Learning Finer-class Networks for Universal Representations