Learning Compositional Representations for Few-Shot Recognition