Model training method, server, and client device

Abstract

The application provides a model training method, a server and a client device to improve the stability of a global shared model fused in a federated multi-task learning scenario. In the federated multi-task learning scenario, multiple client devices train models of multiple related tasks. The method comprises: a client device measuring a locally trained model to obtain model measurement information and reporting the server, the server obtaining first indication information corresponding to each model according to the model measurement information corresponding to the multiple models, the first indication information being used to indicate a training strategy of the client device for training the model, and the server sending the first indication information corresponding to each model to the client device for training the model, so that the client device adjusts the training strategy of the training model according to the first indication information to coordinate the training progress of the multiple models.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a model training method, a server, and a client device. Background Technology

[0002] In a multi-task federated learning scenario, multiple clients upload some or all of the model's weight parameters to the server based on their local tasks. After the server completes the model fusion, it sends the fused global model to the clients, who then replace some or all of their corresponding local models with the global model.

[0003] In multi-task federated learning scenarios, the training difficulty and time of models for different tasks often vary, and the performance of models for multiple tasks is not synchronized, resulting in unstable performance of the fused global model. Summary of the Invention

[0004] This application provides a model training method, server, and client device to manage the training progress of each model in federated multi-task learning, thereby obtaining a more stable globally shared model.

[0005] The first aspect of this application provides a model training method applicable to a federated multi-task learning system. This system includes multiple client devices and a server. The client devices jointly train multiple related task-specific models, where each model includes a shared model. These shared models are identical and can share knowledge through federated learning. Each client device can locally train at least one of the multiple models. The method includes: the server acquiring model measurement information corresponding to multiple models, where the model measurement information for each model is obtained from measurements taken by the client devices and sent to the server; the server obtaining first indication information for each model based on the model measurement information, which indicates the model's training strategy; and the server sending the first indication information for each model to the client devices training the models. By measuring multiple models to obtain model measurement information, the server obtains the first indication information for the training strategy of each model, thereby coordinating the training progress of multiple models and making the performance changes of the fused global model smoother and more stable.

[0006] In one possible implementation, the server obtains first indication information for each model based on model measurement information corresponding to multiple models. This includes: the server obtaining the model performance improvement for each model based on the model measurement information corresponding to multiple models; and the server determining the first indication information for each model based on the model performance improvement. The server can infer the training progress of each model based on the model performance improvement, thereby accurately determining the training strategy for each model.

[0007] In one possible implementation, the training strategy includes stopping training, restarting training, accelerating training, or slowing down training. Thus, the client device can adjust the training strategy of its own model based on the server's initial instructions, allowing models with slower training progress to have their training accelerated, and models with faster training progress to have their training slowed down, thereby keeping the training progress of each model as consistent as possible.

[0008] In one possible implementation, each model includes a shared model, which is the same model among multiple models. The first indication information is also used to indicate the model upload parameters for uploading the shared model to the server. The model upload parameters may include the shared model upload period and the model upload compression ratio, etc. The shared model upload period can also affect the model training progress. For example, a short shared model upload period can accelerate the model training speed to some extent, while a long shared model upload period can slow down the model training speed to some extent.

[0009] In one possible implementation, model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency. Model measurement information can accurately reflect the model's performance status and training progress, thereby accurately determining the initial indication information for each model and achieving precise management of the model's lifecycle.

[0010] In one possible implementation, the method further includes: the server sending second instruction information to multiple client devices, the second instruction information being used to instruct the multiple client devices to measure the locally trained model to obtain model measurement information. The server sends the second instruction information to each client device to instruct model measurement, thereby enabling the server to obtain the model measurement information of each model and, based on the model measurement information, manage the lifecycle of each model.

[0011] In one possible implementation, the second indication information includes at least one of the following: model identifier, model measurement cycle, model measurement information reporting conditions, model measurement information reporting cycle, and measurement indicators.

[0012] In one possible implementation, each model includes a shared model, which is the same model among multiple models. The method further includes: a server receiving shared models corresponding to multiple models uploaded by multiple client devices; the server fusing the shared models corresponding to multiple models according to the weights corresponding to each model to obtain a global shared model, where the weights corresponding to each model are obtained based on model measurement information corresponding to multiple models; and the server sending the global shared model to multiple client devices. By determining the weights for fusing the global shared model through model measurement information, the lifecycle management of the models can be achieved based on the weights of the contributions of each model's shared model to the global shared model, while ensuring balanced model performance.

[0013] A second aspect of this application provides a model training method. The method includes: a client device obtaining first instruction information from a server, the first instruction information being model measurement information corresponding to multiple models, the multiple models being models corresponding to multiple associated tasks, the multiple models including models trained locally by the client device, and the model measurement information being measurements taken by the multiple client devices of the locally trained models and then sent to the server; the client device adjusting the training strategy of the models trained by the client device according to the first instruction information.

[0014] In one possible implementation, training strategies include stopping training, restarting training, speeding up training, or slowing down training.

[0015] In one possible implementation, each model includes a shared model, which is the same model among multiple models, and the first indication information is also used to indicate the model upload parameters of the shared model upload server.

[0016] In one possible implementation, the model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency.

[0017] In one possible implementation, the method further includes: the client device obtaining second instruction information from the server; the client device measuring the model trained by the client device according to the second instruction information to obtain model measurement information; and the client device sending the model measurement information to the server.

[0018] In one possible implementation, the second indication information includes at least one of the following: model identifier, model measurement cycle, model measurement information reporting conditions, model measurement information reporting cycle, and measurement indicators.

[0019] A third aspect of this application provides a server. The server includes a processing module and a transceiver module. The processing module is used to acquire model measurement information corresponding to multiple models, where the multiple models are models corresponding to multiple related tasks, and the multiple models are models trained locally on multiple client devices. The model measurement information is generated by the multiple client devices measuring the locally trained models and then sending the measurements to the server. The processing module is also used to obtain first indication information corresponding to each model based on the model measurement information, whereby the first indication information indicates the model's training strategy. The transceiver module is used to send the first indication information corresponding to each model to the client devices training the models.

[0020] In one possible implementation, the processing module is used to obtain the model performance improvement for each model based on model measurement information corresponding to multiple models. Based on the model performance improvement for each model, the processing module determines the first indication information for each model.

[0021] In one possible implementation, training strategies include stopping training, restarting training, speeding up training, or slowing down training.

[0022] In one possible implementation, each model includes a shared model, which is the same model among multiple models. The first indication information is also used to indicate the model upload parameters of the shared model to the upload server. The model upload parameters may include the shared model upload cycle and the model upload compression rate, etc.

[0023] In one possible implementation, the model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency.

[0024] In one possible implementation, the transceiver module is used to send second instruction information to multiple client devices, which instructs the multiple client devices to measure the locally trained model in order to obtain model measurement information.

[0025] In one possible implementation, the second indication information includes at least one of the following: model identifier, model measurement cycle, model measurement information reporting conditions, model measurement information reporting cycle, and measurement indicators.

[0026] In one possible implementation, each model includes a shared model, which is the same model among multiple models. The transceiver module receives the shared model corresponding to multiple models uploaded by multiple client devices. The processing module merges the shared model corresponding to multiple models according to the weights corresponding to each model to obtain a global shared model. The weights corresponding to each model are obtained based on the model measurement information corresponding to multiple models. The transceiver module then sends the global shared model to multiple client devices.

[0027] A fourth aspect of this application provides a client device. The client device includes a processing module. The processing module is configured to acquire first indication information from a server, wherein the first indication information is obtained from model measurement information corresponding to multiple models, the multiple models are models corresponding to multiple associated tasks, the multiple models include models trained locally by the client device, and the model measurement information is obtained by the multiple client devices measuring the locally trained models and sending the measurements to the server; the processing module is configured to adjust the training strategy of the models trained by the client device according to the first indication information.

[0028] In one possible implementation, training strategies include stopping training, restarting training, speeding up training, or slowing down training.

[0029] In one possible implementation, each model includes a shared model, which is the same model among multiple models, and the first indication information is also used to indicate the model upload parameters of the shared model upload server.

[0030] In one possible implementation, the model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency.

[0031] In one possible implementation, the client device further includes a transceiver module. A processing module is used to acquire second instruction information from the server, and the client device measures the model trained on the client device based on the second instruction information to obtain model measurement information. The transceiver module is used to send the model measurement information to the server.

[0032] In one possible implementation, the second indication information includes at least one of the following: model identifier, model measurement cycle, model measurement information reporting conditions, model measurement information reporting cycle, and measurement indicators.

[0033] This application provides a model training system in a fifth aspect. The system includes a server and multiple client devices. The multiple clients are devices that train models corresponding to multiple related tasks, and these models are models trained locally on the client devices. Each client device is used to send model measurement information of its locally trained model to the server. The server is used to obtain first instruction information corresponding to each model based on the model measurement information corresponding to the multiple models. The first instruction information is used to instruct the model's training strategy. The server is used to send the first instruction information corresponding to each model to the client devices training the models. Each client device is used to adjust the training strategy of the model trained on its client device based on the first instruction information.

[0034] In one possible implementation, the server is specifically used to obtain the model performance improvement amount corresponding to each model based on the model measurement information corresponding to multiple models; the server is specifically used to determine the first indication information corresponding to each model based on the model performance improvement amount corresponding to each model.

[0035] In one possible implementation, training strategies include stopping training, restarting training, speeding up training, or slowing down training.

[0036] In one possible implementation, each model includes a shared model, which is the same model among multiple models, and the first indication information is also used to indicate the model upload parameters of the shared model upload server.

[0037] In one possible implementation, the model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency.

[0038] In one possible implementation, the server is also used to send second instruction information to multiple client devices, the second instruction information being used to instruct the multiple client devices to measure the locally trained model in order to obtain model measurement information.

[0039] In one possible implementation, the client device is used to obtain second instruction information from the server, and the client device measures the model trained by the client device according to the second instruction information to obtain model measurement information; the client device is used to send the model measurement information to the server.

[0040] In one possible implementation, the second indication information includes at least one of the following: model identifier, model measurement cycle, model measurement information reporting conditions, model measurement information reporting cycle, and measurement indicators.

[0041] In one possible implementation, each model includes a shared model, which is the same model among multiple models. The method further includes: a server for receiving the shared model corresponding to multiple models uploaded by multiple client devices; a server for fusing the shared model corresponding to multiple models according to the weights corresponding to each model to obtain a global shared model, where the weights corresponding to each model are obtained based on the model measurement information corresponding to multiple models; and a server for sending the global shared model to multiple client devices. Each client device is used to replace the shared model in its local model with the global shared model.

[0042] A sixth aspect of this application provides a server. The server includes a processor and a communication interface for communicating with multiple client devices, and the processor for executing a set of instructions to perform a model training method as described in the first aspect or any possible implementation thereof.

[0043] A seventh aspect of this application provides a client device. The client device includes a processor and a communication interface, the communication interface being used to communicate with a server, and the processor being used to execute a set of instructions to perform a model training method as described in the second aspect or any possible implementation thereof.

[0044] Eighth aspect of this application: This application provides a computer-readable storage medium. The computer-readable storage medium includes instructions that, when the computer-readable storage medium is executed on a computer, cause the computer to perform a model training method as described in the first aspect or any possible implementation thereof, or in the second aspect or any possible implementation thereof. Attached Figure Description

[0045] Figure 1 A schematic diagram of the architecture of a federated learning system provided for this application;

[0046] Figure 2 A schematic diagram of the lifecycle of the model provided in this application;

[0047] Figure 3 A schematic diagram of the lifecycle of the federal multitasking model provided for this application;

[0048] Figure 4 A schematic diagram of the architecture of a model training system provided in this application;

[0049] Figure 5 A flowchart illustrating an embodiment of the model training method provided in this application;

[0050] Figure 6 A flowchart illustrating another embodiment of the model training method provided in this application;

[0051] Figure 7 A schematic diagram of a scenario for wireless federated multitasking learning provided in this application;

[0052] Figure 8a A schematic diagram of the positioning model provided in this application;

[0053] Figure 8b A schematic diagram of the channel compression model provided in this application;

[0054] Figure 8c This is a schematic diagram of the structure of the LOS / NLOS recognition model provided in this application;

[0055] Figure 9 A schematic diagram of the structure of a server according to an embodiment provided in this application;

[0056] Figure 10This application provides a schematic diagram of the structure of an embodiment of a client device;

[0057] Figure 11 A schematic diagram of another embodiment of the server provided in this application;

[0058] Figure 12 A schematic diagram of another embodiment of the client device provided in this application. Detailed Implementation

[0059] This application provides a model training method, server, and client device to manage the training progress of each model in federated multi-task learning, thereby obtaining a more stable globally shared model.

