I’m trying to set up a controllable GAN arcitecture, but i don’t want to use a class as the conditional input but two floating point variables (i’ts kind of an 2 angle dependend image deformation).
I’ve tried it with a simple DCGAN with conditional inputs, with moderate to good results. To enhance this i want to try more complex arcitecures like BIGGAN.
Unfortunately this structure is designed for classes
Now i could remap my dataset to from floating point to classes, but i would end up with far more than 1000 classes.
Do you think it is worth trying? Or do you have any idea how to modify the BIG-GAN arcitecture?
I’ve also searched for research papers with “no-classes” conditioning but i did not found any. Do you know some?
Does no one have an idea?