Model training method and device

Abstract

The invention discloses a model training method and device, relates to the technical field of communication, and can enable a third node to obtain a training model with better performance, reduce the transmission overhead in the training process and improve the robustness in the model training process. The method comprises the steps that a third node receives a first model from a first node; testing the first model to obtain a first test result; when the first test result meets a preset test condition, sending first indication information to the first node; wherein the first model is a model trained by the first node; the first indication information is used for indicating the first node to switch to the target training mode.

Description

Model training method and device Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a model training method and device. Background Art

[0002] Artificial intelligence (AI) technology can be applied to the communication processing processes of the network layer and physical layer in communication systems. For example, AI technology can be applied to the model training process in communication to make the model training process more efficient, real-time, and effectively protect user data privacy.

[0003] Among them, the AI ​​node can send the model to be trained to one or more AI sub-nodes, and the one or more AI sub-nodes can train the model to be trained according to a preset training mode and upload the trained model to the AI ​​node.

[0004] However, due to the constant changes in channel conditions in communication scenarios, the transmission overhead during model training can be high, affecting the robustness of the model training process. Therefore, how to enable AI nodes to obtain high-performance training models, reduce the transmission overhead during model training, and improve the robustness of the model training process has become an urgent problem to be solved.

[0005] Summary of the Invention

[0006] The present application provides a model training method and device, which can enable nodes to obtain a training model with better performance, reduce the transmission overhead during the model training process, and improve the robustness of the model training process.

[0007] In a first aspect, a model training method is provided. The method can be executed by a third node, or by a component of the third node, such as a processor, chip, or chip system of the third node. It can also be implemented by a logic module or software that implements all or part of the functions of the third node. The method includes: the third node receiving a first model from a first node; testing the first model to obtain a first test result; and when the first test result meets a preset test condition, sending a first indication message to the first node. The first model is a model trained by the first node; and the first indication message is used to instruct the first node to switch to a target training mode.

[0008] Based on the first aspect, when the first test result meets the preset test conditions, it indicates that the current training mode may not match the first model. If the current training mode continues to be used, the transmission overhead in the model training process may increase and the robustness of the model training process may be reduced; therefore, the third node can dynamically adjust the target training mode that is more suitable for the current model training process according to the first test result, and send a first indication information to the first node to instruct the first node to switch to the target training mode, so that the first node can perform model training according to the dynamically adjusted target training mode, which can improve the efficiency of model training, reduce the transmission overhead in the model training process, and improve the robustness of the model training process.

[0009] In one possible implementation, the first test result satisfies the preset test conditions, including one or more of the following situations: the test accuracy of the first test result is less than or equal to a first preset threshold; or, the first test result indicates that the first model is overfitting; or, the test accuracy of the first test result remains unchanged or decreases within multiple test cycles; or, the difference in network traffic associated with the first test result within multiple test cycles is greater than or equal to a second preset threshold, wherein the network traffic is used to indicate the amount of data used to train the first model; or, the difference in the signal-to-noise ratio associated with the first test result within multiple test cycles is greater than or equal to a third preset threshold; or, the difference in the interference associated with the first test result within multiple test cycles is greater than or equal to a fourth preset threshold.

[0010] Based on this possible implementation, the third node can determine that the current training mode does not match the first model based on one or more of the above situations. Furthermore, the third node can determine the target training mode based on one or more of the above situations, and send a first indication information to the first node to instruct the first node to switch to the target training mode, so that the first node can perform model training according to the target training mode, which can reduce the transmission overhead during the model training process and improve the robustness of the model training process.

[0011] In one possible implementation, the third node sends first configuration information to the first node; wherein the first configuration information includes one or more of the following: hyperparameter information, model test cycle information, and model maximum retained version number information; the model test cycle information is used to indicate the number of model training rounds experienced by the first node when reporting the first model, and the model maximum retained version number information is used to indicate the maximum number of first models saved within a first preset time, and different first models are obtained through different training rounds.

[0012] Based on this possible implementation, the model testing cycle can enable the first node to periodically send the first model to the third node, and the third node can periodically test the first model and promptly instruct the first node to switch to the target training mode when the first test result meets the preset test conditions, so that the first node can better train the first model, reduce the transmission overhead during the model training process, and improve the robustness of the model training process; the maximum retained version number information can enable the first node to save multiple first models, so that the first node can fall back to the previous first model with better performance when an abnormality occurs in the model training, which can save the time of retraining the first model and improve the efficiency of model training.

[0013] In one possible implementation, the third node receives a switching request from the first node, and tests the first model according to the switching request to obtain a first test result, wherein the switching request is used to request switching to the training mode.

[0014] Based on this possible implementation, the third node receives a switching request from the first node, and the third node can test the first model in a timely manner without waiting for a predefined time to measure the first model, which can improve the efficiency of model training; at the same time, the third node tests the first model after receiving the switching request, and can further determine whether it is necessary to switch the training mode, thereby avoiding the situation where the first node makes a misjudgment as much as possible, and improving the accuracy of model training.

[0015] In a possible implementation, the third node receives a switching request from the first node; and sends first indication information to the first node according to the switching request.

[0016] Based on this possible implementation, the third node can directly send the first indication information to the first node according to the switching request, which is used to instruct the first node to switch to the target training mode. The third node does not need to test the first model, which can reduce the overhead of the third node in testing the first model, reduce the memory usage of the third node, save the computing power resources of the third node, and improve the efficiency of information interaction between the first node and the third node.

[0017] In one possible implementation, the switching request is used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease of the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease of the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold.

[0018] Based on this possible implementation, the switching request can indicate one or more of the above. The third node can determine the problems existing in the first model based on the switching request, and then better determine the target training mode that the first node can switch to, and avoid the situation of multiple switching training modes as much as possible, so that the first node can better perform model training and improve the efficiency of model training.

[0019] In one possible implementation, the third node receives training status information from the first node; when the training status information satisfies a first preset condition, the third node sends third indication information to the first node. The training status information includes one or more of the following: training loss, number of training rounds, data set status, or second indication information; the second indication information is used to indicate whether the first node should continue to participate in training; the first preset condition is one or more of the following: training loss is greater than or equal to an eighth preset threshold, the number of training rounds does not match the data set, the error in the data set is greater than or equal to a ninth preset threshold, the amount of data in the data set is less than a tenth preset threshold, or the second indication information is used to indicate that the first node should not continue to participate in training; the third indication information is used to instruct the first node to use the target model learning method for model training.

[0020] Based on this possible implementation, the third node can determine whether the first node can continue to participate in training based on the status information, and can also determine the status of model training based on the training status information; when the training status information meets the first preset condition, the third node can send a third indication information to the first node, used to instruct the first node to switch to the target model learning mode, so that the first node can better perform model training based on the target model learning mode, which can improve the efficiency of model training and enhance the robustness of the model training process.

[0021] In one possible implementation, the third node sends fourth indication information to the first node; wherein the fourth indication information is used to indicate frequency domain information when the first node performs model training.

[0022] Based on this possible implementation, when the first node has a low signal-to-noise ratio or large interference when performing model training in the current frequency band, the third node can send a fourth indication message to the first node to instruct the first node to switch the frequency band, which can improve the signal-to-noise ratio of the first node when performing model training, or effectively reduce the noise, reduce the transmission overhead during the model training process, and improve the robustness of the model training process.

[0023] In one possible implementation, the third node obtains multiple first test results; when the multiple first test results all meet the preset test conditions, the second configuration information associated with the first node is sent to the second node; wherein different first test results are associated with different first models, and different first models are associated with different training modes; the second configuration information is used to indicate one or more of the following: a training task associated with the first node, a training mode associated with the first node, or a training model associated with the first node.

