Communication method, communication apparatus and communication system

Abstract

A communication method, a communication apparatus and a communication system. In the method, a model that has been trained by a training function entity and / or a model that is being trained by the training function entity is indicated by means of model availability information, such that whether there is a model meeting a model selection requirement among the model that has been trained or the model that is being trained can be accurately determined; and if there is a model meeting the model selection requirement, the model is selected for training from the model that has been trained or the model that is being trained, without the need to request that the training function entity retrains a model. Therefore, the probability of a model being reused can be improved and the repeated training of the model can be reduced, thereby saving on computing resources for training the model, and improving the use efficiency of the computing resources.

Description

Communication method, communication apparatus, and communication system

[0001] Cross-reference to Related Applications

[0002] This application claims priority to the Chinese Patent Application No. 202410889432.6, filed on July 3, 2024, and entitled "A communication method, a communication apparatus and a communication system", the content of which is incorporated herein by reference in its entirety. TECHNICAL FIELD

[0003] The present application relates to the field of communication technology, and in particular to a communication method, a communication apparatus and a communication system. BACKGROUND

[0004] In order to improve the intelligent and automated level of the network, artificial intelligence (AI) models are used in more and more technical fields. AI models can play an important role in complex network operation and management, network optimization, and improving user experience.

[0005] Models are usually obtained by training. For example, a training function (TF) entity can be configured with a model training task, and use training data to perform the model training task to obtain a usable model. One or more training function entities can be configured in a communication system, such as being deployed in an access network management system, an access network device, a core network management system, or a core network device (such as a network data analytics function (NWDAF) entity). The training tasks in each training function entity can be configured according to the needs.

[0006] Since the network environment and network requirements of different training function entities can be different and dynamically change, the model training tasks of different training function entities will also be different. When training a model, a training function entity needs to use computing resources. The more and more complex the trained model is, the more computing resources are needed.

[0007] During model training, how to improve the use efficiency of computing resources needs to be solved. SUMMARY

[0008] Embodiments of the present application provide a communication method, a communication apparatus and a communication system to improve the use efficiency of computing resources.

[0009] In a first aspect, an embodiment of the present application provides a communication method, which can be applied to a first device, a module (such as a circuit, a chip or a chip system, etc.) in the first device, or a logic node, a logic module or software capable of realizing all or part of the functions of the first device. The method comprises: obtaining model availability information of a training function entity, the model availability information comprising information of a model that has completed training and / or information of a model that is being trained; determining a first model according to the model availability information; and deploying the first model.

[0010] Based on the above scheme, the model availability information indicates the model that has completed training and / or the model that is being trained of the training function entity, so as to accurately determine whether there is a model that meets the model selection requirement in the model that has completed training and / or the model that is being trained, and if there is, a model is selected from the model that has completed training or the model that is being trained for deployment, without the need to request the training function entity to retrain a model, thereby improving the probability of model reuse, reducing repeated training of the model, and further saving the computing resources for training the model and improving the use efficiency of the computing resources.

[0011] In a possible design, the determining of the first model according to the model availability information comprises: determining one model that meets a first model selection requirement in the model that has completed training as the first model; or determining one model that meets the first model selection requirement in the model that is being trained as the first model.

[0012] In a possible design, the model availability information comprises information of the model that has completed training and information of the model that is being trained; and the determining of the first model according to the model availability information comprises: in a case where there is a model that meets the first model selection requirement in the model that has completed training, determining one model that meets the first model selection requirement in the model that has completed training as the first model; or in a case where none of the model that has completed training meets the first model selection requirement, determining one model that meets the first model selection requirement in the model that is being trained as the first model.

[0013] In a possible design, the first model selection requirement comprises an inference name and / or a performance index value.

[0014] Based on the above scheme, according to the first model selection requirement, it can be accurately determined whether there is a model that meets the model selection requirement in the model that has completed training and / or the model that is being trained, which is helpful for correctly selecting a model.

[0015] In a possible design, the information of the trained model includes one or more of an identifier of the trained model, an inference name of the trained model, or a performance indicator value of the trained model; and the information of the model being trained includes one or more of an identifier of the model being trained, an inference name of the model being trained, or an expected performance indicator value of the model being trained.

[0016] In a possible design, the method further includes: sending, to the training function entity, a test request, where the test request includes the information of the first model and information of a test data set, and the test request is used to request that the first model be tested based on the test data set; and receiving a model test result from the training function entity.

[0017] Based on this scheme, different test requirements of different users for the same model are met through model testing, which helps to provide appropriate models for users.

[0018] In a possible design, the deploying the first model includes: in a case where the model test result meets a model test requirement, sending, to the training function entity, a request message, where the request message is used to request that the first model be deployed to a target device.

[0019] Based on the above scheme, the model can be accurately deployed to the target device.

[0020] In a possible design, the method further includes: in a case where the model test result does not meet a model test requirement, sending, to the training function entity, a training request, where the training request includes an incremental training indication, and the incremental training indication is used to indicate that the first model is incrementally trained.

[0021] Based on this scheme, by introducing incremental training, the training time can be reduced and the computing resources can be saved.

[0022] In a possible design, the model availability information further includes information of a model that can be trained by using remaining computing resources of the training function entity; and the determining the first model according to the model availability information includes: in a case where neither the trained model nor the model being trained meets a first model selection requirement, determining, as the first model, a model that meets a second model selection requirement from the model that can be trained by using the remaining computing resources of the training function entity.

[0023] Based on the above scheme, the model availability information indicates the model that has completed training, the model that is being trained, and the model supported by the remaining computing resources of the training function entity for training, so that the first device can accurately determine whether there is a model that meets the model selection requirement in the model that has completed training and / or the model that is being trained, and if not, select a model from the model supported by the remaining computing resources of the training function entity for training for deployment, so that the computing resources of the training function entity are fully utilized, and the use efficiency of the computing resources is improved. Moreover, the first device does not need to re-request a training model from other training function entities, so that the signaling overhead of the first device can be saved, and the speed of the first device to obtain the trained model is improved.

[0024] In a possible design, the information about the model supported by the remaining computing resources of the training function entity for training includes one or more of an identifier of the model supported by the remaining computing resources of the training function entity for training, an inference name of the model supported by the remaining computing resources of the training function entity for training, or an expected training time of the model supported by the remaining computing resources of the training function entity for training.

[0025] In a possible design, the model availability information further includes information about the model supported by the remaining computing resources of the training function entity for training; in a case where neither the model that has completed training nor the model that is being trained meets the first model selection requirement, and the model supported by the remaining computing resources of the training function entity for training does not meet the second model selection requirement, a model is selected from a model set as the first model, the model set including models supported by the training function entity except for the model that has completed training, the model that is being trained, and the model supported by the remaining computing resources of the training function entity for training.

[0026] Based on the above scheme, in a case where neither the model that has completed training nor the model that is being trained meets the first model selection requirement, and the model supported by the remaining computing resources of the training function entity for training does not meet the second model selection requirement, a model is selected from a model set as the first model, so that the first device does not need to re-request a training model from other training function entities, so that the signaling overhead of the first device can be saved, and the speed of the first device to obtain the trained model is improved.

[0027] In a possible design, the method further includes: sending a training request to the training function entity, the training request being used to request training of the first model, the training request including a training requirement, the training requirement including a model performance requirement and / or a model training time requirement.

[0028] Based on the above scheme, by training the first model, a model meeting the training requirement can be obtained.

[0029] In a possible design, the sending of the training request to the training function entity includes: in a case where the training function entity supports distributed training, sending the training request to the training function entity.

[0030] In a possible design, the first model selection requirement includes an inference name and / or a performance indicator value; and the second model selection requirement includes an inference name and / or training time information.

[0031] In a possible design, the obtaining of the model availability information of the training function entity includes: sending a first query message to the training function entity, where the first query message is used for querying the model availability information; and receiving the model availability information from the training function entity.

[0032] In a second aspect, an embodiment of the present application provides a communication method, which can be applied to a training function entity, a module (for example, a circuit, a chip or a chip system, etc.) in the training function entity, or a logic node, a logic module or software capable of realizing all or part of the functions of the training function entity. The method includes: receiving a first query message, where the first query message is used for querying model availability information of a training function entity, and the model availability information includes information of a model that has completed training and / or information of a model that is being trained; and sending the model availability information.

[0033] Based on the above scheme, by indicating the model that has completed training and / or the model that is being trained by the model availability information, it is helpful to accurately determine, based on the model availability information, whether there is a model meeting the model selection requirement in the model that has completed training and / or the model that is being trained, and if there is, to select a model from the model that has completed training or the model that is being trained for deployment, without the need to request the training function entity to retrain a model, so that the probability of reuse of the model can be improved, repeated training of the model is reduced, and therefore, the computing resources used for training the model are saved, and the use efficiency of the computing resources is improved.

[0034] In a possible design, the information of the model that has completed training includes one or more of an identifier of the model that has completed training, an inference name of the model that has completed training or a performance indicator value of the model that has completed training; and the information of the model that is being trained includes one or more of an identifier of the model that is being trained, an inference name of the model that is being trained or an expected performance indicator value of the model that is being trained.

[0035] In a possible design, the model availability information further includes information about a model supported by remaining computing resources of the training function entity for training, and the information about the model supported by the remaining computing resources of the training function entity for training includes one or more of an identifier of the model supported by the remaining computing resources of the training function entity for training, an inference name of the model supported by the remaining computing resources of the training function entity for training, or an expected training time of the model supported by the remaining computing resources of the training function entity for training.

[0036] In a possible design, the method further includes: sending, to the computing resource data center, a second query message, where the second query message is used to query the computing resource usage information of the training function entity; receiving, from the computing resource data center, the computing resource usage information of the training function entity; and determining, according to the computing resource usage information, the information about the model supported by the remaining computing resources of the training function entity for training.

[0037] Based on the foregoing scheme, by querying the computing resource usage information of the training function entity from the computing resource data center, the information about the model supported by the remaining computing resources of the training function entity for training can be accurately determined based on the computing resource usage information.

