Name : Divam Gupta

Andrew ID : divamg

Deep Convolutional GAN

Given width , kernal size, stride and padding the new width of the output is given by w' = (w-k+2p)/2 + 1

hence for the all the layers except last the padding should be 1 and the padding of last layer should be 0

Loss Plots

We can observe that all the losses decrease over time. We also observe that the fluctuation in the loss values are also decreased over time.

Outputs

We can see qualitatively that the outputs of the GAN improve. Initially the GAN only generates noise, then it starts outputting the same image, but towards the end it learns to generate the cat images in different poses / colors.

CycleGAN

Outputs

Without Cyclic Consistency Loss

With cyclic consistency loss:

We observe that the model with the cyclic consistency loss performs better than the model without cyclic consistency loss

Visually the X-Y looks much better than the Y-X.

Although we can see that the poses and positions of the eyes are aligned with each other. Which is learned by the model in an unsupervised way.