dilated convolution with learnable spacings (DCLS) is a recent convolution
method in which the positions of the kernel elements are learned throughout
training by backpropagation. Its interest has recently been demonstrated in
computer vision (ImageNet classification and downstream tas