[0060] In this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0061] In the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "multiple" means two or more. The term "at least one" in this application means one or more, and the term "multiple" in this application means two or more.

[0062] It should be understood that the terminology used in the description of the various examples herein is for the purpose of describing particular examples only and is not intended to be limiting. As used in the description of the various examples and the appended claims, the singular forms “a” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0063] It should also be understood that, in the various embodiments of this application, the sequence number of each process does not imply the 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 application. It should be understood that determining B based on A does not mean determining B solely based on A; B can also be determined based on A and / or other information. It should also be understood that the term "comprising" (also referred to as "includes", "including", "comprises", and / or "comprising"), when used in this specification, specifies the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should also be understood that the term "if" can be interpreted as meaning "when" or "upon" or "in response to determination" or "in response to detection". Similarly, depending on the context, the phrases “if it is determined that…” or “if [the stated condition or event] is detected” can be interpreted as “when it is determined that…” or “in response to determining that…” or “when [the stated condition or event] is detected” or “in response to detecting [the stated condition or event]”.

[0064] It should be understood that the phrases "an embodiment," "an embodiment," and "a possible implementation" used throughout the specification mean that a specific feature, structure, or characteristic related to an embodiment or implementation is included in at least one embodiment of this application. Therefore, the phrases "in an embodiment," "an embodiment," or "a possible implementation" 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.

[0065] The solution provided in this application is applicable to a machine learning system, which includes a server and at least one edge device. This application uses a federated learning system as an example for illustration. To better understand the solution provided in this application, relevant knowledge of federated learning will be introduced below.

[0066] Federated learning is a machine learning method designed to protect user privacy. In some practical applications of machine learning, limitations such as insufficient data features or a small number of samples on a single device make it difficult to train a good machine learning model independently. Therefore, it is necessary to fuse data from multiple devices together for training to obtain a higher-quality model. While fusing data from multiple devices for training, it is also crucial to ensure user data privacy; that is, data cannot be transferred outside the user's device and can only be used locally for model training. Federated learning emerged to address this requirement, effectively helping multiple computing nodes to use data and perform machine learning modeling while meeting the requirements of user privacy protection and data security.

[0067] See Figure 1 , Figure 1This application provides an architectural diagram of a federated learning system. Federated learning is a machine learning technique for protecting user privacy. A federated learning system typically includes a server and multiple client devices as participants. The workflow mainly includes model delivery and model fusion. In the model delivery process, client devices download models from the server, train the downloaded models using locally stored training data, and upload the models to the server after a certain training period (a fixed number of training rounds). In the model fusion process, the server collects the models uploaded by each client device, merges them, and delivers the merged models to each client device. The server can perform fusion processing on multiple models in various ways. For example, the server can average the received models and use the result as a shared model; or the server can weight the received models and use the weighted result as a shared model. The model delivery and model fusion processes iterate repeatedly until the models converge. It should be noted that in some possible implementations, the server can be a cloud server or a cloud-based device; this application does not limit this. In wireless federated learning scenarios, the server can also be a base station, access point, or satellite. This application sometimes refers to the client device as an edge device; the two terms have the same meaning. For example, an edge device can be a terminal, such as a mobile phone, tablet, laptop, drone, augmented reality (AR) / virtual reality (VR) device, vehicle-mounted terminal, security equipment (such as surveillance cameras, smoke alarms, fire extinguishers), smart speaker, or smart home appliance. Furthermore, it should be noted that the models in this application include neural network models. This application does not limit the specific category of the model; for example, the model in this application can be a convolutional neural network model, a recurrent neural network model, a deep residual network model, etc.

[0068] Multi-task learning (MTL) is a machine learning method that contrasts with single-task learning. Its aim is to improve the generalization ability of multiple related tasks through joint learning. Specifically, MTL leverages the correlation between multiple tasks, allowing them to mutually reinforce each other and learn in parallel. In other words, MTL enables the simultaneous training of multiple tasks, allowing them to influence each other. Generally, MTL has better generalization ability than single-task learning. The correlation between multiple tasks can be understood as shared features. MTL requires building a model for each task. The models for multiple tasks have sub-models with the same structure and are trained on the same type of data. These common sub-models can share model parameters. In this application, the common sub-models are called shared models, and each model in the multiple models includes this shared model. The shared model can be the entirety of a model or a part of a model.

[0069] Combining federated learning with multi-task learning results in federated multi-task learning. In this process, the federated learning system trains multiple different but related tasks, and the shared models from these tasks are fused using federated learning methods. During model fusion, the server collects the shared models uploaded by each client device and fuses them to obtain a global shared model. Once the server obtains the global shared model, it distributes it to each client device. The client devices replace their local shared models with the global shared model and continue training or using the model for inference. These two processes—model distribution and model fusion—iterate repeatedly until the shared model converges.

[0070] Because client device data is constantly increasing and updating, the model's performance evolves from high initial performance to increasingly mismatched with the current data distribution, leading to low performance. To address this performance degradation, lifecycle management (LCM) of the model is necessary. Figure 2 As shown, Figure 2 This diagram illustrates the lifecycle of the model provided in this application. From its initial setup, the model begins training and its performance gradually improves. At its peak, the model is considered to have converged, and training ceases. However, as the data in the system changes (the data distribution becomes inconsistent with the distribution used for training), the model and the data become increasingly mismatched, leading to a gradual decline in performance. Monitoring the model's lifecycle allows communication devices to adjust the model's performance in a timely manner; for example, when performance drops to a threshold, training can be restarted using the current data.

[0071] In multi-task federated learning scenarios, multiple clients upload partial or complete model weights to the server based on their local tasks. After the server completes model fusion, it distributes the fused global model to the clients, who then replace part or all of their corresponding local models with the global model. It is evident that different tasks often have varying training difficulties and durations. In multi-task federated learning scenarios, model management for a single task can affect other tasks; therefore, joint multi-task model lifecycle management is necessary. Figure 3 As shown, Model 2, due to its faster training speed, stops training first. Stopping training means that Model 2 will no longer upload updated models. At this time, since only Model 1 is uploading its model, the global model is more biased towards Model 1, causing Model 1's performance to jump. As the data distribution changes, Model 2's performance gradually declines. After the decline reaches a preset threshold, Model 2 restarts training. At this time, Model 1 has completed training and stopped training. Because the global model is now more biased towards Model 2, Model 1's performance declines again.

