Model determination method and apparatus, device, computer storage medium, and computer program product
Patent Information
- Application Number
- PCT/CN2026/090747
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-04-18
- Filing Date
- 2026-04-15
- Publication Date
- 2026-10-22
Smart Images
Figure CN2026090747_22102026_PF_FP_ABST
Abstract
Description
A model determination method, apparatus, device, computer storage medium, and computer program product.
[0001] Cross-reference to related applications
[0002] This disclosure is based on and claims priority to Chinese Patent Application No. 202510497093.1, filed on April 18, 2025, the entire contents of which are incorporated herein by reference. Technical Field
[0003] This disclosure relates to model determination technology in the field of wireless technology, and includes, but is not limited to, a model determination method, apparatus, device, computer storage medium, and computer program product. Background Technology
[0004] Currently, federated learning's distributed training of local models is performed by the client, eliminating the need to transfer training datasets between the server and client, thus ensuring data security. Simultaneously, it effectively utilizes the computing resources of different network elements, reducing the computing resources required for centralized training methods like independent server training. Federated learning has already been introduced on the core network side. Artificial Intelligence (AI) model training and inference on the Radio Access Network (RAN) side are relatively independent; that is, if the AI model is trained / inferred by the RAN side, different RAN nodes train their deployed AI models independently or use their AI models for inference independently. For a split architecture, the next-generation Node B Centralized Unit (gNB-CU) and the next-generation Node B Distributed Unit (gNB-DU) are connected via the F1 interface. The gNB-CU handles higher-layer protocols, data forwarding, and network management functions, while the gNB-DU handles physical layer functions. In scenarios involving prediction using AI / Machine Language (ML), AI / ML training / inference is performed on the gNB-CU. Considering that the CU side is responsible for high-level protocol processing, data forwarding, network management, etc., in addition to AI training / inference, there is a problem that when the CU side has too many tasks, the computing power resources of the CU side cannot support the computing power required for AI model training. Summary of the Invention
[0005] To address the related technical problems, embodiments of this disclosure provide a model determination method, apparatus, device, computer storage medium, and computer program product.
[0006] The technical solution of this disclosure embodiment is implemented as follows:
[0007] A model determination method, the method being applied to a first node in a first base station, includes:
[0008] Obtain the training task for training the model and the required computing resources corresponding to the training task;
[0009] When the first node does not have training data for the training task, and it is determined that the first node cannot process the training task based on its idle computing resources and the required computing resources, a first request is sent to the second node in the first base station; wherein the first request includes the required computing resources.
[0010] Receive first feedback information from the second node in response to the first request; wherein the first feedback information is generated by the second node based on the required computing power resources and the second node's idle computing power resources;
[0011] Based on the first feedback information, a first post-trained model is determined, and based on the first post-trained model, a target model corresponding to the training task is determined.
[0012] In the above scheme, determining the first trained model based on the first feedback information includes:
[0013] If the first feedback information indicates that the second node can handle the training task, the training task is decomposed into multiple training sub-tasks based on the first feedback information and the training data; wherein, the first feedback information is generated by the second node based on the required computing power resources, the idle computing power resources of the second node, and the control strategy;
[0014] Send the training subtask corresponding to each of the second nodes to the second nodes;
[0015] The system receives a first model for the training subtask sent by each of the second nodes, and determines a first trained model based on multiple first models; wherein the first model is determined by the second node based on the training data.
[0016] In the above scheme, determining the first trained model based on the first feedback information includes:
[0017] If the first feedback information indicates that the first target node in the second node can handle the training task, the training task is split into multiple training sub-tasks based on the first feedback information and the training data.
[0018] A third request is sent to the first target node, and a second model for the first training sub-task is received from the first target node; wherein, the third request includes the first training sub-task and the initial model corresponding to the first target node;
[0019] A third model is determined for the second training subtask; wherein the training subtask includes the first training subtask and the second training subtask;
[0020] The first trained model is determined based on the second model and the third model.
[0021] In the above scheme, determining the third model for the second training sub-task includes:
[0022] A fourth request is sent to the second target node; wherein the fourth request includes the required data for the second training subtask and the initial model; the second target node is a node other than the first target node among the second nodes;
[0023] Receive training sub-data for the second training sub-task sent by the second target node; wherein the training sub-data is determined by the second target node based on the requirement data;
[0024] The third model is determined based on the training sub-data, the target training data, and the initial model.
[0025] In the above scheme, receiving the training sub-data for the second training sub-task sent by the second target node includes:
[0026] The training sub-data sent by the second target node within the response time is received; wherein, the fourth request also includes the response time;
[0027] Alternatively, receive the indication information sent by the second target node and extend the response time; wherein the indication information is used to indicate the reason for not returning the training sub-data;
[0028] The response time is updated based on the extended response time, and the training sub-data sent by the second target node within the updated response time is received; wherein, the fourth request also includes the response time.
[0029] In the above scheme, determining the third model based on the training sub-data, the target training data, and the initial model includes:
[0030] If the idle computing resources of the first node can handle the second training subtask, the initial model is trained based on the training sub-data and the target training data to obtain the third model;
[0031] If the idle computing resources of the first node cannot process the second training subtask, a fifth request is sent to the second base station; wherein, the fifth request includes the second training subtask, the initial model, the training sub-data, and the target training data;
[0032] The system receives the third model sent by the second base station for the second training sub-task; wherein the third model is obtained by the second base station after training the initial model based on the training sub-data and the target training data.
[0033] The method in the above scheme further includes:
[0034] If the idle computing resources of the first node cannot process the second training subtask, the second training subtask is decomposed into multiple decomposed training subtasks.
[0035] Based on the training sub-data and the target training data, the initial model is trained sequentially, and the third model is determined based on the multiple sub-models obtained from the training.
[0036] In the above scheme, determining the target model corresponding to the training task based on the first trained model includes:
[0037] If the first trained model meets the target requirements, the first trained model is determined to be the target model.
[0038] If the first trained model does not meet the target requirements, a sixth request is sent to the second node, and the second feedback information sent by the second node in response to the sixth request is received; wherein, the sixth request includes the required computing resources and the first trained model;
[0039] Based on the second feedback information, a second post-trained model is determined until the target model that meets the target requirements is obtained based on the determined Nth post-trained model.
[0040] The method in the above scheme further includes:
[0041] If the first model or the second model is updated, receive the updated model sent by the second node or the first target node;
[0042] The target model is updated based on the updated model.
[0043] A model determining apparatus, the apparatus comprising:
[0044] The acquisition section is configured to acquire the training task used to train the model and the required computing resources corresponding to the training task;
[0045] The sending part is configured to send a first request to a second node in the first base station when the first node does not have the training data for the training task and it is determined that the training task cannot be processed based on the idle computing resources of the first node and the required computing resources; wherein the first request includes the required computing resources;
[0046] The receiving part is configured to receive first feedback information sent by the second node in response to the first request; wherein the first feedback information is generated by the second node based on the required computing power resources and the second node's idle computing power resources;
[0047] The processing section is configured to determine a first post-trained model based on the first feedback information, and to determine the target model corresponding to the training task based on the first post-trained model.
[0048] A first node, the first node comprising: a processor, a memory, and a communication bus;
[0049] The communication bus is used to realize the communication connection between the processor and the memory;
[0050] The processor is used to execute the model determination program in the memory to implement the steps of the model determination method described above.
[0051] A computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the model determination method described above.
[0052] A computer program product comprising a computer program that, when executed by a processor, implements the above-described method.
[0053] The model determination method, apparatus, device, computer storage medium, and computer program product provided in the embodiments of this disclosure can obtain the training task for training the model and the computing power resources required for the training task. When the first node does not have the training data for the training task and it is determined that it cannot process the training task based on the idle computing power resources and required computing power resources of the first node, it sends a first request including the required computing power resources to the second node in the first base station, and receives the first feedback information sent by the second node in response to the first request. The first feedback information is generated by the second node based on the required computing power resources and the idle computing power resources of the second node. Based on the first feedback information, a first post-trained model is determined, and based on the first post-trained model, the target model corresponding to the training task is determined. In this way, when the CU side does not have enough computing power resources required for the training task, it can request the DU to assist in processing and completing the training task and generate the AI model corresponding to the training task, instead of the entire AI model training being completed by the CU as in related technologies. This greatly reduces the demand for computing power resources of the CU, thereby solving the problem that the computing power resources of the CU side cannot support the computing power required for AI model training when there are too many services on the CU side, and improving the generation efficiency of AI models. Attached Figure Description
[0054] The accompanying drawings, which are incorporated herein by reference and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.
[0055] Figure 1 is a schematic flowchart of a model determination method provided in an embodiment of this disclosure;
[0056] Figure 2 is a schematic diagram of an application scenario of a model determination method provided by an embodiment of this disclosure;
[0057] Figure 3 is a schematic diagram of the structure of a model determining device provided in an embodiment of this disclosure;
[0058] Figure 4 is a schematic diagram of the structure of a first node provided in an embodiment of this disclosure. Detailed Implementation
[0059] The technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.
[0060] It should be understood that the phrases "embodiments of this disclosure" or "foreign embodiments" throughout the specification mean that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this disclosure. Therefore, "embodiments of this disclosure" or "in the foreign embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the various embodiments of this disclosure, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this disclosure. The sequence numbers of the above-described embodiments are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0061] Unless otherwise specified, any step in the embodiments of this disclosure may be performed by the processor of the electronic device. It is also worth noting that the embodiments of this disclosure do not limit the order in which the electronic device performs the following steps. Furthermore, the methods used to process data in different embodiments may be the same or different methods. It should also be noted that any step in the embodiments of this disclosure can be performed independently by the electronic device; that is, when the electronic device performs any step in the following embodiments, it may not depend on the execution of other steps.
[0062] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this disclosure.
[0063] This disclosure provides a model determination method, which can be applied to a first node. Referring to FIG1, the method may include the following steps:
[0064] Step 101: Obtain the training task for training the model and the computing resources required for the training task.
[0065] In other embodiments of this disclosure, the training task may refer to a task used to train an AI model; in one feasible implementation, the AI model may refer to a model used for load prediction, traffic prediction, energy consumption prediction, etc.; the required computing resources may refer to information on the computing resources required to complete the training task. It should be noted that the first node may refer to the CU on the first base station side.
[0066] It should be noted that the training task can be requested by other base stations that do not have the ability to train AI models from the first base station, or it can be an AI model training task initiated by the CU on the first base station side, or it can be issued by the core network and instructed to the first base station to perform an AI model training task, or it can be an AI model training task requested by the user equipment (UE) from the first base station when it does not have sufficient computing power resources, or it can be a model training task initiated by the CU by the DU on the first base station side requesting model prediction.
