Hello everyone,
I am training an SRGAN and after each output I see a visual grid on each image. The problem seems to be originating from the Conv2d and is highly related to the kernel size. the higher the kernel size, the thicker the grid is. This grid only appears when working with images that have more than 1 channels, if you use 1 channel grayscale images for example, the grid does not appear. It is as if internally the image gets tiled and retiled after each convolutional layer.
Here is a demonstration of the issue (35 kernel to make the grid easier to see):
The same SRGAN implementation in tensorflow produces none of the above issues. any thoughts?
Thanks in advance.