[0072] See Figure 4 , Figure 4 This is a schematic diagram of the architecture of a model training system provided in this application. The model training system provided in this application includes a server and at least two end-side devices. Specifically, the model training system is, for example, a federated multi-task learning system. One client device can train a model for at least one task, and a model for one task can be trained on at least one client device. The number of tasks and terminals shown in the diagram are for illustrative purposes only and should not be construed as a limitation of this application.

[0073] The server can connect to multiple client devices via wired or wireless means. Multiple client devices train models for multiple related tasks, meaning they collaboratively train multiple models. These models share a common model. Client devices train their local models and report the shared model to the server, allowing the server to fuse the shared model from multiple models. The server then fuses the shared models from each client device to obtain a global shared model and distributes the global shared model to the client devices. Once the client devices receive the global shared model, they replace their local shared model with it.

[0074] Furthermore, in this embodiment, the server is also used to send a second instruction to the client device to instruct the measurement of the model trained locally on the client device, so that the client device can obtain model measurement information. Based on the model measurement information of multiple models reported by the client device, the server determines a first instruction for each model. The first instruction is used to instruct the client device on the training strategy for training the model. Training strategies include, for example, stopping training, restarting training, or continuing training, thereby enabling joint lifecycle management of multiple models. That is, the server has a global understanding of the training status of multiple models. Whether a model stops training, restarts training, or continues training is decided by referring to the training status of other models, rather than considering only the training status of one model itself. This allows the server to coordinate the lifecycles of multiple models as much as possible, reducing the possibility that some models, after convergence, stop training and do not upload the shared model, causing the performance of the global shared model to be biased towards another model that has not yet fully converged and continues training and uploading the shared model. This results in unstable performance of the global shared model, affecting the performance of other models.

[0075] In this embodiment, model measurement information is obtained by measuring multiple models. The server manages the lifecycle of multiple models based on the model measurement information of multiple models, guides the training strategies of each client device, and makes the performance of the fused global shared model more balanced and stable.

[0076] like Figure 5 As shown, Figure 5 This is a flowchart illustrating an embodiment of the model training method provided in this application. The execution entity in this embodiment is the server in the aforementioned model training system. This embodiment includes the following steps:

[0077] S501, the server obtains model measurement information corresponding to multiple models.

[0078] Multiple models refer to models corresponding to multiple related tasks. Multiple models are models trained locally on multiple client devices. Model measurement information is the measurement information of the locally trained models by multiple client devices and then sent to the server.

[0079] Multiple model measurement information can be uploaded to the server simultaneously by multiple client devices. Alternatively, it can be uploaded to the server at different times, allowing the server to obtain the latest model measurement information for each model within a given time window. If a model is trained on multiple client devices, all client devices will report the model's measurement information to the server. The server can then calculate the average or median of the model measurement information reported by these client devices as the model's overall measurement information.

[0080] The client device measures the entire model, and the model measurement information reflects the training state or training stage of the entire model. Since the training state / training stage of the model is related to its performance, the model measurement information can also be considered to reflect the overall performance state of the model. Model measurement information includes at least one of the following metrics: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency. The model performance improvement indicates whether the model's performance is improving or deteriorating, and the degree of improvement or deterioration. When the model performance improvement is positive, a larger value indicates a faster improvement in model performance, suggesting that the model has not yet converged. A negative value indicates a declining trend in model performance; a smaller value indicates a faster decline in model performance. The model performance improvement can be calculated by comprehensively considering at least one of the model loss value, accuracy, communication overhead, latency, and energy efficiency. For example, the model performance improvement can be calculated by weighted summation of at least one of the model loss value, accuracy, communication overhead, latency, and energy efficiency. Of course, the improvement in model performance can also be calculated in other ways.

[0081] The improvement in model performance can be calculated by the client device and sent to the server, or it can be calculated by the server after the client device sends at least one of the following metrics: model loss value, accuracy, communication overhead, latency, and energy efficiency.

[0082] S502, the server obtains the first indication information corresponding to each model based on the model measurement information corresponding to multiple models. The first indication information is used to indicate the training strategy of the model.

[0083] The server determines the first indication information for each model based on model measurement information from multiple models. This first indication information is used to specify the model's training strategy. Training strategies may include, for example, stopping training, restarting training, accelerating training, or slowing down training. Training strategies may also include at least one of the following: model freezing, model invalidation, model selection, model sharing, model fine-tuning, sub-model configuration, model architecture reconfiguration, model reset, model merging, model compression, model decompression, model update, and model splitting.

[0084] The process from the start of model training to the point where the model converges and training stops, and then to the point where model performance deteriorates to the point where training restarts, can be called the model's lifecycle. The performance trends of a model differ at different stages of its lifecycle. For example, model performance improves rapidly at the beginning of training; after a period of training, the model begins to converge, and performance improvement slows down; when the model reaches convergence and training stops, its performance is optimal, and no more shared models are uploaded to the server during the training stoppage; as data distribution changes and globally shared models move away from the model after training stops, model performance declines. Different models may have different lifecycles.

[0085] The server determines the current training progress of each model, i.e., which stage of its lifecycle each model is in, based on measurement information from multiple models. For models that train quickly and converge earlier than others, the server can instruct the corresponding terminal to slow down the training speed to extend the training period and keep pace with the training progress of other models as much as possible. For models that train slowly and converge later than others, the server can instruct the corresponding terminal to speed up the training speed to shorten the training period and catch up with the training progress of other models as much as possible. Thus, by instructing training strategies on client devices, the server can ensure that the lifecycles of each model are kept as consistent as possible, or that the differences in the lifecycles of each model are within an acceptable range. This reduces the likelihood of the globally shared model becoming increasingly biased towards different models due to inconsistent lifecycles, which could lead to performance instability in the globally shared model.

[0086] The server can determine the training progress of each model based on the performance improvement of each model. Further comparison of the training progress of multiple models determines the initial indication information for each model.

