Deep face model learned on big dataset surpasses human for face recognition task on difficult unconstrained face dataset. But in practice, we are often lack of resources to learn such a complex model, or we only have very limited training samples (sometimes only one for each class) for a specific face recognition task. In this paper, we address these problems through transferring an already learned deep face model to specific tasks on hand. We empirically transfer hierarchical representations of deep face model as a source model and then learn higher layer representations on a specific small training set to obtain a final task-specific target model. Experiments on face identification tasks with public small data set and practical real faces verify the effectiveness and efficiency of our approach for transfer learning. We also empirically explore an important open problem - attributes and transferability of different layer features of deep model. We argue that lower layer features are both local and general, while higher layer ones are both global and specific which embraces both intra-class invariance and inter-class discrimination. The results of unsupervised feature visualization and supervised face identification strongly support our view.