[0038] In a third aspect, an embodiment of the present application provides a communication method, which can be applied to a first training function entity, a module (for example, a circuit, a chip or a chip system, etc.) in the first training function entity, or a logic node, a logic module or software capable of realizing all or part of functions of the first training function entity. The method includes: receiving a training request, where the training request includes an identifier of a first model and training requirements, and the training request is used to request training of the first model; in a case where remaining computing resources of the first training function entity do not meet the training requirements, sending, to a resource management function entity, a first message, where the first message includes state information of the first training function entity; receiving, from the resource management function entity, a second message, where the second message includes information about newly added computing resources; and performing model training on the first model according to the newly added computing resources.

[0039] Based on the above scheme, when the first device requests a training model from the first training function entity, if the computing resources for training the model on the first training function entity are insufficient, the first training function entity can request the resource management function entity to add computing resources for the first training function entity, and then the first training function entity can train the model based on the added computing resources and the remaining computing resources, so that the remaining computing resources on the first training function entity are used for training the model instead of being idle all the time, thereby improving the use efficiency of the computing resources. Moreover, the first device does not need to request the training model from other training function entities, thereby saving the signaling overhead of the first device and improving the speed of obtaining the trained model by the first device.

[0040] In a possible design, the state information includes one or more of the following information: location information of the first training function entity, information of the computing resources allocated to the first training function entity, location information of the computing resources allocated to the first training function entity, computing resource demand of the first training function entity, or computing resource usage information of the first training function entity.

[0041] In a possible design, the training demand includes a model performance demand and / or a model training time demand.

[0042] In a fourth aspect, an embodiment of the present application provides a communication method, which can be applied to a resource management function entity, a module (such as a circuit, a chip or a chip system, etc.) in the resource management function entity, or a logic node, a logic module or software capable of realizing all or part of the functions of the resource management function entity. The method includes: receiving a first message from a first training function entity, the first message including state information of the training function entity; sending a first request to a computing resource data center, the first request being used to request to allocate computing resources to the first training function entity; receiving a first response from the computing resource data center, the first response being used to indicate a resource allocation result; and according to the first response, sending a second message to the first training function entity, the second message including information of the added computing resources.

[0043] Based on the above scheme, the computing resources for model training can be dynamically allocated to the first training function entity, which helps to improve the use efficiency of the computing resources and realize fast training of the model.

[0044] In a possible design, the first request includes information of at least one second training function entity, and the first request is used to request to allocate idle computing resources of the at least one second training function entity to the first training function entity.

[0045] Based on the above scheme, idle computing resources on the training function entity can be reasonably allocated, and the use efficiency of computing resources can be improved

[0046] In a possible design, the state information includes one or more of the following information: location information of the first training function entity, information of the allocated computing resources of the first training function entity, location information of the allocated computing resources of the first training function entity, computing resource requirement of the first training function entity, or computing resource usage information of the first training function entity.

[0047] In a possible design, the method further includes: determining the at least one second training function entity according to the location information of the first training function entity and / or the location information of the allocated computing resources of the first training function entity, wherein the idle computing resources of the second training function entity and the computing resources of the first training function entity have high-speed connection.

[0048] Based on the above scheme, appropriate computing resources can be allocated to the first training function entity, so as to improve the model training speed and accuracy.

[0049] In a fifth aspect, the present application provides a communication device, which has the function of realizing the first aspect, for example, the communication device includes a module, unit or means corresponding to the operation related to the first aspect, which can be realized by software, or by hardware, or by the combination of software and hardware.

[0050] In a sixth aspect, the present application provides a communication device, which has the function of realizing the second aspect, for example, the communication device includes a module, unit or means corresponding to the operation related to the second aspect, which can be realized by software, or by hardware, or by the combination of software and hardware.

[0051] In a seventh aspect, the present application provides a communication device, which has the function of realizing the third aspect, for example, the communication device includes a module, unit or means corresponding to the operation related to the third aspect, which can be realized by software, or by hardware, or by the combination of software and hardware.

[0052] In an eighth aspect, the present application provides a communication apparatus, which has the functions of the fourth aspect, and the communication apparatus includes modules, units or means corresponding to the operations of the fourth aspect, which can be implemented in software, or in hardware, or in a combination of software and hardware.

[0053] In a ninth aspect, the present application provides a communication apparatus, which includes an interface circuit and one or more processors. The one or more processors are coupled to a memory. The memory is configured to store part or all of the necessary computer programs or instructions for implementing the functions of the first aspect. The one or more processors are configured to execute the computer programs or instructions, which, when executed, cause the communication apparatus to implement the method in any possible design or implementation manner of the first aspect. The interface circuit is configured to implement the communication function within the communication apparatus and / or the communication function between the communication apparatus and other apparatuses or components.

[0054] The communication apparatus can be the first device, a module (e.g., a circuit, a chip or a chip system, etc.) in the first device, or a logic node, a logic module or software capable of implementing all or part of the functions of the first device.

[0055] In a tenth aspect, the present application provides a communication apparatus, which includes an interface circuit and one or more processors. The one or more processors are coupled to a memory. The memory is configured to store part or all of the necessary computer programs or instructions for implementing the functions of the second aspect. The one or more processors are configured to execute the computer programs or instructions, which, when executed, cause the communication apparatus to implement the method in any possible design or implementation manner of the second aspect. The interface circuit is configured to implement the communication function within the communication apparatus and / or the communication function between the communication apparatus and other apparatuses or components.

[0056] The communication apparatus can be the training function entity, a module (e.g., a circuit, a chip or a chip system, etc.) in the training function entity, or a logic node, a logic module or software capable of implementing all or part of the functions of the training function entity.

[0057] In an eleventh aspect, the present application provides a communication apparatus, which comprises an interface circuit and one or more processors. The one or more processors are coupled with a memory. The memory is configured to store part or all of the computer programs or instructions necessary for implementing the functions related to the third aspect described above. The one or more processors can execute the computer programs or instructions, which, when executed, cause the communication apparatus to implement the method in any possible design or implementation manner of the third aspect described above. The interface circuit is configured to implement the communication function within the communication apparatus and / or the communication function of the communication apparatus with other apparatuses or components.

[0058] The communication apparatus described above can be a first training function entity, a module (e.g., a circuit, a chip or a chip system, etc.) in the first training function entity, or a logic node, a logic module or software capable of implementing all or part of the functions of the first training function entity.

[0059] In a twelfth aspect, the present application provides a communication apparatus, which comprises an interface circuit and one or more processors. The one or more processors are coupled with a memory. The memory is configured to store part or all of the computer programs or instructions necessary for implementing the functions related to the fourth aspect described above. The one or more processors can execute the computer programs or instructions, which, when executed, cause the communication apparatus to implement the method in any possible design or implementation manner of the fourth aspect described above. The interface circuit is configured to implement the communication function within the communication apparatus and / or the communication function of the communication apparatus with other apparatuses or components.

[0060] The communication apparatus described above can be a resource management function entity, a module (e.g., a circuit, a chip or a chip system, etc.) in the resource management function entity, or a logic node, a logic module or software capable of implementing all or part of the functions of the resource management function entity.

[0061] In a thirteenth aspect, the present application provides a computer readable storage medium, which stores computer programs or instructions, which, when executed, implement the method in any possible design of the first aspect to the fourth aspect described above.

[0062] In a fourteenth aspect, the present application provides a computer program product, which comprises computer programs or instructions, which, when executed, implement the method in any possible design of the first aspect to the fourth aspect described above.

[0063] In a fifteenth aspect, the present application provides a communication system, which comprises a first device for executing the method in any possible design of the first aspect described above, and a training function entity for executing the method in any possible design of the second aspect described above. In a fifteenth aspect, the present application provides a communication system, which comprises a first device for executing the method in any possible design of the first aspect described above, and a training function entity for executing the method in any possible design of the second aspect described above.

[0064] In a sixteenth aspect, the present application provides a communication system, comprising a first training function entity configured to perform the method of any possible design of the third aspect above, and a resource management function entity configured to perform the method of any possible design of the fourth aspect above. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 is a possible, non-limiting system diagram;

[0066] Figure 2(a) is a flow diagram of a communication method according to an embodiment of the present application;

[0067] Figure 2(b) is a flow diagram of a communication method according to an embodiment of the present application;

[0068] Figure 3 is a flow diagram of a communication method according to an embodiment of the present application;

[0069] Figure 4 is a flow diagram of a communication method according to an embodiment of the present application;

[0070] Figure 5 is a flow diagram of a communication method according to an embodiment of the present application;

[0071] Figure 6 is a possible exemplary block diagram of a communication apparatus according to an embodiment of the present application;

[0072] Figure 7 is a possible exemplary block diagram of a communication apparatus according to an embodiment of the present application. DETAILED DESCRIPTION

[0073] Figure 1 is a possible, non-limiting system diagram. As shown in Figure 1, the communication system 10 includes a radio access network (RAN) 100 and a core network (CN) 200. Optionally, the communication system further includes an Internet 300. The RAN 100 includes at least one RAN node (e.g., 110a and 110b in Figure 1, collectively referred to as 110) and at least one terminal (e.g., 120a-120j in Figure 1, collectively referred to as 120). The RAN 100 can further include other RAN nodes, such as wireless relay devices and / or wireless backhaul devices (not shown in Figure 1), etc. The terminal 120 is connected to the RAN node 110 in a wireless manner. The RAN node 110 is connected to the core network 200 in a wireless or wired manner. The core network device in the core network 200 and the RAN node 110 in the RAN 100 can be different physical devices respectively, or can be the same physical device integrated with the logical functions of the core network and the logical functions of the radio access network.

[0074] The RAN 100 can be a 3rd generation partnership project (3GPP) related cellular system, e.g., a 4th generation (4G), a 5th generation (5G) mobile communication system, or a future oriented evolutional system, e.g., a 6th generation (6G) mobile communication system. The RAN 100 can also be an open radio access network (O-RAN or ORAN), a cloud radio access network (CRAN), or a wireless fidelity (WiFi) system. The RAN 100 can also be a communication system that combines two or more of the above systems.

