Convolutional neural network data processing method and device based on winograd convolution operation

Abstract

The invention provides a convolutional neural network data processing method to simplify the convolution operation, and the method comprises the steps of 1) calculating an intermediate matrix U for aweight value based on GgGT according to the weight value g of a convolutional neural network model and a conversion matrix G; 2) compressing the intermediate matrix U for the weight to obtain a compression conversion weight matrix Uz used for indicating the numerical value of an effective element in the U and a compression coordinate code S used for indicating the coordinate of the effective element in the U; 3) according to the compressed coordinate encoding S, calculating a compression result for an intermediate matrix V = [BTdB] as a compression conversion feature map Vz, d being an input feature map of the convolutional neural network, and B being the conversion matrix; 4) carrying out point multiplication operation on the compression conversion weight matrix Uz and the compression conversion feature map Vz to obtain a point multiplication matrix Mz, and 5) calculating a convolution operation result F for the input feature map d based on ATMzA according to the point multiplicationmatrix Mz and the conversion matrix A.

Description

technical field [0001] The invention relates to the acceleration of convolution operation, in particular to an acceleration scheme for convolution operation based on win ograd in neural network. Background technique [0002] Deep learning technology has developed rapidly in recent years, and it has been widely used in solving advanced abstract cognitive problems, such as image recognition, speech recognition, natural language understanding, weather prediction, gene expression, content recommendation and intelligent robots. And has excellent performance, so it has become a research hotspot in academia and industry. Deep neural network is one of the perception models with the highest level of development in the field of artificial intelligence. This type of network simulates the neural connection structure of the human brain by building a model, and describes the data features layered through multiple transformation stages, providing image, video and audio Such large-scale da...

AI Technical Summary

Problems solved by technology

[0004] However, this does not mean that the convolution operation method based on winograd is perfect, and there is still room for further use of sparsity to reduce the amount of calculation in traditional winograd

Moreover, there is currently no dedicated processor for sparse winograd convolution operations

Images