Week 3 - Post - GAN

Official Open
Visual Generative
AI Application
Generative Adversarial Networks (GAN)
Week 3
AY 24/25
S PEC IALIST D IPLOMA IN APPLIE D GE NE RATIV E AI ( S D GAI)
Official Open
Official Open
Official Open
Official Open
Objectives
• By the end of this module, learners will be able to
• Describe the concepts of Generative Adversarial Networks (GAN)
• Train GAN model to generate output
Official Open
What is Generative Adversarial Networks?

• A framework where a generative model is pitted against an
adversary:
• A discriminative model learns to determine whether a sample is
generated (“fake”) or not generated (“real”)
• A generative model generates (fake) samples
• Competition between the discriminator and the generator drives both
models to improve until the “fake” are indistinguishable from the “real”
Official Open
Game Theory and GAN

Aim: Generate samples will Aim: Able to tell that the
not be classified as fake generated samples are fake
Generator G(z) Discriminator D(x)
Solution to this game is the Nash Equilibrium: G(z) being drawn from the same
distribution as the training data and D(x) = 0.5
Official Open
Applications: StyleGAN
• Human Face Generation(StyleGAN)
• Go to StyleGAN or StyleGAN3
• Generate a few images on a human face.
Are you able to tell that the

image is “fake”?
Official Open
Applications: CycleGAN
Deep Fake: From

horse to zebras
Turning a horse video into a zebra video (by CycleGAN) (youtube.com)

Official Open
Applications: GauGAN
Translating Scribbles
into Images
How to use NVIDIA GauGAN web demo app. (youtube.com)

Official Open
Generative Adversarial Networks (GAN)

Discriminative Model
• GAN consists of a:
• Discriminative Models D(x): samples
Examine samples to determine if D(x) Real or Fake
they are real or fake
• Generative Models G(z):

Create samples intended to come
from the sample distribution of Generative Model
training data
Noise Generated
G(z)
Samples
Official Open
Generative Vs Discriminative Models

Generator The Generator strives to Discriminator
make D(G(z)) approach 1
Maximize Probability of D(x) to be 1

Generator The discriminator strives to Discriminator
sample, G(z) make D(G(z)) approach 0 Decision D(x)
Generator Discriminator
function, G function, D
Solution to this game is the

Nash Equilibrium: G(z)
Input noise (z) being drawn from the same Input Sample x from
distribution as the training training dataset
data and D(x) = 0.5
Official Open
Discriminator Function
• A discriminator is a classifier model.
• It can be implemented using a Feedforward Neural Network.
• It classifies if a given sample is real or fake.
Prob (y | x)
Samples, x Probability (Fake,
given Sample)
Discriminator D(x)
Official Open
Generator Function
• Given some simple prior distribution (e.g. Gaussian or z), the
generator gives a sample of 𝑥 drawn ideally from the distribution of
the training data.
• Generator can either be Feedforward Neural Networks or
Convolutional Neural Networks
Generator G(z)
Generated
Input, z Samples, 𝑋෠
Sampling
Prior Distribution
Official Open
Understanding Cost Function for GAN

• A binary cross entropy (BCE) function is used to capture the loss
of both the generator and the discriminator.
• The BCE cost function has two parts:
• When the label and the prediction are similar, BCE ~ 0
• When the label and the prediction are different, BCE → ∞
min max − 𝐸 log(𝑑 𝑥 ) + 𝐸 1 − log(𝑑 𝑔 𝑧 )

𝑑 𝑔 Relates to
Discriminator on Relates to Generated Samples
actual samples
Discriminator will
minimize the cost Generator will Captures the Adversarial Implementation
maximize the cost (Zero sum Game)
Official Open
Official Open
Training a GAN Model

