Introduction to GANs

Deep Learning for Images with PyTorch

Michal Oleszak

Machine Learning Engineer

General Adversarial Networks

Generated cat image.

https://thesecatsdonotexist.com
Generative Adversarial Networks (GANs)
Generate new samples based on training data

Pokemon Sprites Dataset

Sample of images from the Pokemon Sprites dataset

Pokemon Sprites Dataset from PokeAPI
About 1300 sprites of animal-like creatures from a Pokemon video game
Goal: Generate new Pokemons!

GANs architecture

GAN workflow diagram.

GANs architecture

GAN workflow diagram.

GANs architecture

GAN workflow diagram.

GANs architecture

GAN workflow diagram.

GANs learning process

GAN workflow diagram.

Generator: learns to produce realistic-looking images
Discriminator: learns to tell the fakes from real images
Conflicting objectives ensure each network gets better at its task
In the end, generator should generate realistic images

Basic Generator

class Generator(nn.Module):
    def __init__(self, in_dim, out_dim):
        super(Generator, self).__init__()

        self.generator = nn.Sequential(
            gen_block(in_dim, 256),
            gen_block(256, 512),
            gen_block(512, 1024),
            nn.Linear(1024, out_dim),
            nn.Sigmoid(),
        )


    def forward(self, x):
        return self.generator(x)

Define Generator class

Sequence of generator blocks, a linear layer, and a sigmoid activation

def gen_block(in_dim, out_dim):
  return nn.Sequential(
      nn.Linear(in_dim, out_dim),
      nn.BatchNorm1d(out_dim),
      nn.ReLU(inplace=True)
  )

Pass input through all layers
Input: noise of size in_dim
Output: image of size out_dim

Basic Discriminator

class Discriminator(nn.Module):
    def __init__(self, im_dim):
        super(Discriminator, self).__init__()

        self.disc = nn.Sequential(
            disc_block(im_dim, 1024),
            disc_block(1024, 512),
            disc_block(512, 256),
            nn.Linear(256, 1),
        )


    def forward(self, x):
        return self.disc(x)

Define Discriminator class

Sequence of discriminator blocks and a linear layer

def disc_block(in_dim, out_dim):
  return nn.Sequential(
      nn.Linear(in_dim, out_dim),
      nn.LeakyReLU(0.2)
  )

Pass input through all layers
Input: image of size in_dim
Output: classification of size 1

Let's practice!

Deep Learning for Images with PyTorch