WO2025007722A1 - 模型的管理方法、系统以及共享节点、网络节点 - Google Patents
模型的管理方法、系统以及共享节点、网络节点 Download PDFInfo
- Publication number
- WO2025007722A1 WO2025007722A1 PCT/CN2024/099249 CN2024099249W WO2025007722A1 WO 2025007722 A1 WO2025007722 A1 WO 2025007722A1 CN 2024099249 W CN2024099249 W CN 2024099249W WO 2025007722 A1 WO2025007722 A1 WO 2025007722A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- model
- node
- request
- requested
- shared
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
Definitions
- the present disclosure relates to the field of communication technology, and in particular to a management method and system for a model, as well as a shared node and a network node.
- Endogenous intelligence is an important research direction for 6G ( 6th Generation) networks, and both access networks and core networks are conducting research in this direction.
- the core network has added intelligent network elements such as NWDAF (Network Data Analytics Function) based on the 3GPP (3rd Generation Partnership Project) architecture, and 3GPP has also started research on the access network side through scenarios such as network energy saving.
- NWDAF Network Data Analytics Function
- the content field includes at least one of a model type identifier and a model function identifier.
- obtaining training data from a network node includes: sending a model training request to the network node; receiving a model training response sent by the network node, wherein the model training response includes: training data and model requirement information, and the model requirement information includes: the number of input parameters, the type of input parameters, the number of output parameters, and at least one of the type of output parameters.
- the method further includes: storing the requested model and the label of the requested model in correspondence.
- the method further includes: when the found model matches the requested model content information and attribute information, sending the found model as the requested model to the network node.
- the method further includes: receiving a model upload request sent by a network node, wherein the model upload request includes the model to be uploaded, attribute information and content information of the model to be uploaded; and storing the model to be uploaded, the attribute information and the content information correspondingly in the shared model library.
- the method also includes: receiving a model update request sent by a network node, wherein the model update request includes an updated model, attribute information and content information of the updated model; searching the shared model library for an original model corresponding to the stored updated model based on the attribute information and content information of the updated model; and storing the updated model in place of the original model.
- the method further includes: receiving a model download request sent by a network node, wherein the model download request includes: attribute information and content information of the model to be updated; searching for the model to be updated in a shared model library based on the attribute information and content information of the model to be updated; and sending a model download response to the network node, wherein the model download response includes the model to be updated.
- a model management method is provided, which is executed by a network node, wherein the network node is a node in an access network domain or a node in a core network domain, and the management method includes: in response to a model inference request sent by an inference-side node, determining whether a model corresponding to the model inference request is stored in the domain to which it belongs; in a case where the model corresponding to the model inference request is not stored, sending a model reuse request to a shared node, wherein the model reuse request includes: attribute information and content information of the requested model, the requested model is the model corresponding to the model inference request, and the attribute information is used to indicate the domain to which the network node belongs; in a case where the model found by the shared node matches the requested model content information, and the attribute information does not match, sending training data and model requirement information to the shared node; receiving the requested model sent by the shared node, wherein the requested model is obtained by adjusting the found model by
- sending training data and model requirement information to a shared node includes: receiving a model training request sent by a shared node; determining model requirement information of the requested model based on attribute information and content information of the requested model, wherein the model requirement information includes: the number of input parameters, the type of input parameters, the number of output parameters, and at least one of the type of output parameters; sending a model training response to the shared node, wherein the model training response includes: training data and model requirement information.
- the method also includes: obtaining training data of the model to be trained; training the model to be trained to obtain the model to be uploaded; storing the model to be uploaded in a model library within the domain to which it belongs; and sending a model upload request to the shared node, wherein the model upload request includes the model to be uploaded, attribute information and content information of the model to be uploaded.
- the method further includes: when the model found by the shared node matches the requested model content information and attribute information, receiving the found model sent by the shared node as the requested model.
- the method further includes: sending a model update request to a shared node, wherein the model update request includes the updated model, attribute information and content information of the updated model, so that the shared node updates the original model corresponding to the stored updated model according to the model update request.
- the method further includes: in response to a model update requirement, determining whether the model to be updated is stored in the domain; if the model to be updated is not stored in the domain, sending a model download request to a shared node, wherein the model download request includes: attribute information and content information of the model to be updated; receiving a model download response sent by the shared node, wherein the model download response includes the model to be updated; and updating the model to be updated according to current training parameters of the model to be updated to obtain an updated model.
- a shared node including: a receiving module, used to receive a model reuse request sent by a network node, wherein the network node is a node in an access network domain or a node in a core network domain, and the model reuse request includes: attribute information and content information of the requested model, and the attribute information is used to indicate the domain to which the network node belongs; a search module, used to search for the requested model in a shared model library according to the model reuse request; an acquisition module, used to obtain training data and model requirement information from the network node when the found model matches the requested model content information and the attribute information does not match; an adjustment module, used to adjust the found model according to the training data and model requirement information as the requested model; and a sending module, used to send the requested model to the network node.
- a network node comprising: a determination module, for determining whether a model corresponding to the model inference request is stored in the domain to which it belongs in response to a model inference request sent by an inference-side node; a sending module, for sending a model reuse request to a shared node when the model corresponding to the model inference request is not stored, wherein the model reuse request comprises: attribute information and content information of the requested model, the requested model is the model corresponding to the model inference request, and the attribute information is used to indicate the domain to which the network node belongs; when the model found by the shared node matches the requested model content information, and when the attribute information does not match, sending training data and model requirement information to the shared node; a receiving module, for receiving the requested model sent by the shared node, wherein the requested model is obtained by the shared node adjusting the found model according to the training data and the model requirement information.
- a model management device including: a processor; and a memory coupled to the processor, for storing instructions, which, when executed by the processor, causes the processor to execute a model management method as described in any of the aforementioned embodiments.
- a non-transitory computer-readable storage medium on which a computer program is stored, wherein when the program is executed by a processor, the model management method of any of the aforementioned embodiments is implemented.
- a model management system comprising: a shared node of any of the aforementioned embodiments and a network node of any of the aforementioned embodiments.
- FIG1 is a schematic flow chart showing a method for managing a model according to some embodiments of the present disclosure.
- FIG2 is a schematic flow chart showing a method for managing a model according to some other embodiments of the present disclosure.
- FIG. 3A shows a schematic diagram of a system architecture according to some embodiments of the present disclosure.
- FIG. 3B is a schematic flowchart of a method for managing a model according to yet other embodiments of the present disclosure.
- FIG. 4 is a schematic diagram showing the structure of a shared node according to some embodiments of the present disclosure.
- FIG5 is a schematic diagram showing the structure of a network node according to some embodiments of the present disclosure.
- FIG6 shows a schematic diagram of the structure of a model management device according to some embodiments of the present disclosure.
- FIG. 7 is a schematic diagram showing the structure of a model management device according to some other embodiments of the present disclosure.
- FIG8 is a schematic diagram showing the structure of a management system of a model according to some embodiments of the present disclosure.
- the inventors found that currently the endogenous intelligence of the access domain and the core network domain are implemented independently within their respective domains, that is, the training, reasoning, and storage of AI (Artificial Intelligence)/ML (Machine Learning) models are performed within their respective domains. There has not been much progress in the research on cross-domain endogenous intelligence in the wireless domain and the core network domain.
- a technical problem to be solved by the present disclosure is to propose a model management method to achieve model sharing between domains and improve resource utilization.
- the present disclosure provides a model management method, which is described below in conjunction with FIGS. 1 to 3B .
- Fig. 1 is a flow chart of some embodiments of the management method of the model disclosed in the present invention. As shown in Fig. 1, the method of this embodiment includes: steps S102 to S110.
- step S102 the network node sends a model reuse request to the shared node, and correspondingly, the shared node receives the model reuse request sent by the network node.
- the system includes shared nodes, which have storage capabilities, connection capabilities, and scheduling capabilities.
- a shared model library is constructed on the shared node side to store shared models, so as to realize cross-domain intercommunication and sharing of models in the access network domain and the core network domain.
- the network node can be a node in the access network domain, such as a base station, a terminal, etc.
- the network node can also be a node in the core network domain, such as NWDAF, etc., not limited to the examples given.
- the network nodes can establish their own in-domain model libraries, such as an access network model library or a core network model library.
