Towards Real-Time DNN Inference on Mobile Platforms with Model Pruning and Compiler Optimization.