CN-116362350-B - Parallel segmentation learning method, device, equipment and storage medium based on clusters

CN116362350BCN 116362350 BCN116362350 BCN 116362350BCN-116362350-B

Abstract

The invention belongs to the technical field of computers, and discloses a parallel segmentation learning method, device and equipment based on clusters and a storage medium. The method comprises the steps of obtaining a plurality of user terminals to be learned and user communication information of the user terminals to be learned, dividing clusters of the user terminals to be learned according to the user communication information of the user terminals to be learned and the user communication information of the user terminals to be learned, determining a plurality of target user clusters, carrying out cluster serial division learning according to an aggregation user terminal model and the target user clusters to obtain a target user terminal model, and carrying out parallel division learning on the aggregation user terminal model in the target user clusters according to target spectrum resources of the user terminals to be learned. By the method, the overall training time delay in the segmentation learning process is effectively reduced, the segmentation learning efficiency is improved, the negative influence caused by network isomerism and dynamic property is restrained, and the convergence and accuracy of the existing segmentation learning technology are ensured.

Inventors

WU WEN
ZHANG SONGGE
LIU SHENGBO
LI SHAOFENG

Assignees

鹏城实验室

Dates

Publication Date: 20260508
Application Date: 20230403

Claims (9)

1. The cluster-based parallel segmentation learning method is characterized by comprising the following steps of: acquiring a plurality of user terminals to be learned and user communication information of each user terminal to be learned; according to the user terminals to be learned and the user communication information of the user terminals to be learned, carrying out cluster division on the user terminals to be learned, and determining a plurality of target user clusters; Performing cluster serial segmentation learning according to the aggregation user end model and each target user cluster to obtain a target user end model, wherein the aggregation user end model is obtained by performing parallel segmentation learning in each target user cluster according to target spectrum resources of each user end to be learned; The step of dividing the clusters of the user ends to be learned according to the user ends to be learned and the user communication information of the user ends to be learned, and determining a plurality of target user clusters comprises the following steps: The method comprises the steps of carrying out cluster division on each user end to be learned according to a random aggregation strategy to obtain a plurality of first user clusters, determining first training time delay according to the plurality of first user clusters and user communication information of each user end to be learned, carrying out random extraction in each first user cluster according to a random extraction strategy to determine an exchange user end and the first user cluster where the exchange user end is located; Performing exploration probability calculation according to the first training time delay and the second training time delay, and determining target exploration probability; comparing the target exploration probability with a preset cluster variable to determine a comparison result, and determining the number of user ends of each second user cluster when the comparison result is that the target exploration probability is larger than the preset cluster variable; when the number of the user ends of the second user clusters is different, taking the second user clusters as a plurality of target user clusters; the method comprises the steps of determining a delay user end in each second user cluster according to user communication information of each user end to be learned when the number of the user ends of each second user cluster is equal, carrying out random association according to the delay user end in each second user cluster and a plurality of second user clusters, determining a plurality of first association results and first association time delays of each first association result, comparing the first association time delays of each first association result, determining a first target result in the plurality of first association results according to comparison results, determining a plurality of third user clusters according to the first target result, and determining a plurality of target user clusters according to the plurality of third user clusters.
2. The cluster-based parallel partition learning method of claim 1, wherein the determining a plurality of target user clusters from a plurality of third user clusters comprises: determining a quick user side in each third user cluster according to the user communication information of each user side to be learned; According to the quick user terminal in each third user cluster and the random association of the plurality of third user clusters, a plurality of second association results and second association time delays of the second association results are determined; Comparing the second association delays of the second association results, and determining a second target result from the plurality of second association results according to the comparison result; and determining a plurality of target user clusters according to the second target result.
3. The cluster-based parallel partition learning method according to claim 1, wherein the performing cluster serial partition learning according to the aggregate client model and each target client cluster to obtain the target client model comprises: determining a first target cluster and a second target cluster according to a preset cluster training sequence; Determining an aggregation user side model according to the first target cluster; And transmitting the aggregate client model to the second target cluster so that the second target cluster feeds back the target client model.
