Super-res predictions not displaying properly... (Lesson 7)

Hello all,

  • This is my first post, thanks for this amazing course and community!

I’m trying to set up the “SuperRes” example from the course with my own image crappifier.
The concept is to create a model that can turn a pencil sketch into a real photo
So something like this would become an actual cat image:

I found a small code snippet that does this kind of image -> pencil effect and ran it on the images.

The training seems to have worked really well, and the generated images during training are surprisingly accurate:

When I tried to test the results with the same code from the example notebook, I am getting a distorted view of some kind.
My assumption is that the setting of the model are not getting propagated, but I couldn’t find a solution.
Would appreciate your help.

Same code as example

Same dog that looked nice during training:

When trying on an out-of-dataset image, it also looks bad:

Just updating that I have solved the problem…
It’s solved by enforcing size as an int (not a tuple).

I believe that using the tuple breaks the model since it was trained with the int. @jeremy I don’t understand how it worked for your example, maybe the example happened to have these exact proportions?

The fix (for my model)

The cat sketch has made up green eyes: