I am trying to implement a gaussian blur as a convolution layer in a network, where the weights do not change. I currently have this:

class GaussianBlur(nn.Module):
    def __init__(self, kernelSize=5, sigma=1):
        self.weight = nn.Parameter(self._calculate_weights(kernelSize, sigma),
                                   requires_grad=False)

    def forward(self, x):
        return F.conv2d(x, self.weight)

This then gives the error optimizing a parameter that doesn't require gradients from the optimizer.

It would be a nice feature to be able to exclude some parameters from the optimizer.

I also tried changing self.weight to a Variable, but then when .cuda() is called, the weights are not transferred, resulting in other errors