- the model can be an AI or ML model implemented by a computer.
- the network node responds to the model inference request sent by the inference side node and determines whether the model corresponding to the model inference request is stored in the domain to which it belongs; if the model corresponding to the model inference request is not stored, Send a model reuse request to the shared node. If the model corresponding to the model inference request is stored, there is no need to initiate a model reuse request.
- the initiator of the model inference request can be a user in the domain of each network node.
- the network node first matches whether there is a trained model in the model library in the domain.
- the matching rule can also use the matching mode of ⁇ attribute: content ⁇ . If the match is successful, the model is directly pushed to the inference side node to complete the model inference based on the inference data.
- the network node starts to send model reuse requests to the shared node to request model collaboration.
- a model reuse request includes: attribute information and content information of the requested model, where the attribute information is used to indicate the domain to which the network node belongs.
- the attribute information and content information can be represented by tags, where the attribute field in the tag indicates the attribute information, and the content field indicates the content information.
- the content field includes: at least one of a model type identifier and a model function identifier.
- a reinforcement learning model used for energy-saving scenarios in an access network has a tag of ⁇ Access Network: Reinforcement Learning_Energy Saving ⁇ , and the specific representation method is not limited to the examples given.
- step S104 the shared node searches for the requested model in the shared model library according to the model reuse request.
- the shared node matches the content fields of the requested model with the content fields of each stored model in the shared model library; for the stored models that match the content fields of the requested model, the attribute fields of the stored model are matched with the attribute fields of the requested model.
- the shared node first matches the content field according to the label of ⁇ attribute: content ⁇ , and then matches the attribute word count after the content field matches successfully. For example, if the label is ⁇ access network: reinforcement learning_energy saving ⁇ , it first matches “reinforcement learning_energy saving”, and then matches "access network” after the content field matches successfully.
- step S106 when the found model matches the requested model content information but the attribute information does not match, the shared node obtains training data and model requirement information from the network node. Correspondingly, the network node sends the training data and model requirement information to the shared node.
- the shared node sends a model training request to the network node, and correspondingly, the network node receives the model training request sent by the shared node.
- the network node determines the model requirement information of the requested model based on the attribute information and content information of the requested model, and the network node sends a model training response to the shared node, and correspondingly, the shared node receives the model training response sent by the network node.
- the model requirement information includes: at least one of: the number of input parameters, the type of input parameters, the number of output parameters, and the type of output parameters.
- the model training response includes: training data and model requirement information. For models with the same function and type, different model requirement information for the core network and the access network can be pre-configured. For example, for the "reinforcement learning_energy saving" model, the number and type of input/output parameters required on the core network side and the number and type of input/output parameters required on the access network side are pre-configured.
- the shared node sends a request for training data and detailed model requirements, namely a model training request, and the network node uploads the training data and model requirement information (for example, detailed input and output parameters of the model, etc.). After receiving the corresponding training data and model requirement information, the shared node begins to modify and retrain the found model.
- the found model when the model found by the shared node matches the requested model content information and attribute information, the found model is sent to the network node as the requested model. Accordingly, the network node receives the found model sent by the shared node as the requested model.
- the shared node directly finds the model requested by the network node, it can directly send the model to the network node.
- step S108 the shared node adjusts the found model according to the training data and the model requirement information as the requested model.
- the shared node can perform secondary training on the found model, and the specific training method can be determined according to the actual model.
- step S110 the shared node sends the requested model to the network node, and correspondingly, the network node receives the requested model sent by the shared node.
- the shared node can send the relevant data of the requested model to the network node, so that the network node can build the model locally and use it.
- the network node tests or evaluates the received model, and if the test or evaluation passes, the model reasoning and other functions are completed. If the test or evaluation fails, the received model can be modified and retrained according to the training data until it passes the test or evaluation, and the model that passes the test or evaluation is updated in the model library in the domain to which it belongs.
- the shared node stores the requested model and the label of the requested model in correspondence.
- the label of the model found by the shared node is ⁇ core network: reinforcement learning_energy saving ⁇
- the requested model is obtained after modification or secondary training
- the label is ⁇ access network: reinforcement learning_energy saving ⁇ , then the requested model and the label are stored in correspondence.
- the network node may be a node in the access network domain or a node in the core network domain.
- Receive a model reuse request sent by a network node search for the requested model in the shared model library, and if the found model matches the requested model content information but the attribute information does not match, obtain training data from the network node, retrain the found model, obtain the requested model and send it to the network node.
- the shared node establishes a shared model library, and can perform secondary training of the model based on the differences between models in different domains, thereby realizing cross-domain model sharing between the access network and the core network, reducing the occupancy of access network and core network resources, improving resource utilization, and providing a basis for cross-domain intelligent endogenization.
- Fig. 2 is a flow chart of some other embodiments of the management method of the model disclosed in the present invention. As shown in Fig. 2, the method of this embodiment includes: steps S202 to S214.
- step S202 the network node obtains training data of the model to be trained, trains the model to be trained, and obtains the model to be uploaded.
- the access domain or core domain first builds a database through the collected data, and then provides the data in the database to the model training for model training.
- step S204 the network node stores the model to be uploaded into the model library in the domain to which it belongs.
- the data collection, training and reasoning of the model can be completed by multiple network nodes in the same domain.
- the database and the model library in the domain can be shared or distributed for multiple network nodes in the same domain, that is, distributed on multiple network nodes, without limitation here.
- step S206 the network node sends a model upload request to the sharing node, and correspondingly, the sharing node receives the model upload request sent by the network node.
- the model upload request includes the model to be uploaded, the attribute information and content information of the model to be uploaded.
- the network node can push the model to be uploaded to the shared node periodically or regularly.
- step S208 the sharing node stores the model to be uploaded, the attribute information and the content information of the model to be uploaded in the shared model library accordingly.
- the shared node can store the models to be uploaded and the corresponding labels in the shared model library.
- the access network model library will periodically push the data of the trained model to the shared model library of the shared node, and store the model according to the label ⁇ attribute: content ⁇ .
- the attributes include: access network and core network.
- the content can be adjusted based on the difference. For example, if a reinforcement learning model for energy-saving scenarios in the access network is uploaded, the stored label will be ⁇ access network: reinforcement learning_energy saving ⁇ .
- the core network node first stores the trained model in the core network model library, and then pushes it to the shared model library on the shared node side.
- the push mode is consistent with the access network side, using a scheduled push mode, and the stored label is also ⁇ attribute: content ⁇ . For example, if the core network side uploads If a supervised learning model for trajectory prediction is used, the storage label is ⁇ core network: supervised learning_trajectory prediction ⁇ .
- the network node can select models with usage frequencies below a threshold based on the usage frequencies of each model within a preset time and send them to a shared node for storage, and delete models with usage frequencies below the threshold from the model library within the domain to which they belong, thereby saving storage resources.
- step S210 the network node sends a model update request to the shared node, and correspondingly, the shared node receives the model update request sent by the network node.
- the model update request includes the updated model, attribute information and content information of the updated model.
- step S212 the sharing node searches the shared model library for the original model corresponding to the updated model that has been stored according to the attribute information and content information of the updated model.
- the shared node can search for the original model in the shared model library based on the label of the updated model.
- step S214 the shared node replaces the original model with the updated model for storage.
- the updated model is first stored in the model library in the domain to which it belongs, and then pushed to the shared model library of the shared node to replace the original model.
- the models of the access network and core network do not support the regression of the model version, but only support the update of the model version. If the old version of the model is needed, it needs to be retrained.
- the model to be updated may not be stored in the model library within the domain to which the network node belongs, but only in the shared model library. The network node needs to download the model to be updated from the shared node side before updating.
- the network node determines whether the model to be updated is stored in the domain to which it belongs; if the model to be updated is stored in the domain to which it belongs, the model to be updated is updated; if the model to be updated is not stored in the domain to which it belongs, the network node sends a model download request to the shared node, and correspondingly, the shared node receives the model download request sent by the network node.
- the model download request includes: attribute information and content information of the model to be updated; the shared node searches for the model to be updated in the shared model library according to the attribute information and content information of the model to be updated.
- the shared node sends a model download response to the network node, and correspondingly, the network node receives the model download response sent by the shared node, wherein the model download response includes the model to be updated.
