Assignment #3 - When Cats meet GANs

Sanil Pande (spande)

PART 1: DCGAN

DCGAN Discriminator Part 1: Padding

The padding should be 1. The formula to calculate this is:

Wout = [(Win−K+2P)/S]+1

where W=resolution, K=kernel size, P=padding and S=stride.

DCGAN Experiment: Loss Curves

If the GAN manages to train properly, the discriminator loss will start between 0 and 1, but stay closer to 1. This happens for both the real and fake losses. For the Generator, initially the loss is higher, around 2, but then reduces close to 0.5, staying slightly above it.

DCGAN Experiment: Results

Iteration 200:                                                   Iteration 9000;

PART 2: CycleGAN

1. Without Cycle Consistency Loss: 600 iters

2. With Cycle Consistency Loss: 600 iters

3. With Cycle Consistency Loss: 10000 iters

4. Difference with/without consistency

It is immediately apparent that without cycle consistency loss, the structure between the images is lost, or not as good as it is when translating using cycle consistency. This is because the cycle consistency encourages the translated image to retain the structure of the original image so that it can be recovered using another translation, which would not be possible if the original structure is lost.

Bells and Whistles

Different Datasets:

1. Pokemon Dataset using DCGAN

This was trained for ~100 epochs on the given pokemon dataset.