[0024] Based on this possible implementation, when the first node switches the training mode multiple times and the first test results associated with different training modes all meet the preset test conditions, the third node may consider that the first node is not suitable for model training, or the first node is unable to perform model training; the third node can schedule the second node and send the second configuration information associated with the first node to the second node, so that the second node replaces the first node to continue model training, which can improve the efficiency of model training and enhance the robustness of the model training process.

[0025] In a second aspect, a model training method is provided. The method can be executed by a first node, or by a component of the first node, such as a processor, chip, or chip system of the first node. It can also be implemented by a logic module or software that implements all or part of the first node's functions. The method includes: the first node sending a first model to a third node; receiving first indication information from the third node; and switching a training mode to a target training mode based on the first indication information. The first model is a model trained by the first node; and the first indication information is used to instruct the first node to switch to the target training mode.

[0026] Based on the second aspect, the first node can switch to the target training mode according to the first indication information from the third node, and the first node can perform model training according to the target training mode, which can improve the efficiency of model training, reduce the transmission overhead during model training, and improve the robustness of the model training process.

[0027] In one possible implementation, a first node receives first configuration information from a third node and trains a model based on the first configuration information to obtain a first model. The first configuration information includes one or more of the following: hyperparameter information, model testing cycle information, and information about the maximum number of retained model versions. The model testing cycle information indicates the number of model training rounds the first node undergoes when reporting the first model, and the maximum number of retained model versions indicates the maximum number of first models that can be saved within a first preset time period. Different first models are obtained through different training rounds.

[0028] Based on this possible implementation, the first node can periodically send the first model to the third node according to the test cycle of the model, so that the third node can periodically test the first model and promptly instruct the first node to switch to the target training mode when the first test result meets the preset test conditions. The first node can better train the first model, reduce the transmission overhead during the model training process, and improve the robustness of the model training process; the first node can save multiple first models according to the maximum retained version number information. When the first node encounters an abnormality during model training, it can fall back to the first model with better performance before, which can save the time of retraining the first model and improve the efficiency of model training.

[0029] In one possible implementation, the first node measures one or more of the following to obtain a first measurement result: a channel or a training gradient; and sends a handover request to the third node when the first measurement result satisfies a second preset condition. The second preset condition includes one or more of the following: a decrease in the training accuracy of the first model, a change in the training loss of the first model within a second preset time, an increase or decrease in the uplink signal-to-noise ratio within a third preset time that is greater than or equal to a fifth preset threshold, an increase or decrease in the downlink signal-to-noise ratio within a third preset time that is greater than or equal to a fifth preset threshold, an increase or decrease rate of the uplink signal-to-noise ratio within a fourth preset time that is greater than or equal to a sixth preset threshold, an increase or decrease rate of the downlink signal-to-noise ratio within a fourth preset time that is greater than or equal to a sixth preset threshold, or a gradient variance of the first model that is greater than or equal to a seventh preset threshold.

[0030] Based on this possible implementation, the first node can measure the channel or training gradient, monitor the changes in training accuracy and gradient variance in real time, and monitor the impact of noise and interference on model training in real time; when the first measurement result meets the second preset condition, the first node can send a switching request to the third node, requesting to switch the training mode. The first node can perform model training based on the switched training model, which can improve the efficiency of model training and improve the robustness of the model training process.

[0031] In one possible implementation, the switching request is used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease of the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease of the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold.

[0032] Based on this possible implementation, the switching request can indicate one or more of the above, so that the third node can determine the problems existing in the first model according to the switching request, and then better determine the target training mode that the first node can switch to, and avoid the situation of multiple switching training modes as much as possible. The first node can better perform model training according to the target training mode, which can improve the efficiency of model training.

[0033] In one possible implementation, the first node sends training status information to the third node; wherein the training status information includes one or more of the following: training loss, number of training rounds, data set status, or second indication information; the second indication information is used to indicate whether the first node continues to participate in training.

[0034] Based on this possible implementation, the first node sends training status information to the third node, which can enable the third node to determine whether the first node can continue to participate in training based on the training status information, or enable the third node to determine the status of model training based on the training status information; further, the third node can instruct the first node to switch to the target model learning mode based on the training status information, and the first node can perform model training according to the target model learning mode, which can improve the efficiency of model training and enhance the robustness of the model training process.

[0035] In one possible implementation, the first node receives third instruction information from the third node and switches the model learning mode to the target model learning mode according to the third instruction information, wherein the third instruction information is used to instruct the first node to use the target model learning mode for model training.

[0036] Based on this possible implementation, the first node can switch to the target model learning mode according to the third indication information from the third node. The first node can better perform model training based on the target model learning mode and improve the robustness of the model training process.

[0037] In one possible implementation, fourth indication information is received from a third node, and frequency domain information for model training is determined based on the fourth indication information, wherein the fourth indication information is used to indicate the frequency domain information for model training by the first node.

[0038] Based on this possible implementation, the first node can switch the frequency band of model training according to the fourth indication information from the third node, which can improve the signal-to-noise ratio during model training, or can effectively reduce noise, reduce the transmission overhead during model training, and improve the robustness of the model training process.

[0039] In a third aspect, a communication device is provided for implementing the method described in the first aspect. The communication device may be the third node described in the first aspect, or a device or component included in the third node, such as a chip.

[0040] The communication device includes modules, units, or means corresponding to the above-mentioned method, which can be implemented by hardware, software, or hardware executing corresponding software implementation. The hardware or software includes one or more modules or units corresponding to the above-mentioned functions.

[0041] In some possible implementations, the communication device may include a processing module and a transceiver module. The transceiver module may include a sending module and a receiving module, respectively used to implement the sending and receiving functions of the first aspect and any possible implementation thereof. The transceiver module is used to receive a first model from a first node; wherein the first model is a model trained by the first node. The processing module is used to test the first model and obtain a first test result. The transceiver module is also used to send a first indication message to the first node when the first test result meets a preset test condition; wherein the first indication message is used to instruct the first node to switch to a target training mode.

[0042] Optionally, the transceiver module and processing module of the communication device in the third aspect can also perform the corresponding functions in the above-mentioned first aspect or any possible implementation of the first aspect. Please refer to the detailed description in the method example for details. The beneficial effects that can be achieved can also be referred to the above-mentioned related content.

[0043] In a fourth aspect, a communication device is provided, configured to implement the method described in the second aspect. The communication device may be the first node described in the second aspect, or a device or component included in the first node, such as a chip.

[0044] The communication device includes modules, units, or means corresponding to the above-mentioned method, which can be implemented by hardware, software, or hardware executing corresponding software implementation. The hardware or software includes one or more modules or units corresponding to the above-mentioned functions.

[0045] In some possible implementations, the communication device may include a processing module and a transceiver module. The transceiver module may include a sending module and a receiving module, respectively used to implement the sending and receiving functions in the above-mentioned second aspect and any possible implementation thereof. The processing module may be used to implement the processing functions in the above-mentioned second aspect and any possible implementation thereof. Exemplarily, the transceiver module is used to send a first model to a third node; wherein the first model is a model trained by the first node. The transceiver module is also used to receive first indication information from the third node; wherein the first indication information is used to instruct the first node to switch to a target training mode. The processing module is used to switch the training mode to the target training mode according to the first indication information.

[0046] Optionally, the transceiver module and processing module of the communication device in the fourth aspect can also perform the corresponding functions in the above-mentioned second aspect or any possible implementation of the second aspect. Please refer to the detailed description in the method example for details. The beneficial effects that can be achieved can also be referred to the above-mentioned related content.

[0047] In a fifth aspect, a communication device is provided, comprising: at least one processor configured to cause the communication device to perform the method described in any one of the above aspects or any possible implementations of any one of the aspects by executing computer instructions stored in a memory or through a logic circuit. The communication device may be the third node in the first aspect or any possible implementation of the first aspect, or a device or component included in the third node, such as a chip; or the communication device may be the first node in the second aspect or any possible implementation of the second aspect, or a device or component included in the first node, such as a chip.

