Comment by perfobotto

is this approach viable on training? wat about kernels that require different grids?