[0087] In one implementation, the training progress of each model can be determined by the trend of its performance improvement over the most recent M iterations. For example, a model whose performance improvement has been decreasing overall over the most recent M iterations can be considered to have converged. Here, M is an integer greater than or equal to 2. A model whose performance improvement has not decreased significantly over the most recent M iterations can be considered not to have converged. The server compares the training progress of each model. If some models are converging while others have not, it can send a first instruction to the client devices corresponding to the converging models to indicate a reduction in training speed, and send a first instruction to the client devices corresponding to the non-converging models to indicate an acceleration in training speed.

[0088] In another implementation, the server can determine the weight of each model based on model measurement information corresponding to multiple models. Then, the server determines the first indication information for each model based on its weight. In this embodiment, the weight can be related to the improvement in model performance. Specifically, the weight can be positively correlated with the improvement in model performance.

[0089] The server calculates the ratio between the absolute value of the performance improvement of each model and the sum of the absolute values ​​of the performance improvements of all models to obtain the weight of each model. For example, if model A has a performance improvement of 0.4, model B has a performance improvement of 0.6, and model C has a performance improvement of 1, then model A has a weight of 0.2, model B has a weight of 0.3, and model C has a weight of 0.5. Of course, the weights can also be obtained through other normalization methods, which are not limited here.

[0090] Weights can indicate the training progress of each model relative to other models. If the weights of a certain model are significantly greater than those of other models, it indicates that the model has not yet converged. The training progress of each model can be further determined by combining historical weight data. For example, if the weights of a certain model have generally increased in the most recent M model measurements, while the weights of other models have generally decreased, it can be said that the model with the increased weights has not yet converged, while the model with the decreased weights has begun to converge.

[0091] The training progress of each model can also be determined by combining historical model performance improvements. For example, a model whose weights have decreased overall and whose overall performance improvement has also decreased in the most recent M model measurements can be considered to be converging, and the server can send a first instruction to the client device corresponding to that model to slow down the training speed. A model whose weights have increased overall and whose overall performance improvement has also increased or not decreased significantly in the most recent M model measurements can be considered not to have converged, and the server can send a first instruction to the client device corresponding to that model to speed up the training speed.

[0092] In this embodiment, the first indication information may include the aforementioned weights, which indicate the model training strategy executed by the client device. Of course, the first indication information can also take other forms, such as using different strings to indicate different training strategies.

[0093] The first indication information can also be used to instruct the model upload parameters of the shared model upload server. Model upload parameters include, for example, the model upload period. When the first indication information instructs a reduction in training speed, it can also instruct a reduction in the model upload period. Model upload parameters may also include, for example, the model upload compression ratio, and the first indication information carries the model upload compression ratio.

[0094] S503, the server sends the first instruction information corresponding to each model to the client device that trained the model.

[0095] After the server receives the first instruction information corresponding to each model, it sends the first instruction information corresponding to each model to the client device that trained the model in order to dynamically adjust the life cycle of each model and keep the life cycle of each model as consistent as possible. This allows the global shared model to be integrated into more shared models during fusion, making the performance of the global shared model more stable.

[0096] like Figure 6 As shown, Figure 6 This is a flowchart illustrating another embodiment of the model training method provided in this application. This embodiment is implemented by the aforementioned model training system. This embodiment includes the following steps:

[0097] S601, the server sends a second instruction message to multiple client devices.

[0098] The server sends a second instruction to multiple client devices, instructing them to measure the locally trained models for each task and report the measurement information. These client devices are those participating in federated multi-task learning, and each client device trains a model for at least one task.

[0099] The second set of information includes at least one of the following: the identifier of the model trained on the corresponding client device, the model measurement cycle, the model measurement information reporting conditions, the model measurement information reporting cycle, and the measurement index. The model measurement cycle, model measurement information reporting conditions, model measurement information reporting cycle, or measurement index may be the same or different for different models.

[0100] The server can periodically send a second indication message to the client device to periodically obtain model measurement information reported by the client device. Alternatively, the server can send a second indication message once, which carries parameters such as the model measurement cycle, model measurement information reporting conditions, or model measurement information reporting period, so that the client device can periodically report model measurement information to the server according to the parameters in the second indication message or when the reporting conditions are met.

[0101] S602, multiple client devices measure the locally trained model according to the second instruction information to obtain the model measurement information of each model.

[0102] After receiving the second instruction information from the server, multiple client devices measure the locally trained models according to the second instruction information to obtain model measurement information for each model.

[0103] The second instruction information includes, for example, measurement metrics, which the client device uses to measure the model. Measurement metrics include, for example, at least one of model loss value, accuracy, communication overhead, latency, and energy efficiency. Therefore, the model measurement information includes at least one of the measured values, such as model loss value, accuracy, communication overhead, latency, and energy efficiency. The client device can send the model measurement information, including at least one of the measured values, to the server.

[0104] In another implementation, the client device can calculate the model performance improvement based on at least one of the measured values ​​of model loss, accuracy, communication overhead, latency, and energy efficiency, and send the model performance improvement as model measurement information to the server. Alternatively, the client device can send the model performance improvement along with at least one of the measured values ​​of model loss, accuracy, communication overhead, latency, and energy efficiency as model measurement information to the server.

[0105] S603: Multiple client devices send model measurement information to the server.

[0106] If the second instruction information includes a model measurement information reporting cycle, the client device sends model measurement information to the server according to the reporting cycle indicated in the second instruction information. If the second instruction information includes model measurement information reporting conditions, the client device sends model measurement information to the server when the reporting conditions are met. Model measurement information reporting conditions may include, for example, the model having completed convergence, the model performance having decreased to a specified threshold, or the client device having excessive computational load.

[0107] S604, the server obtains model measurement information corresponding to multiple models.

[0108] This step can be found in the relevant content of S501, so it will not be repeated here.

[0109] S605, the server obtains the first indication information corresponding to each model based on the model measurement information corresponding to multiple models. The first indication information is used to indicate the training strategy of the model.

[0110] This step can be found in the relevant content of S502, so it will not be repeated here.

[0111] S606, the server sends the first instruction information corresponding to each model to the client device that trained the model.

[0112] This step can be found in the relevant content of S503, so it will not be repeated here.

[0113] S607, multiple client devices adjust the training strategy of the trained model according to the first instruction information.

