Convolutional neural network cutting method based on adjacent layer weight

Abstract

The invention discloses a convolutional neural network cutting method based on weights of adjacent layers. The method comprises the following steps: calculating the weight of each convolution kernel in each convolution layer in a convolutional neural network and the weight of a corresponding channel of the next layer; for each convolution layer, calculating the sum of the weight absolute values of each convolution kernel and the sum of the weight absolute values of the corresponding channels of the next layer, multiplying the sum of the weight absolute values of each convolution kernel and the sum of the weight absolute values of the channels of the next layer by corresponding convolution operation times and then adding to obtain a parameter C for measuring the importance of each convolution kernel in the convolution layer, and cutting off m convolution kernels of which the C values are smaller than a threshold value and the corresponding feature map to obtain a parameter C; cutting off a channel, which is subjected to convolution with the cut-off feature map, in the next convolution layer; and retraining the clipped network to complete network clipping. The convolutional neural network can be clipped under the condition that the precision is not lost by comprehensively considering three elements including the weight of the current convolutional layer, the weight of the next convolutional layer and the number of convolution operations of the convolutional neural network for face recognition and the like.

Description

technical field [0001] The invention relates to a convolutional neural network clipping method. Background technique [0002] Convolutional Neural Networks (Convolutional Neural Networks) is a feed-forward neural network that includes convolutional calculations and has a deep structure. It has a wide range of applications in image classification, image detection, semantic segmentation, and video tracking. [0003] With the increasing application of artificial intelligence technology in life, the edge computing of convolutional neural network has also begun to rise. However, convolutional neural network with excellent effect often has a lot of parameters, which limits its application on mobile terminals or embedded chips. running locally. Therefore, it is often necessary to cut the original neural network to a certain extent. Therefore, how to tailor convolutional neural networks without loss of accuracy has become a technique with practical value. [0004] "Pruning Filter...

AI Technical Summary

Problems solved by technology

[0005] However, the above method actually has a flaw: since the channel corresponding to the next layer needs to be removed when removing the convolution kernel of the current layer, the weight of the next layer is not considered when performing model clipping.

If the channel weight of the next layer to be removed is large, it may cause a large loss of accuracy

Images