Hello,
I am training on FCN using caffe-master branch.
1. I am trying to figure out the behaviour of crop layer. I am using the crop function provided by caffe. But I don't understand how it works.
In the deconvolution layer, if I keep filter_size = 2xstride, sometimes it works, sometimes it throws the error that offset is a fraction. Then I have to keep filter_size = 2xStride - 1
Can anyone explain the logic behind it or how to adjust the filter_size?
2. When I give an offset to a crop layer in FCN, is it always fixed for any input image size (which means it is not a center-cropping)?
Best Regards,
ARK