[0048] In some possible implementations, the communication device further includes a memory for storing computer instructions and / or configuration files of logic circuits. Optionally, the memory is integrated with the processor, or the memory is independent of the processor.

[0049] In a sixth aspect, a communication device is provided, comprising: a processor and a communication interface; the communication interface is configured to input and / or output signals; and the processor is configured to execute a computer program or instruction to cause the communication device to perform the method described in any of the above aspects. The communication device may be a third node in the first aspect or any possible implementation of the first aspect, or a device or component included in the third node, such as a chip; or the communication device may be a first node in the second aspect or any possible implementation of the second aspect, or a device or component included in the first node, such as a chip.

[0050] In some possible implementations, the communication interface is an interface circuit for reading and writing computer instructions. For example, the interface circuit is used to receive computer execution instructions (computer execution instructions are stored in a memory, may be read directly from the memory, or may pass through other devices) and transmit them to the processor.

[0051] In some possible implementations, the communication interface is used to communicate with a module outside the communication device.

[0052] In some possible implementations, the communication device may be a chip or a chip system. When the device is a chip system, the chip system may include the chip, or may include the chip and other discrete devices.

[0053] In a seventh aspect, a communication device is provided, comprising: a logic circuit and an interface circuit; the interface circuit is configured to input and / or output information; and the logic circuit is configured to execute the method described in any of the above aspects, processing the input information and / or generating output information. The communication device may be a third node in the first aspect or any possible implementation of the first aspect, or a device or component included in the third node, such as a chip; or the communication device may be a first node in the second aspect or any possible implementation of the second aspect, or a device or component included in the first node, such as a chip.

[0054] In an eighth aspect, a computer-readable storage medium is provided, in which a computer program or instruction is stored. When the computer program or instruction is executed by a processor, the method described in any one of the above aspects is executed.

[0055] In a ninth aspect, a computer program product is provided, which, when executed by a processor, enables the method described in any one of the above aspects to be executed.

[0056] It can be understood that when the communication device provided in any one of the third to seventh aspects is a chip, the above-mentioned sending action / function can be understood as output information, and the above-mentioned receiving action / function can be understood as input information.

[0057] Among them, the technical effects brought about by any implementation method from the third aspect to the ninth aspect can refer to the technical effects brought about by the above-mentioned first aspect or any possible implementation of the first aspect, or refer to the technical effects brought about by the above-mentioned second aspect or any possible implementation of the second aspect, and will not be repeated here.

[0058] In a tenth aspect, a communication system is provided, which includes the third node described in the first aspect or any possible implementation of the first aspect and the first node described in the second aspect or any possible implementation of the second aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] FIG1 is a schematic diagram of a gradient transparent transmission provided in an embodiment of the present application;

[0060] FIG2 is a schematic diagram of a communication system provided in an embodiment of the present application;

[0061] FIG3 is a schematic structural diagram of a communication device provided in an embodiment of the present application;

[0062] FIG4 is an interactive schematic diagram of a model training method provided in an embodiment of the present application;

[0063] FIG5 is an interactive diagram of a model training method provided in an embodiment of the present application;

[0064] FIG6 is an interactive diagram of a model training method provided in an embodiment of the present application;

[0065] FIG7 is an interactive diagram of a model training method provided in an embodiment of the present application;

[0066] FIG8 is an interactive schematic diagram of a model training method provided in an embodiment of the present application;

[0067] FIG9 is an interactive diagram of a model training method provided in an embodiment of the present application;

[0068] FIG10 is an interactive diagram of a model training method provided in an embodiment of the present application;

[0069] FIG11 is a schematic structural diagram of a third node provided in an embodiment of the present application;

[0070] FIG12 is a schematic structural diagram of a first node provided in an embodiment of the present application;

[0071] FIG13 is a schematic structural diagram of another communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0072] The following describes in detail the implementation of the embodiments of the present application in conjunction with the accompanying drawings.

[0073] In the description of this application, unless otherwise specified, " / " indicates that the objects associated before and after are in an "or" relationship, for example, A / B can represent A or B; "and / or" in this application is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural.

[0074] In the description of this application, unless otherwise specified, "plurality" means two or more than two. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or plural.

[0075] In addition, to facilitate the clear description of the technical solutions of the embodiments of the present application, in the embodiments of the present application, the words "first" and "second" are used to distinguish between identical or similar items with substantially the same functions and effects. Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or execution order, and the words "first" and "second" do not necessarily mean different.

[0076] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner to facilitate understanding.

[0077] It will be understood that the “embodiment” mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, the various embodiments throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It will be understood that in the various embodiments of the present application, the size of the sequence number of each process does not mean 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 embodiment of the present application.

[0078] It is understood that some optional features in the embodiments of the present application may, in certain scenarios, be implemented independently of other features, such as the solution on which they are currently based, to solve corresponding technical problems and achieve corresponding effects. They may also be combined with other features in certain scenarios as needed. Accordingly, the devices provided in the embodiments of the present application may also implement these features or functions accordingly, which will not be described in detail here.

[0079] In this application, unless otherwise specified, the same or similar parts between the various embodiments can be referenced to each other. In the various embodiments of this application, unless otherwise specified and there is no logical conflict, the terms and / or descriptions between different embodiments are consistent and can be referenced to each other. The technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships. The following description of the embodiments of this application does not constitute a limitation on the scope of protection of this application.

[0080] In order to facilitate understanding of the technical solutions of the embodiments of the present application, a brief introduction to the relevant technologies of the present application is first given as follows.

[0081] 1) AI technology

[0082] AI technology is widely used in image processing and natural language processing, and has been very successful in these areas. The increasing maturity of AI technology can play a significant role in promoting the evolution of future mobile communication network technologies.

[0083] On the one hand, AI technology can be applied to the network layer (such as network optimization, mobility management, resource allocation, etc.) and physical layer (such as channel coding, channel prediction, receiver, etc.) in communication networks.

[0084] On the other hand, with the deepening of research on deep learning models in AI technology, deep learning models can complete more complex tasks and achieve better performance.

[0085] For example, the deep learning model can be a bidirectional encoder representation from transformers (BERT), which is mainly used for natural language processing; the deep learning model can also be a generation pre-trained transformer (GPT), which is mainly used in the deep learning model architecture of the self-attention mechanism.

[0086] However, deep learning models may include a large number of parameters, and during model training or model inference, a large amount of computing resources (such as computer resources for calculation and processing tasks) are required. When general nodes undertake the training or inference of such models, the timeliness may be low.

[0087] Therefore, for the next generation of wireless communications, communication networks can expand traditional access services to the computing power, data, and algorithm levels, and can provide a variety of support capabilities for AI technology, making model training or model reasoning based on AI technology more efficient, real-time, and able to effectively protect user data privacy.

[0088] 2) Gradient Transmission

[0089] In the patent document with application number 202110412895.X (A gradient transmission method and related device), gradient transparent transmission is a method of air interface online training, which is a training mode for model training. As shown in Figure 1 below, gradient transparent transmission can transmit gradient information between two or more AI sub-nodes, thereby achieving end-to-end optimization.

[0090] Among them, AI sub-node 1 sends training data to AI sub-node 2 based on forward reasoning; accordingly, AI sub-node 2 can determine the reverse gradient based on the training data from AI sub-node 1, and send the reverse gradient and control information to AI sub-node 1. At the same time, AI sub-node 2 can update the parameters of the training model based on the training data, so as to better perform model training; AI sub-node 1 can perform channel estimation based on the control information to obtain channel information, and then based on the channel information and reverse gradient, it can more accurately update the parameters of the training model, so as to better perform model training.

