How to Transform Kernels for Scale-Convolutions