Neural network mixed precision quantization method and system for memristors

Abstract

The application provides a neural network mixed precision quantization method and system of a memristor, comprising: obtaining the number of convolution kernels, obtaining the final output through a grouping quantization strategy on the number of convolution kernels; defining a signed uniform symmetric quantization of a quantizer; obtaining the interval between any two adjacent quantization levels according to the signed uniform symmetric quantization of the quantizer; wherein the interval is obtained through a threshold value and a bit width; obtaining the optimal threshold value and the optimal bit width through a network architecture search algorithm; analyzing the resource consumption of each module under the quantization bit width under an accelerator, and incorporating it into a loss function; through the grouping quantization strategy, the group size can be dynamically adjusted flexibly according to the cross array size; the quantizer design, the network architecture search algorithm and the cross array-aware regularizer can obtain the trade-off between the recognition rate and the resource consumption.

Images