I am trying to understand what is happening in the scenario where I have a 2d
convolutional layer with the following parameters:
2d_conv = nn.Conv2d(3, 3, kernel_size=1, stride=1,
padding=0, bias=False, dilation=2)
Is this just the same as a regular 1x1 2-d convolution, since I’m not sure what effect
the dilation would have.
I am asking since in the paper describing the a-trous algorithm modification of a resnet-101 backbone in R-FCN the authors write:
All layers before and on the conv4 stage [9] (stride=16) are unchanged; the stride=2 operations in the first conv5 block is modified to have stride=1, and all convolutional filters on the conv5 stage are modified by the “hole algorithm” [15, 2] (“Algorithme à trous”) to compensate for the reduced stride