Neural network low-bit quantization method

Abstract

The invention relates to a neural network low-bit quantification method. The weight value of each channel of a neural network is quantified to a low-bit fixed-point weight. And the method also includes obtaining an input quantization coefficient of the current convolution layer according to the target quantization threshold and the floating point quantization interval; taking the input quantization coefficient of the next convolution layer as the output quantization coefficient of the current convolution layer; quantizing the input floating point data to obtain input fixed point data; and quantizing the output floating point data to obtain output fixed point data; and converting the scaling factor and the bias into a scaling factor fixed-point value and a bias fixed-point value respectively; according to the low-bit fixed-point weight, the input fixed-point data, the output fixed-point data, the scaling factor fixed-point value and the offset fixed-point value, obtaining a neural network, and applying the quantized neural network model to embedded equipment.

Description

Technical field[0001]The invention relates to the field of data compression, in particular to a neural network low-bit quantization method.Background technique[0002]Neural network technology has achieved good results in tasks including image classification, target detection, and natural language processing. In order to improve recognition accuracy, the scale of neural network models has been increasing, and the complexity of the model has been increasing.[0003]Subsequently, higher requirements have been placed on the computing performance of the equipment, and the increasing network scale and power consumption have gradually become the main obstacles limiting the application of neural networks. The ever-increasing neural network has led to the need for more and more memory when computing the neural network. At the same time, the increase of the network model also requires greater bandwidth, which greatly limits the application of neural networks in embedded devices.Summary of the in...

AI Technical Summary

Problems solved by technology

[0003] As a result, higher requirements are placed on the computing performance of the device, and the increasing network scale and power consumption have gradually become the main obstacles that limit the application of neural networks.

The ever-increasing neural network leads to more and more memory required for computing the neural network, and the increase of the network model also requires greater bandwidth, which greatly limits the application of neural networks in embedded devices

Images