12-02 (MLSys 2020) TQT - Trained Quantization Thresholds for Accurate and Efficient Fix-point Inference of Deep Neural Networks