[0067] Step 102: When the first node does not have training data for the training task, and it is determined that the first node cannot process the training task based on its idle computing resources and required computing resources, a first request is sent to the second node in the first base station.
[0068] The first request includes the computing resources required for the training task.
[0069] In this embodiment of the disclosure, the CU on the first base station side can determine whether its current idle computing resources can meet the computing resources required for the training task. If the idle computing resources currently available to the CU on the first base station side (i.e., gNB-CU) cannot meet the computing resources required for the training task, it means that the CU on the first base station side cannot process the training task. At this time, a first request is generated to request the second node on the first base station side to process the training task and sent to the second node.
[0070] In other embodiments of this disclosure, training data may refer to the data required to complete the training task; the second node may refer to the DU (i.e., gNB-DU) on the first base station side. It should be noted that the first request is used to instruct the DU to join the joint training / federated learning.
[0071] Step 103: Receive the first feedback information sent by the second node in response to the first request.
[0072] The first feedback information is generated by the second node based on the required computing power resources and the second node's idle computing power resources.
[0073] In this embodiment of the disclosure, after receiving the first request, the second node can generate first feedback information for the training task based on its current idle computing resources and the computing resources required for the training task, and send the first feedback information to the DU on the first base station side.
[0074] It should be noted that the first feedback information can be a characterization of the DU's ability to process training tasks on the first base station side.
[0075] Step 104: Determine the first post-trained model based on the first feedback information, and determine the target model corresponding to the training task based on the first post-trained model.
[0076] In other embodiments of this disclosure, the CU on the first base station side can determine the node that has completed the training task based on the first feedback information and receive the model sent by the node; then, based on the sent model, determine the first post-trained model corresponding to the training task.
[0077] It should be noted that the CU on the first base station side can determine the final target model of the training task based on the relationship between the determined first post-trained model and the target requirements.
[0078] Based on the foregoing embodiments, in other embodiments of this disclosure, the step 104 above, "determining the first trained model based on the first feedback information," can be implemented in the following ways:
[0079] A1. If the first feedback information indicates that the second node can handle the training task, the training task is decomposed into multiple training sub-tasks based on the first feedback information and training data.
[0080] The first feedback information is generated by the second node based on the required computing resources, the second node's idle computing resources, and the control strategy. It should be noted that the control strategy can refer to the DU's local control strategy for managing computing resources. After receiving the first request, the second node can generate the first feedback information for the training task based on its current idle computing resources, the computing resources required for the training task, and its local control strategy.
[0081] In this embodiment of the disclosure, if the first feedback information indicates that each DU on the first base station side can handle the training task, then as shown in FIG2, the training task can be split according to the number of DUs added to this joint training, the idle computing resources of each DU and the training data, thereby obtaining multiple training sub-tasks.
[0082] A2. Send a second request to the second node. The second request includes the training subtask and initial model corresponding to each second node.
[0083] In this embodiment of the disclosure, as shown in Figure 2, the CU on the first base station side can carry the training sub-tasks and initial models that each gNB-DU needs to complete after splitting in a second request and send them to the corresponding gNB-DU. It should be noted that the second request may also include the required data corresponding to the training sub-tasks and a first maximum feedback time. The first maximum feedback time may refer to the pre-set longest waiting time for obtaining the first model, that is, the latest time that the DU is allowed to report the first model; the required data may refer to the data required by the training sub-tasks.
[0084] A3. Receive the first model for the training subtask sent by each second node, and determine the first post-trained model based on multiple first models.
[0085] The first model is determined by the second node based on the training data.
[0086] In this embodiment, after receiving their respective training subtasks, if a pre-trained local model identical to the training subtask exists in the gNB-DU (indicating that the gNB-DU has previously trained the model for the training subtask and the initial model is consistent), the gNB-DU reuses this pre-trained model and updates it based on the training data to obtain the first model. If no pre-trained local model identical to the training subtask exists in the gNB-DU, the gNB-DU uses the training data to train the initial model to obtain the first model. Furthermore, as shown in Figure 2, each gNB-DU can send the trained first model to the gNB-CU.
[0087] It should be noted that after receiving multiple first models from gNB-DU, gNB-CU can aggregate the multiple first models to obtain the first trained model.
[0088] In other embodiments of this disclosure, the step 104 above, "determining the first trained model based on the first feedback information," can also be implemented in the following ways:
[0089] B1. If the first feedback information indicates that the first target node in the second node can handle the training task, the training task is split into multiple training sub-tasks based on the first feedback information and training data.
[0090] In this embodiment of the disclosure, if the first feedback information indicates that the first target DU in the DU on the first base station side can handle the training task, then in response to the first feedback information, the training task can be split according to the number of DUs added to this joint training, the idle computing resources of the DUs added to the joint training, and the training data, thereby obtaining multiple training sub-tasks. Here, the first target node can refer to the first target DU in the gNB-DU that can handle the training task.
[0091] B2. Send a third request to the first target node and receive the second model for the first training subtask from the first target node. The third request includes the first training subtask and the initial model corresponding to the first target node.
[0092] It should be noted that the third request may also include the required data corresponding to the first training subtask and the second maximum feedback time. The second maximum feedback time may refer to the pre-set longest waiting time for obtaining the second model, that is, the latest time that the first target DU is allowed to report the second model.
