I want a convolution layer for an input image, with each sliding window gets rotated by a distinct angle (distinct angles for distinct windows) before convolution.
For example, the input image has the size 100×100, and the kernel is 10×10. At a certain pixel, the coordinates of all pixels in the sliding window should rotate 10 deg first (needs padding if the rotated window is not a rectangle), and then calculate the convolution.
How to achieve this using torch? Thank you.
