16-726 Learning-Based Image Synthesis
Project 3: When Cats meet GANs
Trung Nguyen
Objective
The goal of this assignment is to generate images with GAN networks. Specifically, we start with a vanilla DCGAN and then proceed with CycleGan that sample from two different domains. We will use DCGAN on grumpy cats images and CycleGan on cats and Pokemons images
Padding Formula
The formula to compute the size for output is given below
DCGAN
Architecture
The architecture of the DC GAN is decribed as below
Discriminator
Generator
Training steps
Data Augmentation
Two type of data augmentation is implemented in this experiment
Basic: With only image normalization with mean = 0.5 and std = 0.5
Deluxe: With random crop, random horizontal flip and image normalization
DCGAN Resutls
Basic Data Augmentation
Iteration 200
Iteration 1000
Iteration 5000
Iteration 10000
Deluxe Data Augmentation
Iteration 200
Iteration 1000
Iteration 5000
Iteration 10000
The curves in all of the four graphs look as GAN manages to train. The D loss decreases as the G loss oscilates around 1 with smaller oscillation as it learns.
CycleGAN
For this experiment, only deluxe dat augmentation is used
Architecture
The architecture of the CycleGAN is decribed as below
Training steps
CycleGAN Resutls
Cat dataset
Without Cycle Loss
Iteration 200
X to Y
Y to X
Iteration 600
X to Y
Iteration 1000
X to Y
Iteration 5000
X to Y
Y to X
Iteration 10000
X to Y
Y to X
With Cycle Loss
Iteration 200
X to Y
Y to X
Iteration 600
X to Y
Iteration 1000
X to Y
Iteration 5000
X to Y
Y to X
Iteration 10000
X to Y
Y to X
Pokemon dataset between Fire and Water pokemons
Without Cycle Loss
Iteration 200
Fire to Water
Water to Fire
Iteration 1000
Fire to Water
Water to Fire
Iteration 5000
Fire to Water
Water to Fire
Iteration 10000
Fire to Water
Water to Fire
Iteration 20000
Fire to Water
Water to Fire
Iteration 40000
Fire to Water
Water to Fire
Iteration 55000
Fire to Water
Water to Fire
With Cycle Loss
Iteration 200
Fire to Water
Water to Fire
Iteration 1000
Fire to Water
Water to Fire
Iteration 5000
Fire to Water
Water to Fire
Iteration 10000
Fire to Water
Water to Fire
Iteration 20000
Fire to Water
Water to Fire
Iteration 40000
Fire to Water
Water to Fire
Iteration 55000
Fire to Water
Water to Fire