[0093] As shown in Figure 2, the CU on the first base station side can send the first training sub-task that each first target DU needs to complete in the split training sub-tasks to the corresponding first target DU. Furthermore, after receiving their respective first training sub-tasks, the first target DUs can determine the second model corresponding to the first training sub-task based on the training data. It should be noted that the first target DUs can determine the second model using the same method as determining the first model.
[0094] B3. Determine the third model for the second training subtask; wherein the training subtask includes the first training subtask and the second training subtask.
[0095] In this embodiment of the disclosure, the third model may be the model corresponding to the second training subtask in the split training subtasks. It should be noted that the third model may be obtained by training the model based on the training sub-data sent by the second target DU (excluding the first target DU) in gNB-DU.
[0096] B4. Determine the first post-trained model based on the second and third models.
[0097] In this embodiment of the disclosure, gNB-CU can aggregate the determined second model and third model to obtain the first post-trained model.
[0098] In other embodiments of this disclosure, determining the third model for the second training subtask includes:
[0099] Send a fourth request to the second target node; wherein the fourth request includes the required data and initial model for the second training subtask; the second target node is a node other than the first target node among the second nodes.
[0100] In this embodiment of the disclosure, the required data may include UE website traffic and physical layer measurement information; specifically, the measurement information may include information required for CU training models, such as Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ), and Signal-to-Interference Plus Noise Ratio (SINR). It should be noted that the specific information in the required data may vary depending on the functionality of the AI model.
[0101] Receive training sub-data for the second training sub-task sent by the second target node; wherein the training sub-data is determined by the second target node based on the requirement data.
[0102] In this embodiment of the disclosure, the CU on the first base station side can determine the training sub-data for the second training sub-task based on the targeted feedback information received from the second target DU. Furthermore, the training sub-data is obtained by the second target DU based on the received requirement data and then sent to the gNB-CU.
[0103] Based on the training sub-data, target training data, and initial model, the third model is determined.
[0104] It should be noted that the gNB-CU can determine whether to handle the second training subtask itself or request the second base station to handle it, based on its current available resources. In other words, the second training subtask can be determined by the gNB-CU based on the training sub-data, target training data, and initial model, or it can be determined by the second base station based on the training sub-data, target training data, and initial model. The target training data can be the data stored locally by the gNB-CU that is needed to train the initial model.
[0105] In other embodiments of this disclosure, receiving the training sub-data for the second training sub-task sent by the second target node includes:
[0106] Receive training sub-data sent by the second target node within the response time.
[0107] The fourth request also includes a response time; the response time can refer to the maximum waiting time that can be set in advance to obtain training sub-data, that is, the latest time that the second target DU is allowed to report training sub-data.
[0108] If the second target DU acquires the training sub-data within the response time and sends the training sub-data to the gNB-CU, then the gNB-CU determines that the training sub-data received at this moment is the final training sub-data.
[0109] Alternatively, receive the indication information sent by the second target node and extend the response time; wherein, the indication information is used to indicate the reason for not returning the training sub-data;
[0110] The response time is updated based on the extended response time, and the training sub-data sent by the second target node within the updated response time is received. The fourth request also includes the response time.
[0111] If the second target DU fails to acquire training sub-data within the response time and sends an indication message and requests an extension of the response time to the gNB-CU, then if the gNB-CU determines that the response time should be extended, the response time can be extended accordingly to obtain an updated response time. If the second target DU acquires training sub-data within the updated response time and sends the training sub-data to the gNB-CU, then the gNB-CU determines that the training sub-data is the final training sub-data.
[0112] It should be noted that the indication information can be a reason value indicating that the second target DU has not returned training sub-data. Furthermore, each data requested by the gNB-CU corresponds to a reason value. In addition, the indication information may also include the time point at which the training sub-data becomes available, or the target time remaining until the response time is reached; thus, the gNB-CU can assess whether to extend the response time to wait for the training sub-data to be acquired based on this time point or target time.
[0113] In other embodiments of this disclosure, determining the third model based on training sub-data, target training data, and the initial model includes:
[0114] If the idle computing resources of the first node can handle the second training sub-task, the initial model is trained based on the training sub-data and the target training data to obtain the third model.
[0115] In this embodiment of the disclosure, if the current idle computing resources of gNB-CU are sufficient to process the second training sub-task, then gNB-CU can use the training sub-data and the target training data to train the initial model and obtain the third model.
[0116] If the idle computing resources of the first node cannot process the second training subtask, a fifth request is sent to the second base station; wherein, the fifth request includes the second training subtask, the initial model, the training sub-data and the target training data;
[0117] Receive the third model sent by the second base station for the second training sub-task; wherein the third model is obtained by the second base station after training the initial model based on the training sub-data and the target training data.
[0118] In this embodiment, if the current idle computing resources of the gNB-CU are insufficient to handle the second training subtask, indicating that the gNB-CU cannot handle the second training subtask, then the gNB-CU can generate a fifth request to request the second base station to assist in handling the second training subtask, and send the fifth request to the second base station. Subsequently, the second base station will use the training sub-data and target training data to train the initial model to obtain a third model, and send the third model to the gNB-CU. Of course, the fifth request may also include a third maximum feedback time. The third maximum feedback time can refer to a pre-set maximum waiting time for obtaining the third model, that is, the latest time allowed for the second base station to report the third model.
[0119] It should be noted that when the training sub-data and target training data do not involve security / privacy issues, a second base station can be requested to assist in processing the second training sub-task. The second base station can be a base station adjacent to the first base station.
