Learning Efficient Convolutional Networks Through Network Slimming