4. The parallel partition learning method based on cluster of claim 1, the parallel segmentation learning method based on the clusters is characterized by further comprising the following steps: issuing a user terminal model to each user terminal to be learned in a target user cluster, so that each user terminal to be learned in each target user cluster performs data sampling in parallel to generate crushing data, and feeding back the crushing data and a sampling data label according to target spectrum resources of each user terminal to be learned in the target user cluster; determining a shredding data gradient according to the shredding data and the labels of the sampling data; Sending the crushed data gradient to each user end to be learned in each target user cluster, so that each user end to be learned in each target user cluster feeds back and updates a user end model according to the crushed data gradient; And determining an aggregation user end model of each target user cluster according to each updated user end model.
5. The cluster-based parallel partition learning method of claim 4, wherein determining an aggregate client model for each target user cluster based on each updated client model comprises: acquiring the number of user ends of each target user cluster and the number of samples of the user ends to be learned in each target user cluster; And carrying out weighted aggregation according to the number of the user ends of each target user cluster, the number of the samples of the user ends to be learned in each target user cluster and each user update model to obtain an aggregation user end model of each target user cluster.
6. The cluster-based parallel partition learning method according to any one of claims 1 to 5, wherein before the cluster serial partition learning is performed according to the aggregate client model and each target client cluster to obtain the target client model, the method further comprises: According to the resource allocation strategy, spectrum resource allocation is carried out for each user end to be learned of each target user cluster, and initial resources of each user end to be learned are determined; calculating the training time delay of the user end according to the user communication information of each user end to be learned and the initial resources of each user end to be learned, and obtaining the training time delay of the user end of each user end to be learned; sequencing the training time delays of all the user terminals, and determining a target learning user terminal according to the sequencing result; Performing spectrum resource allocation on the target learning user side, and determining allocation resources of the target learning user side; determining the quantity of residual resources according to the initial resources of each user side to be learned and the allocated resources of the target learning user side; And when the number of the residual resources is the preset number of resources, updating the initial resources of each user end to be learned according to the allocated resources of the target learning user end, and determining the target spectrum resources of each user end to be learned of each target user cluster.
7. A cluster-based parallel division learning device, characterized in that the cluster-based parallel division learning device comprises: the acquisition module is used for acquiring a plurality of user terminals to be learned and user communication information of each user terminal to be learned; the dividing module is used for dividing the clusters of the user ends to be learned according to the user ends to be learned and the user communication information of the user ends to be learned and determining a plurality of target user clusters; The learning module is used for carrying out cluster serial segmentation learning according to the aggregation user end model and each target user cluster to obtain a target user end model, and the aggregation user end model is obtained by carrying out parallel segmentation learning in each target user cluster according to the target spectrum resources of each user end to be learned; the dividing module is further used for carrying out cluster division on each user end to be learned according to a random aggregation strategy to obtain a plurality of first user clusters; Determining a first training time delay according to the plurality of first user clusters and the user communication information of each user end to be learned; randomly extracting in each first user cluster according to a random extraction strategy, and determining an exchange user side and the first user cluster where the exchange user side is located; according to the exchange user terminal, the first user cluster where the exchange user terminal is located and the plurality of first user clusters, carrying out random association to obtain a plurality of second user clusters; Determining a first training time delay and a second training time delay according to the plurality of second user clusters and the user communication information of each user end to be learned; Performing exploration probability calculation according to the first training time delay and the second training time delay, and determining target exploration probability; comparing the target exploration probability with a preset cluster variable to determine a comparison result; When the comparison result shows that the target exploration probability is larger than the preset cluster variable, determining the number of the user ends of each second user cluster; when the number of the user ends of the second user clusters is different, taking the second user clusters as a plurality of target user clusters; the dividing module is further configured to determine a hold-off user end in each second user cluster according to user communication information of each user end to be learned when the number of user ends of each second user cluster is the same; According to the delay user end in each second user cluster and the plurality of second user clusters, carrying out random association, and determining a plurality of first association results and first association time delays of each first association result; comparing the first association delays of the first association results, and determining a first target result from the plurality of first association results according to the comparison result; And determining a plurality of target user clusters according to the plurality of third user clusters.
8. A cluster-based parallel partition learning device comprising a memory, a processor, and a cluster-based parallel partition learning program stored on the memory and executable on the processor, the cluster-based parallel partition learning program configured to implement the cluster-based parallel partition learning method of any one of claims 1-6.
9. A storage medium having stored thereon a cluster-based parallel segmentation learning program which, when executed by a processor, implements the cluster-based parallel segmentation learning method according to any one of claims 1 to 6.