[0075] The RAN nodes 110, which can also be referred to as access network devices, RAN entities or access nodes, etc., form part of the communication system 10 and are configured to facilitate wireless access to the communication system. The RAN nodes 110 in the communication system 10 can be of the same type or of different types. In some scenarios, the roles of the RAN nodes 110 and the terminals 120 are relative, e.g., the network element 120i in Figure 1 can be a helicopter or a drone, which can be configured to be a mobile base station, to the terminals 120j that access the RAN 100 through the network element 120i, the network element 120i is a base station; but to the base station 110a, the network element 120i is a terminal. The RAN nodes 110 and the terminals 120 are sometimes referred to as communication apparatuses, e.g., the network elements 110a and 110b in Figure 1 can be understood as communication apparatuses with base station functionalities, and the network elements 120a-120j can be understood as communication apparatuses with terminal functionalities.

[0076] In a possible scenario, the RAN node can be a base station, an evolved NodeB (eNodeB), an access point (AP), a transmission reception point (TRP), a next generation NodeB (gNB), a next generation base station in a 6G mobile communication system, a base station in a future mobile communication system, or an access node in a WiFi system, etc. A terminal can be a device or module with corresponding communication functions for accessing the above communication system. The terminal can also be referred to as a terminal device, user equipment (UE), mobile station, mobile terminal, etc. The terminal can be widely applied to various scenarios, such as device-to-device (D2D), vehicle to everything (V2X) communication, machine-type communication (MTC), internet of things (IOT), virtual reality, augmented reality, industrial control, automatic driving, remote medical treatment, smart power grid, smart furniture, smart office, smart wear, smart transportation, smart city, etc. The terminal can be a mobile phone, tablet computer, computer with wireless transceiver function, wearable device, vehicle, unmanned aerial vehicle, helicopter, airplane, ship, robot, mechanical arm, smart home device, transport vehicle with wireless communication function, communication module, etc. Embodiments of the present application do not limit the device form of the terminal. The terminal is usually provided with a communication module, circuit or chip for executing corresponding communication functions. The terminal is also configured with program instructions for executing corresponding communication functions.

[0077] In order to improve the intelligent and automatic level of the network, AI models are applied to more and more technical fields. AI models can play an important role in complex network operation and management, network optimization, and improvement of user experience.

[0078] Currently, 3GPP standards define AI model management standards in the field of network management, including model management life cycle processes, as follows:

[0079] (1) Model training: Model training includes initial training and retraining of one or a group of models, and also includes model verification to evaluate the performance of the model on training data and validation data. If the verification result does not meet the expectation (for example, variance or accuracy is unacceptable), the model needs to be retrained.

[0080] (2) Model testing: testing a validated model to evaluate the performance of the trained model on test data. If the test result is not as expected, the model needs to be retrained.

[0081] (3) Inference emulation: running a model for inference in an emulation environment, which aims to evaluate the inference performance of the model in the emulation environment before applying the model to a target network or system.

[0082] (4) Model deployment: model deployment includes a model loading process (also known as a series of atomic operations) to make the trained model available for target inference functions. In some cases, the model may not need to be deployed, for example, when the training function and the inference function are located in the same place, and the model can be used after the model training is completed without further model deployment.

[0083] (5) Inference: through the inference function, the trained model is used for inference. When performing inference, model retraining or updating can also be triggered based on performance monitoring and evaluation.

[0084] Models are usually obtained through training, such as a training function entity can be configured with a model training task, and the training data is used to perform the model training task to obtain a usable model. One or more training function entities can be configured in a communication system, such as deploying one or more training function entities in an access network management system, an access network device, a core network management system, or a core network device (such as a NWDAF entity). The training task in each training function entity can be configured according to the needs.

[0085] Since the network environment and network demand of different training function entities may be different and dynamically change, the model training tasks of different training function entities will also be different. When training a model, a training function entity needs to use computing resources, and the more and more complex the trained model is, the more computing resources it needs.

[0086] During model training, how to improve the use efficiency of computing resources needs to be solved.

[0087] To solve the above problems, the present application provides corresponding solutions.

[0088] The communication method, the communication apparatus and the communication system provided in the present application are described below with reference to the drawings. It can be understood that the first device, the training function entity and the resource management function entity are taken as examples of the execution subject of the interaction in the present application, but the present application does not limit the execution subject of the interaction. For example, the method executed by the first device in the present application can also be implemented by a module (such as a circuit, a chip or a chip system, etc.) in the first device, or a logical node, a logical module or software capable of realizing all or part of the function of the first device. The method executed by the training function entity in the present application can also be implemented by a module (such as a circuit, a chip or a chip system, etc.) in the training function entity, or a logical node, a logical module or software capable of realizing all or part of the function of the training function entity. The method executed by the resource management function entity in the present application can also be implemented by a module (such as a circuit, a chip or a chip system, etc.) in the resource management function entity, or a logical node, a logical module or software capable of realizing all or part of the function of the resource management function entity.

[0089] In the present application, the first device refers to a model using device or a model requesting device, which can request the training function entity to obtain the information of the model, or request the deployment of the model to a target device, etc. For example, the first device can be an operations support system (OSS) or a terminal, etc.

[0090] In the present application, the training function entity refers to an entity with a model training function. The training function entity can be independently deployed, or deployed on other devices (such as access network devices, core network devices, etc.) in other fields or fields as a functional unit of other devices.

[0091] In the present application, the resource management function entity can manage the computing resources on different training function entities, for example, the resource management function entity can allocate or release computing resources for different training function entities. The computing resources can be used for training models. The computing resources can be deployed in a computing resource data center and managed by a corresponding system (such as a network function virtualization (NFV) management and orchestration (MANO) system). The resource management function entity can be independently deployed, or deployed on other devices (such as access network devices, core network devices, etc.) as a functional unit of other devices.

[0092] FIG. 2(a) is a flowchart of a communication method provided in an embodiment of the present application. The method includes the following steps:

[0093] At step 201a, the first device acquires model availability information of the training function entity.

[0094] The model availability information includes one or more of the following information: information of a model that has completed training, information of a model that is being trained, or information of a remaining computing resource of the training function entity supporting a model that is being trained.

[0095] The information of the model that has completed training includes one or more of an identification of the model that has completed training, an inference name of the model that has completed training, or a performance indicator value of the model that has completed training. The inference name of the model that has completed training is used to indicate an inference scenario or an inference type of the model, for example, the inference name can be an MDA type, etc. The performance indicator value of the model that has completed training is used to indicate a performance of the model, for example, can be a model inference accuracy, etc. Wherein, the number of the model that has completed training can be one or more.

[0096] The information of the model that is being trained includes one or more of an identification of the model that is being trained, an inference name of the model that is being trained, or an expected performance indicator value of the model that is being trained. The inference name of the model that is being trained is used to indicate an inference scenario or an inference type of the model, for example, the inference name can be an MDA type, etc. The expected performance indicator value of the model that is being trained is used to indicate an expected performance of the model, for example, can be a model inference accuracy, etc. Wherein, the number of the model that is being trained can be one or more.

[0097] The information of the remaining computing resource of the training function entity supporting the model that is being trained includes one or more of an identification of the remaining computing resource of the training function entity supporting the model that is being trained, an inference name of the remaining computing resource of the training function entity supporting the model that is being trained, or an expected training time of the remaining computing resource of the training function entity supporting the model that is being trained. The remaining computing resource of the training function entity supporting the model that is being trained is a model that is supported by the training function entity for training but has not been trained out. The inference name of the remaining computing resource of the training function entity supporting the model that is being trained is used to indicate an inference scenario or an inference type of the model, for example, the inference name can be an MDA type, etc. Wherein, the number of the remaining computing resource of the training function entity supporting the model that is being trained can be one or more.

[0098] Exemplarily, the following gives multiple different implementation methods of the first device acquiring the model availability information of the training function entity.

[0099] In implementation method one, the first device locally stores the model availability information of the training function entity, and then the first device acquires the model availability information of the training function entity from the local.

[0100] Method two is implemented, the training function entity actively sends the model availability information of the training function entity to the first device.

[0101] Method three is implemented, the first device requests the training function entity to obtain the model availability information of the training function entity.

[0102] For example, the first device sends a first query message to the training function entity, the first query message is used to query the model availability information, and then the training function entity sends the model availability information of the training function entity to the first device based on the first query message.

[0103] For example, in the case where the model availability information includes the information that the remaining computing resources of the training function entity support the trained model, after receiving the first query message, the training function entity can send a second query message to the computing resource data center, the second query message is used to query the computing resource usage information of the training function entity, then the computing resource data center sends the computing resource usage information of the training function entity to the training function entity, and then the training function entity determines the information that the remaining computing resources of the training function entity support the trained model according to the computing resource usage information of the training function entity, and carries the information that the remaining computing resources of the training function entity support the trained model in the model availability information sent to the first device.

[0104] For example, the computing resource usage information includes one or more of the computing resource allocation amount, the computing resource usage amount or the computing resource remaining amount of the training function entity. The computing resource includes one or more of the CPU resource, the graphics processing unit (GPU) resource, the memory resource or the bandwidth resource.

[0105] Step 202a, the first device determines the first model according to the model availability information.

[0106] The following describes multiple cases of determining the first model according to the model availability information by the first device.

[0107] Case one, the first device determines one of the trained models that meets the first model selection requirement as the first model; or the first device determines one of the models being trained that meets the first model selection requirement as the first model.

[0108] For example, the first device first determines whether there is a model meeting the first model selection requirement in the trained models. If there is, the first device selects a model from the trained models as the first model. If there is not, the first device further determines whether there is a model meeting the first model selection requirement in the training models. If there is, the first device selects a model from the training models as the first model.

