Neural network cutting method and device, equipment and storage medium

Abstract

The invention provides a neural network cutting method and device, equipment and a storage medium, and the method comprises the steps: determining the correlation between convolution kernel parameters of a to-be-cut convolution kernel in a to-be-cut neural network; determining a clipping mask corresponding to the to-be-clipped convolution kernel according to the correlation between the convolution kernel parameters of the to-be-clipped convolution kernel; and according to the cutting mask corresponding to the to-be-cut convolution kernel, cutting the convolution kernel parameter of the to-be-cut convolution kernel. The cutting method provided by the invention does not destroy the structured information of the network parameters of the original neural network, so that the neural network cut by the cutting method provided by the invention can be deployed in most hardware platforms without being supported by additional special hardware, and in addition, the convolution kernel parameters of the to-be-clipped convolution kernel are clipped according to the clipping mask corresponding to the to-be-clipped convolution kernel instead of directly clipping the whole convolution kernel or the whole channel, so that refined clipping of the neural network is realized.

Description

technical field [0001] The present application relates to the technical field of network pruning, and in particular to a neural network pruning method, device, equipment and storage medium. Background technique [0002] In recent years, deep learning technology has achieved remarkable results in image recognition, speech recognition, natural language processing and other fields. However, due to the huge amount of parameters of the deep neural network, it is difficult to use it in many scenarios, such as embedded systems and mobile phones. In order to solve this problem, currently popular methods include network model search, network model quantization, and network model clipping. Among them, network model clipping is widely used because of its simplicity, reliability, and maturity. [0003] The current network model pruning is mainly weight pruning. Weight pruning is a refined pruning method, which prunes the independent parameters whose median value of the network paramete...

AI Technical Summary

Problems solved by technology

[0004] In view of this, the present application provides a neural network clipping method, device, device and storage medium to solve the problem that the existing neural network clipping method destroys the structure of the network parameters of the original neural network, causing most hardware platforms to It is difficult to support the problem of the tailored network, and the technical solution is as follows:

Images