Convolutional neural network quantification method and device, computer and storage medium

Abstract

The invention provides a convolutional neural network quantification method, which comprises the steps of training a full-precision model of a convolutional neural network to be quantified, and calculating standard deviation of weight and response distribution of each layer of the full-precision model; estimating scale factors of parameters and features of the full-precision model according to thestandard deviation and hyper-parameters of the weight and response distribution of each layer of the full-precision model; for the to-be-optimized convolutional neural network, establishing a quantization module containing scaling factor-based forward calculation and backward gradient propagation functions to obtain a corresponding quantization network; carrying out fine tuning training on the quantization network, and determining an optimal scale factor; and retraining the quantization network generated by the optimal scaling factor to obtain a final quantization neural network model. The invention further provides a convolutional neural network quantization device, a computer and a storage medium. According to the invention, the problems of complex realization and high calculation complexity of the existing model quantification method are improved.

Description

technical field [0001] The present invention relates to a deep neural network compression method, in particular to a method, device, computer equipment and storage medium for compressing and accelerating a convolutional neural network through model quantization. Background technique [0002] In today's computer vision and other technical fields, deep learning has proven to be a very useful method, and has achieved good results in tasks such as image classification, object detection, and semantic segmentation. At present, with the continuous improvement of the theory, the deep neural network model has a tendency to develop in the direction of more parameters, deeper network, and greater calculation. At the same time, the industry is gradually applying deep learning technology to specific scenarios, which puts strict requirements on the size of the model, computing performance, power consumption and other indicators. [0003] In recent years, the application of neural network...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to provide a convolutional neural network quantization method, device and computer equipment, to improve the existing convolutional neural network quantization method complex implementation, high computational complexity

Images