An 8-bit Fixed Point Quantization Method for Sparse MobileNetV2
Jinxiang Yu, Ximing Yu, Yilin Liu, Liansheng Liu, Xiyuan Peng · 2021 China Automation Congress (CAC) · 2021
Sparse MobileNetV2 has potential for embedded image processing applications due to its low computation complexity. However, it uses 32-bit float point number system for computing and storage. As a result, its computing efficiency can be further improved. In this paper, an 8-bit fixed point quantization method for Sparse MobileNetV2 is proposed for model size reduction and computing operation simplification, which improves throughput and shrinks memory consumption. Firstly, quantization objects are confirmed by analyzing the computation pattern of Sparse MobileNetV2. Then, after the statistical analysis of the distribution characteristic of parameters, the 8-bit fixed point quantization method is applied. Experimental results of image classification on remote sensing image dataset show that the memory consumption is reduced to 18.7% and the throughput increases by l.6x, while there is only a slight accuracy loss.