- the network node updates the model to be updated according to the current training parameters of the model to be updated to obtain an updated model.
- the network node obtains the updated model, and then uploads it to the shared node for storage according to the method of the aforementioned embodiment.
- the method of the above embodiment realizes cross-domain model storage and updating, reduces the occupation of access network and core network resources, improves resource utilization, and provides a basis for cross-domain intelligent endogenization.
- the access network side can implement model reasoning, training, and storage, store the trained model in the access network model library, and push the trained model to the shared model library of the shared node. Inference request, get the corresponding model to perform inference. If there is no corresponding model in the access network model library, request the trained model from the shared node, receive the trained model sent by the shared node, and then evaluate the model. After the evaluation is passed, perform inference.
- the implementation scheme on the core network side is the same or similar, so it will not be repeated.
- the present disclosure also provides a shared node, which is described below in conjunction with FIG. 4 .
- the receiving module 410 is used to receive a model reuse request sent by a network node, wherein the network node is a node in an access network domain or a node in a core network domain, and the model reuse request includes: attribute information and content information of the requested model, and the attribute information is used to indicate the domain to which the network node belongs.
- the search module 420 is used to search for the requested model in the shared model library according to the model reuse request.
- attribute information and content information are represented by tags, the attribute field in the tag represents the attribute information, and the content field in the tag represents the content information.
- the search module 420 is used to match the content field of the requested model with the content field of each stored model in the shared model library; for the stored model that matches the content field of the requested model, the attribute field of the stored model is matched with the attribute field of the requested model.
- the content field includes at least one of a model type identifier and a model function identifier.
- the acquisition module 430 is used to send a model training request to a network node; receive a model training response sent by the network node, wherein the model training response includes: training data and model requirement information, and the model requirement information includes: the number of input parameters, the type of input parameters, the number of output parameters, and at least one of the type of output parameters.
- the adjustment module 440 is used to adjust the found model according to the training data and model requirement information as the requested model.
- the sharing node 40 further includes a storage module 460 for storing the requested model and the label of the requested model in correspondence.
- the sending module 450 is further configured to send the found model as the requested model to the network node when the found model matches the requested model content information and attribute information.
- the receiving module 410 is also used to receive a model update request sent by a network node, wherein the model update request includes an updated model, attribute information and content information of the updated model; the searching module 420 is also used to search for the original model corresponding to the stored updated model in the shared model library according to the attribute information and content information of the updated model; the storage module 460 is also used to replace the original model with the updated model for storage.
- the receiving module 410 is also used to receive a model download request sent by a network node, wherein the model download request includes: attribute information and content information of the model to be updated; the searching module 420 is also used to search for the model to be updated in the shared model library according to the attribute information and content information of the model to be updated; the sending module 450 is also used to send a model download response to the network node, wherein the model download response includes the model to be updated.
- the present disclosure also provides a network node, which is a node in an access network domain or a node in a core network domain, and is described below in conjunction with FIG. 5 .
- FIG5 is a structural diagram of some embodiments of the network node of the present disclosure.
- the network node 50 of this embodiment includes: a determination module 510 , a sending module 520 , and a receiving module 530 .
- the receiving module 530 is used to receive a requested model sent by a shared node, wherein the requested model is obtained by adjusting the found model by the shared node according to the training data and model requirement information.
- the model inference request includes: a label of the model corresponding to the model inference request, an attribute field in the label represents attribute information, and a content field in the label represents content information.
- the determination module 510 is used to search for the model corresponding to the model inference request in the model library within the domain according to the label of the model corresponding to the model inference request.
- the receiving module 530 is further used to receive a model training request sent by a shared node; the determining module 510 is further used to determine the model requirements of the requested model based on the attribute information and content information of the requested model.
- the receiving module 530 is further configured to receive the found model sent by the shared node as the requested model when the model found by the shared node matches the requested model content information and attribute information.
- the sending module 520 is also used to send a model update request to the shared node, wherein the model update request includes the updated model, attribute information and content information of the updated model, so that the shared node updates the original model corresponding to the stored updated model according to the model update request.
- the determination module 510 is also used to determine whether the model to be updated is stored in the domain in response to the model update demand; the sending module 520 is also used to send a model download request to the shared node when the model to be updated is not stored in the domain, wherein the model download request includes: attribute information and content information of the model to be updated; the receiving module 530 is also used to receive a model download response sent by the shared node, wherein the model download response includes the model to be updated; the training module 540 is also used to update the model to be updated according to the current training parameters of the model to be updated to obtain an updated model.
- the management device (shared node or network node) of the model in the embodiment of the present disclosure can be implemented by various computing devices or computer systems, which will be described below in conjunction with Figures 6 and 7.
- Fig. 6 is a structural diagram of some embodiments of the management device of the model of the present disclosure.
- the device 60 of this embodiment includes: a memory 610 and a processor 620 coupled to the memory 610, and the processor 620 is configured to execute the management method of the model in any of the embodiments of the present disclosure based on the instructions stored in the memory 610.
- the memory 610 may include, for example, a system memory, a fixed non-volatile storage medium, etc.
- the system memory may store, for example, an operating system, an application program, a boot loader, a database, and other programs.
- Fig. 7 is a structural diagram of some other embodiments of the management device of the disclosed model.
- the device 70 of this embodiment includes: a memory 710 and a processor 720, which are similar to the memory 610 and the processor 620 respectively. It may also include an input/output interface 730, a network interface 740, a storage interface 750, etc. These interfaces 730, 740, 750, the memory 710, and the processor 720 may be connected, for example, via a bus 760.
- the input/output interface 730 provides a connection interface for input/output devices such as a display, a mouse, a keyboard, and a touch screen.
- the network interface 740 provides a connection interface for various networked devices, such as a database server or a cloud storage server.
- the storage interface 750 provides a connection interface for external storage devices such as an SD card and a USB flash drive.
- the model management system 80 includes: a shared node 40 and a network node 50 .
- the network node may be an access domain node and/or a core domain node.
- the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present disclosure may take the form of a computer program product implemented on one or more computer-usable non-transient storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
- a computer-usable non-transient storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
- These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.
- These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.
