Integer Quantization for Deep Learning Inference : Principles and Empirical Evaluation