How to write Deconvolution upsamling layer for big scale? How's stride related to filter size?

[Ilya Zhenin]

I want to upsamlpe my blob by known factor, for example from (1,1,2,2) to (1,1,80,80).

I understand that my strdie value can be seen as scale factor. So it is supposed to be stride=40 for this case?

I also do not understand how Deconvolution filter size relates to stride. How big filter size should be in this case?

I looked at FCN example, they use stride=16 and filter size 32 (FCN Voc-32). It seems that filter size should be 2  * stride? So each filter gets multipled by real value from blob, besides zeros. But I'm not sure that I understand deconvolution that well, though I read a lot of available resources.

[Przemek D]

I wrote an explanation for this once, hope it helps. Also see this answer if you're interested in creating a bilinear upsampling layer.