Sorry for the naive question but I am really struggling to understand how to create the Validation dataset in the case of GZSL. For Zero Shot learning, I can imagine we pick certain classes from the training set and use them for validation. However, for GZSL, I am not sure if this is the best way since we have both seen and unseen examples at test time. As such, should I also have same number of unseen classes as the seen classes? To make matters a bit more complicated, I am referring to this paper and there they have their validation set created only based on the seen classes. Any sort of clarification would be really appreciated.