Neural network operation optimization and data processing method and device and storage medium

A technology of operation optimization and neural network, applied in the computer field, can solve the problems of popularization and application of multi-core computing equipment, low computing efficiency, etc., and achieve the effect of improving computing efficiency

Active Publication Date: 2019-08-30
SHENZHEN UNIV
View PDF4 Cites 4 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0003] The purpose of the present invention is to provide a neural network operation optimization and data processing method, equipment and storage medium, aiming to solve the existing problems in the prior art, such as low

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Neural network operation optimization and data processing method and device and storage medium
  • Neural network operation optimization and data processing method and device and storage medium
  • Neural network operation optimization and data processing method and device and storage medium

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0036] figure 1 The implementation flow of the neural network operation optimization method provided by the first embodiment of the present invention is shown. For the convenience of description, only the parts related to the embodiment of the present invention are shown, and the details are as follows:

[0037] The forward graph of the neural network corresponds to a set of at least two paths between the input and the output. Each path uses a feature map (Feature Map) as a node and a calculator as an edge, and the calculator corresponds to at least one network layer. .

[0038] In this embodiment, the neural network is similar to the Inception-Net, and the computational sub-connection structure between the input and output of the neural network is composed of multiple paths, which is a multi-branch structure . When the calculation sub-combination of this type of neural network is more complex, the network calculation accuracy is higher, and accordingly, parallel acceleratio...

Embodiment 2

[0051] On the basis of Embodiment 1, this embodiment further provides the following content:

[0052] Such as figure 2 As shown, in this embodiment, step S101 may mainly include:

[0053] In step S201, a topological sorting algorithm is used to convert the forward graph to obtain a topological sequence.

[0054] In this embodiment, the topological sorting algorithm can mainly perform topological sorting on the forward graph, and arrange all the nodes in it into a linear sequence satisfying the topological order, so that any pair of nodes (u, v) is in the linear sequence, and u is in v Before.

[0055] In step S202, the critical path is determined from the set according to the topology sequence.

[0056] In this embodiment, step S202 may include such as image 3 The flow shown:

[0057] In step S301, the activity duration of the path is determined according to the amount of floating-point calculation of the network layer.

[0058] In this embodiment, network layers such ...

Embodiment 3

[0064] This embodiment further provides the following content on the basis of embodiment one or two:

[0065] In this embodiment, the number of parallel processing threads is preset as N, and N is a natural number greater than 1. Then, step S102 is specifically:

[0066] When the real-time in-degree data of the node at the start position of the key edge is zero, determine the non-key edges that can be processed in parallel with the key edge and located on at most N-1 non-key paths; when the node at the start position of the key edge When the real-time in-degree data is not zero, non-critical edges that can be processed in parallel and located on at least two and at most N non-critical paths are determined. Wherein, the real-time in-degree data is obtained based on changes in node in-degree statistical data.

[0067] The parallelism of N threads can be regarded as a sliding window on the calculation sub-queue. Whenever a certain thread completes the current calculation sub-tas...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

The method is suitable for the technical field of computers, and provides a neural network operation optimization and data processing method and device and a storage medium. The method comprises the steps of during a forward calculation process of a neural network, obtaining the node in-degree statistical data and determining a key path from a path set; if the key edges of the key paths meet the parallel processing conditions, determining the non-key edges which can be processed with the key edges in parallel and are located on the non-key paths, and if the key edges do not meet the parallel processing conditions, determining at least two non-key edges which can be processed in parallel to form a calculation sub-thread distribution model which can be processed by the parallel processing threads. Therefore, the multi-core parallel accelerated optimization on the hierarchical structure of the neural network can be realized, the computing efficiency of the neural network is effectively improved, and the popularization and application of the large-scale neural retention on the computing equipment using the multi-core computing resources are facilitated.

Description

technical field [0001] The invention belongs to the technical field of computers, and in particular relates to a neural network operation optimization and data processing method, device and storage medium. Background technique [0002] After the deep learning neural network is trained, the neural network will be deployed to the actual project for application. This application process mainly uses the forward calculation results of the neural network. However, the accuracy of the network trained by neural networks with different structures is different. Generally speaking, the more complex the structure of the neural network, the higher the network accuracy. Therefore, the more complex neural network can be deployed and its forward calculation time can be reduced, the calculation efficiency will be effectively improved, and the more beneficial it is to practical applications. Most of the current mainstream deep learning neural network deployments are based on open source fram...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
IPC IPC(8): G06F9/50G06N3/04G06N3/08
CPCG06N3/08G06F9/5022G06N3/045
Inventor 解为成刘源张喜沈琳琳
Owner SHENZHEN UNIV
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap