Can you explain a bit how atrous mode combines with the pretrained weights?
I'm aware of how dilated/atrous convolutions work, however, I wonder how effective is it to copy weights of kernels that did not use dilation into kernels that do.
If you can elaborate on this and how empirically it worked for you that would be great :)
Can you explain a bit how atrous mode combines with the pretrained weights?
I'm aware of how dilated/atrous convolutions work, however, I wonder how effective is it to copy weights of kernels that did not use dilation into kernels that do.
If you can elaborate on this and how empirically it worked for you that would be great :)