Description

Parallel segmentation learning method, device, equipment and storage medium based on clusters Technical Field The present invention relates to the field of computer technologies, and in particular, to a cluster-based parallel partition learning method, apparatus, device, and storage medium. Background Segmentation learning is a mainstream distributed learning scheme, and can train an AI (ARTIFICIAL INTELLIGENCE ) model between clients (such as vehicles, mobile terminals, etc.) with the aid of an internet of vehicles edge server, without sharing local data of the clients. When the number of users is large in the existing segmentation learning scheme, the edge server needs to train one user first and then sequentially move to the next user, the whole model training process is controlled by the network controller located at the edge server, and long model training delay exists. Because the users train the model sequentially, the training delay in the existing segmentation learning scheme is accumulated and is proportional to the number of users. When the number of users is large, the model training can lead to long training delay, and the overall training time is greatly increased, so that a segmentation learning method when multiple users exist is needed to be provided, the time delay of the training process is reduced, and the training time is shortened. Disclosure of Invention The invention mainly aims to provide a parallel segmentation learning method, device, equipment and storage medium based on clusters, and aims to solve the technical problem of how to reduce training time delay of segmentation learning and shorten training time when a multi-user end exists in the prior art. In order to achieve the above object, the present invention provides a cluster-based parallel division learning method, including: acquiring a plurality of user terminals to be learned and user communication information of each user terminal to be learned; according to the user terminals to be learned and the user communication information of the user terminals to be learned, carrying out cluster division on the user terminals to be learned, and determining a plurality of target user clusters; And carrying out cluster serial segmentation learning according to the aggregation user end model and each target user cluster to obtain a target user end model, wherein the aggregation user end model is obtained by carrying out parallel segmentation learning in each target user cluster according to the target spectrum resources of each user end to be learned. Optionally, the grouping the to-be-learned clients according to the to-be-learned clients and the user communication information of the to-be-learned clients, and determining a plurality of target user clusters includes: Carrying out cluster division on each user end to be learned according to a random aggregation strategy to obtain a plurality of first user clusters; Determining a first training time delay according to the plurality of first user clusters and the user communication information of each user end to be learned; randomly extracting in each first user cluster according to a random extraction strategy, and determining an exchange user side and the first user cluster where the exchange user side is located; according to the interactive user terminal, the first user cluster where the exchange user terminal is located and the plurality of first user clusters, carrying out random association to obtain a plurality of second user clusters; Determining a first training time delay and a second training time delay according to the plurality of second user clusters and the user communication information of each user end to be learned; And determining a plurality of target user clusters according to the first training time delay and the second training time delay. Optionally, the determining a plurality of target user clusters according to the first training delay and the second training delay includes: Performing exploration probability calculation according to the first training time delay and the second training time delay, and determining target exploration probability; comparing the target exploration probability with a preset cluster variable to determine a comparison result; When the comparison result shows that the target exploration probability is larger than the preset cluster variable, determining the number of the user ends of each second user cluster; And when the number of the user ends of the second user clusters is different, taking the second user clusters as a plurality of target user clusters. Optionally, after the obtaining the number of the user ends of each second user cluster when the comparison result is that the target exploration probability is greater than the preset cluster variable, the method further includes: When the number of the user ends of the second user clusters is the same, determining a delay user end in each second user cluster according to the