[0091] The process in which AI sub-node 1 sends data and AI sub-node 2 receives and recovers the data is called the forward reasoning process of the neural network. AI sub-node 1 can use the beamforming method to assign training data to a beam and send it to AI sub-node 2. Similarly, AI sub-node 2 can use the same method to send reverse gradients to AI sub-node 1.

[0092] Optionally, gradient passthrough can exploit the reciprocity of the channel to simultaneously perform gradient calculation and backpropagation.

[0093] It is understandable that gradient passthrough can send complex symbols during backpropagation, without quantizing, encoding, modulating, inserting pilot signals, and other operations on the sent data / gradients. This can simplify the design of the physical layer and reduce the transmission overhead during model training. However, the function of gradient passthrough is a loss function, which may reduce the accuracy of the training model.

[0094] In a communication system, an AI node can send a model to be trained to one or more AI sub-nodes. The one or more AI sub-nodes can train the model to be trained according to a preset training mode (such as gradient passthrough) and upload the trained model to the AI ​​node.

[0095] When the channel state changes, it may cause large transmission overhead during training and affect the robustness of the model training process.

[0096] In summary, how to enable AI nodes to obtain training models with better performance, reduce transmission overhead during training, and improve the robustness of the model training process has become an urgent problem to be solved.

[0097] In order to solve the above technical problems, the present application provides a model training method, in which a third node can receive a first model from a first node; test the first model to obtain a first test result; when the first test result meets a preset test condition threshold, send a first indication information to the first node; wherein the first model is a model trained by the first node; the first indication information is used to instruct the first node to switch to a target training mode.

[0098] In an embodiment of the present application, when the first test result meets the preset test conditions, it indicates that the current training mode may not match the first model. If the current training mode continues to be used, the transmission overhead in the model training process may increase and the robustness of the model training process may be reduced; therefore, the third node can dynamically adjust the target training mode that is more suitable for the current model training process according to the first test result, and send a first indication information to the first node to instruct the first node to switch to the target training mode, so that the first node can perform model training according to the dynamically adjusted target training mode, which can improve the efficiency of model training, reduce the transmission overhead in the model training process, and improve the robustness of the model training process.

[0099] The technical solutions of the embodiments of the present application can be used in various communication systems, which may be third generation partnership project (3GPP) communication systems, for example, fourth generation (4G), long term evolution (LTE), fifth generation (5G) mobile communication systems, new radio / new air interface (NR), or a system of hybrid networking of LTE and 5G, or a non-terrestrial network (NTN) system, or a mobile communication system evolved after 5G such as the sixth generation (6G), a vehicle to everything (V2X) system, or a device-to-device (D2D) communication system, a machine to machine (M2M) communication system, the Internet of Things (IoT), a narrowband Internet of Things (NB-IoT), other next-generation communication systems, integrated perception and communication systems, satellite communication systems, and the like. The communication system may also be a non-3GPP communication system, such as a wireless local area network (WLAN) system such as wireless fidelity (Wi-Fi), without limitation.

[0100] For example, as shown in Figure 2, a schematic diagram of the structure of a communication system provided by the present application is shown. The communication system may include one or more AI nodes and one or more AI sub-nodes.

[0101] Among them, the AI ​​node can send the model to be trained to one or more AI sub-nodes, and the one or more AI sub-nodes can train the model to be trained and upload the trained model to the AI ​​node.

[0102] Among them, the AI ​​node or AI sub-node can be a terminal device, or a network device or other device that can implement the model training process, without restriction.

[0103] This application can be applied to various communication scenarios, such as synchronization, channel estimation, signal detection, beam measurement, etc.

[0104] The above-mentioned communication systems and communication scenarios applicable to the present application are merely examples. The communication systems and communication scenarios applicable to the present application are not limited thereto, and the above description does not impose any limitation on the solutions of the present application.

[0105] Among them, the terminal device in the embodiment of the present application can be located within the beam / cell coverage of the network device, and the network device can provide communication services for the terminal device.

[0106] The terminal device in the embodiments of the present application may be a device with wireless transceiver functions or a chip or chip system that can be set in the device, which can allow a user to access the network and is a device for providing voice and / or data connectivity to the user. The terminal device may also be referred to as user equipment (UE), subscriber unit (subscriber unit), terminal (terminal), mobile station (MS), or mobile terminal (MT).

[0107] Optionally, the terminal device in the embodiment of the present application may be a user-side device for implementing wireless communication functions, such as a terminal or a chip that can be used in a terminal. The terminal may be a user equipment (UE), an access terminal, a terminal unit, a terminal station, a mobile station, a mobile station, a remote station, a remote terminal, a mobile device, a wireless communication device, a terminal agent, or a terminal device in a 5G network or a public land mobile network (PLMN) evolved after 5G. The access terminal can be a cellular phone, a cordless phone, a Session Initiation Protocol (SIP) phone, a wireless local loop (WLL) station, a personal digital assistant (PDA), a handheld device with wireless communication capabilities, a computing device or other processing device connected to a wireless modem, an in-vehicle device, a drone, a robot, a smart point of sale (POS) machine, customer-premises equipment (CPE) or a wearable device, a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, a wireless terminal in industrial control, a wireless terminal in self-driving, a wireless terminal in remote medical, a wireless terminal in a smart grid, a wireless terminal in transportation safety, a wireless terminal in a smart city, a wireless terminal in a smart home, etc. Alternatively, the terminal may be a terminal with communication functionality in IoT, such as a terminal in V2X (e.g., a vehicle-to-everything (V2X) device), a terminal in D2D communication, or a terminal in M2M communication. The terminal may be mobile or fixed.

[0108] The network device in the embodiments of the present application can be any device deployed in an access network that can communicate wirelessly with a terminal device, or a chip or chip system that can be provided in the above-mentioned device, or a logical node or a logical module or a function implemented in software, and can be used to implement wireless physical control functions, resource scheduling and wireless resource management, wireless access control, and mobility management functions. Specifically, the network device can be a device that supports wired access or a device that supports wireless access.

[0109] Optionally, the network device in the embodiment of the present application is a device that connects a terminal device to a wireless network. The network device may be a node in a radio access network (RAN), or may be a base station, which may be referred to as a radio access network node (or device).

[0110] For example, the network device may include an evolved NodeB (eNB) or e-NodeB in an LTE system or an enhanced LTE (LTE-Advanced, LTE-A) system, such as a traditional macro eNB and a micro eNB in ​​a heterogeneous network scenario. Alternatively, it may include a next-generation NodeB (gNB) in an NR system. Alternatively, it may include a transmission reception point (TRP), a home base station (e.g., a home evolved NodeB, or HNB), a baseband unit (BBU), a baseband pool (BBU pool), or a wireless fidelity (WiFi) access point (AP). Alternatively, it may include a base station in an NTN, which may be deployed on an aircraft or a satellite. In the NTN, the network device may function as a Layer 1 (L1) relay, a base station, or an integrated access and backhaul (IAB) node. Alternatively, the network device may be a device that implements a base station function in IoT, such as a device that implements a base station function in drone communications, V2X, D2D, or machine to machine (M2M).

[0111] A network device may also be a module or unit that implements some of the functions of a base station. For example, a network device may be a centralized unit (CU), a distributed unit (DU), a CU-control plane (CP), a CU-user plane (UP), or a radio unit (RU). The CU and DU may be configured separately or included in the same network element, such as a baseband unit (BBU). The RU may be included in a radio frequency device or radio frequency unit, such as a remote radio unit (RRU), an active antenna unit (AAU), or a remote radio head (RRH).

[0112] In different systems, CU (or CU-CP and CU-UP), DU or RU may also have different names, but those skilled in the art can understand their meanings. For example, the network device may be a network device or a module of a network device in an open radio access network (open RAN, ORAN) system. In the ORAN system, CU may also be referred to as open (open, O)-CU, DU may also be referred to as O-DU, CU-CP may also be referred to as O-CU-CP, CU-UP may also be referred to as O-CU-UP, and RU may also be referred to as O-RU. Any of the CU (or CU-CP, CU-UP), DU and RU in this application may be implemented by a software module, a hardware module, or a combination of a software module and a hardware module.

