Incremental Network Quantization: Towards Lossless CNNs with Low-Precision Weights