Project3 When Cats meet GANs

Lena Du

Part 1 DCGAN

1.1 Implement Data Augmentation

Deluxe data augmentation helps the model to be more robust.

1.2 DCGAN - Discriminator

Padding

The calculation of padding is:

, where m is the output size and n is the input size. p is the padding, K is the kernel size, and S is the stride. Given the size is downsampled by scale 2, we know n = 2m. With K = 4 and S = 2, we will have

Then,

which means padding is 1.

1.3 DCGAN - Generator

The design of the first layer in DCGenerator is using conv, instead of up_conv. The idea is to use padding 3, kernel size 4, and stride 1 to obtain a 4x4 output. I also replaced nn.ReLU with nn.LeakyReLU for its better performance.

1.4 Result

As we can see, the result of the Deluxe data augmentation + full diffaug configuration with more iterations has better quality and resolution.

Config	Real Image	1000 Iterations	7000 Iterations
Basic
Deluxe
Deluxe + diffaug (cutout)
Deluxe + diffaug (color, translation)
Deluxe + diffaug (color, cutout)
Deluxe + diffaug (color, translation, cutout)

Config	Loss
Basic
Deluxe
Deluxe + diffaug (all)

Part 2 CycleGAN

Observations:

• Generally, with more iteration, the result is better.
• If the Russian blue cat image has eye color other than those of most of the blue cat images, then the model is hard to generate the eyes of grumpy cat result.
• The quality of results is improved with Cycle Consistency.
• Patch discriminator has better performance than the DC discriminator.

2.1 Cat

2.1.1 DC Discriminator /wo Cycle Consistency

Direction	1000 Iterations	5000 Iterations	10000 Iterations
X -> Y
Y -> X

2.1.2 DC Discriminator /w Cycle Consistency

Direction	1000 Iterations	5000 Iterations	10000 Iterations
X -> Y
Y -> X

2.1.3 Patch Discriminator /wo Cycle Consistency

Direction	1000 Iterations	5000 Iterations	10000 Iterations
X -> Y
Y -> X

2.1.4 Patch Discriminator /w Cycle Consistency

Direction	1000 Iterations	5000 Iterations	10000 Iterations
X -> Y
Y -> X

2.2 Fruits

2.2.1 DC Discriminator /wo Cycle Consistency

Direction	1000 Iterations	5000 Iterations	10000 Iterations
X -> Y
Y -> X

2.2.2 DC Discriminator /w Cycle Consistency

Direction	1000 Iterations	5000 Iterations	10000 Iterations
X -> Y
Y -> X

2.2.3 Patch Discriminator /wo Cycle Consistency

Direction	1000 Iterations	5000 Iterations	10000 Iterations
X -> Y
Y -> X

2.2.4 Patch Discriminator /w Cycle Consistency

Direction	1000 Iterations	5000 Iterations	10000 Iterations
X -> Y
Y -> X

2.3 Loss

Discriminator	Smoothness 0.972
DC /wo Cycle Consistency
DC /w Cycle Consistency
Patch /wo Cycle Consistency
Patch /w Cycle Consistency

Discriminator	Smoothness 0.999
DC /wo Cycle Consistency
DC /w Cycle Consistency
Patch /wo Cycle Consistency
Patch /w Cycle Consistency

Part 3 B&W

Spectral loss

Config	Real Image	1000 Iterations	7000 Iterations
Deluxe + Instance
Deluxe + Spectral