• GANs are trained in alternating (Discriminator – Generator) fashion
• The 2 models should always have similar “skill” level, progressing
and improving at approximately similar degree.
Use SGD-like algorithm of choice
(Adam) on two minibatches
simultaneously:
• A minibatch of training examples Real, 𝑋 Output, 𝑌෠
• A minibatch of generated
samples Discriminator
Aim: Distinguish the fakes from
reals
Noise
Fakes, 𝑋෠
Generator
Aim: Make the fakes look reals
Official Open
Training a Discriminator
Both real and fake examples Discriminator
are used in training Parameters
Update
Real, 𝑋 Output, 𝑌෠ Cost
Binary Cross Entropy

Discriminator
(Cost function) with
Aim: Distinguish the fakes from
labels from real and fake
reals
Noise
Fakes, 𝑋෠
Generator
Aim: Make the fakes look reals
Official Open
Training a Generator
Generator
Parameters
Update
Noise Output, 𝑌෠ Cost

Fakes, 𝑋෠
Binary Cross Entropy
Generator Discriminator
(Cost function) with all
Aim: Make the fakes Aim: Distinguish the fakes from
labels equals to real
look reals reals
Only fake examples are used

in training
Official Open
Issues with Training GAN Models

• GANs should be trained for multi-mode output.
• Modes are peaks in distribution of features, and are typical with
real-world datasets
• Mode collapse happens when generator gets stuck in one of the
modes
Official Open
Issues with Training GAN Models

• GAN attempts to make the
generated distribution and the
real distribution looks similar.
• If the discriminator becomes
too good at distinguishing
between real and fake data
early in the training, it provides
very small gradients to the
generator. The generator then
receives insufficient information
to improve, leading to slow or
stalled learning.
Official Open
Practical 3
Generative Adversarial Networks
• Vanila GAN
• Deep Convolutional GAN
Official Open
Hand-digits Generation
• In this practical we will learn to generate a 28 X 28 grayscale image
of handwritten digits (numbered from 0 to 9).
• We will train a GAN model using the MNIST dataset
Official Open
Vanilla GAN
• Generator model: Noise vector z as input and
repeating blocks that transform and scale up
the vector to the required dimensions.
• Each block consists of a dense layer
• Leaky ReLU activation
• Batch-normalization layer.
• We simply reshape the output from the final block
to transform it into the required output image size
• The discriminator is a simple feedforward
network.
• This model takes an image as input (a real image or
the fake output from the generator) and
• Classifies it as real or fake.
Official Open
Deep Convolutional GAN

• Generator with Convolutional
Neural Network architecture.
• Noise vector as input and then
passes it through a repeating
setup of up-sampling layers,
convolutional layers, and batch
normalization layers to stabilize
the training.
• Modifications
• Use batch-norm in both the
generator and the discriminator.
• Use ReLU activation in generator
for all layers except for the
output, which uses Tanh.
• Use LeakyReLU activation in the
discriminator for all layers.
Alec Radford et. al, Unsupervised Representation learning with Deep convolutional Generative Adversarial Networks
Official Open
Official Open
Official Open
Official Open
Official Open
Official Open
Official Open
Official Open
Official Open
Official Open
Official Open
Official Open
Official Open

Week 3 - Post - GAN

Uploaded by

Copyright:

Available Formats

Week 3 - Post - GAN

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Week 3 - Post - GAN

Uploaded by

Copyright:

Available Formats

Official Open

What is Generative Adversarial Networks?

Game Theory and GAN

Generator G(z) Discriminator D(x)

Are you able to tell that the

Deep Fake: From

Turning a horse video into a zebra video (by CycleGAN) (youtube.com)

How to use NVIDIA GauGAN web demo app. (youtube.com)

Generative Adversarial Networks (GAN)

• Generative Models G(z):

Generative Vs Discriminative Models

Maximize Probability of D(x) to be 1

Solution to this game is the

Understanding Cost Function for GAN

min max − 𝐸 log(𝑑 𝑥 ) + 𝐸 1 − log(𝑑 𝑔 𝑧 )

Training a GAN Model

Real, 𝑋 Output, 𝑌෠ Cost

Binary Cross Entropy

Noise Output, 𝑌෠ Cost

Only fake examples are used

Issues with Training GAN Models

Issues with Training GAN Models

Deep Convolutional GAN

You might also like