Intro to GAN's
AkulMehra
Agenda
- Introduction to GAN
- Architecture of GAN
- Introduction to CNN
- Applications of GAN's
Hello world of Machine learning
Classifying Handwritten digits using image processing and machine learning
MNIST Dataset
5 0 4 1
But with GAN's we can generate new images
Generative Adversarial Networks
Generative Adversarial Networks (GAN's) are a class of neural networks which allow a network to generate data with the same internal structure as other data.
A GAN has two parts in it:
- the generator that generates images
- the discriminator that classifies real and fake images
Generative Adversarial Networks
- The generator trying to maximize the probability of making the discriminator mistakes its inputs as real.
- And the discriminator guiding the generator to produce more realistic images.
- In the perfect equilibrium, the generator would capture the general training data distribution. As a result, the discriminator would be always unsure of whether its inputs are real or not.
Structure of a GAN
The generator that generates images and the discriminator that classifies real and fake images.
Convolutional Neural Network(CNN)
Convolutional Neural Networks are made up of neurons that have learnable weights and biases. Each neuron receives some inputs, performs a dot product and optionally follows it with a non-linearity.
How we goanna use CNN in GAN's?
-
Generator :- It uses De-CNN
-
Discriminator:- It uses CNN
CNN Basic Architecture
How CNN works
1st Step: Filtering/convolution
Step 2nd : Pooling
Step 3rd : Normalization/ReLU’s
Step 4th : Fully Connected Layer
Highest probability is the selected value.
Applications of GAN's
- Text to Image generation
- Style Transfer
- Increasing Image Resolution
- Image inpainting
- Video generation
- Face Swap
- Image to Image Translation
Text to Image generation (Link)
Style Transfer
Style Transfer
Increasing Image Resolution (Link)
Image Inpainting (Link)
Face Swap
DiscoGAN's
And why they rock !!
DiscoGAN stands for
Learning to Discover Cross-Domain Relations with Generative Adversarial Networks (Link)
Image to Image Translation
(a) Translation of gender in Facescrub dataset and CelebA dataset
(b) Blond to black and black to blond hair color conversion in CelebA dataset.
(c) Wearing eyeglasses conversion in CelebA dataset
DiscoGAN
After iteration: 11,000
Handbags -> Shoes -> Handbags
After iteration: 22000
Image -> Segmentation -> Image
Resources
Thank You
AkulMehra
akul08
akul08
introduction to gans
By Akul Mehra
introduction to gans
- 457