BilinearFiller kernel size for interpolation

[Tim Meinhardt]

Hello,

today I came across the BilinearFiller for the deconvolution layer and was pretty happy

that somebody already implemented that. Until now I always filled in the weights manually.

But I found the given formula for the kernel_size somewhat irritating:

According to here the kernel for factor 2 should have size 3 which I found plausible looking at the example. But if I follow the documentation the kernel size would be 4.

As far as I understood the relationship between factor and kernel size should be: kernel_size = (factor - 1) * 2 + 1

Using a different kernel size with different weights might be negligible in a lot of cases, but since the deconvolutution layer

just sums up the resulting windows (see deconv layer) the result definitely differs when the kernel is bigger.

Am I getting something wrong or is the difference in this case just "negligible"?

Thank you for any help or tip that might solve my confusion!

Best,

Tim