[0109] For another example, the first device first determines whether there is a model meeting the first model selection requirement in the training models. If there is, the first device selects a model from the training models as the first model. If there is not, the first device further determines whether there is a model meeting the first model selection requirement in the trained models. If there is, the first device selects a model from the trained models as the first model.

[0110] For another example, the first device considers the trained models and the training models as a whole, and does not distinguish between them. The first device determines whether there is a model meeting the first model selection requirement in the trained models and the training models. If there is, the first device selects a model from the trained models and the training models as the first model. The first model can be a model in the trained models or a model in the training models.

[0111] For example, the first model meeting the first model selection requirement can be that the inference name of the first model is the same as the inference name in the first model selection requirement. The same inference name indicates that the scenario to which the first model is applied meets the demand of the first device. For another example, the first model meeting the first model selection requirement can be that the performance index value of the first model matches the performance index value in the first model selection requirement. The matching means that the performance index indicated by the performance index value of the first model is better than the performance index indicated by the performance index value in the first model selection requirement, or the performance index indicated by the performance index value of the first model is the same as the performance index indicated by the performance index value in the first model selection requirement. For another example, the first model meeting the first model selection requirement can be that the inference name of the first model is the same as the inference name in the first model selection requirement, and the performance index value of the first model matches the performance index value in the first model selection requirement.

[0112] In a possible implementation, if the first device selects a model from the trained models as the first model, when there are multiple models in the trained models that meet the first model selection requirement, the first device can randomly select a model from the multiple models as the first model, or the first device can select a model with the optimal performance indicator from the multiple models as the first model.

[0113] In a possible implementation, if the first device selects a model from the trained models as the first model, when there are multiple models in the trained models that meet the first model selection requirement, the first device can randomly select a model from the multiple models as the first model, or the first device can select a model with the optimal performance indicator from the multiple models as the first model.

[0114] Based on the case one, in a possible implementation, after the first device determines the first model from the trained models or the trained models, the first device can send a test request to the training function entity, where the test request includes information of the first model and information of a test data set, and the test request is used to request testing of the first model based on the test data set. For example, the information of the test data set can be address information or identification information of the model test data set. The training function entity acquires the model test data set based on the information of the model test data set, tests the first model using the model test data set, and sends a model test result to the first device. The model test is needed because the inference name and / or the performance indicator of the first model can meet the requirements of the first device, but the training data used in the training process of the first model can not match the requirements of the first device. In order to ensure that the first device can meet the requirements of the first device to the greatest extent, the first device can request the training function entity to complete the model test on the specified test data set, and determine whether the first model meets the model test requirements of the first device according to the model test result. Based on this scheme, different test requirements of different users for the same model are realized through the model test, which helps to provide appropriate models for users.

[0115] When the model test result does not meet the model test requirement, the first device can send a training request to the training function entity, where the training request includes an incremental training indication, and the incremental training indication is used to indicate that the first model is subjected to incremental training, so that the training function entity performs incremental training on the first model based on the incremental training indication. The incremental training refers to performing further training on the trained model or the trained model using the training data. Based on this scheme, the training time can be reduced and the computing resources can be saved by introducing the incremental training.

[0116] In a second case, when neither the trained model nor the model being trained meets the first model selection requirement, the first device determines that the remaining computing resources of the training function entity support one model of the trained models that meets the second model selection requirement, and selects the one model as the first model.

[0117] Exemplarily, the second model selection requirement includes an inference name and / or training time information, the training time information being used to indicate a training time requirement of the first device for the first model, for example, the training time information is used to indicate a maximum training time, i.e., the first device expects to complete the model training within the maximum training time. For example, the first model meets the second model selection requirement when the inference name of the first model is the same as the inference name in the second model selection requirement. For another example, the first model meets the second model selection requirement when the expected training time of the first model is less than or equal to the maximum training time indicated by the training time information in the second model selection requirement. For another example, the first model meets the second model selection requirement when the inference name of the first model is the same as the inference name in the second model selection requirement, and the expected training time of the first model is less than or equal to the maximum training time indicated by the training time information in the second model selection requirement.

[0118] In a possible implementation method, if the first device selects one model from the trained models supported by the remaining computing resources of the training function entity as the first model, when there are multiple models in the trained models supported by the remaining computing resources of the training function entity that meet the first model selection requirement, the first device can randomly select one model from the multiple models as the first model, or the first device can select the model with the shortest expected training time from the multiple models as the first model.

[0119] In the second case, in a possible implementation method, after determining the first model, the first device can send a training request to the training function entity, the training request being used to request training of the first model. After receiving the training request, the training function entity trains the first model to obtain a trained first model.

[0120] In a third case, when neither the trained model nor the model being trained meets the first model selection requirement, and the trained model supported by the remaining computing resources of the training function entity does not meet the second model selection requirement, the first device selects one model from a model set as the first model, the model set including models supported by the training function entity excluding the trained model, the model being trained, and the trained model supported by the remaining computing resources of the training function entity.

[0121] The model in the model set can be understood as a model that the training function entity can train but the remaining computing resources currently do not support training.

[0122] Exemplarily, the first device can obtain information about the models supported by the training function entity before, after or at the same time as step 201a. For example, the training function entity supports models 1 to 100, that is, supports training models 1, 2, …, and 100. The information about the models supported by the training function entity can include the identifiers of the models 1 to 100.

[0123] Exemplarily, it is assumed that the models supported by the training function entity include models 1 to 100, the models that the training function entity has completed training include models 1 to 10, the models that the training function entity is training include models 11 to 20, the models that the remaining computing resources of the training function entity support training include models 21 to 40, and neither models 1 to 20 nor models 21 to 40 meet the first model selection requirement and the second model selection requirement. The training function entity can select a model from a model set as the first model, the model set including the identifiers of models 41 to 100.

[0124] For this case three, in a possible implementation method, after determining the first model, the first device can send a training request to the training function entity, the training request being used to request training of the first model, the training request including a training requirement, the training requirement including a model performance requirement and / or a model training time requirement. The model performance requirement is used to indicate the performance requirement of the first device on the trained first model. The model training time requirement is used to indicate the training time requirement of the first device on the first model, for example, the training time requirement is used to indicate the maximum training time, that is, the first device expects to complete model training within the maximum training time. After receiving the training request, the training function entity trains the first model based on the training requirement to obtain the trained first model.

[0125] Optionally, before sending the training request to the training function entity, the first device also determines that the training function entity supports distributed training, that is, the first device sends the training request to the training function entity in the case that the training function entity supports distributed training. The training function entity supporting distributed training means that when the remaining computing resources of the training function entity do not support training of the first model, the training function entity can request the resource management function entity to increase computing resources for the training function entity, so that the training function entity can complete training of the first model.

[0126] Step 203a, the first device deploys the first model.

[0127] The first device deploys the first model, which can be that the first device requests the training function entity to deploy the first model on the first device, or can be that the first device requests the training function entity to deploy the first model on a target device, which is different from the first device.

[0128] In a possible implementation method, for the case one described in step 202a, if the first device sends a test request to the training function entity and receives a model test result from the training function entity, the first device sends a request message to the training function entity in a case where the model test result meets a model test requirement, and the request message is used to request to deploy the first model to the target device or the first device.

[0129] In another possible implementation method, for the case two or the case three described in step 202a, if the first device sends a training request to the training function entity and receives a training completion message from the training function entity, the training completion message indicating that the training is completed, the first device can send a request message to the training function entity after receiving the training completion message, and the request message is used to request to deploy the first model to the target device or the first device.

[0130] In another possible implementation method, for the case two or the case three described in step 202a, if the first device sends a training request to the training function entity and receives a training completion message from the training function entity, the training completion message indicating that the training is completed, and the training result is included in the training completion message, the first device sends a request message to the training function entity in a case where the training result meets a training requirement, and the request message is used to request to deploy the first model to the target device or the first device. The training requirement can include, for example, a model performance requirement and / or a model training time requirement.

[0131] Based on the above scheme, the training function entity has completed the model and / or the model being trained is indicated by the model availability information, so that the first device can accurately determine whether there is a model meeting the model selection requirement in the model that has completed the training and / or the model being trained, and if there is, the model is selected from the model that has completed the training or the model being trained for deployment, without the need to request the training function entity to retrain a model, so that the probability of reusing the model can be improved, the repeated training of the model is reduced, and the computing resources for training the model are saved, and the use efficiency of the computing resources is improved.

[0132] FIG. 2(b) is a flow diagram of a communication method provided by an embodiment of the present application. The method includes the following steps:

[0133] In step 201b, the first device sends a training request to the first training function entity. Correspondingly, the first training function entity receives the training request.

[0134] The training request comprises an identification of the first model and a training requirement, and is used to request training of the first model.

[0135] The training requirement comprises a model performance requirement (e.g. model accuracy, etc.) and / or a model training time requirement.

[0136] The model performance requirement is used to indicate a performance requirement of the first device on the trained first model.

[0137] The model training time requirement is used to indicate a requirement of the first device on a training time of the first model, e.g. the model training time requirement is used to indicate a maximum training time, i.e. the first device expects to complete model training within the maximum training time.

[0138] At step 202b, the first training function entity sends a first message to the resource management function entity in a case where the remaining computing resources of the first training function entity do not meet the training requirement. Accordingly, the resource management function entity receives the first message.

[0139] In one possible implementation method, after receiving the training request, the first training function entity determines whether the first model exists in the models that have been completed by the first training function entity and the models that are being trained by the first training function entity. If not, the first training function entity continues to determine whether the remaining computing resources of the first training function entity meet the training requirement. If yes, the first device completes training of the first model based on the remaining computing resources. If not, the first device sends the first message to the resource management function entity.

[0140] In another possible implementation method, after receiving the training request, the first training function entity determines whether the remaining computing resources of the first training function entity meet the training requirement. If yes, the first device completes training of the first model based on the remaining computing resources. If not, the first device sends the first message to the resource management function entity.

