Method and system for edge filling and filter expansion operation in convolutional neural network

Abstract

The invention provides an edge filling and filter expansion operation method and system in a convolutional neural network, including: according to the input data block and the control signal of the filter data block, the input data block after the edge filling operation and the expansion operation Carry out zero-filling judgment on the filter data block to obtain the data control signal; according to the data control signal, perform zero-filling operation on the first input data vector and the first filter data vector to obtain the second input data vector and the second filter data vector; performing the logical operation of the convolutional neural network on the second input data vector and the second filter data vector. The method provided by the present invention judges whether the current operation is in the position of input zero padding or filter zero padding according to the index vectors provided in the program to the filter data block and the output data block, and realizes with hardware structure that originally needs software programming to realize The judging and jumping function realizes the acceleration of input edge filling operation and filter expansion operation.

Description

Technical field[0001]The embodiment of the present invention relates to the field of computer technology, in particular to a method and system for edge filling and filter expansion in a convolutional neural network.Background technique[0002]In recent years, convolutional deep neural network algorithms have been widely used in image recognition, video surveillance and other fields. The algorithm can be implemented on the CPU or GPU. In order to accelerate the computing speed of the convolutional deep neural network algorithm, the current There are also many dedicated accelerator implementations.[0003]Convolutional deep neural networks have two special operations: input edge filling and filter expansion. The input edge filling operation means that in order to control the size of the output matrix and prevent the loss of edge information, before the operation, the edges of the rows and columns of the input data matrix are expanded by one row and one column, or multiple rows and multipl...

AI Technical Summary

Problems solved by technology

[0005] The embodiment of the present invention provides an edge filling and filter expansion operation method and system in a convolutional neural network, which is used to solve the data that needs to be stored and transmitted after the zero-filling preprocessing of the input matrix and the filter matrix in the prior art The amount will become larger, which will reduce storage efficiency and increase access power consumption; and if the original matrix is ​​not changed, adding a judgment method to the program will introduce a lot of branch jumping processes in the program, wasting computing time and reducing the algorithm The problem of computing speed

Images