[0120] In other embodiments of this disclosure, the method further includes:
[0121] If the idle computing resources of the first node cannot process the second training subtask, the second training subtask is decomposed into multiple decomposed training subtasks.
[0122] The initial model is trained sequentially based on the training sub-data and the target training data, and the third model is determined based on the multiple sub-models obtained from the training.
[0123] It should be noted that if the current idle computing resources of gNB-CU are insufficient to process the second training subtask, it means that gNB-CU cannot process the second training subtask. In this case, if the training sub-data and the target training data involve security / privacy issues, then gNB-CU can split the second training subtask and process the multiple decomposed training subtasks in a time-sharing manner.
[0124] In other embodiments of this disclosure, the initial model can be trained sequentially using training sub-data and target training data at different times to obtain multiple sub-models; then, the multiple sub-models can be aggregated to obtain a third model.
[0125] In one feasible implementation, the second training subtask can be split into three decomposed training subtasks: training the first decomposed training subtask 1 between time T1 and T2 to obtain the first sub-model; training the second decomposed training subtask 2 between T2 and T3 to obtain the second sub-model; and training the third decomposed training subtask 3 between T3 and T4 to obtain the third sub-model. After completing all the decomposed training subtasks, the first, second, and third sub-models are aggregated to obtain the third model.
[0126] In other embodiments of this disclosure, the above-mentioned determination of the target model corresponding to the training task based on the first trained model includes:
[0127] If the first trained model meets the target requirements, then the first trained model is determined as the target model.
[0128] Here, the target requirement can refer to the trainer's requirements for the model, or pre-configured conditions (such as the accuracy the model needs to achieve). Furthermore, if the first trained model meets the target requirement, it means that the first trained model has reached its optimal state, and this first trained model is determined as the target model.
[0129] If the first trained model does not meet the target requirements, a sixth request is sent to the second node, and the second feedback information sent by the second node in response to the sixth request is received; wherein, the sixth request includes the required computing resources and the first trained model;
[0130] The second post-trained model is determined based on the second feedback information, until the target model that meets the target requirements is obtained based on the determined Nth post-trained model.
[0131] In this embodiment of the disclosure, if the first post-trained model does not meet the target requirements, it means that the first post-trained model has not yet reached its optimal state. At this time, it is necessary to iteratively execute the process of training the first post-trained model, that is, send a sixth request to gNB-DU to request gNB-DU to complete the training task, and continue to train the first post-trained model to determine the second post-trained model according to the method for determining the first post-trained model. If the second post-trained model does not meet the target requirements, the method for determining the first post-trained model is continued to determine the third post-trained model, until the Nth post-trained model that meets the target requirements is determined, and the Nth post-trained model is determined as the final target model.
[0132] It should be noted that once gNB-CU has determined the target model, it can send the target model to the requesting party.
[0133] In other embodiments of this disclosure, the method further includes:
[0134] If the first model or the second model is updated, receive the updated model sent by the second node or the first target node;
[0135] Update the target model based on the updated model.
[0136] In this embodiment, if a gNB-DU performing local model training discovers that received real-time data (i.e., data used for model training) affects model performance (e.g., a new UE moves into the coverage area of this gNB), the data may not have a significant impact on the target model and will not trigger an update of the target model. However, considering the correlation between the gNB-DU's local model and the target model, as well as UE mobility, this data may subsequently affect the target model on the gNB-CU side. To avoid model performance degradation, when the first model or the second model is updated, the second base station or the first target DU sends the updated model to the gNB-CU, and the gNB-CU updates the target model based on the updated model.
[0137] It should be noted that gNB-CU splits training tasks, distributing AI models across gNB-CU and gNB-DU. gNB-CU then aggregates / summarizes the AI models to generate the final AI model, reducing the transmission of training data between gNB-CU and gNB-DU. Furthermore, when the computing resources required for the split training sub-tasks on the gNB-CU side still exceed available computing resources, neighboring base stations can be requested to join the joint training task, or training can be performed using time-sharing methods, further reducing the computing resources required for model training on the gNB-CU side.
[0138] The model determination method provided in the embodiments of this disclosure can request the DU to assist in completing the training task and generate the AI model corresponding to the training task when the CU side does not have sufficient computing power resources required for the training task. Instead of the CU completing the training of the entire AI model as in related technologies, this greatly reduces the demand for computing power resources of the CU. This solves the problem that the computing power resources of the CU side cannot support the computing power required for AI model training when there are too many services on the CU side, and improves the generation efficiency of AI models.
[0139] Based on the foregoing embodiments, the present disclosure provides a model determination apparatus that can be applied to the model determination method provided in the embodiment corresponding to FIG1. Referring to FIG3, the model determination apparatus 2 may include: an acquisition part 21, a transmission part 22, a receiving part 23, and a processing part 24, wherein:
[0140] The acquisition section 21 is configured to acquire the training task used to train the model and the computing resources required for the training task;
[0141] The sending part 22 is configured to send a first request to the second node in the first base station when the first node does not have training data for the training task and it is determined that the first node cannot process the training task based on its idle computing resources and required computing resources; wherein the first request includes the required computing resources.
[0142] The receiving part 23 is configured to receive first feedback information sent by the second node in response to the first request; wherein the first feedback information is generated by the second node based on the required computing power resources and the idle computing power resources of the second node;
[0143] The processing unit 24 is configured to determine the first post-trained model based on the first feedback information, and to determine the target model corresponding to the training task based on the first post-trained model.
