Convolutional neural network compression and transplantation method and device, equipment and storage medium

Abstract

The invention discloses a convolutional neural network compression and transplantation method and device, equipment and a storage medium. The convolutional neural network compression method comprises the following steps: acquiring an initial convolutional neural network to be compressed, and adding a sparse convolutional layer behind each convolutional layer to be compressed in the initial convolutional neural network; obtaining a weight parameter of each sparse convolutional layer, and performing pruning and decomposition processing on each to-be-compressed convolutional layer based on the weight parameter to obtain a sparse convolutional neural network; training the sparse convolutional neural network, and taking the trained convolutional neural network as a compressed convolutional neural network. According to the scheme, model parameters can be effectively reduced.

Description

technical field [0001] The present application relates to the technical field of artificial intelligence, in particular to a convolutional neural network compression and transplantation method, device, equipment and storage medium. Background technique [0002] Over the past few years, convolutional neural networks have achieved state-of-the-art performance in various computer vision tasks. However, millions of parameters and heavy computational burdens, which are essential for new advances in this field, are not practical for deploying neural network solutions on edge devices and mobile devices. [0003] Therefore, it is necessary to propose a solution that can compress the neural network model and facilitate the deployment and operation of the mobile terminal, so as to reduce the size of the original model without reducing the accuracy. Although both pruning-based and decomposition-based methods can maintain model compression and accuracy, both pruning-based and decomposi...

AI Technical Summary

Problems solved by technology

The pruning method can only take effect in the feature output channel. It does not prevent the convolution kernel parameters from participating in the convolution. At the same time, the pruning operation causes some key features to be unable to propagate downwards, which affects the network accuracy; for the convolution kernel decomposition, another 1*1 convolutional layer, which means additional GPU calculations will be added

Images