CNN model compression method and device based on DS structure, and storage medium

Abstract

The invention relates to the technical field of neural networks, and provides a CNN model compression method and device based on a DS structure, and a storage medium, wherein the method comprises thesteps: S110, enabling DW convolution and SE Module to form DS convolution blocks through common convolution and batch standardization BN operation, wherein the DS convolution blocks comprise a convolution Conv-1, a batch standardization BN, an activation function, a DW convolution, a batch standardization BN, an activation function, an SE Module, a convolution Conv-2 and a batch standardization BN; S120, stacking the DS convolution blocks in order to form a neural network structure; and S130, adding an input layer, a pooling layer, a full connection layer and a classification layer to the neural network structure to form a neural network model. According to the method, the DS convolution block structure is applied to the neural network, and the number of parameters of the neural network isgreatly reduced while the picture feature extraction capability is ensured.

Description

technical field [0001] The invention relates to the technical field of neural networks, in particular to a CNN model compression method, device and storage medium based on a DS structure. Background technique [0002] Although Convolutional Neural Network (CNN) has achieved outstanding results in fields such as computer vision and natural language processing, its number of parameters exceeding 100 million makes many practical applications (especially those based on embedded devices) prohibitive. Taking the classic VGG16 network as an example, the number of parameters has reached more than 130 million, occupying more than 500MB of storage space, and requires 30.9 billion floating-point operations to complete the recognition task of an image. Such a huge storage cost and computing overhead severely restrict the application of deep networks on small devices such as mobile terminals. [0003] In view of this, the compression of neural networks has gradually become a hot researc...

AI Technical Summary

Problems solved by technology

[0005] 1. The effect of compression is uneven;

[0006] 2. There is a serious loss of accuracy in the compressed model

Images