[0141] The first message comprises state information of the first training function entity. The first message is used to feed back or indicate that the computing resources of the first training function entity are insufficient to the resource management function entity, so as to trigger the resource management function entity to allocate more computing resources to the first training function entity.

[0142] Exemplarily, the state information of the first training function entity comprises one or more of the following information:

[0143] (1) Position information of the first training function entity.

[0144] The location information can be address information of a virtual machine (VM) where software running on the first training function entity is located, or address information of a container where the software running on the first training function entity is located.

[0145] (2) Information of the computing resources allocated to the first training function entity.

[0146] For example, the information of the computing resources includes information such as the number and model of the GPU cards allocated.

[0147] (3) Location information of the computing resources allocated to the first training function entity.

[0148] For example, the location information of the computing resources includes location information (such as region, cabinet, etc.) of the computing resources in the computing resource data center.

[0149] (4) Computing resource requirement of the first training function entity.

[0150] For example, the computing resource requirement information includes 6 GPU cards, etc.

[0151] (5) Computing resource usage information of the first training function entity.

[0152] The computing resource usage information is used to indicate the usage of the computing resources on the first training function entity, that is, the proportion of the used computing resources in the allocated computing resources.

[0153] In step 203b, the resource management function entity sends a first request to the computing resource data center. Accordingly, the computing resource data center receives the first request.

[0154] The first request is used to request to allocate computing resources for the first training function entity.

[0155] In one possible implementation, after receiving the first message, the resource management function entity determines the amount of computing resources that need to be allocated to the first training function entity according to the computing resource requirement of the first training function entity and / or the computing resource usage information of the first training function entity, and then sends a first request to the computing resource data center, where the first request includes the amount of computing resources and is used to request the computing resource data center to allocate or increase computing resources to the first training function entity. Accordingly, the computing resource data center allocates the idle computing resources (i.e., the remaining computing resources) on the computing resource data center to the first training function entity according to the amount of computing resources. Optionally, the first request further includes at least one of the location information of the first training function entity, the information of the computing resources allocated to the first training function entity, or the location information of the computing resources allocated to the first training function entity, so that the computing resource data center can allocate the target computing resources to the first training function entity, where the target computing resources refer to the computing resources among the idle computing resources that have high-speed connection with the remaining computing resources of the first training function entity. Based on the method, the training speed and accuracy of the first model can be improved.

[0156] In another possible implementation, after receiving the first message, the resource management function entity determines the amount of computing resources that need to be allocated to the first training function entity according to the computing resource requirement of the first training function entity and / or the computing resource usage information of the first training function entity, and further determines at least one second training function entity according to at least one of the location information of the first training function entity, the information of the computing resources allocated to the first training function entity, or the location information of the computing resources allocated to the first training function entity, where the idle computing resources of the second training function entity have high-speed connection with the computing resources of the first training function entity. Then the resource management function entity sends a first request to the computing resource data center, where the first request includes the amount of computing resources and the information of the at least one second training function entity, and is used to request the computing resource data center to allocate the idle computing resources of the at least one second training function entity to the first training function entity. Accordingly, the computing resource data center allocates the target computing resources to the first training function entity, where the target computing resources refer to the computing resources among the idle computing resources of the at least one second training function entity that have high-speed connection with the remaining computing resources of the first training function entity. Based on the method, the training speed and accuracy of the first model can be improved.

[0157] In step 204b, the computing resource data center sends a first response to the resource management function entity. Accordingly, the resource management function entity receives the first response.

[0158] The first response is used to indicate a resource allocation result. The resource allocation result is allocation success or allocation failure.

[0159] In step 205b, the resource management function entity sends a second message to the first training function entity according to the first response. Correspondingly, the first training function entity receives the second message.

[0160] When the first response indicates that the resource allocation is successful, the resource management function entity sends a second message to the first training function entity, and the second message includes information of newly added computing resources for the first training function entity. The information of the newly added computing resources is used to indicate one or more of a location, a size, or a type of the newly added computing resources.

[0161] In step 206b, the first training function entity performs model training on the first model based on the newly added computing resources.

[0162] Since the computing resources are newly added for the first training function entity, the first training function entity can perform model training on the first model based on the newly added computing resources and the remaining computing resources, and complete the model training.

[0163] In one possible implementation method, after the model training is completed, the first training function entity can send a training completion message to the first device. After receiving the training completion message, the first device can request the first training function entity to deploy the trained first model to the first device or a target device.

[0164] In another possible implementation method, after the model training is completed, the first training function entity can send a training completion message to the first device, and the training completion message indicates that the training is completed. The training completion message includes a training result. In a case where the training result meets a training requirement, the first device can request the first training function entity to deploy the trained first model to the target device or the first device. The training requirement includes, for example, a model performance requirement and / or a model training time requirement.

[0165] Based on the above scheme, when the first device requests the first training function entity to train a model, if the computing resources for training the model on the first training function entity are insufficient, the first training function entity can request the resource management function entity to newly add computing resources for the first training function entity. Then, the first training function entity can perform model training based on the newly added computing resources and the remaining computing resources, so that the remaining computing resources on the first training function entity are used for training the model instead of being idle all the time, thereby improving the use efficiency of the computing resources. In addition, the first device does not need to request other training function entities to train the model again, thereby saving signaling overhead of the first device and improving the speed of the first device to obtain the trained model.

[0166] The embodiments of FIG. 2(a) and FIG. 2(b) are described in detail below with reference to the accompanying drawings. The following embodiments of FIG. 3 and FIG. 4 are specific examples of the embodiment of FIG. 2(a), and the following embodiment of FIG. 5 is a specific example of the embodiment of FIG. 2(b).

[0167] FIG. 3 is a flow diagram of a communication method according to an embodiment of the present application. The method comprises the following steps:

[0168] In step 301, the first device sends a first query message to the training function entity. Correspondingly, the training function entity receives the first query message.

[0169] The first query message is used to query model availability information.

[0170] For example, the first device can send the first query message to the training function entity based on the getMOIAttributes operation defined in 3GPP TS 28.532.

[0171] In step 302, the training function entity sends a second query message to the computing resource data center. Correspondingly, the computing resource data center receives the second query message.

[0172] The second query message is used to query the computing resource usage information of the training function entity. For example, if the computing resource is managed by the NFV MANO system in the computing resource data center, the training function entity sends the second query message to the NFV MANO system in this step.

[0173] In step 303, the computing resource data center sends a second query response to the training function entity. Correspondingly, the training function entity receives the second query response.

[0174] The second query response comprises the computing resource usage information of the training function entity.

[0175] In step 304, the training function entity sends a first query response to the first device. Correspondingly, the first device receives the first query response.

[0176] The first query response comprises the model availability information of the training function entity.

[0177] In a possible implementation method, the first query response further comprises information of the model supported by the training function entity. For example, the information of the model supported by the training function entity can comprise the identification of the model supported by the training function entity for training. For example, the model supported by the training function entity for training comprises model 1 to model 100, and the information of the model supported by the training function entity comprises the identification of model 1 to model 100.

[0178] Step 305, the first device determines the first model.

[0179] The first model is one of the models that have completed training, or one of the models that are being trained. And the first model meets the first model selection requirement. For specific implementation process of this step, please refer to the case one in the aforementioned step 202a.

[0180] Step 306, the first device sends a test request to the training function entity. Correspondingly, the training function entity receives the test request.

[0181] The test request contains the identification of the first model and the information of the test data set (for example, it can be the address information or identification information of the model test data set, etc.).

[0182] Step 307, the training function entity sends a test response to the first device. Correspondingly, the first device receives the test response.

[0183] The test response contains the model test result.

[0184] The training function entity judges whether the model test result meets the model test requirement. If it meets the model test requirement, the following step 308 is executed; if it does not meet the model test requirement, the following steps 309 to 310 are executed.

[0185] Step 308, the first device sends a request message to the training function entity. Correspondingly, the training function entity receives the request message.

[0186] The request message is used to request the address information of the first model. Correspondingly, the training function entity returns the address information of the first model to the first device, and the first device can subsequently obtain the related information of the first model from the address information and deploy the first model.

[0187] Or, the request message is used to request to deploy the first model to a target device. For example, the request message includes the identification or address of the target device, so that the training function entity deploys the first model to the target device according to the identification or address of the target device, and the target device can subsequently use the first model.

[0188] Step 309, the first device sends a training request to the training function entity. Correspondingly, the training function entity receives the training request.

[0189] The training request includes the identification of the first model and an incremental training indication, which is used to indicate that the first model is incrementally trained.

[0190] At step 310, the training function entity sends the training result to the first device. Accordingly, the first device receives the training result.

[0191] The first device determines whether the training result meets the training requirement, for example, whether the performance indicator of the first model meets the model performance requirement of the first device.

[0192] If the training requirement is met, the first device sends a request message to the training function entity to request the address information of the first model, or to request the deployment of the second model to the target device.

[0193] If the training requirement is not met, the first device can continue to request the training function entity to perform incremental training on the model or request a brand-new model training until the trained first model meets the training requirement.

[0194] Based on the above scheme, the model availability information indicates the model that has been trained and / or the model that is being trained by the training function entity, so that the first device can accurately determine whether there is a model that meets the model selection requirement in the model that has been trained and / or the model that is being trained, and if there is, the model is selected from the model that has been trained or the model that is being trained for deployment, without the need to request the training function entity to retrain a model, thereby improving the probability of model reuse, reducing the repeated training of the model, and further saving the computing resources for training the model and improving the use efficiency of the computing resources.

[0195] FIG. 4 is a flow diagram of a communication method according to an embodiment of the present application. The method includes the following steps:

[0196] Steps 401 to 404 are the same as steps 301 to 304 in the embodiment of FIG. 3.

[0197] At step 405, the first device determines the first model.

[0198] The first model is one of the models that can be trained by the remaining computing resources of the training function entity.

