A CNN-based low-precision training and 8-bit integer quantitative reasoning method

Abstract

The invention provides a CNN-based low-precision training and 8-bit integer quantization reasoning method. The method mainly comprises the steps of carryin gout low-precision model training; Performing model training by using a 16-bit floating point type low-precision fixed point algorithm to obtain a model for target detection; Quantifying the weight; Proposing an 8-bit integer quantization scheme, and quantizing the weight parameters of the convolutional neural network from 16-bit floating point type to 8-bit integer according to layers; carrying out 8-bit integer quantitative reasoning; quantizing the activation value into 8-bit integer data, i.e., each layer of the CNN accepts an int8 type quantization input and generates an int8 quantization output. According to the invention, a 16-bit floating point type low-precision fixed point algorithm is used to train a model to obtain a weight; Compared with a 32-bit floating point type algorithm, the method has the advantages that the 8-bit integer quantization reasoning is directly carried out on the weight obtained by training the model, the reasoning process of the convolutional layer is optimized, and the precision loss caused by the low-bit fixed point quantization reasoning is effectively reduced.

Description

technical field [0001] The invention belongs to the technical field of convolutional neural network, and in particular relates to a CNN-based low-precision training and 8-bit integer quantization reasoning method. Background technique [0002] Convolutional Neural Networks (CNNs) have achieved excellent results in the fields of image classification, object detection, face recognition, etc. To realize real-time forward reasoning of CNNs on the embedded platform, it is necessary to compress the model size of the neural network and improve the computational efficiency of the model under the condition of controlling the loss of accuracy. [0003] The current commonly used method is to quantize the weight and (or) activation value of CNN, and convert the data from 32-bit floating point type to lower integer type. However, the current quantization methods still have shortcomings in the trade-off between accuracy and computational efficiency. Many quantization methods perform netw...

AI Technical Summary

Problems solved by technology

[0005] The problem to be solved by the present invention is to overcome the defects in the above-mentioned prior art, and propose a CNN-based low-precision training and 8-bit integer quantization reasoning method to solve the large loss of precision existing in the existing quantization method , Calculation efficiency is not high enough

Images