[0114] After receiving the first instruction from the server, multiple clients determine the training strategy for their locally trained models based on the first instruction and execute the corresponding training strategy. Training strategies may include, for example, stopping training, restarting training, accelerating training, or slowing down training. Training strategies may also include at least one of the following: model freezing, model invalidation, model selection, model sharing, model fine-tuning, sub-model configuration, model architecture reconfiguration, model reset, model merging, model compression, model decompression, model update, and model splitting.

[0115] When the first indication information includes the weights of the model trained by the client device, the client device can determine the training strategy corresponding to its own trained model based on the weights. For example, when the weights are greater than a first threshold, the training speed can be increased; when the weights are less than or equal to the first threshold but greater than a second threshold, the training speed can be decreased; when the weights are less than or equal to the second threshold, training can be stopped. When the client device's local model is already in a stopped training state, any non-zero weight received by the client device can be considered an indication to restart training.

[0116] The first instruction received is to stop training on the client device, halt the training of the local model, and refrain from sending the local shared model to the server until training is restarted.

[0117] The first instruction received is to instruct the client device to restart training, restart the training of the local model, and resend the locally trained shared model to the server.

[0118] The first instruction received is to instruct the client device to speed up the training process, which can be achieved by increasing the learning rate or increasing the batch size.

[0119] The first instruction received is from the client device to reduce training speed. This can be achieved by lowering model accuracy, learning rate, or the number of local training epochs. Alternatively, the client device can also reduce training speed by increasing the upload cycle of the locally trained shared model. Specifically, after increasing the upload cycle, when the client device does not send the shared model to the server, the server does not have a shared model from that client device when fusing the global shared model; that is, the client device's shared model does not contribute to the global shared model. When the client device receives the global shared model, it replaces its local shared model with it. Since the replaced shared model did not participate in the fusion of the global shared model, the performance of the client device's local model may decrease to some extent after the replacement, thus extending the model convergence speed and prolonging the lifespan of the client device's local model. However, the client device will still send the shared model to the server. When the client device sends the shared model to the server, the shared model will still be merged into the global shared model. Therefore, the model trained by the client device will not experience a continuous performance decline and can maintain a relatively stable state overall during the period of reduced training speed.

[0120] S608: Multiple client devices send a shared model to the server.

[0121] Except for client devices that have stopped training their models, each client device sends a shared model to the server according to its own shared model upload cycle, so that the server can merge the shared models from all client devices to obtain a global model. Client devices can send the shared model to the server simultaneously or at different times; there is no restriction on this.

[0122] S609 is a server that integrates multiple models into a shared model to obtain a globally shared model.

[0123] In one implementation, the server can average the received multiple shared models and use the result of averaging the multiple shared models as the global shared model.

[0124] In another implementation, the server weights the received multiple shared models according to their weights, and uses the weighted result as the global shared model. The weights can be derived from the model measurement information of each model. For example, the weights could be those obtained in S605 based on the performance improvement of each model. Therefore, models with high weights that have not yet converged contribute more to the global shared model. When client devices using models with high weights that have not yet converged continue training based on the global shared model, the high degree of matching with local training data accelerates model convergence, thereby increasing model training speed and shortening the model training cycle. Conversely, models with low weights that have begun to converge contribute less to the global shared model. When client devices using models with low weights that have begun to converge continue training based on the global shared model, the slightly lower degree of matching with local training data slows down model convergence, thereby reducing model training speed and extending the model training cycle. However, since the global shared model still includes shared models with low weights that have begun to converge, it will not cause the global shared model to deviate too much from the models with low weights that have begun to converge. Instead, it reduces the training speed while ensuring the stability of the models with low weights that have begun to converge, so that the performance of the global shared model and the models trained locally on each client device is controllable.

[0125] S610: The server sends a globally shared model to each client device.

[0126] After obtaining the globally shared model, the server sends it to each client device so that the client devices can replace their local shared models with the globally shared model and continue training or inference based on the replaced model.

[0127] S611 allows multiple client devices to replace the shared model of their local model with a globally shared model.

[0128] After receiving the globally shared model from the server, each client device updates the shared model of its locally trained model to the globally shared model, and performs training or inference based on the updated model.

[0129] In this embodiment, the server sends second instruction information to multiple client devices participating in federated multi-task learning to instruct on model measurement. Each client device measures its own trained model according to the second instruction information and reports the measured model measurement information to the server. The server determines the current training progress of each model based on the model measurement information, and thus determines the corresponding training strategy based on the training progress of each model, achieving lifecycle management of multiple models and ensuring that the lifecycles of each model are as consistent as possible. Consequently, the globally shared model obtained by the server by fusing the shared model of each model has more stable performance.

[0130] To make the technical solution provided in this application easier to understand, the following description uses a wireless federation multi-task learning scenario as an example.

[0131] like Figure 7 As shown, Figure 7 This is a schematic diagram of a scenario for wireless federated multitasking learning provided in this application. Figure 7 This example uses three user equipment (UEs) to train models for one of three tasks: channel compression, localization, and line-of-sight (LOS) / non-line-of-sight (NLOS) recognition. UE1 trains a localization model, UE2 trains a channel compression model, and UE3 trains a LOS / NLOS recognition model. The base station, acting as a server, manages the lifecycle of the three models trained by UE1 to UE3, and merges the shared model of the three models trained by UE1 to UE3 to obtain a globally shared model, which is then distributed to each UE.

[0132] like Figures 8a to 8c As shown, Figure 8a A schematic diagram of the positioning model provided in this application; Figure 8b A schematic diagram of the channel compression model provided in this application; Figure 8c This is a schematic diagram of the LOS / NLOS recognition model provided in this application. The localization model includes convolutional neural networks (CNN)1, CNN2, and a multi-layer perceptron (MLP). The channel compression model includes CNN1. The LOS / NLOS recognition model includes CNN1. It can be seen that the three models have different levels of complexity, and therefore their training cycles are also different. The CNN1 network in all three models has the same structure, used to process channel data. Since all three tasks utilize channel features, this part of the network can be shared; that is, CNN1 is a shared model. The parameters of CNN1 can be fused at the base station to obtain a globally shared model and then distributed to the three UEs.

[0133] The base station can carry first or second indication information through the information element (IE) in the radio resource control (RRC) message.