[0113] Optionally, the base station in the embodiment of the present application may include various forms of base stations, such as: macro base stations, micro base stations (also called small stations), relay stations, access points, home base stations, TRPs, transmitting points (TPs), or mobile switching centers, etc. The embodiment of the present application does not make specific limitations on this.

[0114] It should be noted that the communication system described in the embodiment of the present application is intended to more clearly illustrate the technical solution of the embodiment of the present application, and does not constitute a limitation on the technical solution provided in the embodiment of the present application. Ordinary technicians in this field can know that with the evolution of network architecture and the emergence of new business scenarios, the technical solution provided in the embodiment of the present application is also applicable to similar technical problems.

[0115] In specific implementations, the AI ​​node and one or more AI sub-nodes shown in Figure 2 can adopt the structure shown in Figure 3, or include the components shown in Figure 3. Figure 3 is a schematic diagram of the structure of a communication device 30 provided in an embodiment of the present application. The communication device 30 can be a chip or system-on-chip of an AI node and an AI sub-node; it can also be a chip or system-on-chip in an AI node and an AI sub-node; it can also be a chip or system-on-chip in an AI node and an AI sub-node.

[0116] As shown in FIG3 , the communication device 30 includes one or more processors 301. Furthermore, the communication device 30 may also include a communication bus 302 and at least one communication interface ( FIG3 is merely exemplary, illustrating the communication device 30 including a communication interface 304 and one processor 301). Optionally, the communication device 30 may also include a memory 303.

[0117] Processor 301 can be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the program of the present application, or a processing core for processing data (e.g., computer program instructions). The processor can be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor.

[0118] In a specific implementation, as an embodiment, the processor 301 may include one or more CPUs, such as CPU0 and CPU1 in FIG. 3 .

[0119] Communication bus 302 may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, for example. This bus may be classified as an address bus, a data bus, a control bus, or the like. For ease of illustration, FIG3 shows only one thick line, but this does not imply a single bus or type of bus. Communication bus 302 is used to connect the various components within communication device 30, enabling communication and interaction between the various components within communication device 30.

[0120] The communication interface 304 may be a transceiver module for communicating with other devices or a communication network, such as Ethernet, a radio access network (RAN), or a wireless local area network (WLAN). For example, the communication interface 304 may be a device such as a transceiver or a transceiver. Alternatively, the communication interface 304 may be a transceiver circuit within the processor 301, configured to implement signal input and output to the processor.

[0121] The memory 303 may be a device having a storage function. For example, it may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory may exist independently and be connected to the processor via the communication bus 302. The memory may also be integrated with the processor.

[0122] Exemplarily, the memory 303 is used to store computer-executable instructions for executing the solution of the present application, and the execution is controlled by the processor 301. The processor 301 is used to execute the computer-executable instructions stored in the memory 303, thereby implementing the method provided in the embodiment of the present application.

[0123] Alternatively, optionally, in an embodiment of the present application, the processor 301 may also perform processing-related functions in the method provided in the following embodiments of the present application, and the communication interface 304 is responsible for communicating with other devices or communication networks, which is not specifically limited in the embodiments of the present application.

[0124] Optionally, the computer-executable instructions in the embodiments of the present application may also be referred to as application code, which is not specifically limited in the embodiments of the present application.

[0125] In a specific implementation, as an embodiment, the communication device 30 may further include an output device 305 and an input device 306. The output device 305 communicates with the processor 301 and can display information in a variety of ways. For example, the output device 305 can be a liquid crystal display (LCD), a light emitting diode (LED) display device, a cathode ray tube (CRT) display device, or a projector. The input device 306 communicates with the processor 301 and can receive user input in a variety of ways. For example, the input device 306 can be a mouse, a keyboard, a touch screen device, or a sensor device.

[0126] It should be noted that the composition structure shown in Figure 3 does not constitute a limitation on the communication device. In addition to the components shown in Figure 3, the communication device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0127] The model training method provided by the embodiment of the present application will be described below in conjunction with the accompanying drawings. It will be understood that in the embodiment of the present application, the third node (any AI node that issues the model training task) and the first node (any AI sub-node participating in the model training process) can perform some or all of the steps in the embodiment of the present application, and these steps or operations are only examples. The embodiment of the present application can also perform other operations or variations of various operations. In addition, the various steps can be performed in a different order presented in the embodiment of the present application, and it is possible that not all operations in the embodiment of the present application need to be performed.

[0128] As shown in Figure 4, it is an interactive diagram of a model training method provided by this application. Referring to Figure 4, the model training method includes the following steps:

[0129] S401: A first node sends a first model to a third node; correspondingly, the third node receives the first model from the first node.

[0130] The first model is the model after the first node is trained.

[0131] Among them, the third node can be a node used to send model training tasks to one or more first nodes, such as the above-mentioned AI node; the first node can be a node that trains the model according to the model training task, such as any AI sub-node participating in model training.

[0132] It is understandable that different first models can be obtained by the first node through different training rounds and different training modes.

[0133] Among them, the third node can send a model training task to the first node, and the first node performs one or more rounds of training on the model according to the model training task. Each time the first node performs a round of model training, a first model can be obtained.

[0134] Optionally, the third node may further indicate a training mode to the first node, and the first node performs model training according to the training mode indicated by the third node.

[0135] The training modes corresponding to different rounds can be the same or different without restriction.

[0136] Optionally, before the first node sends the first model to the third node, the third node may send first configuration information to the first node, and the first node may report the first model to the third node based on the received first configuration information.

[0137] Among them, the first configuration information may include one or more of the following: hyperparameter information, model test cycle information, and maximum retained version number information of the model.

[0138] The super-parameter information may be a super-parameter shared by or unique to each training mode.

[0139] Exemplarily, hyperparameter information may include one or more of the following: batch size (the larger the batch size, the faster the model training speed, the larger the memory usage, and the faster the convergence speed), number of training rounds, or learning rate (the speed at which the model weights are updated).

[0140] The model test cycle information may be used to indicate the number of model training rounds that the first node experienced when reporting the first model.

[0141] For example, taking the model test cycle information as 3 as an example, the first node can report the first model obtained after the third round of model training to the third node after three rounds of model training; or, taking the model test cycle information as 5 as an example, the first node can report the first model obtained after the fifth round of model training to the third node after five rounds of model training.

[0142] Optionally, the specific number of rounds of model training that the first node indicated by the model test cycle information undergoes when reporting the first model can be determined according to an actual communication scenario or actual communication requirement.

[0143] The maximum retained version number information of the model can be used to indicate the maximum number of first models saved within the first preset time, and different first models are obtained through different training rounds.

[0144] The first preset time may be determined according to actual communication scenarios or actual communication requirements.

[0145] Optionally, the performance of the saved first model is better than the performance of the unsaved first model, or the generalization of the saved first model is better than the generalization of the unsaved first model.

[0146] In a first possible implementation, the first node can obtain a first model when each round of model training is completed. The first node can save each obtained first model according to the model performance (or the generalization of the model), such as giving priority to saving the first model with better performance (or better generalization).

[0147] For example, taking the maximum number of retained versions of the model indicating that the maximum number of first models saved within the first preset time is 3 as an example, the first node can save three first models during the first three rounds of model training. When the first node completes the fourth round of model training, the first node can compare the performance of the fourth first model obtained with the three saved first models. If the performance of the fourth first model is better than at least one of the three saved first models, the fourth first model can replace the first model with the worst performance among the three saved first models to complete the saving. Similarly, each first model obtained by the first node during subsequent model training can be saved accordingly with reference to the saving operation of the fourth first model, which will not be elaborated on.

