A convolution neural network optimization method based on a winograd algorithm

Abstract

The invention discloses a convolution neural network optimization method based on a winograd algorithm, comprising: (1) determining a convolution neural network to be optimized, wherein the convolution neural network to be optimized comprises at least one convolution layer, each convolution layer comprises a plurality of filters, each filter has a plurality of channels, and each channel corresponds to a convolution core; (2) in each convolution layer, sorting according to the sum of absolute values of each convolution kernel, selecting convolution kernel with small sum of absolute values according to proportion for cutting, and updating the mask matrix of each convolution layer; (3) determining an output characteristic map corresponding to each filter using a winograd algorithm based on the mask matrix. The invention adds a mask matrix to identify the clipped convolution kernel when the convolution kernel is converted based on winograd algorithm. The preset matrices W, X, M are used tocompute the clipping model, so as to avoid the computational disadvantages caused by local unstructuring.

Description

Technical field [0001] The present invention relates to the technical field of artificial intelligence, and more specifically, to a convolutional neural network optimization method based on a winograd algorithm. Background technique [0002] After AlphaGo successfully defeated humans, the deep learning model has attracted more and more people's attention, especially the convolutional neural network model has shown surprising achievements in the business and research fields. Because of the world-renowned achievements of deep learning models, the need to deploy them on mobile devices such as mobile phones has emerged. However, the CPU and memory resources of these mobile devices are far from enough to run huge deep learning models. In order to solve such problems, the optimization and acceleration methods of deep learning models have quickly become popular. [0003] In the past two years, there have been many inference and optimization acceleration methods for deep models. The first...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the defects of the prior art, the purpose of the present invention is to solve the existing convolutional neural network optimization method, unstructured sparse will destroy the structure of the convolution kernel, seriously affect the calculation speed, and coarse-grained structured sparse coarse-grained Structuring will miscut many important parameters during the sparse process

Images