Convolution optimization method based on distributed coding

Abstract

The invention relates to a convolution optimization method based on distributed coding. The method comprises the steps of dividing an input initial matrix into a preset number of sub-matrixes, whereinthe size of each sub-matrix is a * a; respectively encoding each sub-matrix to obtain each encoding matrix, and respectively sending each encoding matrix to the corresponding sub-node; receiving a sub-node convolution operation result obtained by carrying out convolution operation on the coding matrix by each sub-node by adopting a Winograd algorithm; when the number of the received sub-node convolution operation results reaches a preset value, decoding each received sub-node convolution operation result to obtain each convolution result; and splicing the convolution results to obtain a finalconvolution result. Convolution is calculated by a Winograd algorithm, calculation efficiency can be effectively improved, the influence of the straggler node on the calculation delay of the system can be reduced through distributed coding, the situation that the straggler node slows down the task execution time of the system is relieved, and therefore the operation speed of the distributed system is increased.

Description

technical field [0001] The present application relates to the technical field of distributed computing architecture, in particular to a convolution optimization method based on distributed coding. Background technique [0002] Convolution is a mathematical method of integral transformation, which is defined as the sum of two variables multiplied within a certain range. Convolution has a wide range of applications in statistics, probability theory, signal processing, and computer science, and plays a vital role. In today's more popular field of deep learning, convolution also plays an important role. When an input containing a lot of redundant and useless information enters the neural network, how to remove these useless information and keep only useful information input into the network becomes a problem. A key issue, convolutional filtering plays a key role in it. The method of using convolutional operations to filter input information and then extract features is the basi...

AI Technical Summary

Problems solved by technology

[0003] Convolution is widely used in deep learning, image processing and other fields. How to optimize the operation of convolution in the computer and accelerate its operation speed has become a problem that can be discussed. At the same time, with the development of distributed computing, its use is also It is becoming more and more popular, in which the straggler nodes in the distributed system will slow down the time for the entire system to perform tasks, resulting in a long time-consuming calculation speed of the distributed system

Images