Thanks for sharing. This was great. I wonder how much of a limitation it poses that the gradient requires some massaging to preserve efficient training; does t

Thanks for sharing. This was great. I wonder how much of a limitation it poses that the gradient requires some massaging to preserve efficient training; does that only work for some kernels, or can it be automated for arbitrary kernels?