Landscapes
- Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Information Transfer Between Computers (AREA)
Abstract
本公开涉及一种模型的管理方法、系统以及共享节点、网络节点,涉及通信技术领域。本公开的方法包括:接收网络节点发送的模型复用请求,其中,网络节点为接入网域内的节点或核心网域内的节点,模型复用请求包括:请求的模型的属性信息和内容信息,属性信息用于表示网络节点所属的域;根据模型复用请求,在共享模型库中查找请求的模型;在查找到的模型与请求的模型内容信息相匹配,属性信息不匹配的情况下,向网络节点获取训练数据和模型需求信息;根据训练数据和模型需求信息对查找到的模型进行调整,作为请求的模型;将请求的模型发送至网络节点。
Description
相关申请的交叉引用
本申请是以CN申请号为202310817613.3,申请日为2023年7月5日的申请为基础,并主张其优先权,该CN申请的公开内容在此作为整体引入本申请中。
本公开涉及通信技术领域,特别涉及一种模型的管理方法、系统以及共享节点、网络节点。
智能内生是6G(6th Generation,第六代移动通信技术)网络的重要研究方向,接入网和核心网均针对智能内生的方向开展的研究。核心网在基于3GPP(3rd Generation Partnership Project,第三代合作伙伴计划)的架构基础上加入了如NWDAF(Network Data Analytics Function,网络数据分析功能)等智能化网元,接入网侧的研究3GPP也通过网络节能等场景入手展开。
发明内容
根据本公开的一些实施例,提供的一种模型的管理方法,由共享节点执行,包括:接收网络节点发送的模型复用请求,其中,网络节点为接入网域内的节点或核心网域内的节点,模型复用请求包括:请求的模型的属性信息和内容信息,属性信息用于表示网络节点所属的域;根据模型复用请求,在共享模型库中查找请求的模型;在查找到的模型与请求的模型内容信息相匹配,属性信息不匹配的情况下,向网络节点获取训练数据和模型需求信息;根据训练数据和模型需求信息对查找到的模型进行调整,作为请求的模型;将请求的模型发送至网络节点。
在一些实施例中,属性信息和内容信息采用标签表示,标签中的属性字段表示属性信息,标签中的内容字段表示内容信息,根据模型复用请求,在共享模型库中查找请求的模型包括:将请求的模型的内容字段与共享模型库中的每个已存储的模型的内容字段进行匹配;针对与请求的模型的内容字段相匹配的已存储的模型,将该已存储的模型的属性字段与请求的模型的属性字段进行匹配。
在一些实施例中,内容字段包括:模型类型标识和模型功能标识中至少一项。
在一些实施例中,向网络节点获取训练数据包括:向网络节点发送模型训练请求;接收网络节点发送的模型训练响应,其中,模型训练响应包括:训练数据和模型需求信息,模型需求信息包括:输入参数的个数,输入参数的类型,输出参数的个数,输出参数的类型中至少一项。
在一些实施例中,该方法还包括:将请求的模型和请求的模型的标签进行对应存储。
在一些实施例中,该方法还包括:在查找到的模型与请求的模型内容信息和属性信息相匹配的情况下,将查找到的模型作为请求的模型,发送至网络节点。
在一些实施例中,该方法还包括:接收网络节点发送的模型上传请求,其中,模型上传请求包括待上传模型、待上传模型的属性信息和内容信息;将待上传模型、属性信息和内容信息对应存储至共享模型库的。
在一些实施例中,该方法还包括:接收网络节点发送的模型更新请求,其中,模型更新请求包括更新后的模型、更新后的模型的属性信息和内容信息;根据更新后的模型的属性信息和内容信息,在共享模型库中查找已存储的更新后的模型对应的原模型;将更新后的模型替换原模型进行存储。
在一些实施例中,该方法还包括:接收网络节点发送模型下载请求,其中,模型下载请求包括:待更新模型的属性信息和内容信息;根据待更新模型的属性信息和内容信息,在共享模型库中查找待更新模型;向网络节点发送的模型下载响应,其中,模型下载响应包括待更新模型。
根据本公开的另一些实施例,提供的一种模型的管理方法,由网络节点执行,其中,网络节点为接入网域内的节点或核心网域内的节点,管理方法包括:响应于推理侧节点发送的模型推理请求,确定所属域内是否存储模型推理请求对应的模型;在未存储模型推理请求对应的模型的情况下,向共享节点发送模型复用请求,其中,模型复用请求包括:请求的模型的属性信息和内容信息,请求的模型为模型推理请求对应的模型,属性信息用于表示网络节点所属的域;在共享节点查找到的模型与请求的模型内容信息相匹配,属性信息不匹配的情况下,向共享节点发送训练数据和模型需求信息;接收共享节点发送的请求的模型,其中,请求的模型是共享节点根据训练数据和模型需求信息对查找到的模型进行调整得到的。
在一些实施例中,模型推理请求包括:模型推理请求对应的模型的标签,标签中
的属性字段表示属性信息,标签中的内容字段表示内容信息,确定所属域内是否存储模型推理请求对应的模型包括:根据模型推理请求对应的模型的标签,在所属域内的模型库中查找模型推理请求对应的模型。
在一些实施例中,向共享节点发送训练数据和模型需求信息包括:接收共享节点发送的模型训练请求;根据请求的模型的属性信息和内容信息,确定请求的模型的模型需求信息,其中,模型需求信息包括:输入参数的个数,输入参数的类型,输出参数的个数,输出参数的类型中至少一项;向共享节点发送模型训练响应,其中,模型训练响应包括:训练数据和模型需求信息。
在一些实施例中,该方法还包括:获取待训练的模型的训练数据;对待训练的模型进行训练,得到待上传模型;将待上传模型存储至所属域内的模型库;向共享节点发送模型上传请求,其中,模型上传请求包括待上传模型、待上传模型的属性信息和内容信息。
在一些实施例中,该方法还包括:在共享节点查找到的模型与请求的模型内容信息和属性信息相匹配的情况下,接收共享节点发送的查找到的模型作为请求的模型。
在一些实施例中,该方法还包括:向共享节点发送模型更新请求,其中,模型更新请求包括更新后的模型、更新后的模型的属性信息和内容信息,以便共享节点根据模型更新请求对已存储的更新后的模型对应的原模型进行更新。
在一些实施例中,该方法还包括:响应于模型更新需求,确定所属域内是否存储待更新模型;在所属域内未存储待更新模型的情况下,向共享节点发送模型下载请求,其中,模型下载请求包括:待更新模型的属性信息和内容信息;接收共享节点发送的模型下载响应,其中,模型下载响应包括待更新模型;根据待更新模型的当前训练参数,对待更新模型进行更新,得到更新后的模型。
根据本公开的又一些实施例,提供的一种共享节点,包括:接收模块,用于接收网络节点发送的模型复用请求,其中,网络节点为接入网域内的节点或核心网域内的节点,模型复用请求包括:请求的模型的属性信息和内容信息,属性信息用于表示网络节点所属的域;查找模块,用于根据模型复用请求,在共享模型库中查找请求的模型;获取模块,用于在查找到的模型与请求的模型内容信息相匹配,属性信息不匹配的情况下,向网络节点获取训练数据和模型需求信息;调整模块,用于根据训练数据和模型需求信息对查找到的模型进行调整,作为请求的模型;发送模块,用于将请求的模型发送至网络节点。
根据本公开的再一些实施例,提供的一种网络节点,包括:确定模块,用于响应于推理侧节点发送的模型推理请求,确定所属域内是否存储模型推理请求对应的模型;发送模块,用于在未存储模型推理请求对应的模型的情况下,向共享节点发送模型复用请求,其中,模型复用请求包括:请求的模型的属性信息和内容信息,请求的模型为模型推理请求对应的模型,属性信息用于表示网络节点所属的域;在共享节点查找到的模型与请求的模型内容信息相匹配,属性信息不匹配的情况下,向共享节点发送训练数据和模型需求信息;接收模块,用于接收共享节点发送的请求的模型,其中,请求的模型是共享节点根据训练数据和模型需求信息对查找到的模型进行调整得到的。
根据本公开的又一些实施例,提供的一种模型的管理装置,包括:处理器;以及耦接至处理器的存储器,用于存储指令,指令被处理器执行时,使处理器执行如前述任意实施例的模型的管理方法。
根据本公开的再一些实施例,提供的一种非瞬时性计算机可读存储介质,其上存储有计算机程序,其中,该程序被处理器执行时实现前述任意实施例的模型的管理方法。
根据本公开的又一些实施例,提供的一种模型的管理系统,包括:前述任意实施例的共享节点和前述任意实施例的网络节点。
通过以下参照附图对本公开的示例性实施例的详细描述,本公开的其它特征及其优点将会变得清楚。
为了更清楚地说明本公开实施例或现有技术中的技术方案,下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本公开的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1示出本公开的一些实施例的模型的管理方法的流程示意图。
图2示出本公开的另一些实施例的模型的管理方法的流程示意图。
图3A示出本公开的一些实施例的系统架构的示意图。
图3B示出本公开的又一些实施例的模型的管理方法的流程示意图。
图4示出本公开的一些实施例的共享节点的结构示意图。
图5示出本公开的一些实施例的网络节点的结构示意图。
图6示出本公开的一些实施例的模型的管理装置的结构示意图。
图7示出本公开的另一些实施例的模型的管理装置的结构示意图。
图8示出本公开的一些实施例的模型的管理系统的结构示意图。