[0134] The base station sends RRC messages carrying second indication information to UE1 through UE3 respectively. The second indication information includes, for example, at least one of the following: the model identifier of the specified model to be measured, the time for conducting the model measurement, the specified test dataset, measurement indicators, and the period or conditions for reporting model measurement information. After receiving the second indication information, UE1 performs measurements on the positioning model according to the second indication information; UE2 performs measurements on the channel compression model according to the second indication information; and UE3 performs measurements on the LOS / NLOS identification model according to the second indication information. After completing the model measurements, UE1 through UE3 send their respective model measurement information to the base station according to the reporting period or conditions indicated in the second indication information.

[0135] After receiving model measurement information for the positioning model reported by UE1, the channel compression model reported by UE2, and the LOS / NLOS identification model reported by UE3, the base station determines the training progress of these three models based on this model measurement information. If the training progress of the channel compression model and the LOS / NLOS identification model is faster than that of the positioning model, the base station can send a first instruction to UE2 to instruct UE2 to reduce the training speed of the channel compression model, send a first instruction to UE3 to instruct UE3 to reduce the training speed of the LOS / NLOS identification model, and send a first instruction to UE1 to instruct UE1 to accelerate the training speed of the positioning model, thereby making the lifecycles of these three models as close as possible.

[0136] The base station also determines the weights of the shared model among the three models based on the model measurement information of the three models. After receiving CNN1 from UE1 to UE3, the base station fuses the global CNN1 according to the weights corresponding to each CNN1. After obtaining the global CNN1, the base station sends it to UE1 to UE3 respectively. UE1 replaces the CNN1 in its localization model with the global CNN1, UE2 replaces the CNN1 in its channel compression model with the global CNN1, and UE3 replaces the CNN1 in its LOS / NLOS recognition model with the global CNN1. After completing the replacement, they continue training or inference.

[0137] like Figure 9As shown, based on the same technical concept, this application also provides a server 900. The server 900 includes a processing module 901 and a transceiver module 902. The processing module 901 is used to acquire model measurement information corresponding to multiple models, where the multiple models are models corresponding to multiple related tasks, and the multiple models are models trained locally on multiple client devices. The model measurement information is sent to the server by the multiple client devices after measuring the locally trained models. The processing module 901 is used to obtain first indication information corresponding to each model based on the model measurement information corresponding to the multiple models. The first indication information is used to indicate the training strategy of the model. The transceiver module 902 is used to send the first indication information corresponding to each model to the client device training the model.

[0138] In one possible implementation, the processing module 901 is used to obtain the model performance improvement for each model based on the model measurement information corresponding to multiple models. The processing module 901 then determines the first indication information for each model based on the model performance improvement for each model.

[0139] In one possible implementation, training strategies include stopping training, restarting training, speeding up training, or slowing down training.

[0140] In one possible implementation, each model includes a shared model, which is the same model among multiple models. The first indication information is also used to indicate the model upload parameters of the shared model to the upload server. The model upload parameters may include the shared model upload cycle and the model upload compression rate, etc.

[0141] In one possible implementation, the model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency.

[0142] In one possible implementation, the transceiver module is used to send second instruction information to multiple client devices, which instructs the multiple client devices to measure the locally trained model in order to obtain model measurement information.

[0143] In one possible implementation, the second indication information includes at least one of the following: model identifier, model measurement cycle, model measurement information reporting conditions, model measurement information reporting cycle, and measurement indicators.

[0144] In one possible implementation, each model includes a shared model, which is the same model among multiple models. The transceiver module 902 receives the shared model corresponding to multiple models uploaded by multiple client devices. The processing module 901 merges the shared model corresponding to multiple models according to the weights corresponding to each model to obtain a global shared model. The weights corresponding to each model are obtained based on the model measurement information corresponding to multiple models. The transceiver module 902 sends the global shared model to multiple client devices.

[0145] like Figure 10 As shown, Figure 10 This application provides a schematic diagram of the structure of a client device according to an embodiment. The client device 1000 includes a processing module 1001. The processing module 1001 is used to obtain first indication information from a server, wherein the first indication information is obtained from model measurement information corresponding to multiple models, the multiple models are models corresponding to multiple associated tasks, the multiple models include models trained locally by the client device, and the model measurement information is sent to the server after the multiple client devices have measured the locally trained models; the processing module 1001 is used to adjust the training strategy of the models trained by the client device according to the first indication information.

[0146] In one possible implementation, training strategies include stopping training, restarting training, speeding up training, or slowing down training.

[0147] In one possible implementation, each model includes a shared model, which is the same model among multiple models, and the first indication information is also used to indicate the model upload parameters of the shared model upload server.

[0148] In one possible implementation, the model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency.

[0149] In one possible implementation, the client device further includes a transceiver module 1002. The processing module 1001 is used to acquire second instruction information from the server, and the client device measures the model trained by the client device according to the second instruction information to obtain model measurement information. The transceiver module 1002 is used to send the model measurement information to the server.

[0150] In one possible implementation, the second indication information includes at least one of the following: model identifier, model measurement cycle, model measurement information reporting conditions, model measurement information reporting cycle, and measurement indicators.

[0151] like Figure 11 As shown, Figure 11This is a schematic diagram of another embodiment of the server provided in this application. The server 1100 includes a communication interface 1101 and a processor 1102. The communication interface 1101 is used to communicate with multiple client devices. Based on the communication interface 1101 and the processor 1102, the operations performed by the server in any of the above-described model training method embodiments are implemented.

[0152] like Figure 12 As shown, Figure 12 This is a schematic diagram of another embodiment of the client device provided in this application. The client device 1200 includes a communication interface 1201 and a processor 1202. The communication interface 1201 is used to communicate with the server. Based on the communication interface 1201 and the processor 1202, the operations performed by the client device in any of the above-described model training method embodiments are implemented.

[0153] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a computer, implements the process implemented by the server or client device in the model training method of any of the above method embodiments.

[0154] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0155] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between devices or units, and may be electrical or other forms.

