I am a researcher in the field of machine learning and deep learning, and I’m wondering if there are development plans for including local convolutions in cudnn?
local convolutions are convolutions with unshared weights, meaning you have a learnable kernel at every location of the image, instead of a kernel that have shared weights meaning it is unchanged over the entire image. Local convolutions have been employed in DeepFace for face detection, and we’re trying to implement something similar. However, cudnn doesn’t seem to support local convolution right now, and therefore our cuda version is running very slowly…
it would be great if cudnn supports local convolution, does anyone know if that’s part of the development plan in cudnn?