As we all know, zero shot learning involves a model predicting classes that it has not seen. But we are given all the attributes each class might have.
Is it fair to assume that we are "aware" of all the class labels a dataset might have ? (Including the test set)
The formal definition of zero-shot learning is that given labeled training instances $D_{tr}$ belonging to the seen classes $S$, the aim is to learn a classifier $f^u(·):X→U$ that can classify testing instances $X_{te}$ (i.e., to predict $Y_{te}$) belonging to the unseen classes $U$.
The general idea of zero-shot learning is to transfer the knowledge contained in the training instances $D_{tr}$ to the task of testing instance classification. So it is considered a transfer learning method, and, more specifically, heterogeneous transfer learning with different label spaces.
However, since no labeled instances belonging to the unseen classes are available, to solve the zero-shot learning problem, some auxiliary information is necessary. Such auxiliary information should contain information about all of the unseen classes. This auxiliary information usually contains some semantic information about the unseen classes, and its representation belongs to a space that is often referred as $semantic$ $space$. In the semantic space, each class has a corresponding vector representation, which is referred to as the class prototype or $prototype$ for short.
So, you should be aware of classes of your problem, in order to construct the semantic space and their class prototypes. I also suggest you read the survey paper "A survey of zero-shot learning: settings, methods and applications".
