ACGAN Architectural Design

ACGAN Model

Conditional Image Synthesis with Auxiliary Classifier GANs paper

• CNN have brought advances in image classification

• CNN can also be reversed to generate images from scratch (generative models)

• One type of generative model are generative adversarial networks (GANs)

• Special type of GAN is Auxiliary Classifier GANs (ACGAN)

• GAN with class-label conditioning to generate images

Example of Generated Images Conditioned on Label

Introduction to Generative Adversarial Networks (GANs)

• GAN is composed of two competing neural network models (often CNNs)

• Generator: takes noise input and generate a realistic image

• Discriminator: takes real and fake images and has to distinguish the fake from the real

• Two networks play an adversarial game

• generator learns to produce more and more realistic samples

• discriminator learns to get better and better at distinguishing generated data from real data.

• networks are trained simultaneously

Training on real image:

• Discriminator should classify real image as real

• Ouput probability close to 1

Training on fake image:

• Generator generate fake image from noise

• Discriminator should classify fake image as fake

• Ouput probability close to 0

Auxiliary Classifier ACGAN

Proposed a new method for improved training of GANs by conditioning input with class labels.

Multi-input multi-output network:

• Inputs: class embedding and noise vector

• Outputs: binary classifier (fake/real images) and multi-class classifier (image classes)

Next Lesson

Implementation of ACGAN model

• Generator

• Discriminator

• dual loss functions