[0156] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0157] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0158] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the technical solution of this application can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network testing device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A model training method, characterized in that, The method includes: The server obtains model measurement information corresponding to multiple models, which are models corresponding to multiple related tasks. The multiple models are models trained locally on multiple client devices, and the model measurement information is sent to the server by the multiple client devices after measuring the locally trained models. The server obtains first indication information corresponding to each of the multiple models based on the model measurement information corresponding to the multiple models. The first indication information is used to indicate the training strategy of the model. The server sends the first indication information corresponding to each model to the client device that trains the model; The server obtains first indication information corresponding to each of the multiple models based on the model measurement information corresponding to the multiple models, including: The server determines the training progress of each model based on the model measurement information corresponding to the multiple models; the training progress of each model is determined based on the model performance improvement of each model or the weight of each model. The server obtains the first indication information corresponding to each model based on the training progress of each model.

2. The method according to claim 1, characterized in that, The server obtains first indication information corresponding to each of the multiple models based on the model measurement information corresponding to the multiple models, including: The server obtains the model performance improvement for each model based on the model measurement information corresponding to the multiple models. The server determines the first indication information corresponding to each model based on the model performance improvement amount corresponding to each model.

3. The method according to claim 1, characterized in that, The training strategies include stopping training, restarting training, increasing training speed, or decreasing training speed.

4. The method according to claim 1, characterized in that, Each of the models includes a shared model, which is the same model among the multiple models, and the first indication information is also used to instruct the shared model to upload model upload parameters to the server.

5. The method according to claim 1, characterized in that, The model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency.

6. The method according to claim 1, characterized in that, The method further includes: The server sends a second instruction to the plurality of client devices, the second instruction being used to instruct the plurality of client devices to measure the locally trained model in order to obtain the model measurement information.

7. The method according to claim 6, characterized in that, The second indication information includes at least one of the following: the model identifier, the model measurement cycle, the model measurement information reporting conditions, the model measurement information reporting cycle, and the measurement index.

8. The method according to any one of claims 1 to 7, characterized in that, Each of the models includes a shared model, which is the same model among the multiple models, and the method further includes: The server receives the shared model corresponding to the multiple models uploaded by the multiple client devices; The server fuses the shared model corresponding to the multiple models according to the weights corresponding to each model to obtain a global shared model, wherein the weights corresponding to each model are obtained based on the model measurement information corresponding to the multiple models. The server sends the globally shared model to the multiple client devices.

9. A model training method, characterized in that, The method includes: The client device receives first indication information from the server. This first indication information is obtained from model measurement information corresponding to multiple models. The multiple models are models corresponding to multiple associated tasks, including models trained locally by the client device. The model measurement information is obtained by the client device measuring the locally trained models and then sending the measurements to the server. The first indication information corresponding to each model is obtained based on the training progress of each model, which is also obtained based on the model measurement information corresponding to the multiple models. The training progress of each model is determined based on the performance improvement of each model or the weights of each model. The client device adjusts the training strategy of the model it is training based on the first instruction information.

10. The method according to claim 9, characterized in that, The training strategies include stopping training, restarting training, increasing training speed, or decreasing training speed.

11. The method according to claim 9, characterized in that, Each of the models includes a shared model, which is the same model among the multiple models, and the first indication information is also used to instruct the shared model to upload model upload parameters to the server.

12. The method according to claim 9, characterized in that, The model measurement information includes at least one of the following: model performance improvement, model loss value, accuracy, communication overhead, latency, and energy efficiency.

13. The method according to any one of claims 9 to 12, characterized in that, The method further includes: The client device obtains second instruction information from the server. The client device measures the model trained by the client device according to the second instruction information to obtain the model measurement information; The client device sends the model measurement information to the server.

14. The method according to claim 13, characterized in that, The second indication information includes at least one of the following: the model identifier, the model measurement cycle, the model measurement information reporting conditions, the model measurement information reporting cycle, and the measurement index.

15. A server, characterized in that, The server includes a processing module and a transceiver module; The processing module is used to obtain model measurement information corresponding to multiple models, wherein the multiple models are models corresponding to multiple related tasks, the multiple models are models trained locally on multiple client devices, and the model measurement information is sent to the server by the multiple client devices after measuring the locally trained models. The processing module is used to obtain first indication information corresponding to each of the multiple models based on the model measurement information corresponding to the multiple models. The first indication information is used to indicate the training strategy of the model. The transceiver module is used to send the first indication information corresponding to each model to the client device that trains the model; The processing module is used for: The training progress of each model is determined based on the model measurement information corresponding to the multiple models; the training progress of each model is determined based on the model performance improvement of each model or the weight of each model. Based on the training progress of each model, obtain the first indication information corresponding to each model.

16. A client device, characterized in that, The client device includes a processing module; The processing module is used to obtain first indication information from the server. The first indication information is obtained from model measurement information corresponding to multiple models. The multiple models are models corresponding to multiple associated tasks. The multiple models include models trained locally by the client device. The model measurement information is sent to the server by the multiple client devices after measuring the locally trained models. The first indication information corresponding to each model is obtained based on the training progress of each model, and the training progress of each model is obtained based on the model measurement information corresponding to the plurality of models; The training progress of each model is determined based on the improvement in model performance or the weights of each model. The processing module is used to adjust the training strategy of the model trained by the client device according to the first instruction information.

17. A model training system, characterized in that, The system includes a server and multiple client devices. The multiple clients are models that are trained on multiple related tasks. The multiple models are models trained locally on multiple client devices. The multiple client devices are respectively used to send model measurement information of the locally trained model to the server; The server is configured to obtain first indication information corresponding to each of the plurality of models based on the model measurement information corresponding to the plurality of models, wherein the first indication information is used to indicate the training strategy of the model; The server is configured to send the first indication information corresponding to each model to the client device that trains the model; The plurality of client devices are respectively used to adjust the training strategy of the model trained by the client device according to the first instruction information; The server is used for: The training progress of each model is determined based on the model measurement information corresponding to the multiple models; the training progress of each model is determined based on the model performance improvement of each model or the weight of each model. Based on the training progress of each model, obtain the first indication information corresponding to each model.

18. A server, characterized in that, The server includes a processor and a communication interface for communicating with multiple client devices, and implements the model training method as described in any one of claims 1 to 8 based on the processor and the communication interface.

19. A client device, characterized in that, The client device includes a processor and a communication interface, the communication interface being used to communicate with the server, and the model training method as described in any one of claims 9 to 14 being implemented according to the processor and the communication interface.

20. A computer-readable storage medium, characterized in that, Includes instructions that, when the computer-readable storage medium is run on a computer, cause the computer to perform the model training method as described in any one of claims 1 to 8 or 9 to 14.

Images