[0148] In a second possible implementation, the first node may also report the first model to the third node with reference to S401, and the third node may test the first model with reference to the following S402 to obtain a first test result. If the first test result does not meet the preset test condition, the third node may instruct the first node to save the reported first model. When the first node saves the first model indicated by the third node, if the number of saved first models does not exceed the maximum number of first models saved within the first preset time, the first model indicated by the third node may be directly saved. If the number of saved first models is equal to the maximum number of first models saved within the first preset time, it may be determined whether the performance (or generalization) of the first model indicated by the third node is better than at least one of the saved first models. If so, the first model indicated by the third node may replace the first model with the worst performance (or worst generalization) among the saved first models. Otherwise, the first model indicated by the third node is not saved.

[0149] The description of the third node determining the first test result according to the first model and whether the first test result meets the preset test condition can refer to the specific description of the following S402 and the following S403, which will not be repeated here.

[0150] For example, taking the example where the maximum number of retained versions of a model indicates that the maximum number of first models saved within the first preset time is 3, when the first node saves the first model indicated by the third node, if the number of saved first models does not exceed 3, the first node can directly save the first model indicated by the third node; if the number of saved first models is equal to 3, and the first node determines that the performance of the first model indicated by the third node is better than at least one of the saved first models, the first node can replace the first model indicated by the third node with the worst performance among the saved first models. Similarly, when the number of saved first models is equal to 3, when the third node indicates the first model saved by the first node, the first node can judge the model performance and determine whether to save the first model indicated by the third node based on the judgment result. This will not be elaborated on here.

[0151] S402: The third node tests the first model to obtain a first test result.

[0152] The third node may test the first model according to the test set to obtain a first test result.

[0153] Among them, the test set can be used to evaluate the prediction performance of the test model.

[0154] Illustratively, the first test result may be one or more of the following: test accuracy, signal-to-noise ratio, interference, network traffic, or model complexity.

[0155] S403. When the first test result meets the preset test condition, the third node sends first indication information to the first node; correspondingly, the first node receives the first indication information from the third node.

[0156] The first indication information is used to instruct the first node to switch to the target training mode.

[0157] S404: The first node switches the training mode to the target training mode according to the first instruction information.

[0158] Optionally, the first test result meeting the preset test condition may include one or more of the following situations:

[0159] 1) The test accuracy of the first test result is less than or equal to a first preset threshold.

[0160] The first preset threshold may be determined according to actual communication scenarios or actual communication requirements.

[0161] Exemplarily, when the test accuracy of the first test result is less than or equal to the first preset threshold, it indicates that the performance of the trained first model has not reached the expected performance. If continuing to use the current training mode cannot obtain a first model with better performance, at this time, the third node can instruct the first node to switch the target training mode to improve the performance of the trained first model and improve the model training efficiency.

[0162] 2) The first test result indicates that the first model is overfitting.

[0163] Among them, when the order of magnitude of the training set does not match the complexity of the first model, or the feature distribution of the training set and the test set is inconsistent, the first model may be overfitted.

[0164] Among them, the training set can be used to estimate the parameters in the training model, so that the training model can reflect reality and then predict the future or other unknown information.

[0165] 3) The test accuracy of the first test result remains unchanged or decreases over multiple test cycles.

[0166] Among them, if the test accuracy of the first test result remains unchanged within multiple test cycles, it indicates that the performance of the first model after training has not improved. If continuing to use the current training mode cannot obtain a first model with better performance, at this time, the third node can instruct the first node to switch the target training mode to improve the performance of the first model after training and improve the model training efficiency.

[0167] Alternatively, if the test accuracy of the first test result decreases within multiple test cycles, it indicates that the performance of the trained first model is getting worse and worse. If continuing to use the current training mode cannot obtain a first model with better performance, at this time, the third node can instruct the first node to switch the target training mode to improve the performance of the trained first model and improve the model training efficiency.

[0168] Alternatively, if the test accuracy of the first test result is lower than a preset threshold within multiple test cycles, it indicates that the performance of the trained first model has not reached the expected performance. If continuing to use the current training mode cannot obtain a first model with better performance, at this time, the third node can instruct the first node to switch the target training mode to improve the performance of the trained first model and improve the model training efficiency.

[0169] Exemplarily, taking three test cycles as an example, the third node tests the received first model within the three test cycles and obtains three first test results. When the accuracy of the three first test results remains unchanged, decreases, or is lower than a preset threshold, the third node can send a first indication message to the first node; accordingly, after receiving the first indication message, the first node can switch the training mode to the target training mode, and further, the first node can perform model training according to the target training mode.

[0170] 4) The difference in network traffic associated with the first test result within multiple test cycles is greater than or equal to a second preset threshold.

[0171] The network traffic is used to indicate the amount of data used to train the first model, and the second preset threshold can be determined according to the actual communication scenario or actual communication requirements.

[0172] Among them, if the difference in the network traffic in the first model within multiple test cycles is greater than or equal to the second preset threshold, it indicates that the network traffic of the first model has changed significantly. When the network traffic becomes larger, if the current training mode is continued to be used, it may not be able to support the network traffic, affecting the performance of the first model; when the network traffic becomes smaller, if the current training mode is continued to be used, resources may be wasted. Therefore, the third node can instruct the first node to switch the target training mode so that the target training model can be suitable for situations with large network traffic to improve the performance of the trained first model, or, make the target training mode more suitable for situations with small network traffic to save resources and improve the efficiency of model training.

[0173] Exemplarily, taking three test cycles as an example, the third node can test the received first model within three cycles, and can determine the network traffic of the first model associated with the three test cycles. When the difference between the network traffic of any two test cycles in the three test cycles is greater than or equal to the second preset threshold (such as the difference between the network traffic of the first model associated with the first test cycle and the network traffic of the first model associated with the second test cycle, or the difference between the network traffic of the first model associated with the first test cycle and the network traffic of the first model associated with the third test cycle, or the difference between the network traffic of the first model associated with the second test cycle and the network traffic of the first model associated with the third test cycle), the third node can instruct the first node to switch the target training mode to improve the performance of the trained first model and improve the model training efficiency.

[0174] 5) A difference in the signal-to-noise ratio associated with the first test result within a plurality of test cycles is greater than or equal to a third preset threshold.

[0175] The third preset threshold may be determined according to actual communication scenarios or actual communication requirements.

[0176] Among them, if the difference in the signal-to-noise ratio in the first model within multiple test cycles is greater than or equal to the third preset threshold, it indicates that the noise is large, which may affect the performance of the first model. If the current training mode is continued to be used, it may not be possible to obtain a first model with better performance. At this time, the third node can instruct the first node to switch the target training mode to reduce noise and improve the signal-to-noise ratio, thereby improving the performance of the trained first model and improving the model training efficiency.

[0177] Exemplarily, taking three test cycles as an example, when the increase in the signal-to-noise ratio of any two test cycles is greater than or equal to the third preset threshold, the third node can instruct the first node to switch the target training mode to reduce noise and improve the signal-to-noise ratio, thereby improving the performance of the trained first model and improving the model training efficiency.

[0178] 6) A difference in interference associated with the first test result within multiple test cycles is greater than or equal to a fourth preset threshold.

[0179] The interference may be interference in the network, such as interference caused by other nodes to the first node.

[0180] The fourth preset threshold may be determined according to actual communication scenarios or actual communication requirements.

[0181] Among them, if the difference in the interference associated with the first test result within multiple test cycles is greater than or equal to the fourth preset threshold, it indicates that the interference changes greatly during the model training process, which may affect the performance of the first model. If the current training mode continues to be used, it may not be possible to obtain a first model with better performance. At this time, the third node can instruct the first node to switch the target training mode to improve the performance of the trained first model and improve the efficiency of model training.

