Block convolution optimization method and device for convolutional neural network

Abstract

The present invention relates to the field of deep neural networks, and proposes a block convolution optimization method and device for convolutional neural networks, aiming to solve the processing bottleneck problem of convolution operations in neural networks in hardware processing systems. The optimization method includes: Select the convolutional layer to be divided into blocks and determine the upper limit of its block size; determine the number of blocks and block size of the input feature map according to the upper limit of the block size; based on the number of blocks, block size, convolution Kernel size, input feature map size, input feature map boundary padding size, calculate the block boundary padding size of the block feature map; based on the number of blocks, block size, and block boundary padding size, construct a volume based on block boundary padding product, and replace the original convolution. The invention greatly alleviates the resource limitation problem of the convolutional neural network running on the embedded hardware platform, and maximizes the burst length when reading and writing the memory, improves the throughput, reduces the delay, and improves the efficiency.

Description

technical field [0001] The invention relates to the technical field of deep neural networks, in particular to a block convolution optimization method and device for convolutional neural networks. Background technique [0002] Deep learning, as a cutting-edge branch of machine learning, has developed rapidly in theory and application in recent years. Driven by deep learning, traditional fields such as computer vision and speech and language processing are developing rapidly. Computers can even surpass humans in recognizing images, videos, and speech and text. A number of emerging industries and applications have emerged in the wave of deep learning development, such as self-driving cars, chat robots, smart monitoring, smart homes, etc. Intelligent applications can be seen almost everywhere in people's daily lives. Driven by big data and deep learning, traditional retail, banking, and insurance industries have entered a new era of Internet development. [0003] Deep convolut...

AI Technical Summary

Problems solved by technology

[0009] In order to solve the above problems in the prior art, that is, in order to solve the bottleneck problem of convolution processing in the neural network in the hardware processing system, one aspect of the present invention proposes a block convolution optimization method for convolutional neural networks , including the following steps:

Images