[0144] In other embodiments of this disclosure, the processing section 24 is further configured to perform the following steps:
[0145] If the first feedback information indicates that the second node can handle the training task, the training task is decomposed into multiple training sub-tasks based on the first feedback information and training data; wherein, the first feedback information is generated by the second node based on the required computing power resources, the idle computing power resources of the second node, and the control strategy.
[0146] Send a second request to the second node; wherein the second request includes the training subtask and the initial model corresponding to each second node;
[0147] Receive the first model for the training subtask sent by each second node, and determine the first trained model based on multiple first models; wherein the first model is determined by the second node based on the training data.
[0148] In other embodiments of this disclosure, the processing section 24 is further configured to perform the following steps:
[0149] If the first feedback information indicates that the first target node in the second node can handle the training task, the training task is split into multiple training sub-tasks based on the first feedback information and training data.
[0150] Send a third request to the first target node and receive a second model for the first training subtask from the first target node; wherein the third request includes the first training subtask and the initial model corresponding to the first target node;
[0151] Determine a third model for the second training subtask; wherein the training subtask includes the first training subtask and the second training subtask.
[0152] The first post-trained model is determined based on the second and third models.
[0153] In other embodiments of this disclosure, the processing section 24 is further configured to perform the following steps:
[0154] Send a fourth request to the second target node; wherein the fourth request includes the required data and initial model for the second training subtask; the second target node is a node in the second node group other than the first target node;
[0155] Receive training sub-data for the second training sub-task sent by the second target node; wherein the training sub-data is determined by the second target node based on the requirement data;
[0156] Based on the training sub-data, target training data, and initial model, the third model is determined.
[0157] In other embodiments of this disclosure, the processing section 24 is further configured to perform the following steps:
[0158] Receive the training sub-data sent by the second target node within the response time; the fourth request also includes the response time.
[0159] Alternatively, receive the indication information sent by the second target node and extend the response time; wherein, the indication information is used to indicate the reason for not returning the training sub-data;
[0160] The response time is updated based on the extended response time, and the training sub-data sent by the second target node within the updated response time is received; the fourth request also includes the response time.
[0161] In other embodiments of this disclosure, the processing section 24 is further configured to perform the following steps:
[0162] If the idle computing resources of the first node can handle the second training subtask, the initial model is trained based on the training sub-data and the target training data to obtain the third model;
[0163] If the idle computing resources of the first node cannot process the second training subtask, a fifth request is sent to the second base station; wherein, the fifth request includes the second training subtask, the initial model, the training sub-data and the target training data;
[0164] Receive the third model sent by the second base station for the second training sub-task; wherein the third model is obtained by the second base station after training the initial model based on the training sub-data and the target training data.
[0165] In other embodiments of this disclosure, the processing section 24 is further configured to perform the following steps:
[0166] If the idle computing resources of the first node cannot process the second training subtask, the second training subtask is decomposed into multiple decomposed training subtasks.
[0167] The initial model is trained sequentially based on the training sub-data and the target training data, and the third model is determined based on the multiple sub-models obtained from the training.
[0168] In other embodiments of this disclosure, the processing section 24 is further configured to perform the following steps:
[0169] If the first trained model meets the target requirements, then the first trained model is determined as the target model.
[0170] If the first trained model does not meet the target requirements, a sixth request is sent to the second node, and the second feedback information sent by the second node in response to the sixth request is received; wherein, the sixth request includes the required computing resources and the first trained model;
[0171] The second post-trained model is determined based on the second feedback information, until the target model that meets the target requirements is obtained based on the determined Nth post-trained model.
[0172] In other embodiments of this disclosure, the processing section 24 is further configured to perform the following steps:
[0173] If the first model or the second model is updated, receive the updated model sent by the second node or the first target node;
[0174] Update the target model based on the updated model.
[0175] It should be noted that the specific details of the steps performed in each part can be found in the model determination method provided in the corresponding embodiment of Figure 1, and will not be repeated here.
[0176] The model determination apparatus provided in the embodiments of this disclosure can request the DU to assist in completing the training task and generate the AI model corresponding to the training task when the CU side does not have sufficient computing power resources required for the training task. Instead of the CU completing the training of the entire AI model as in related technologies, this greatly reduces the demand for computing power resources of the CU. This solves the problem that the computing power resources of the CU side cannot support the computing power required for AI model training when there are too many services on the CU side, and improves the generation efficiency of AI models.
[0177] Based on the foregoing embodiments, the embodiments of this disclosure provide a first node, which can be applied to the model determination method provided in the embodiment corresponding to FIG1. Referring to FIG4, the first node 3 may include: a processor 31, a memory 32, and a communication bus 33, wherein:
[0178] Communication bus 33 is used to realize the communication connection between processor 31 and memory 32;
[0179] The processor 31 is used to execute the model determination program in the memory 32 to perform the following steps:
[0180] Obtain the training tasks for training the model and the corresponding computing resources required for the training tasks;
[0181] When the first node does not have training data for the training task, and it is determined that the first node cannot process the training task based on its idle computing resources and required computing resources, a first request is sent to the second node in the first base station; wherein the first request includes the required computing resources.
[0182] Receive first feedback information from the second node in response to the first request; wherein the first feedback information is generated by the second node based on the required computing power resources and the second node's idle computing power resources;
[0183] The first post-trained model is determined based on the first feedback information, and the target model corresponding to the training task is determined based on the first post-trained model.
