Trained Uniform Quantization for Accurate and Efficient Neural Network Inference on Fixed-Point Hardware.