I read an interesting paper called Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.
I am having some difficulty understanding the concept of "Pixel Shuffle", which uses sub-pixel convolution to upscale the input images in the final layer.

If anybody could explain how exactly the sub-pixel and transpose convolution operations are performed, I'll be very thankful.

P.S. I referred to this document too, but it didn't help much.