Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference