Convolutional neural network channel pruning method based on characteristic variance ratio

Abstract

The invention relates to a convolutional neural network channel pruning method based on a characteristic variance ratio. Firstly, a data set is trained; estimating the variance of the primary featuremap of each input channel in each layer after being filtered by the convolution kernel corresponding to the input channel in each filter and the variance of the output feature map corresponding to each filter; according to the sum of the ratio of the variance of each primary characteristic pattern of each input channel in each layer after being filtered by different filters to the variance of theoutput characteristic pattern of the filter corresponding to the primary characteristic pattern; and finally, performing global channel pruning on the convolutional neural network according to the importance parameter of each input channel of each layer. Compared with a traditional channel pruning method, the method has the advantages that interpretability is achieved, extra hyper-parameters are not introduced, and the pruned network structure does not need to be defined manually.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and in particular relates to a convolutional neural network channel pruning method based on feature variance ratio. Background technique [0002] Deep convolutional neural networks have achieved success in many areas of computer vision, such as image classification, object detection, semantic segmentation, etc. However, the performance improvement comes at the cost of huge storage and computing resources. The current deep convolutional neural networks are often accompanied by a huge amount of parameters and calculations, which makes their deployment on resource-constrained devices such as mobile terminals and embedded boards extremely difficult. In order to solve this problem, researchers have proposed many deep neural network compression and acceleration methods, such as channel pruning, knowledge distillation, tensor decomposition, etc. Among them, channel pruning has become a mainstrea...

AI Technical Summary

Problems solved by technology

Although these methods are simple to implement, they only consider local connection information, but do not consider the connection relationship between layers, which makes them unable to fully compress the network.

Images