[0199] For example, the first device first determines whether there is a model that meets the first model selection requirement in the model that has been trained and the model that is being trained. If there is not, the first device further determines whether there is a model that meets the second model selection requirement in the models that can be trained by the remaining computing resources of the training function entity, and finally selects a model that meets the second model selection requirement from the models that can be trained by the remaining computing resources of the training function entity as the first model. For the specific implementation process of this step, please refer to the case two in the aforementioned step 202a.

[0200] Step 406, the first device sends a training request to the training function entity. Correspondingly, the training function entity receives the training request.

[0201] The training request comprises an identifier of the first model, and the training request is used to request training of the first model.

[0202] Step 407, the training function entity sends a training result to the first device. Correspondingly, the first device receives the training result.

[0203] The first device judges whether the training result meets the training requirement, for example, judges whether a performance index of the first model meets a model performance requirement of the first device.

[0204] If the training requirement is met, the first device sends a request message to the training function entity to request address information of the first model, or to request deployment of a second model to a target device.

[0205] If the training requirement is not met, the first device can continue to request the training function entity to perform incremental training of the model or to perform training of a new model until the trained first model meets the training requirement.

[0206] Based on the above scheme, the model availability information indicates the model that has been completed training, the model that is being trained, and the model that can be trained based on the remaining computing resources of the training function entity, so that the first device can accurately determine whether there is a model that meets the model selection requirement in the model that has been completed training and / or the model that is being trained, and if not, selects a model from the model that can be trained based on the remaining computing resources of the training function entity for deployment, so that the computing resources of the training function entity are fully utilized, and the use efficiency of the computing resources is improved. Moreover, the first device does not need to request training of the model from other training function entities, so that the signaling overhead of the first device can be saved, and the speed of the first device to obtain the trained model is improved.

[0207] FIG. 5 is a flow diagram of a communication method provided by an embodiment of the present application. The method comprises the following steps:

[0208] Step 501, the first device sends a first query message to the training function entity. Correspondingly, the training function entity receives the first query message.

[0209] The first query message is used to query model availability information and to query distributed training capability information of the first training function entity, and the distributed training capability information is used to indicate whether the first training function entity supports distributed training.

[0210] Exemplarily, the first device can send the first query message to the first training function entity based on a getMOIAttributes operation defined in 3GPP TS 28.532.

[0211] In step 502, the first training function entity sends a second query message to the computing resource data center. Correspondingly, the computing resource data center receives the second query message.

[0212] The second query message is used to query the computing resource usage information of the first training function entity

[0213] Exemplarily, if the computing resource is managed by an NFV MANO system in the computing resource data center, the first training function entity sends the second query message to the NFV MANO system in this step.

[0214] In step 503, the computing resource data center sends a second query response to the first training function entity. Correspondingly, the first training function entity receives the second query response.

[0215] The second query response includes the computing resource usage information of the first training function entity.

[0216] In step 504, the first training function entity sends a first query response to the first device. Correspondingly, the first device receives the first query response.

[0217] The first query response includes the model availability information and the distributed training capability information of the first training function entity. In this embodiment, the distributed training capability information is used to indicate that the first training function entity supports distributed training.

[0218] In a possible implementation method, the first query response further includes information of a model supported by the first training function entity. Exemplarily, the information of the model supported by the first training function entity can include an identification of a model trained by the first training function entity. For example, the model trained by the first training function entity includes model 1 to model 100, and the information of the model supported by the first training function entity includes the identifications of model 1 to model 100.

[0219] In step 505, the first device sends a training request to the first training function entity. Correspondingly, the first training function entity receives the training request.

[0220] The training request includes an identification of a first model and a training requirement, and the training request is used to request training of the first model.

[0221] The training requirement includes a model performance requirement and / or a model training time requirement.

[0222] The model performance requirement is used to indicate a performance requirement of the first device on the trained first model.

[0223] The model training time requirement is used to indicate a training time requirement of the first device on the first model, for example, the model training time requirement is used to indicate a maximum training time, i.e., the first device expects to complete the model training within the maximum training time.

[0224] In step 506, the first training function entity sends a first message to the resource management function entity. Accordingly, the resource management function entity receives the first message.

[0225] The first message includes the state information of the first training function entity.

[0226] The first training function entity determines that the computing resource of the first training function entity is insufficient according to the training requirement, and then executes the step 506.

[0227] Optionally, before the step 506, the resource management function entity sends a subscription notification to the first training function entity, to subscribe to the state information of the first training function entity.

[0228] In step 507, the resource management function entity sends a first request to the computing resource data center. Accordingly, the computing resource data center receives the first request.

[0229] In a possible implementation method, the resource management function entity determines that there is idle computing resource on the computing resource data center, and the idle computing resource can meet the model training requirement of the first training function entity, and then the request message is used to request the computing resource data center to allocate the idle computing resource to the first training function entity.

[0230] In another possible implementation method, the resource management function entity determines that at least one second training function entity other than the first training function entity has idle computing resource, and the idle computing resource can meet the model training requirement of the first training function entity, and then the request message is used to request the computing resource data center to allocate the idle computing resource of the at least one second training function entity to the first training function entity. For example, one or more training function entities (e.g., at least one second training function entity) other than the first training function entity can report state information to the resource management function entity based on the subscription of the resource management function entity, and the state information of the one or more training function entities is similar to the state information of the first training function entity, which can be referred to the foregoing description, so that the resource management function entity can determine which training function entity (e.g., at least one second training function entity) has idle computing resource to meet the model training requirement of the first training function entity based on the state information of the one or more training function entities.

[0231] At step 508, the computing resource data center allocates idle computing resources to the first training function entity.

[0232] The idle computing resources allocated by the computing resource data center to the first training function entity have high-speed connection with the remaining computing resources on the first training function entity, so that after the idle resources are allocated to the first training function entity, the first training function entity can efficiently use the idle resources and the remaining computing resources on the first training function entity in combination to improve the model training efficiency.

[0233] The idle computing resources allocated by the computing resource data center to the first training function entity can be idle computing resources on the computing resource data center, or idle computing resources on at least one second training function entity.

[0234] At step 509, the computing resource data center sends a first response to the resource management function entity. Correspondingly, the resource management function entity receives the first response.

[0235] The first response is used to indicate a resource allocation result. The resource allocation result is a success or a failure.

[0236] At step 510, the resource management function entity sends a second message to the first training function entity. Correspondingly, the first training function entity receives the second message.

[0237] When the first response indicates a successful resource allocation, the resource management function entity sends a second message to the first training function entity, and the second message includes information of the newly added computing resources for the first training function entity. The information of the newly added computing resources is used to indicate one or more of the location, size, or type of the newly added computing resources.

[0238] At step 511, the first training function entity performs model training.

[0239] Since the first training function entity has newly added computing resources, the first training function entity can perform model training on the first model based on the newly added computing resources and the remaining computing resources, and complete the model training.

[0240] At step 512, the first training function entity sends a training result to the first device. Correspondingly, the first device receives the training result.

[0241] After receiving the training result, if the training result indicates that the training is completed and the training result meets the training requirement, the first device can request the first training function entity to deploy the trained first model to the first device or a target device.

[0242] Based on the above scheme, when the first device requests a training model from the first training function entity, if the computing resources for training the model on the first training function entity are insufficient, the first training function entity can request the resource management function entity to add computing resources for the first training function entity, and then the first training function entity can train the model based on the added computing resources and the remaining computing resources, so that the remaining computing resources on the first training function entity are used for training the model instead of being idle all the time, thereby improving the use efficiency of the computing resources. Moreover, the first device does not need to request the training model from other training function entities, thereby saving the signaling overhead of the first device and improving the speed of the first device to obtain the trained model.

[0243] FIG. 6 shows a possible exemplary block diagram of a communication apparatus involved in the embodiments of the present application. As shown in FIG. 6, the communication apparatus 600 can include modules or units for implementing the above-mentioned method embodiments. In one possible design, the communication apparatus 600 includes a processing unit 602 and a communication unit 603. Optionally, the communication apparatus 600 can further include a storage unit 601 for storing apparatus program codes and / or data.

[0244] The communication apparatus 600 can also be the first device, a module (such as a circuit, a chip or a chip system, etc.) in the first device, or a logic node, a logic module or software capable of realizing all or part of the functions of the first device.

[0245] For example, in one embodiment, the processing unit 602 is configured to obtain model availability information of a training function entity, the model availability information including information of a model that has completed training and / or information of a model that is being trained; determine a first model according to the model availability information; and deploy the first model.

[0246] In one possible design, the processing unit 602 is configured to determine the first model according to the model availability information, specifically including: determining one model in the models that have completed training and that meets a first model selection requirement as the first model; or determining one model in the models that are being trained and that meets the first model selection requirement as the first model.

[0247] In a possible design, the model availability information includes information of the trained models and information of the training-in-progress models; the processing unit 602 is configured to determine the first model according to the model availability information, specifically including: determining, in a case where there is a model in the trained models that meets a first model selection requirement, the model in the trained models that meets the first model selection requirement as the first model; or, in a case where none of the trained models meets the first model selection requirement, determining a model in the training-in-progress models that meets the first model selection requirement as the first model.

[0248] In a possible design, the first model selection requirement includes an inference name and / or a performance indicator value.

[0249] In a possible design, the information of the trained models includes one or more of an identifier of the trained model, an inference name of the trained model, or a performance indicator value of the trained model; and the information of the training-in-progress models includes one or more of an identifier of the training-in-progress model, an inference name of the training-in-progress model, or an expected performance indicator value of the training-in-progress model.

[0250] In a possible design, the communication unit 603 is configured to send, to the training function entity, a test request including information of the first model and information of a test data set, where the test request is used to request testing of the first model based on the test data set; and receive a model test result from the training function entity.

[0251] In a possible design, the processing unit 602 is configured to deploy the first model, specifically including: in a case where the model test result meets a model test requirement, sending, to the training function entity via the communication unit 603, a request message used to request deployment of the first model to a target device.

[0252] In a possible design, the communication unit 603 is configured to, in a case where the model test result does not meet the model test requirement, send, to the training function entity, a training request including an incremental training indication used to indicate incremental training of the first model.

