Trained Quantization Thresholds for Accurate and Efficient Fixed-Point Inference of Deep Neural Networks.