output size does not match input size when padding=True

[yi-li-yang]

Hi, I input a image (1646,1646) into the unet, and assigned padding=True, but the output is (1632,1632), may I ask why and how to make the size unified?

[jvanvugt]

It is because the input sizes are divided by 2 (and floored) in each pooling layer. 1646 is divisible by 2 only once. You should either crop or pad your input images such that they are a multiple of 2^(depth-1). By the way 1646 is already ridiculously large. It might be a good idea to downsample or train on crops of the original image.