Neural network inference method based on software and hardware cooperative acceleration

Abstract

A neural network inference method based on software and hardware cooperative acceleration comprises the steps that firstly, a convolution operation module, a pooling operation module and a full connection operation module are built and configured, then, the configured modules are used for building a neural network inference system, and a neural network comprises convolution layers and pooling layers which are alternately connected and a full connection layer which is in pipeline connection; the trained parameters and input data of the neural network are stored into a neural network inference system; and finally, the neural network inference system starts to work. A first input signal is input into the first convolution layer in a first small work period of each large work period, a secondinput signal is input into the first convolution layer in a second small work period of each large work period, one layer of operation is complted by each small work period neural network, and an operation result is input into the next layer; and the first input signal and the second input signal are sequentially subjected to each layer of operation of the neural network to obtain inference results of the first input signal and the second input signal.

Description

Technical field [0001] The invention belongs to the field of computer technology, and specifically relates to a neural network inference method based on software and hardware cooperative acceleration, which can be applied to a general convolutional neural network. Background technique [0002] Convolutional neural network is the most important model in the field of artificial intelligence deep learning. It is widely used in image classification, recognition, and understanding scenarios, and has achieved high accuracy. In 1998, LeCun et al. proposed a convolutional neural network (LeNet) composed of a convolutional layer and a downsampling layer, forming the prototype of a modern convolutional neural network. In 2012, in the ImageNet large-scale recognition challenge, Krizhevsky and others used ReLU as the activation function to build an AlexNet convolutional neural network, and obtained an excellent result of 84.7% classification accuracy, which became an important turning point ...

AI Technical Summary

Problems solved by technology

[0007] At this stage, most of the designs for accelerating the inference process of convolutional neural networks focus on hardware acceleration on the FPGA side. Few acceleration designs propose software acceleration, let alone collaborative acceleration of the software and hardware system as a whole.

Images