下面将结合本公开实施例中的附图,对本公开实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本公开一部分实施例,而不是全部的实施例。以下对至少一个示例性实施例的描述实际上仅仅是说明性的,决不作为对本公开及其应用或使用的任何限制。基于本公开中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本公开保护的范围。
发明人发现:目前接入域和核心网域的内生智能均是在各自域内独立实现,即在各自域内进行AI(Artificial Intelligence,人工智能)/ML(Machine Learning,机器学习)模型的训练、推理、存储等,针对无线域和核心网域的跨域智能内生研究还未有较大的进展。
本公开所要解决的一个技术问题是:提出一种模型的管理方法,实现域间模型共享,提高资源利用率。
本公开提供一种模型的管理方法,下面结合图1~3B进行描述。
图1为本公开模型的管理方法一些实施例的流程图。如图1所示,该实施例的方法包括:步骤S102~S110。
在步骤S102中,网络节点向共享节点发送模型复用请求,相应的,共享节点接收网络节点发送的模型复用请求。
如图3A所示,系统中包括共享节点,共享节点具备存储能力、连接能力和调度能力。共享节点侧构建共享模型库,存储共享的模型,实现接入网域和核心网域的模型跨域互通、共享。网络节点可以是接入网域内的节点,例如,基站、终端等。网络节点也可以是核心网域内的节点,例如,NWDAF等,不限于所举示例。网络节点可以建立各自的域内模型库,例如,接入网模型库或者核心网模型库。
模型可以是通过计算机实现的AI或ML等模型。
在一些实施例中,网络节点响应于推理侧节点发送的模型推理请求,确定所属域内是否存储模型推理请求对应的模型;在未存储模型推理请求对应的模型的情况下,
向共享节点发送模型复用请求。在存储模型推理请求对应的模型的情况下,则不需要发起模型复用请求。
模型推理请求的发起者可以是网络节点各自域内的用户,网络节点首先在所属域内模型库中匹配是否有已经训练完成的模型,匹配的规则也可以采用{属性:内容}的匹配模式,若匹配成功,则直接推送模型至推理侧节点,根据推理的数据完成模型推理。当域内无模型推理请求对应的模型后,网络节点开始向共享节点的发送模型复用请求,请求模型协同。
例如,模型复用请求包括:请求的模型的属性信息和内容信息,属性信息用于表示网络节点所属的域。属性信息和内容信息可以采用标签表示,标签中的属性字段表示属性信息,内容字段表示内容信息。内容字段包括:模型类型标识和模型功能标识中至少一项。例如,接入网中用于节能场景的强化学习模型,标签为{接入网:强化学习_节能},具体表示方式不限于所举示例。
在步骤S104中,共享节点根据模型复用请求,在共享模型库中查找请求的模型。
在一些实施例中,共享节点将请求的模型的内容字段与共享模型库中的每个已存储的模型的内容字段进行匹配;针对与请求的模型的内容字段相匹配的已存储的模型,将该已存储的模型的属性字段与请求的模型的属性字段进行匹配。
共享节点在共享模型库中根据{属性:内容}的标签,首先匹配内容字段,内容字段匹配成功后再匹配属性字数。例如,标签为{接入网:强化学习_节能},首先匹配“强化学习_节能”,内容字段匹配成功后再匹配“接入网”。
在步骤S106中,共享节点在查找到的模型与请求的模型内容信息相匹配,属性信息不匹配的情况下,向网络节点获取训练数据和模型需求信息,相应的,网络节点向共享节点发送训练数据和模型需求信息。
如果属性字段没有匹配到“接入网”,但可以匹配到{核心网:强化学习_节能},则选择{核心网:强化学习_节能}的模型实现跨域模型协同。但是由于核心网和接入网的设计模型的差异性、训练数据和需求的不同,需要对查找到的模型进行调整或重新训练。
在一些实施例中,共享节点向网络节点发送模型训练请求,相应的,网络节点接收共享节点发送的模型训练请求。网络节点根据请求的模型的属性信息和内容信息,确定请求的模型的模型需求信息,网络节点向共享节点发送模型训练响应,相应的,共享节点接收网络节点发送的模型训练响应。
例如,模型需求信息包括:输入参数的个数,输入参数的类型,输出参数的个数,输出参数的类型中至少一项。模型训练响应包括:训练数据和模型需求信息。针对功能和类型相同的模型,可以预先配置核心网和接入网的不同模型需求信息,例如,针对“强化学习_节能”模型,预先配置核心网侧所需的输入/输出参数的个数和类型,接入网侧所需的输入/输出参数的个数和类型。
共享节点下发训练数据和模型详细需求的请求即模型训练请求,网络节点上传训练数据和模型需求信息(例如,模型详细输入输出参数等),共享节点在接收到对应的训练数据和模型需求信息后,开始对查找到的模型进行修改和二次训练。
在一些实施例中,在共享节点查找到的模型与请求的模型内容信息和属性信息相匹配的情况下,将查找到的模型作为请求的模型,发送至网络节点。相应的,网络节点接收共享节点发送的查找到的模型作为请求的模型。
共享节点如果直接查找到网络节点请求的模型,则直接将该模型发送至网络节点即可。
在步骤S108中,共享节点根据训练数据和模型需求信息对查找到的模型进行调整,作为请求的模型。
共享节点可以对查找到的模型进行二次训练,具体训练方法可以根据实际模型确定。
在步骤S110中,共享节点将请求的模型发送至网络节点,相应的,网络节点接收共享节点发送的请求的模型。
共享节点可以将请求的模型的相关数据发送至网络节点,以使网络节点可以在本地构建该模型并进行使用。
在一些实施例中,网络节点对接收到的模型进行测试或评估,在测试或评估通过的情况下,完成模型推理等功能。在测试或评估未通过的情况下,可以根据训练数据对接接收到的模型修改和再次训练,直至通过测试或评估,并将测试或评估通过的模型在所属域内的模型库中更新。
在一些实施例中,共享节点将请求的模型和请求的模型的标签进行对应存储。例如,共享节点查找到的模型的标签为{核心网:强化学习_节能},经过修改或二次训练得到请求的模型,标签为{接入网:强化学习_节能},则将请求的模型和该标签对应存储。
上述实施例中网络节点可以是接入网域内的节点或核心网域内的节点,共享节点
接收网络节点发送的模型复用请求,在共享模型库中查找请求的模型,如果查找到的模型与请求的模型内容信息相匹配,属性信息不匹配,则从网络节点获取训练数据,对查找到的模型重新训练,得到请求的模型发送至网络节点。上述实施例中共享节点建立共享模型库,并可以针对不同域模型的差异进行模型的二次训练,实现接入网和核心网之间跨域的模型共享,降低了接入网和核心网资源的占用量,提高资源利用率,为跨域智能内生提供了基础。
下面结合图2描述本公开模型的管理方法的另一些实施例。
图2为本公开模型的管理方法另一些实施例的流程图。如图2所示,该实施例的方法包括:步骤S202~S214。
在步骤S202中,网络节点获取待训练的模型的训练数据,对待训练的模型进行训练,得到待上传模型。
接入网域或核心网域首先通过采集到的数据,构建数据库,然后将数据库中的数据提供给模型训练使用,进行模型的训练。
在步骤S204中,网络节点将待上传模型存储至所属域内的模型库。
模型的数据采集、训练和推理等可以是同一域内多个网络节点共同完成的。数据库、域内的模型库对于同一域内的多个网络节点可以是共享的,也可以是分布式的,即分布在多个网络节点侧,在此不作限制。
在步骤S206中,网络节点向共享节点发送模型上传请求,相应的,共享节点接收网络节点发送的模型上传请求。
模型上传请求包括待上传模型、待上传模型的属性信息和内容信息。网络节点可以周期性或定时向共享节点推送待上传的模型。
在步骤S208中,共享节点将待上传模型、待上传模型的属性信息和内容信息对应存储至共享模型库的。
共享节点可以将待上传模型和对应的标签存储至共享模型库。
例如,接入网模型库因为存储空间的限制,会定时将已经训练完成的模型的数据推送至共享节点的共享模型库,会根据{属性:内容}的标签存储模型,属性包括:接入网和核心网,内容可以根据差异调整,如上传了接入网中用于节能场景的强化学习模型,则存储的标签为{接入网:强化学习_节能}。核心网节点首先将已经训练的模型存放于核心网模型库中,然后推送至共享节点侧的共享模型库,推送的模式跟接入网侧保持一致,采用定时推送的模式,存储的标签也为{属性:内容},如核心网侧上传
了用于轨迹预测场景的监督学习模型,则存储标签为{核心网:监督学习_轨迹预测}。
在一些实施例中,网络节点可以根据各个模型在预设时间内的使用频率,选取使用频率低于阈值的模型发送到共享节点进行存储,在所属域内的模型库中删除使用频率低于阈值的模型,节省存储资源。
在步骤S210中,网络节点向共享节点发送模型更新请求,相应的,共享节点接收网络节点发送的模型更新请求。
模型更新请求包括更新后的模型、更新后的模型的属性信息和内容信息。
在步骤S212中,共享节点根据更新后的模型的属性信息和内容信息,在共享模型库中查找已存储的更新后的模型对应的原模型。
共享节点可以根据更新后的模型的标签在共享模型库中查找原模型。
在步骤S214中,共享节点将更新后的模型替换原模型进行存储。
如果核心网或接入网有训练后更新的模型,则更新后的模型首先存储在所属域内的模型库中,然后推送至共享节点的共享模型库中替换原模型。接入网和核心网的模型可以不支持模型版本的倒退,只支持模型版本的更新,如果需要用使用旧版本的模型,则需要重新进行训练。
在一些实施例中,待更新模型可能没有存储在网络节点所属域内的模型库中,仅存储在共享模型库中,网络节点则需要先从共享节点侧下载待更新模型再进行更新。
