This assignment seeks to explore DCGAN and CycleGAN.
Part 1: Deep Convolutional GAN
Data Augmentation
The methods of Data Augmentation I used were the following:
Resize: resize width and height by 1.1
Random Horizontal Flip: randomly decide to flip horizontally
Random Crop: Randomly decide to crop or not to the size of the original image
DCGAN Discriminator padding
I used the formula for calculating the padding. $$P = (S(O-1) + K - I)/2$$ where P, S, O, K and I each stand for
padding size, stride, output size, kernel size and input size, respectively. Hence, $$P = (2(32-1) + 4 - 64)/2 = 1$$
Differentiable Augmentation
Here in this section I report the differentiable augmentation samples from DCGAN from basic and deluxe transforms.
This is the sample result without differentiable augmentation with basic transforms.
This is the sample result with differentiable augmentation with basic transforms.
This is the sample result without differentiable augmentation with deluxe transforms.
This is the sample result with differentiable augmentation with deluxe transforms.
Experiment Loss Curve
Discriminator Loss for basic without differentiable agumentation
Generator Loss for basic without differentiable agumentation
Discriminator Loss for basic with differentiable agumentation
Generator Loss for basic with differentiable agumentation
Discriminator Loss for deluxe without differentiable agumentation
Generator Loss for basic without differentiable agumentation
Discriminator Loss for basic with differentiable agumentation
Generator Loss for basic with differentiable agumentation
Part 2: CycleGAN
These are the images I got from running the CycleGan on the grumpy cat.
X-Y after 1000 iterations without cycle consistency loss
Y-X after 1000 iterations without cycle consistency loss
X-Y after 6500 iterations without cycle consistency loss
Y-X after 6500 iterations without cycle consistency loss
X-Y after 1000 iterations with cycle consistency loss
Y-X after 1000 iterations with cycle consistency loss
X-Y after 6500 iterations with cycle consistency loss
Y-X after 6500 iterations with cycle consistency loss
The following are the results from apple to cat.
X-Y after 1000 iterations without cycle consistency loss
Y-X after 1000 iterations without cycle consistency loss
X-Y after 6500 iterations without cycle consistency loss
Y-X after 6500 iterations without cycle consistency loss
X-Y after 1000 iterations with cycle consistency loss
Y-X after 1000 iterations with cycle consistency loss
X-Y after 6500 iterations with cycle consistency loss
Y-X after 6500 iterations with cycle consistency loss
High Resolution Results
I also experimented with high resolution images as inputs. These are the results for the same dataset with 128x128 resolution.
X-Y after 1000 iterations without cycle consistency loss
Y-X after 1000 iterations without cycle consistency loss
X-Y after 6500 iterations without cycle consistency loss
Y-X after 6500 iterations without cycle consistency loss
X-Y after 1000 iterations with cycle consistency loss
Y-X after 1000 iterations with cycle consistency loss
X-Y after 6500 iterations with cycle consistency loss
Y-X after 6500 iterations with cycle consistency loss
These are the results for the same dataset with 256x256 resolution.
X-Y after 1000 iterations without cycle consistency loss
Y-X after 1000 iterations without cycle consistency loss
X-Y after 6500 iterations without cycle consistency loss
Y-X after 6500 iterations without cycle consistency loss
X-Y after 1000 iterations with cycle consistency loss
Y-X after 1000 iterations with cycle consistency loss
X-Y after 6500 iterations with cycle consistency loss
Y-X after 6500 iterations with cycle consistency loss