[0182] Exemplarily, taking three test cycles as an example, when the difference in interference between any two test cycles is greater than or equal to a fourth preset threshold, the third node can instruct the first node to switch the target training mode to improve the performance of the trained first model and improve the model training efficiency.

[0183] It can be understood that when the first test result satisfies any one of the above conditions, or any several conditions, the third node may send the first indication information to the first node.

[0184] For example, when the first test result indicates that the first model is overfitting, and the test accuracy of the first test result remains unchanged or decreases within multiple test cycles, the third node sends a first indication message to the first node, instructing the first node to switch to the gradient pass-through training mode for model training, which can reduce the transmission overhead during the training process. Alternatively, when the difference in the signal-to-noise ratio associated with the first test result within multiple test cycles is greater than or equal to a third preset threshold, and the difference in the interference associated with the first test result within multiple test cycles is greater than or equal to a fourth preset threshold, the third node sends a first indication message to the first node, instructing the first node to switch to the traditional training mode (that is, the forward propagation data and the backward propagation gradient can be quantized into digital bit streams, and the digital bit streams can be sent in a coded modulation manner) for model training.

[0185] Optionally, the first node may send capability information and the training modes supported by the first node to the third node. During initial configuration, the third node may determine the training mode based on the capability information and the training modes supported by the first node, and instruct the first node to use the training mode for model training. When the first test result meets the preset test conditions, the third node may determine the target training mode based on the first test result, the capability information, and the training modes supported by the first node, and instruct the first node to switch to the target training mode; or, when the first test result meets the preset test conditions, the third node may select any other training mode from the training modes supported by the first node as the target training mode, and instruct the first node to switch to the target training mode.

[0186] Among them, the capability information is used to indicate the memory, computing power resources, etc. of the first node; the training mode supported by the first node is used to indicate the training mode that the first node can adopt, and the third node can select an appropriate training mode as the target training mode based on this.

[0187] Optionally, the first indication information may also instruct the first node to roll back the model version, which may save the first node time in retraining the first model and improve the efficiency of model training.

[0188] For example, when the third node finds that the first model has an abnormality (such as abnormal factor distribution, violent fluctuations in loss value, etc.), it can send fifth indication information to the first node to instruct the first node to adopt the saved first model.

[0189] When the third node sends the fifth indication information to the first node, the fifth indication information may be carried in the first indication information, or may be sent independently.

[0190] For example, the first node can adopt the first model with the best performance among the stored first models and continue model training; or, the first node can send the stored first model to the third node, the third node measures the first model stored by the first node, determines the target first model based on the measurement results, and instructs the first node to adopt the target first model, and the first node continues model training using the target first model.

[0191] It should be noted that the above method can be used for one first node or for multiple first nodes at the same time without limitation.

[0192] Based on the method shown in Figure 4 above, when the first test result meets the preset test conditions, it indicates that the current training mode may not match the first model. If the current training mode continues to be used, the transmission overhead in the model training process may increase and the robustness of the model training process may be reduced; therefore, the third node can dynamically adjust the target training mode that is more suitable for the current model training process according to the first test result, and send a first indication information to the first node to instruct the first node to switch to the target training mode, so that the first node can perform model training according to the dynamically adjusted target training mode, which can improve the efficiency of model training, reduce the transmission overhead in the model training process, and improve the robustness of the model training process.

[0193] Based on the method described in Figure 4 above, when the training task occupies a large amount of computing power resources or the training model is more complex, the third node can split the training task into multiple training tasks, schedule multiple first nodes, and send the multiple training tasks to the scheduled multiple first nodes, which will perform model training to reduce the occupancy of the computing power resources of each first node, reduce the complexity of model training, and improve the training speed.

[0194] In one possible embodiment, the method for pattern training provided in the embodiment of the present application is described by taking the third node splitting the training task into the first task and the second task and scheduling the first node 1 and the first node 2 as an example. The specific steps are shown in FIG5 below:

[0195] S501. The third node splits the training task.

[0196] For example, the third node can split the training task into the first task 1 and the first task 2

[0197] S502 : The third node sends the first task 1 to the first node 1 ; correspondingly, the first node 1 receives the first task 1 from the third node.

[0198] S503 : The third node sends the first task 2 to the first node 2 ; correspondingly, the first node 2 receives the first task 2 from the first node.

[0199] Among them, the third node can schedule multiple first nodes. Taking the multiple first nodes including first node 1 and first node 2 as an example, first task 1 can be sent to first node 1, and first task 2 can be sent to first node 2.

[0200] S504 : The third node sends first configuration information 1 to the first node 1 ; correspondingly, the first node 1 receives the first configuration information 1 from the third node.

[0201] Among them, the first configuration information 1 may include one or more of the following: hyperparameter information, model test cycle information, and maximum retained version number information of the model.

[0202] S505 : The third node sends first configuration information 2 to the first node 2 ; correspondingly, the first node 2 receives the first configuration information 2 from the third node.

[0203] Among them, the first configuration information 2 may include one or more of the following: hyperparameter information, model test cycle information, and maximum retained version number information of the model.

[0204] Exemplarily, the first node 1 and the first node 2 may perform model training through online training 1 .

[0205] For example, online training 1 may be a process of performing model training based on a gradient pass-through training mode.

[0206] S506 : The first node 1 sends the first model 1 to the third node; correspondingly, the third node receives the first model 1 from the first node 1 .

[0207] The first node may periodically send the first model to the third node based on the test period of the model in the first configuration information 1 .

[0208] S507 : The first node 2 sends the first model 2 to the third node; correspondingly, the third node receives the first model 2 from the first node 2 .

[0209] Among them, the first model 2 is the model trained by the first node 2.

[0210] The first node 2 may periodically send the first model 2 to the third node based on the test period of the model in the first configuration information 2 .

[0211] S508 : The third node tests the first model 1 and obtains a first test result 1 .

[0212] S509 : The third node tests the first model 2 and obtains a first test result 2 .

[0213] The description of S508 and S509 may refer to the description of S402 in FIG. 4 , and will not be repeated here.

[0214] S510. When the first test result 1 meets the preset test condition, the third node sends the first indication information 1 to the first node 1; correspondingly, the first node 1 receives the first indication information 1 from the third node.

[0215] S511. When the first test result 2 meets the preset test condition, the third node sends the first indication information 2 to the first node 2; correspondingly, the first node 2 receives the first indication information 2 from the third node.

[0216] The description of S510 and S511 may refer to the description of S403 in FIG. 4 , and will not be repeated here.

[0217] Exemplarily, the third node may send the first indication information 1 to the first node 1 and the first indication information 2 to the first node 2 at the same time; or, the third node may not send the first indication information 1 to the first node and the first indication information 2 to the first node 2 at the same time (such as the third node may only send the first indication information 1 to the first node 1, or the third node may only send the second indication information 2 to the first node 2).

[0218] S512: The first node 1 switches the training mode to the target training mode 1 according to the first instruction information 1.

[0219] The description of S512 may refer to the description of S404 in FIG. 4 , and will not be repeated here.

[0220] S513 : The first node 2 switches the training mode to the target training mode 2 according to the first instruction information 2 .

[0221] The description of S513 may refer to the description of S404 in FIG. 4 , and will not be repeated here.

[0222] Among them, after switching to the target training mode, the first node 1 and the first node 2 can perform model training based on the online training 2.

[0223] Based on the method shown in FIG. 5 , as shown in S601 in FIG. 6 , before the third node sends the training task to each first node, the third node may receive the training task from the terminal device.

[0224] After each first node has trained the first model, as shown in S602 in Figure 6, the first node 1 can send the first model 1 to the third node; as shown in S603 in Figure 6, the first node 2 can send the first model 2 to the third node; as shown in S604 in Figure 6, the third node can also return the trained model to the terminal device.

[0225] The trained model is obtained by the third node fusing the first model 1 and the first model 2.

