In order to solve the shortcomings of traditional face attribute recognition, such as low training efficiency and huge model parameters, a multi network learning framework is proposed.The framework consists of two sub networks:face region location network and attribute classification network, so as to dynamically select different face regions for face attribute prediction In addition, this paper proposes a hint based model compression technique, which greatly compresses the network parameters without significantly reducing the accuracy.Finally, the performance of the proposed method is analyzed by simulation on celeba face attribute classification data set.The average classification accuracy of the proposed network model can reach 90.89%, and the network parameter is only 0.27M.