Compression method, system and device of neural network and medium

Abstract

The invention discloses a compression method of a neural network. The compression method comprises the following steps: respectively calculating gray level co-occurrence matrixes of feature maps output by a plurality of preset layers of a to-be-compressed network in a plurality of directions; respectively averaging a plurality of texture features corresponding to the gray level co-occurrence matrixes of each preset layer in multiple directions, so as to form a texture feature vector by utilizing the average values corresponding to the plurality of texture features; splicing the texture featurevectors of each preset layer into a texture feature tensor; and inputting the texture feature tensor into the trained primary meta-network to obtain a compression ratio of each layer of the to-be-compressed network, and compressing the to-be-compressed network by using the compression ratio. The invention further discloses a system, computer equipment and a readable storage medium. The inventionprovides a neural network compression method based on multi-scale feature map texture analysis, which can reduce the calculation amount and accelerate the flexible compression process of a network tobe compressed under different Flops constraint conditions.

Description

technical field [0001] The invention relates to the field of neural networks, in particular to a neural network compression method, system, device and storage medium. Background technique [0002] The accuracy of the deep neural network gradually increases with the deepening of the network and the expansion of the width. For machine vision tasks such as detection and image segmentation, the structure of the neural network needs to be compressed and adjusted with reference to the constraints of the operating platform to reduce the number of parameters of the neural network, storage space consumption, FLOPs overhead, etc., and ultimately ensure that the neural network can operate on a specific Implement deployment on the running platform. However, the amount of calculation and training delay required by the neural network compression process itself is already very considerable. If the iterative processing of compression and fine-tuning is performed each time it is deployed to...

AI Technical Summary

Problems solved by technology

The search method based on the evolutionary algorithm needs to iteratively perform structure prediction and forward reasoning to complete the structure search and performance evaluation until the maximum number of iterations is reached. This process still takes a lot of time and computation.

Images