[0184] In other embodiments of this disclosure, processor 31 is configured to execute a model determination program in memory 32 to determine a first trained model based on first feedback information, in order to implement the following steps:
[0185] If the first feedback information indicates that the second node can handle the training task, the training task is decomposed into multiple training sub-tasks based on the first feedback information and training data; wherein, the first feedback information is generated by the second node based on the required computing power resources, the idle computing power resources of the second node, and the control strategy.
[0186] Send a second request to the second node; wherein the second request includes the training subtask and the initial model corresponding to each second node;
[0187] Receive the first model for the training subtask sent by each second node, and determine the first trained model based on multiple first models; wherein the first model is determined by the second node based on the training data.
[0188] In other embodiments of this disclosure, processor 31 is configured to execute a model determination program in memory 32 to determine a first trained model based on first feedback information, in order to implement the following steps:
[0189] If the first feedback information indicates that the first target node in the second node can handle the training task, the training task is split into multiple training sub-tasks based on the first feedback information and training data.
[0190] Send a third request to the first target node and receive a second model for the first training subtask from the first target node; wherein the third request includes the first training subtask and the initial model corresponding to the first target node;
[0191] Determine a third model for the second training subtask; wherein the training subtask includes the first training subtask and the second training subtask.
[0192] The first post-trained model is determined based on the second and third models.
[0193] In other embodiments of this disclosure, processor 31 is configured to execute the model determination program in memory 32 to determine a third model for the second training subtask, in order to perform the following steps:
[0194] Send a fourth request to the second target node; wherein the fourth request includes the required data and initial model for the second training subtask; the second target node is a node in the second node group other than the first target node;
[0195] Receive training sub-data for the second training sub-task sent by the second target node; wherein the training sub-data is determined by the second target node based on the requirement data;
[0196] Based on the training sub-data, target training data, and initial model, the third model is determined.
[0197] In other embodiments of this disclosure, processor 31 is configured to execute the model determination program in memory 32 to receive training sub-data sent by the second target node for the second training sub-task, in order to perform the following steps:
[0198] Receive the training sub-data sent by the second target node within the response time; the fourth request also includes the response time.
[0199] Alternatively, receive the indication information sent by the second target node and extend the response time; wherein, the indication information is used to indicate the reason for not returning the training sub-data;
[0200] The response time is updated based on the extended response time, and the training sub-data sent by the second target node within the updated response time is received; the fourth request also includes the response time.
[0201] In other embodiments of this disclosure, processor 31 is configured to execute a model determination program in memory 32 to determine a third model based on training sub-data, target training data, and an initial model, to perform the following steps:
[0202] If the idle computing resources of the first node can handle the second training subtask, the initial model is trained based on the training sub-data and the target training data to obtain the third model;
[0203] If the idle computing resources of the first node cannot process the second training subtask, a fifth request is sent to the second base station; wherein, the fifth request includes the second training subtask, the initial model, the training sub-data and the target training data;
[0204] Receive the third model sent by the second base station for the second training sub-task; wherein the third model is obtained by the second base station after training the initial model based on the training sub-data and the target training data.
[0205] In other embodiments of this disclosure, the processor 31 is used to execute the model determination program in the memory 32, and may also perform the following steps:
[0206] If the idle computing resources of the first node cannot process the second training subtask, the second training subtask is decomposed into multiple decomposed training subtasks.
[0207] The initial model is trained sequentially based on the training sub-data and the target training data, and the third model is determined based on the multiple sub-models obtained from the training.
[0208] In other embodiments of this disclosure, the processor 31 is configured to execute the model determination program in the memory 32 to determine the target model corresponding to the first post-trained model training task, in order to implement the following steps:
[0209] If the first trained model meets the target requirements, then the first trained model is determined as the target model.
[0210] If the first trained model does not meet the target requirements, a sixth request is sent to the second node, and the second feedback information sent by the second node in response to the sixth request is received; wherein, the sixth request includes the required computing resources and the first trained model;
[0211] The second post-trained model is determined based on the second feedback information, until the target model that meets the target requirements is obtained based on the determined Nth post-trained model.
[0212] In other embodiments of this disclosure, the processor 31 is used to execute the model determination program in the memory 32, and may also perform the following steps:
[0213] If the first model or the second model is updated, receive the updated model sent by the second node or the first target node;
[0214] Update the target model based on the updated model.
[0215] It should be noted that the specific details of the steps performed by the processor can be found in the model determination method provided in the embodiment corresponding to Figure 1, and will not be repeated here.
[0216] The first node provided in the embodiments of this disclosure can request the DU to assist in completing the training task and generate the AI model corresponding to the training task when the CU side does not have sufficient computing power resources required for the training task. Instead of the CU completing the training of the entire AI model as in related technologies, this greatly reduces the demand for computing power resources of the CU. This solves the problem that the computing power resources of the CU side cannot support the computing power required for AI model training when there are too many services on the CU side, and improves the generation efficiency of AI models.
[0217] Based on the foregoing embodiments, embodiments of this disclosure provide a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the model determination method provided in the above embodiments.
[0218] Based on the foregoing embodiments, embodiments of this disclosure provide a computer program product, including a computer program that can be executed by a processor 31 to complete the steps of the model determination method provided in the above embodiments.