网络节点响应于模型更新需求,确定所属域内是否存储待更新模型;在所属域内存储待更新模型的情况下,对待更新的模型进行更新;在所属域内未存储待更新模型的情况下,网络节点向共享节点发送模型下载请求,相应的,共享节点接收网络节点发送模型下载请求。其中,模型下载请求包括:待更新模型的属性信息和内容信息;共享节点根据待更新模型的属性信息和内容信息,在共享模型库中查找待更新模型。共享节点向网络节点发送的模型下载响应,相应的,网络节点接收共享节点发送的模型下载响应,其中,模型下载响应包括待更新模型。网络节点根据待更新模型的当前训练参数,对待更新模型进行更新,得到更新后的模型。网络节点得到更新后的模型,再根据前述实施例的方法上传至共享节点进行存储。
上述实施例的方法实现了跨域的模型存储和更新,降低了接入网和核心网资源的占用量,提高资源利用率,为跨域智能内生提供了基础。
如图3B所示,接入网侧可以实现模型的推理、训练和存储,将已训练模型存储至接入网模型库中,同时将已训练模型推送至共享节点的共享模型库中。响应于模型
推理请求,获取相应的模型执行推理。如果接入网模型库中没有对应的模型,则向共享节点请求已训练模型,接收共享节点下发已训练模型,再对模型进行评估,评估通过后执行推理。核心网侧的实现方案相同或相似,不再赘述。
本公开还提供一种共享节点,下面结合图4进行描述。
图4为本公开共享节点的一些实施例的结构图。如图4所示,该实施例的共享节点40包括:接收模块410,查找模块420,获取模块430,调整模块440,发送模块450。
接收模块410用于接收网络节点发送的模型复用请求,其中,网络节点为接入网域内的节点或核心网域内的节点,模型复用请求包括:请求的模型的属性信息和内容信息,属性信息用于表示网络节点所属的域。
查找模块420用于根据模型复用请求,在共享模型库中查找请求的模型。
在一些实施例中,属性信息和内容信息采用标签表示,标签中的属性字段表示属性信息,标签中的内容字段表示内容信息,查找模块420用于将请求的模型的内容字段与共享模型库中的每个已存储的模型的内容字段进行匹配;针对与请求的模型的内容字段相匹配的已存储的模型,将该已存储的模型的属性字段与请求的模型的属性字段进行匹配。
在一些实施例中,内容字段包括:模型类型标识和模型功能标识中至少一项。
获取模块430用于在查找到的模型与请求的模型内容信息相匹配,属性信息不匹配的情况下,向网络节点获取训练数据和模型需求信息。
在一些实施例中,获取模块430用于向网络节点发送模型训练请求;接收网络节点发送的模型训练响应,其中,模型训练响应包括:训练数据和模型需求信息,模型需求信息包括:输入参数的个数,输入参数的类型,输出参数的个数,输出参数的类型中至少一项。
调整模块440用于根据训练数据和模型需求信息对查找到的模型进行调整,作为请求的模型。
发送模块450用于将请求的模型发送至网络节点。
在一些实施例中,共享节点40还包括存储模块460,用于将请求的模型和请求的模型的标签进行对应存储。
在一些实施例中,发送模块450还用于在查找到的模型与请求的模型内容信息和属性信息相匹配的情况下,将查找到的模型作为请求的模型,发送至网络节点。
在一些实施例中,接收模块410还用于接收网络节点发送的模型上传请求,其中,模型上传请求包括待上传模型、待上传模型的属性信息和内容信息;存储模块460还用于将待上传模型、待上传模型的属性信息和内容信息对应存储至共享模型库的。
在一些实施例中,接收模块410还用于接收网络节点发送的模型更新请求,其中,模型更新请求包括更新后的模型、更新后的模型的属性信息和内容信息;查找模块420还用于根据更新后的模型的属性信息和内容信息,在共享模型库中查找已存储的更新后的模型对应的原模型;存储模块460还用于将更新后的模型替换原模型进行存储。
在一些实施例中,接收模块410还用于接收网络节点发送模型下载请求,其中,模型下载请求包括:待更新模型的属性信息和内容信息;查找模块420还用于根据待更新模型的属性信息和内容信息,在共享模型库中查找待更新模型;发送模块450还用于向网络节点发送的模型下载响应,其中,模型下载响应包括待更新模型。
本公开还提供一种网络节点,网络节点为接入网域内的节点或核心网域内的节点,下面结合图5进行描述。
图5为本公开网络节点的一些实施例的结构图。如图5所示,该实施例的网络节点50包括:确定模块510,发送模块520,接收模块530。
确定模块510用于响应于推理侧节点发送的模型推理请求,确定所属域内是否存储模型推理请求对应的模型。
发送模块520用于在未存储模型推理请求对应的模型的情况下,向共享节点发送模型复用请求,其中,模型复用请求包括:请求的模型的属性信息和内容信息,请求的模型为模型推理请求对应的模型,属性信息用于表示网络节点所属的域;在共享节点查找到的模型与请求的模型内容信息相匹配,属性信息不匹配的情况下,向共享节点发送训练数据和模型需求信息。
接收模块530用于接收共享节点发送的请求的模型,其中,请求的模型是共享节点根据训练数据和模型需求信息对查找到的模型进行调整得到的。
在一些实施例中,模型推理请求包括:模型推理请求对应的模型的标签,标签中的属性字段表示属性信息,标签中的内容字段表示内容信息,确定模块510用于根据模型推理请求对应的模型的标签,在所属域内的模型库中查找模型推理请求对应的模型。
在一些实施例中,接收模块530还用于接收共享节点发送的模型训练请求;确定模块510还用于根据请求的模型的属性信息和内容信息,确定请求的模型的模型需求
信息,其中,模型需求信息包括:输入参数的个数,输入参数的类型,输出参数的个数,输出参数的类型中至少一项;发送模块520还用于向共享节点发送模型训练响应,其中,模型训练响应包括:训练数据和模型需求信息。
在一些实施例中,网络节点50还包括训练模块540,用于获取待训练的模型的训练数据;对待训练的模型进行训练,得到待上传模型;存储模块550,用于将待上传模型存储至所属域内的模型库;发送模块520还用于向共享节点发送模型上传请求,其中,模型上传请求包括待上传模型、待上传模型的属性信息和内容信息。
在一些实施例中,接收模块530还用于在共享节点查找到的模型与请求的模型内容信息和属性信息相匹配的情况下,接收共享节点发送的查找到的模型作为请求的模型。
在一些实施例中,发送模块520还用于向共享节点发送模型更新请求,其中,模型更新请求包括更新后的模型、更新后的模型的属性信息和内容信息,以便共享节点根据模型更新请求对已存储的更新后的模型对应的原模型进行更新。
在一些实施例中,确定模块510还用于响应于模型更新需求,确定所属域内是否存储待更新模型;发送模块520还用于在所属域内未存储待更新模型的情况下,向共享节点发送模型下载请求,其中,模型下载请求包括:待更新模型的属性信息和内容信息;接收模块530还用于接收共享节点发送的模型下载响应,其中,模型下载响应包括待更新模型;训练模块540还用于根据待更新模型的当前训练参数,对待更新模型进行更新,得到更新后的模型。
本公开的实施例中的模型的管理装置(共享节点或网络节点)可各由各种计算设备或计算机系统来实现,下面结合图6以及图7进行描述。
图6为本公开模型的管理装置的一些实施例的结构图。如图6所示,该实施例的装置60包括:存储器610以及耦接至该存储器610的处理器620,处理器620被配置为基于存储在存储器610中的指令,执行本公开中任意一些实施例中的模型的管理方法。
其中,存储器610例如可以包括系统存储器、固定非易失性存储介质等。系统存储器例如存储有操作系统、应用程序、引导装载程序(Boot Loader)、数据库以及其他程序等。
图7为本公开模型的管理装置的另一些实施例的结构图。如图7所示,该实施例的装置70包括:存储器710以及处理器720,分别与存储器610以及处理器620类似。
还可以包括输入输出接口730、网络接口740、存储接口750等。这些接口730,740,750以及存储器710和处理器720之间例如可以通过总线760连接。其中,输入输出接口730为显示器、鼠标、键盘、触摸屏等输入输出设备提供连接接口。网络接口740为各种联网设备提供连接接口,例如可以连接到数据库服务器或者云端存储服务器等。存储接口750为SD卡、U盘等外置存储设备提供连接接口。
本公开还提供一种模型的管理系统,如图8所示,模型的管理系统80包括:共享节点40和网络节点50,网络节点可以是接入网域节点和/或核心网域节点。
本领域内的技术人员应当明白,本公开的实施例可提供为方法、系统、或计算机程序产品。因此,本公开可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且,本公开可采用在一个或多个其中包含有计算机可用程序代码的计算机可用非瞬时性存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本公开是参照根据本公开实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解为可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器,使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中,使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品,该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上,使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理,从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
以上所述仅为本公开的较佳实施例,并不用以限制本公开,凡在本公开的精神和原则之内,所作的任何修改、等同替换、改进等,均应包含在本公开的保护范围之内。
Claims (22)
