Given what we saw in class about the checkerboard patterns that appear when stride and conv layers aren’t divisible is there a reason why the neural net conv and deconv layers don’t use stride 3?
I understand that we can get around this by using an upsampling layer instead, but what’s the motivation for stride 2? Is it just the additional pixel information?
A larger stride means that more information is lost at each step, so we’d rather avoid it if possible. Using upsampling works great, so I see no reason to use stride=3