When working with classifiers, a class imbalance is a huge issue for our models. If we have too many images of
class 1 and too few images from
class 2, then there is a class imbalance and our classifier could consistently pick
Class 1 and be correct without ever looking at dataset features.
When looking at a Style GAN to map images from one style to another (photographs to paintings), my model looks prone to having the same problem.
Intuitively, I feel that a Wasserstein Loss may be better, since it only cares about comparing probability densities rather than class representation. But, realistically, does the disparity between the number of images really lead to a problem with generalization?