[0219] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0220] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more flowchart illustrations and / or one or more block diagrams.
[0221] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.
[0222] These computer program instructions may 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 one or more flowcharts and / or one or more block diagrams.
[0223] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.
Claims
1. A model determination method, applied to a first node in a first base station, comprising: obtaining a training task for training a model and required computing resource corresponding to the training task; when the first node does not have training data of the training task and it is determined that the first node cannot handle the training task based on idle computing resource of the first node and the required computing resource, sending a first request to a second node in the first base station; wherein the required computing resource is included in the first request; receiving first feedback information sent by the second node for the first request; wherein the first feedback information is generated by the second node based on the required computing resource and idle computing resource of the second node; determining a first trained model based on the first feedback information, and determining a target model corresponding to the training task based on the first trained model.
2. The method of claim 1, wherein, The determination of the first trained model based on the first feedback information comprises: if the first feedback information indicates that the second node can handle the training task, decomposing the training task based on the first feedback information and the training data to obtain a plurality of training sub-tasks; wherein the first feedback information is generated by the second node based on the required computing resource, idle computing resource of the second node and a control strategy; sending a second request to the second node; wherein the second request includes a training sub-task corresponding to each of the second node and an initial model; receiving a first model sent by each of the second node for the training sub-task, and determining the first trained model based on a plurality of the first models; wherein the first model is determined by the second node based on the training data.
3. The method of claim 1, wherein, The determination of the first trained model based on the first feedback information comprises: if the first feedback information indicates that a first target node in the second node can handle the training task, splitting the training task based on the first feedback information and the training data to obtain a plurality of training sub-tasks; sending a third request to the first target node, and receiving a second model sent by the first target node for the first training sub-task; wherein the third request includes the first training sub-task corresponding to the first target node and an initial model; determining a third model for a second training sub-task; wherein the training sub-tasks include the first training sub-task and the second training sub-task; determining the first trained model based on the second model and the third model.
4. The method of claim 3, wherein, The determination of the third model for the second training sub-task comprises: sending a fourth request to a second target node; wherein the fourth request includes demand data of the second training sub-task and the initial model; the second target node is a node in the second node other than the first target node; receiving training sub-data sent by the second target node for the second training sub-task; wherein the training sub-data is determined by the second target node based on the demand data. determine the third model based on the training sub-data, target training data, and the initial model.
5. The method of claim 4, wherein, The receiving the training sub-data sent by the second target node for the second training sub-task comprises: receiving the training sub-data sent by the second target node within a response time; wherein the fourth request further comprises the response time; or, receiving indication information and an extended response time sent by the second target node; wherein the indication information is used to indicate the reason why the training sub-data is not returned; updating the response time based on the extended response time, and receiving the training sub-data sent by the second target node within the updated response time; wherein the fourth request further comprises the response time.
6. The method of claim 4, wherein, The determining the third model based on the training sub-data, target training data, and the initial model comprises: if the idle computing resource of the first node can process the second training sub-task, training the initial model based on the training sub-data and the target training data to obtain the third model; if the idle computing resource of the first node cannot process the second training sub-task, sending a fifth request to a second base station; wherein the fifth request comprises the second training sub-task, the initial model, the training sub-data, and the target training data; receiving the third model for the second training sub-task sent by the second base station; wherein the third model is obtained by training the initial model based on the training sub-data and target training data by the second base station.
7. The method of claim 6, wherein, The method further comprises: if the idle computing resource of the first node cannot process the second training sub-task, decomposing the second training sub-task to obtain a plurality of decomposed training sub-tasks; training the initial model based on the training sub-data and the target training data in sequence, and determining the third model based on the plurality of sub-models obtained by training.
8. The method of claim 1, wherein, The determining the target model corresponding to the training task based on the first trained model comprises: if the first trained model meets the target demand, determining the first trained model as the target model; if the first trained model does not meet the target demand, sending a sixth request to the second node, and receiving second feedback information sent by the second node for the sixth request; wherein the sixth request comprises the required computing resource and the first trained model; determining a second trained model based on the second feedback information, until the target model meeting the target demand is obtained based on the determined Nth trained model.
9. The method of claim 2 or 3, wherein, The method further comprises: if the first model or the second model is updated, receiving the updated model sent by the second node or the first target node; updating the target model based on the updated model.
10. A model determination apparatus, wherein, The device comprises: an acquisition part configured to acquire a training task for training a model and required computing resource corresponding to the training task; The sending part is configured to send a first request to a second node in the first base station when the first node does not have training data of the training task and it is determined that the training task cannot be processed based on the idle computing resource of the first node and the required computing resource; wherein the first request comprises the required computing resource; The receiving part is configured to receive first feedback information sent by the second node for the first request; wherein the first feedback information is generated by the second node based on the required computing resource and the idle computing resource of the second node; The processing part is configured to determine a first trained model based on the first feedback information, and determine a target model corresponding to the training task based on the first trained model.
11. A first node, wherein, The first node comprises a processor, a memory and a communication bus; The communication bus is used to realize the communication connection between the processor and the memory; The processor is used to execute the model determination program in the memory to realize the steps of the model determination method according to any one of claims 1-9.
12. A computer readable storage medium, wherein, The computer readable storage medium stores one or more programs which can be executed by one or more processors to realize the steps of the model determination method according to any one of claims 1-9.
13. A computer program product comprising a computer program, wherein, The computer program realizes the method according to any one of claims 1-9 when executed by the processor.