- 一种模型的管理方法,由共享节点执行,包括:接收网络节点发送的模型复用请求,其中,所述网络节点为接入网域内的节点或核心网域内的节点,所述模型复用请求包括请求的模型的属性信息和内容信息,所述属性信息用于表示所述网络节点所属的域;根据所述模型复用请求,在共享模型库中查找所述请求的模型;在查找到的模型与所述请求的模型内容信息相匹配,属性信息不匹配的情况下,向所述网络节点获取训练数据和模型需求信息;根据所述训练数据和模型需求信息对所述查找到的模型进行调整,作为所述请求的模型;将所述请求的模型发送至所述网络节点。
- 根据权利要求1所述的管理方法,其中,所述属性信息和所述内容信息采用标签表示,所述标签中的属性字段表示所述属性信息,所述标签中的内容字段表示所述内容信息,所述根据所述模型复用请求,在共享模型库中查找所述请求的模型包括:将所述请求的模型的内容字段与所述共享模型库中的每个已存储的模型的内容字段进行匹配;针对与所述请求的模型的内容字段相匹配的已存储的模型,将该已存储的模型的属性字段与所述请求的模型的属性字段进行匹配。
- 根据权利要求2所述的管理方法,其中,所述内容字段包括:模型类型标识和模型功能标识中至少一项。
- 根据权利要求1-3任一项所述的管理方法,其中,所述向所述网络节点获取训练数据包括:向所述网络节点发送模型训练请求;接收所述网络节点发送的模型训练响应,其中,所述模型训练响应包括所述训练数据和所述模型需求信息,所述模型需求信息包括:输入参数的个数,输入参数的类型,输出参数的个数,输出参数的类型中至少一项。
- 根据权利要求1-4任一项所述的管理方法,还包括:将所述请求的模型和所述请求的模型的标签进行对应存储。
- 根据权利要求1-5任一项所述的管理方法,还包括:在查找到的模型与所述请求的模型内容信息和属性信息相匹配的情况下,将查找到的模型作为所述请求的模型,发送至所述网络节点。
- 根据权利要求1-6任一项所述的管理方法,还包括:接收所述网络节点发送的模型上传请求,其中,所述模型上传请求包括待上传模型、所述待上传模型的属性信息和内容信息;将所述待上传模型、所述待上传模型的属性信息和内容信息对应存储至所述共享模型库的。
- 根据权利要求1-7任一项所述的管理方法,还包括:接收所述网络节点发送的模型更新请求,其中,所述模型更新请求包括更新后的模型、所述更新后的模型的属性信息和内容信息;根据所述更新后的模型的属性信息和内容信息,在所述共享模型库中查找已存储的所述更新后的模型对应的原模型;将所述更新后的模型替换所述原模型进行存储。
- 根据权利要求1-8任一项所述的管理方法,还包括:接收所述网络节点发送模型下载请求,其中,所述模型下载请求包括待更新模型的属性信息和内容信息;根据所述待更新模型的属性信息和内容信息,在所述共享模型库中查找所述待更新模型;向所述网络节点发送的模型下载响应,其中,所述模型下载响应包括所述待更新模型。
- 一种模型的管理方法,由网络节点执行,其中,所述网络节点为接入网域内的节点或核心网域内的节点,所述管理方法包括:响应于推理侧节点发送的模型推理请求,确定所属域内是否存储所述模型推理请求对应的模型;在未存储所述模型推理请求对应的模型的情况下,向共享节点发送模型复用请求,其中,所述模型复用请求包括请求的模型的属性信息和内容信息,所述请求的模型为所述模型推理请求对应的模型,所述属性信息用于表示所述网络节点所属的域;在所述共享节点查找到的模型与所述请求的模型内容信息相匹配,属性信息不匹配的情况下,向所述共享节点发送训练数据和模型需求信息;接收所述共享节点发送的所述请求的模型,其中,所述请求的模型是所述共享节 点根据所述训练数据和所述模型需求信息对所述查找到的模型进行调整得到的。
- 根据权利要求10所述的管理方法,其中,所述模型推理请求包括所述模型推理请求对应的模型的标签,所述标签中的属性字段表示属性信息,所述标签中的内容字段表示内容信息,所述确定所属域内是否存储所述模型推理请求对应的模型包括:根据所述模型推理请求对应的模型的标签,在所属域内的模型库中查找所述模型推理请求对应的模型。
- 根据权利要求10或11所述的管理方法,其中,所述向所述共享节点发送训练数据和模型需求信息包括:接收所述共享节点发送的模型训练请求;根据所述请求的模型的属性信息和内容信息,确定所述请求的模型的所述模型需求信息,其中,所述模型需求信息包括输入参数的个数,输入参数的类型,输出参数的个数,输出参数的类型中至少一项;向所述共享节点发送模型训练响应,其中,所述模型训练响应包括所述训练数据和所述模型需求信息。
- 根据权利要求10-12任一项所述的管理方法,还包括:获取待训练的模型的训练数据;对所述待训练的模型进行训练,得到待上传模型;将所述待上传模型存储至所属域内的模型库;向所述共享节点发送模型上传请求,其中,所述模型上传请求包括所述待上传模型、所述待上传模型的属性信息和内容信息。
- 根据权利要求10-13任一项所述的管理方法,还包括:在所述共享节点查找到的模型与所述请求的模型内容信息和属性信息相匹配的情况下,接收所述共享节点发送的查找到的模型作为所述请求的模型。
- 根据权利要求10-14任一项所述的管理方法,还包括:向所述共享节点发送模型更新请求,其中,所述模型更新请求包括更新后的模型、所述更新后的模型的属性信息和内容信息,以便所述共享节点根据所述模型更新请求对已存储的所述更新后的模型对应的原模型进行更新。
- 根据权利要求15所述的管理方法,还包括:响应于模型更新需求,确定所属域内是否存储待更新模型;在所属域内未存储待更新模型的情况下,向所述共享节点发送模型下载请求,其 中,所述模型下载请求包括所述待更新模型的属性信息和内容信息;接收所述共享节点发送的模型下载响应,其中,所述模型下载响应包括所述待更新模型;根据所述待更新模型的当前训练参数,对所述待更新模型进行更新,得到所述更新后的模型。
- 一种共享节点,包括:接收模块,用于接收网络节点发送的模型复用请求,其中,所述网络节点为接入网域内的节点或核心网域内的节点,所述模型复用请求包括请求的模型的属性信息和内容信息,所述属性信息用于表示所述网络节点所属的域;查找模块,用于根据所述模型复用请求,在共享模型库中查找所述请求的模型;获取模块,用于在查找到的模型与所述请求的模型内容信息相匹配,属性信息不匹配的情况下,向所述网络节点获取训练数据和模型需求信息;调整模块,用于根据所述训练数据和模型需求信息对所述查找到的模型进行调整,作为所述请求的模型;发送模块,用于将所述请求的模型发送至所述网络节点。
- 一种网络节点,包括:确定模块,用于响应于推理侧节点发送的模型推理请求,确定所属域内是否存储所述模型推理请求对应的模型;发送模块,用于在未存储所述模型推理请求对应的模型的情况下,向共享节点发送模型复用请求,其中,所述模型复用请求包括请求的模型的属性信息和内容信息,所述请求的模型为所述模型推理请求对应的模型,所述属性信息用于表示所述网络节点所属的域;在所述共享节点查找到的模型与所述请求的模型内容信息相匹配,属性信息不匹配的情况下,向所述共享节点发送训练数据和模型需求信息;接收模块,用于接收所述共享节点发送的所述请求的模型,其中,所述请求的模型是所述共享节点根据所述训练数据和所述模型需求信息对所述查找到的模型进行调整得到的。
- 一种模型的管理装置,包括:处理器;以及耦接至所述处理器的存储器,用于存储指令,所述指令被所述处理器执行时,使所述处理器执行如权利要求1-16任一项所述的模型的管理方法。
- 一种计算机可读存储介质,其上存储有计算机程序,其中,该程序被处理器执行时实现权利要求1-16任一项所述方法的步骤。
- 一种模型的管理系统,包括:权利要求17所述的共享节点和权利要求18所述的网络节点。
- 一种计算机程序,包括:指令,所述指令被所述处理器执行时,使所述处理器执行如权利要求1-16任一项所述的模型的管理方法。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310817613.3A CN116566846B (zh) | 2023-07-05 | 2023-07-05 | 模型的管理方法、系统以及共享节点、网络节点 |
| CN202310817613.3 | 2023-07-05 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2025007722A1 true WO2025007722A1 (zh) | 2025-01-09 |
Family
ID=87491850
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2024/099249 Ceased WO2025007722A1 (zh) | 2023-07-05 | 2024-06-14 | 模型的管理方法、系统以及共享节点、网络节点 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN116566846B (zh) |
| WO (1) | WO2025007722A1 (zh) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116566846B (zh) * | 2023-07-05 | 2023-09-22 | 中国电信股份有限公司 | 模型的管理方法、系统以及共享节点、网络节点 |
| CN120238457A (zh) * | 2023-12-29 | 2025-07-01 | 华为技术有限公司 | 通信方法和通信装置 |
