Test size is smaller than train size

[eranharpaz]

Hi

We trained our networks on 256x256 images.
During test we (accidentally) applied the generator on 128x128 images.
The result for 128x128 images were drastically degraded in comparison to 256x256.

Is this something you'd expect? and if so what is the reason for that? After all it is a fully convolutional network.

Thanks