[0253] In a possible design, the model availability information further includes information about a model trained by the remaining computing resources of the training function entity; the processing unit 602 is configured to determine the first model according to the model availability information, specifically including: in the case where neither the trained model nor the training model meets the first model selection requirement, determining, as the first model, a model that meets the second model selection requirement from the models trained by the remaining computing resources of the training function entity.

[0254] In a possible design, the information about the models trained by the remaining computing resources of the training function entity includes one or more of an identifier of a model trained by the remaining computing resources of the training function entity, an inference name of a model trained by the remaining computing resources of the training function entity, or an expected training time of a model trained by the remaining computing resources of the training function entity.

[0255] In a possible design, the model availability information further includes information about a model trained by the remaining computing resources of the training function entity; the processing unit 602 is further configured to, in the case where neither the trained model nor the training model meets the first model selection requirement, and the models trained by the remaining computing resources of the training function entity do not meet the second model selection requirement, select a model from a model set as the first model, the model set including models supported by the training function entity excluding the trained model, the training model, and the models trained by the remaining computing resources of the training function entity.

[0256] In a possible design, the communication unit 603 is configured to send a training request to the training function entity, where the training request is used to request training of the first model, and the training request includes a training requirement, where the training requirement includes a model performance requirement and / or a model training time requirement.

[0257] In a possible design, the communication unit 603 is configured to send a training request to the training function entity, specifically including: in the case where the training function entity supports distributed training, sending the training request to the training function entity.

[0258] In a possible design, the first model selection requirement includes an inference name and / or a performance indicator value; and the second model selection requirement includes an inference name and / or training time information.

[0259] In a possible design, the processing unit 602, configured to acquire model availability information of a training function entity, specifically includes: a communication unit 603, configured to send a first query message to the training function entity, where the first query message is used to query the model availability information; and receive the model availability information from the training function entity.

[0260] The communication apparatus 600 can also be a training function entity in the above-described embodiments, a module (for example, a circuit, a chip, or a chip system, etc.) in the training function entity, or a logic node, a logic module, or software capable of implementing all or part of the functions of the training function entity.

[0261] For example, in an embodiment, the communication unit 603, configured to receive a first query message, where the first query message is used to query model availability information of a training function entity, and the model availability information includes information of a model that has completed training and / or information of a model that is being trained; and send the model availability information.

[0262] In a possible design, the information of the model that has completed training includes one or more of an identifier of the model that has completed training, an inference name of the model that has completed training, or a performance indicator value of the model that has completed training; and the information of the model that is being trained includes one or more of an identifier of the model that is being trained, an inference name of the model that is being trained, or an expected performance indicator value of the model that is being trained.

[0263] In a possible design, the model availability information further includes information of a model that can be trained by using remaining computing resources of the training function entity; and the information of the model that can be trained by using the remaining computing resources of the training function entity includes one or more of an identifier of the model that can be trained by using the remaining computing resources of the training function entity, an inference name of the model that can be trained by using the remaining computing resources of the training function entity, or an expected training time of the model that can be trained by using the remaining computing resources of the training function entity.

[0264] In a possible design, the communication unit 603 is further configured to send a second query message to a computing resource data center, where the second query message is used to query computing resource usage information of the training function entity; and receive the computing resource usage information of the training function entity from the computing resource data center; and the processing unit 602 is configured to determine information of a model that can be trained by using remaining computing resources of the training function entity according to the computing resource usage information.

[0265] The communication apparatus 600 can also be the first training function entity in the above embodiments, a module (such as a circuit, a chip or a chip system, etc.) in the first training function entity, or a logic node, a logic module or software capable of realizing all or part of the functions of the first training function entity.

[0266] For example, in an embodiment, the communication unit 603 is configured to receive a training request, the training request comprising an identification of a first model and a training requirement, the training request being used to request training of the first model; in a case where remaining computing resources of the first training function entity do not satisfy the training requirement, send a first message to a resource management function entity, the first message comprising state information of the first training function entity; receive a second message from the resource management function entity, the second message comprising information of newly added computing resources; and the processing unit 602 is configured to perform model training on the first model according to the newly added computing resources.

[0267] In a possible design, the state information comprises one or more of the following information: location information of the first training function entity, information of computing resources allocated to the first training function entity, location information of the computing resources allocated to the first training function entity, computing resource requirement of the first training function entity or computing resource usage information of the first training function entity.

[0268] In a possible design, the training requirement comprises a model performance requirement and / or a model training time requirement.

[0269] The communication apparatus 600 can also be the resource management function entity in the above embodiments, a module (such as a circuit, a chip or a chip system, etc.) in the resource management function entity, or a logic node, a logic module or software capable of realizing all or part of the functions of the resource management function entity.

[0270] For example, in an embodiment, the communication unit 603 is configured to receive a first message from a first training function entity, the first message comprising state information of the training function entity; send a first request to a computing resource data center, the first request being used to request allocation of computing resources to the first training function entity; receive a first response from the computing resource data center, the first response being used to indicate a resource allocation result; and the processing unit 602 is configured to send, according to the first response, a second message to the first training function entity through the communication unit 603, the second message comprising information of newly added computing resources.

[0271] In a possible design, the first request includes information of at least one second training function entity, and the first request is used to request to allocate idle computing resources of the at least one second training function entity to the first training function entity.

[0272] In a possible design, the state information includes one or more of the following: location information of the first training function entity, information of computing resources allocated to the first training function entity, location information of the computing resources allocated to the first training function entity, computing resource requirement of the first training function entity, or computing resource usage information of the first training function entity.

[0273] In a possible design, the processing unit 602 is further configured to determine the at least one second training function entity according to the location information of the first training function entity and / or the location information of the computing resources allocated to the first training function entity, where the idle computing resources of the second training function entity and the computing resources of the first training function entity have a high-speed connection.

[0274] It can be understood that the division of units in the above apparatus is merely a logical division of functions, and one function can correspond to one functional unit, or two or more functions can be integrated into one functional unit. In actual implementation, all or part of the units can be integrated into one physical entity, or can be distributed on different physical entities. In addition, the above functional units can be implemented in the form of hardware, or in the form of software, or in the form of combination of hardware and software. Whether a certain function is implemented in the form of hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for specific applications, but such implementation should not be considered beyond the scope of the present application.

[0275] In one example, the functional units in any of the apparatuses described above can be one or more integrated circuits configured to implement the above methods, for example, one or more application specific integrated circuits (ASICs), or one or more central processing units (CPUs), one or more microcontroller Units (MCUs), one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs), or a combination of at least two of these integrated circuit forms.

[0276] In one example, the storage unit 601 can include random access memory, flash memory, read only memory, programmable read only memory, or electrically erasable programmable memory, and / or registers, etc.

[0277] FIG. 7 shows a possible exemplary block diagram of a communication apparatus involved in the embodiments of the present application. The communication apparatus 700 shown in FIG. 7 includes a processor 710 and an interface circuit 720. The processor 710 and the interface circuit 720 are coupled to each other. It can be understood that the interface circuit 720 can be a transceiver or an input / output interface. Optionally, the communication apparatus 700 can further include a memory 730 for storing instructions executed by the processor 710 or storing input data required by the processor 710 to run instructions or storing data generated after the processor 710 runs instructions.

[0278] When the communication apparatus 700 is used to implement the method embodiments described above, the processor 710 is configured to implement the functions of the processing unit 602 described above, and the interface circuit 720 is configured to implement the functions of the communication unit 603 described above.

[0279] It can be understood that the processor in the embodiments of the present application can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor can be a microprocessor, or any conventional processor.

[0280] The present application provides a computer readable storage medium, which stores a computer program or instructions, when the computer program or instructions are executed, the method in any of the embodiments described above is implemented.

[0281] The present application provides a computer program product, which includes a computer program or instructions, when the computer program or instructions are executed, the method in any of the embodiments described above is implemented.

[0282] The present application provides a communication system, which includes a first device and a training function entity. The first device is configured to perform the operations performed by the first device in the embodiment of FIG. 2(a). The training function entity is configured to send model availability information of the training function entity to the first device.

[0283] The present application provides a communication system including a first device and a first training function entity. The first device is configured to perform the operations performed by the first device in the embodiment of Figure 2(b). The first training function entity is configured to perform the operations performed by the first training function entity in the embodiment of Figure 2(b).

[0284] The present application provides a communication system including a first device and a training function entity. The first device is configured to perform the operations performed by the first device in the embodiment of Figure 3. The training function entity is configured to perform the operations performed by the training function entity in the embodiment of Figure 3.

[0285] The present application provides a communication system including a first device and a training function entity. The first device is configured to perform the operations performed by the first device in the embodiment of Figure 4. The training function entity is configured to perform the operations performed by the training function entity in the embodiment of Figure 4.

[0286] The present application provides a communication system including a first device and a first training function entity. The first device is configured to perform the operations performed by the first device in the embodiment of Figure 5. The first training function entity is configured to perform the operations performed by the first training function entity in the embodiment of Figure 5.

[0287] The method steps in the embodiments of the present application can be implemented by hardware, or can be implemented by a combination of software and essential hardware. The software instructions can be stored in a memory, and can be read and executed by a processor of the terminal device. The memory can be a random access memory (RAM), a flash memory, a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a register, a removable disk, a compact disc read-only memory (CD-ROM) or any other form of storage medium known in the art. One exemplary storage medium is coupled to the processor, such that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can be a component of the processor. The processor and the storage medium can be located in an ASIC. In addition, the ASIC can be located in the first network element or the store-and-forward ground function network element. Of course, the processor and the storage medium can also exist as a discrete component in the access network device or the terminal device.

