. 3D Object Generation (from 2D images)
THANK YOU!

You might also like