Aiming at the problem that convolutional neural networks are difficult to deploy and apply in edge devices with limited memory and computing resources,a mixed precision quantization method based on loss variation was proposed.Resources were reduced using a lower bit fixed-point number instead of float precision.The bit widths were adjusted based on the first and second order information of each quantization layer,and the layers were clustered into blocks using K-means to reduce the search space.An adaptive strategy search method was proposed,the search state was adjusted according to the historical strategy results.The process of quantization training was reintegrated to reduce computation.Experimental results show that the proposed method can effectively compress the CNN with a small inference loss accuracy.
维普
万方数据