| CN121890116A (zh) * | 2024-07-24 | 2026-04-17 | 北京小米移动软件有限公司 | 通信方法、网络设备、通信系统和存储介质 |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113408634A (zh) * | 2021-06-29 | 2021-09-17 | 深圳市商汤科技有限公司 | 模型推荐方法及装置、设备、计算机存储介质 |
| US20220108214A1 (en) * | 2020-08-13 | 2022-04-07 | Electronics And Telecommunications Research Institute | Management method of machine learning model for network data analytics function device |
| WO2023017102A1 (en) * | 2021-08-13 | 2023-02-16 | Telefonaktiebolaget Lm Ericsson (Publ) | Systems and methods to optimize training of ai/ml models and algorithms |
| US20230060071A1 (en) * | 2021-08-20 | 2023-02-23 | Nokia Solutions And Networks Oy | Inference-aware ml model provisioning |
| CN116566846A (zh) * | 2023-07-05 | 2023-08-08 | 中国电信股份有限公司 | 模型的管理方法、系统以及共享节点、网络节点 |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111612153B (zh) * | 2019-02-22 | 2024-06-14 | 华为技术有限公司 | 训练模型的方法和装置 |
| CN110490305B (zh) * | 2019-08-22 | 2021-06-04 | 腾讯科技(深圳)有限公司 | 基于区块链网络的机器学习模型处理方法及节点 |
| US11902104B2 (en) * | 2020-03-04 | 2024-02-13 | Intel Corporation | Data-centric service-based network architecture |
| CN114339821A (zh) * | 2020-09-30 | 2022-04-12 | 英特尔公司 | 用于分布式nwdaf之间的机器学习模型共享的方法和装置 |
| CN113570397B (zh) * | 2021-01-27 | 2025-04-18 | 腾讯科技(深圳)有限公司 | 一种模型训练设备、方法、装置及介质 |
| CN113535808B (zh) * | 2021-06-24 | 2022-09-16 | 复旦大学 | 基于安全多方计算的键值对模型安全训练与推理方法 |
| CN114916015B (zh) * | 2022-05-12 | 2024-12-27 | 东南大学 | 雾无线接入网中基于量化的联邦强化学习的协作缓存方法 |
| CN115622777B (zh) * | 2022-10-09 | 2024-10-29 | 郑州大学 | 一种基于联盟链的多中心联邦学习数据共享方法 |
-
2023
- 2023-07-05 CN CN202310817613.3A patent/CN116566846B/zh active Active
-
2024
- 2024-06-14 WO PCT/CN2024/099249 patent/WO2025007722A1/zh not_active Ceased
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20220108214A1 (en) * | 2020-08-13 | 2022-04-07 | Electronics And Telecommunications Research Institute | Management method of machine learning model for network data analytics function device |
| CN113408634A (zh) * | 2021-06-29 | 2021-09-17 | 深圳市商汤科技有限公司 | 模型推荐方法及装置、设备、计算机存储介质 |
| WO2023017102A1 (en) * | 2021-08-13 | 2023-02-16 | Telefonaktiebolaget Lm Ericsson (Publ) | Systems and methods to optimize training of ai/ml models and algorithms |
| US20230060071A1 (en) * | 2021-08-20 | 2023-02-23 | Nokia Solutions And Networks Oy | Inference-aware ml model provisioning |
| CN116566846A (zh) * | 2023-07-05 | 2023-08-08 | 中国电信股份有限公司 | 模型的管理方法、系统以及共享节点、网络节点 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116566846A (zh) | 2023-08-08 |
| CN116566846B (zh) | 2023-09-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2025007722A1 (zh) | 模型的管理方法、系统以及共享节点、网络节点 | |
| EP3958606B1 (en) | Model pushing method and device, model requesting method and device, storage medium and electronic device | |
| Liu et al. | A task scheduling algorithm based on classification mining in fog computing environment | |
| CN111008521B (zh) | 生成宽表的方法、装置及计算机存储介质 | |
| CN114077602B (zh) | 数据迁移方法和装置、电子设备、存储介质 | |
| US11563650B2 (en) | Determining optimum software update transmission parameters | |
| CN106874281B (zh) | 实现数据库读写分离的方法和装置 | |
| CN110427386B (zh) | 数据处理方法、装置及计算机存储介质 | |
| CN110276074B (zh) | 自然语言处理的分布式训练方法、装置、设备及存储介质 | |
| US20240095529A1 (en) | Neural Network Optimization Method and Apparatus | |
| US20200379821A1 (en) | Method and apparatus for executing workflow including functions written in heterogeneous programing language | |
| CN112118133B (zh) | 基于自定义结构数据便于以太坊智能合约快速升级方法 | |
| CN110389766A (zh) | HBase容器集群部署方法、系统、设备及计算机可读存储介质 | |
| CN115391009A (zh) | 一种批量业务的资源调度方法、系统、设备及存储介质 | |
| CN115730507A (zh) | 模型引擎的构建、核函数处理方法、设备及存储介质 | |
| CN114693220B (zh) | 基于数字孪生DaaS平台的算法仓库管理方法及系统 | |
| WO2025107811A9 (zh) | 一种信息查询方法、装置及系统 | |
| CN112948406A (zh) | 一种配置变更数据存储和同步的方法、系统和装置 | |
| CN119781837B (zh) | 一种操作系统内核的配置方法及相关产品 | |
| WO2021089983A1 (en) | Validating distributed source content | |
| CN119376930B (zh) | 面向资源受限终端的深度模型计算图峰值内存优化方法 | |
| CN114651236A (zh) | 资源删除方法、装置、设备及存储介质 | |
| CN120448121B (zh) | 面向无服务器计算的任务动态分配方法、装置及设备 | |
| CN116109381B (zh) | 电商平台数据处理方法及系统 | |
| CN115017166B (zh) | 垂类数据构建方法、装置、电子设备及存储介质 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 24835205 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |