Perhaps the idea of not having a training set means that one should use other datasets such as image net and use transfer learning.