I was wondering if the Lesson 7: SuperRes tutorial could be modified in order to take input images that are noise and utilized the pretrained resnet model as a starting point. I used the basic generator and basic critic in the wgan tutorial but the results are not that good. But I am looking at data that is quite diverse. Pixel monsters that are floating blobs, quadrupeds, bipeds, etc.
When I tried to do rework Lesson 7 SuperRes to use noise as input and a gen_learner, I get an error regarding the input size being incorrect. This has to do with the fact that the GANItemList() uses 100 int random input and not an image of the correct input size.
Is there a step that allows one to use the GANItemList() with a generator learner?
Does transfer learning in this way work for generating new content and not simply increasing the resolution of the image?