[0288] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer programs or instructions. The computer program refers to a set of instructions for instructing an electronic computer or other devices with message processing capability to perform each step. The computer program is usually written in a certain programming language and runs on a certain target architecture. When the computer program or instructions are loaded and executed on the computer, all or part of the processes or functions described in the embodiments of the present application are performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices. The computer program or instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another computer readable storage medium, for example, the computer program or instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired or wireless mode. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center and the like integrated with one or more available media. The available medium can be a magnetic medium, such as a floppy disk, a hard disk, a magnetic tape; or an optical medium, such as a digital video disc; or a semiconductor medium, such as a solid state disk. The computer readable storage medium can be a volatile or non-volatile storage medium, or can include both volatile and non-volatile storage media.

[0289] In various embodiments of the present application, the terms and / or descriptions of different embodiments are consistent and can be referred to each other if there is no special description and logical conflict. The technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0290] In the present application, "at least one" means one or more, and "multiple" means two or more. "And / or" describes the relationship between the associated objects, which means that there can be three kinds of relationships, for example, A and / or B, which can represent the following three cases: A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. In the literal description of the present application, the character " / ", generally represents that the associated objects before and after are in an "or" relationship; in the formula of the present application, the character " / ", represents that the associated objects before and after are in a "division" relationship.

[0291] It can be understood that various numerical numbers involved in the embodiments of the present application are only distinguished for convenience of description, and are not used to limit the scope of the embodiments of the present application. The size of the serial numbers of the above processes does not mean the order of execution, and the execution order of the processes should be determined according to their functions and inherent logic.

[0292] The terms "system" and "network" in the embodiments of the present application can be used interchangeably. "At least one" means one or more, and "multiple" means two or more. "And / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the cases of A alone, A and B together, and B alone, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the front and rear associated objects. "At least one of the following" or similar expressions means any combination of these items, including any combination of single item or multiple items. For example, "at least one of A, B or C" includes A, B, C, AB, AC, BC or ABC, and "at least one of A, B and C" can also be understood to include A, B, C, AB, AC, BC or ABC. In addition, unless otherwise specified, the ordinal numbers "first", "second" and the like mentioned in the embodiments of the present application are used to distinguish a plurality of objects, and are not used to limit the order, time sequence, priority or importance of the plurality of objects.

[0293] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer usable program code.

[0294] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more flows in the flowcharts and / or one or more blocks in the block diagrams.

[0295] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart or flowsheets and / or block or blocks of the block diagrams.

[0296] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flowsheets and / or block or blocks of the block diagrams.

[0297] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A communication method characterized by comprising: The method comprises: obtaining model availability information of a training function entity, the model availability information comprising information of a model that has completed training and / or information of a model that is being trained; determining a first model according to the model availability information; and deploying the first model.

2. The method of claim 1, wherein, The determining of the first model according to the model availability information comprises: determining one of the models that has completed training and that meets first model selection requirements as the first model; or determining one of the models that is being trained and that meets the first model selection requirements as the first model.

3. The method of claim 1, wherein, The model availability information comprises information of the models that have completed training and information of the models that are being trained. The determining of the first model according to the model availability information comprises: determining one of the models that has completed training and that meets first model selection requirements as the first model, in a case where there is a model that meets the first model selection requirements among the models that have completed training; or determining one of the models that is being trained and that meets the first model selection requirements as the first model, in a case where none of the models that have completed training meets the first model selection requirements.

4. The method of claim 2 or 3, wherein, The first model selection requirements comprise an inference name and / or a performance indicator value.

5. The method of any one of claims 1 to 4, wherein, The information of the models that have completed training comprises one or more of an identifier of the model that has completed training, an inference name of the model that has completed training, or a performance indicator value of the model that has completed training. The information of the models that are being trained comprises one or more of an identifier of the model that is being trained, an inference name of the model that is being trained, or an expected performance indicator value of the model that is being trained.

6. The method of any one of claims 1 to 5, wherein, The method further comprises: sending a test request to the training function entity, the test request comprising information of the first model and information of a test data set, the test request being used to request testing of the first model based on the test data set; and receiving a model test result from the training function entity.

7. The method of claim 6, wherein, The deploying of the first model comprises: in a case where the model test result meets model test requirements, sending a request message to the training function entity, the request message being used to request deployment of the first model to a target device.

8. The method of claim 6, wherein, The method further comprises: in a case where the model test result does not meet model test requirements, sending a training request to the training function entity, the training request comprising an incremental training indication, the incremental training indication being used to indicate incremental training of the first model.

9. The method of claim 1, wherein, The model availability information further comprises information of a model that can be trained using remaining computing resources of the training function entity; The determining of the first model according to the model availability information comprises: in a case where none of the models that have completed training and none of the models that are being trained meets first model selection requirements, determining one of the models that can be trained using remaining computing resources of the training function entity and that meets second model selection requirements as the first model.

10. The method of claim 9, wherein, The information of the model supported by the remaining computing resources of the training function entity includes one or more of an identifier of the model supported by the remaining computing resources of the training function entity, an inference name of the model supported by the remaining computing resources of the training function entity, or an expected training time of the model supported by the remaining computing resources of the training function entity.

11. The method of claim 1, wherein, The model availability information further includes information of the model supported by the remaining computing resources of the training function entity; The method further includes: In a case where neither the trained model nor the model being trained meets the first model selection requirement, and the model supported by the remaining computing resources of the training function entity does not meet the second model selection requirement, selecting a model from a model set as the first model, the model set including models supported by the training function entity excluding the trained model, the model being trained, and the model supported by the remaining computing resources of the training function entity.

12. The method of claim 11, wherein, Further includes: sending a training request to the training function entity, the training request being used to request training of the first model, the training request including a training requirement, the training requirement including a model performance requirement and / or a model training time requirement.

13. The method of claim 12, wherein, The sending of the training request to the training function entity includes: In a case where the training function entity supports distributed training, sending the training request to the training function entity.

14. The method of any one of claims 9 to 13, wherein, The first model selection requirement includes an inference name and / or a performance indicator value; The second model selection requirement includes an inference name and / or training time information.

15. The method of any one of claims 1 to 14, wherein, The obtaining of the model availability information of the training function entity includes: sending a first query message to the training function entity, the first query message being used to query the model availability information; receiving the model availability information from the training function entity.

16. A method of communication, comprising: The method includes: receiving a first query message, the first query message being used to query model availability information of a training function entity, the model availability information including information of a trained model and / or information of a model being trained; sending the model availability information.

17. The method of claim 16, wherein, The information of the trained model includes one or more of an identifier of the trained model, an inference name of the trained model, or a performance indicator value of the trained model; The information of the model being trained includes one or more of an identifier of the model being trained, an inference name of the model being trained, or an expected performance indicator value of the model being trained.

18. The method of claim 16 or 17, wherein, The model availability information further includes information of the model supported by the remaining computing resources of the training function entity; The information of the remaining computing resources of the training function entity supporting the trained model includes one or more of an identifier of the trained model supported by the remaining computing resources of the training function entity, an inference name of the trained model supported by the remaining computing resources of the training function entity, or an expected training time of the trained model supported by the remaining computing resources of the training function entity.

19. The method of claim 18, wherein, Further comprising: sending a second query message to a computing resource data center, the second query message being used to query the computing resource usage information of the training function entity; receiving the computing resource usage information of the training function entity from the computing resource data center; determining the information of the remaining computing resources of the training function entity supporting the trained model according to the computing resource usage information.

20. A method of communication, comprising: The method comprises: receiving a training request, the training request including an identifier of a first model and a training requirement, the training request being used to request training of the first model; in a case where the remaining computing resources of a first training function entity do not meet the training requirement, sending a first message to a resource management function entity, the first message including state information of the first training function entity; receiving a second message from the resource management function entity, the second message including information of newly added computing resources; performing model training on the first model according to the newly added computing resources.

21. The method of claim 20, wherein, The state information includes one or more of the following information: location information of the first training function entity, information of computing resources already allocated to the first training function entity, location information of the computing resources already allocated to the first training function entity, computing resource requirement of the first training function entity, or computing resource usage rate information of the first training function entity.

22. The method of claim 20 or 21, wherein, The training requirement includes a model performance requirement and / or a model training time requirement.

23. A method of communication, comprising: The method comprises: receiving a first message from a first training function entity, the first message including state information of the training function entity; sending a first request to a computing resource data center, the first request being used to request allocation of computing resources to the first training function entity; receiving a first response from the computing resource data center, the first response being used to indicate a resource allocation result; according to the first response, sending a second message to the first training function entity, the second message including information of newly added computing resources.

24. The method of claim 23, wherein, The first request includes information of at least one second training function entity, the first request being used to request allocation of idle computing resources of the at least one second training function entity to the first training function entity.

25. The method of claim 24, wherein, The state information includes one or more of the following information: location information of the first training function entity, information of computing resources already allocated to the first training function entity, location information of the computing resources already allocated to the first training function entity, computing resource requirement of the first training function entity, or computing resource usage rate information of the first training function entity.

26. The method of claim 25, wherein, Further comprising: determining the at least one second training function entity according to the location information of the first training function entity and / or the location information of the computing resource allocated to the first training function entity; wherein there is a high-speed connection between the idle computing resource of the second training function entity and the computing resource of the first training function entity.

27. A communications device, characterized by comprising a module for performing the method of any one of claims 1 to 15, or a module for performing the method of any one of claims 16 to 19, or a module for performing the method of any one of claims 20 to 22, or a module for performing the method of any one of claims 23 to 26.

28. A computer program product, characterised in that, The computer program product comprises a computer program or instructions, which, when executed, implement the method of any one of claims 1 to 15, or the method of any one of claims 16 to 19, or the method of any one of claims 20 to 22, or the method of any one of claims 23 to 26.

29. A computer-readable storage medium, characterized in that, The storage medium stores a computer program or instructions, which, when executed, implement the method of any one of claims 1 to 15, or the method of any one of claims 16 to 19, or the method of any one of claims 20 to 22, or the method of any one of claims 23 to 26.

30. A communication system, characterized by comprising a training function entity for implementing the method of any one of claims 20 to 22, and a resource management function entity for implementing the method of any one of claims 23 to 26.

Images