Could you please explain why we are unrolling the images (reshaping from 3d to 2d) per training example?
What is the benefit of this transformation? It has not been explained in the lectures by Andrew Sir
It is explained in the notebook for the Neural Style Transfer assignment. They are using the Gram Matrix as a metric to compute the similarity of images. You can think of it as a variation on the notion of a covariance matrix. In order to compute that, you need to unroll the images.