[0226] The terminal device can obtain the training model through the method in Figure 6, which can effectively protect the data privacy of the terminal device; the third node can make the model training more efficient and speed up the model training by splitting the training task.

[0227] Different from the above-mentioned third node instructing the first node to switch to the target training mode according to the first test result, as shown in Figure 7 below, the first node can also actively request to switch to the training mode according to the first measurement result. The specific steps can be shown in Figure 7 below:

[0228] S701: A first node measures one or more of the following to obtain a first measurement result: a channel or a training gradient.

[0229] The first node may measure the channel according to a traditional pilot sequence to obtain a first measurement result.

[0230] The first node may measure the training gradient according to predefined training and / or test data to obtain a first measurement result.

[0231] S702: When the first measurement result meets a second preset condition, the first node sends a handover request to the third node; correspondingly, the third node receives the handover request from the first node.

[0232] The switching request is used to request switching to the training mode.

[0233] Optionally, the second preset condition may include one or more of the following:

[0234] 1) The training accuracy of the first model decreases.

[0235] Optionally, when the training accuracy of the first model decreases (such as the training accuracy decreases within a certain period of time), the first node may send a switching request to the third node; or, when the training accuracy of the first model is less than a preset threshold, the first node may also send a switching request to the third node; or, when the training accuracy of the first model decreases by a value less than a preset threshold within a period of time, the first node may also send a switching request to the third node, without restriction.

[0236] 2) The training loss of the first model remains unchanged or increases within the second preset time.

[0237] The second preset time may be determined according to actual communication scenarios or actual communication requirements.

[0238] Optionally, when the training loss of the first model remains unchanged or increases within the second preset time, the first node may send a switching request to the third node; when the training loss of the first model is less than a preset threshold within the second preset time, the first node may also send a switching request to the third node without restriction.

[0239] 3) The value of the uplink signal-to-noise ratio increased or decreased within the third preset time is greater than or equal to the fifth preset threshold.

[0240] Among them, the third preset time and the fifth preset threshold can be determined according to the actual communication scenario or actual communication needs.

[0241] 4) The value by which the downlink signal-to-noise ratio increases or decreases within the third preset time is greater than or equal to a fifth preset threshold.

[0242] It should be noted that the third preset time in 3) and 4) can be the same or different, without limitation.

[0243] Similarly, the fifth preset thresholds in 3) and 4) may be the same or different, without limitation.

[0244] Similarly, the third preset time may be the same or different in different situations (such as when the uplink signal-to-noise ratio value increases, when the downlink signal-to-noise ratio value decreases, etc.), and is not limited.

[0245] For example, the third preset time associated with increasing the uplink signal-to-noise ratio and the third preset time associated with increasing the downlink signal-to-noise ratio may be the same or different.

[0246] Similarly, the fifth preset threshold may be the same or different in different situations, and is not limited.

[0247] For example, the fifth preset threshold associated with the value of the uplink signal-to-noise ratio increase and the fifth preset threshold associated with the value of the uplink signal-to-noise ratio decrease may be the same or different.

[0248] 5) The rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to a sixth preset threshold.

[0249] Among them, the fourth preset time and the sixth preset threshold can be determined according to the actual communication scenario or actual communication needs.

[0250] 6) The rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to a sixth preset threshold.

[0251] It should be noted that the fourth preset time in 5) and 6) can be the same or different, without limitation.

[0252] Similarly, the sixth preset thresholds in 5) and 6) may be the same or different, without limitation.

[0253] Similarly, the fourth preset time may be the same or different in different situations (such as the rate at which the uplink signal-to-noise ratio increases, or the rate at which the downlink signal-to-noise ratio decreases), and is not limited thereto.

[0254] Similarly, the sixth preset threshold may be the same or different in different situations, without limitation.

[0255] 7) The gradient variance of the first model is greater than or equal to a seventh preset threshold.

[0256] Among them, the seventh preset threshold can be determined according to the actual communication scenario or actual communication requirements.

[0257] Based on the above description of the second preset condition, when the first measurement result meets the second preset condition, the first node may send a handover request to the third node.

[0258] The switching request is used to request switching to the training mode.

[0259] Exemplarily, the switching request may be 1-bit information. When the bit value is 1, it indicates that the first node requests to switch to the training mode; or, when the bit value is 0, it indicates that the first node requests to switch to the training mode.

[0260] Optionally, the switching request can also be used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease in the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease in the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold.

[0261] Based on the above description of the handover request, when the third node receives the handover request sent by the first node, it may refer to S703 to S705 below and determine, based on the first test result, whether to send the first indication information to the first node:

[0262] S703: The first node sends the first model to the third node; correspondingly, the third node receives the first model from the first node.

[0263] The first node does not need to periodically report the first model according to the test period of the model in the first configuration information, but may send the first model in real time according to the first measurement result.

[0264] S704: The third node tests the first model to obtain a first test result.

[0265] S705: When the first test result meets the preset test condition, the third node sends first indication information to the first node.

[0266] The first indication information is used to instruct the first node to switch to the target training mode.

[0267] Furthermore, the first node may perform model training according to a target training mode.

[0268] Alternatively, unlike the above-mentioned method in which the third node determines whether to send the first indication information to the first node based on the first test result, the third node may also refer to S706 to S707 below and directly send the first indication information to the first node upon receiving the handover request:

[0269] S706: The first node sends the first model to the third node; correspondingly, the third node receives the first model from the first node.

[0270] The first node does not need to periodically report the first model according to the test period of the model in the first configuration information, but may send the first model in real time according to the first measurement result.

[0271] Among them, when the switching request can also be used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease of the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease of the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold, the third node can determine the target training mode according to the switching request and send the first indication information to the first node, instead of determining the target training mode according to the first test result.

[0272] S707: The third node sends first indication information to the first node; correspondingly, the first node receives the first indication information from the third node.

[0273] The first indication information is used to instruct the first node to switch to the target training mode.

[0274] Furthermore, the first node may perform model training according to a target training mode.

[0275] Alternatively, unlike the above-mentioned case where the third node receives the handover request and the first model, the third node may only receive the handover request and not the first model. Referring to S708 below, upon receiving the handover request, the third node directly sends the first indication information to the first node:

[0276] S708. The third node sends first indication information to the first node; correspondingly, the first node receives the first indication information from the third node.

[0277] The first indication information is used to instruct the first node to switch to the target training mode.

[0278] Among them, when the switching request can also be used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease of the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease of the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold, the third node can determine the target training mode according to the switching request and send the first indication information to the first node.

[0279] Furthermore, the first node may perform model training according to a target training mode.

[0280] Based on the method shown in Figure 7 above, when the model training process requires the participation of multiple first nodes, each first node can send a switching request to the third node with reference to the method shown in Figure 7. The third node, based on the method shown in Figure 7 above, instructs each first node to switch to the target training mode according to the switching request sent by each first node.

[0281] Exemplarily, as shown in FIG7 , taking the case where the multiple first nodes include the first node 1 and the first node 2 , both the first node 1 and the first node 2 may switch the target training mode with reference to the method shown in FIG7 .

[0282] Based on the method shown in FIG7 , the third node receives a switching request from the first node, and the third node can test the first model in a timely manner without having to wait until a predefined time to measure the first model, thereby improving the efficiency of model training. After receiving the switching request, the third node tests the first model and can further determine whether it is necessary to switch the training mode, thereby avoiding the occurrence of misjudgment by the first node as much as possible, thereby improving the accuracy of model training; or, the third node can directly send a first indication message to the first node according to the switching request, for instructing the first node to switch to the target training mode. The third node does not need to test the first model, thereby reducing the overhead of the third node in testing the first model, reducing the memory usage of the third node, saving the computing resources of the third node, and improving the efficiency of information exchange between the first node and the third node.

[0283] It is understandable that the method shown in FIG. 7 can be used alone or in combination with the method shown in FIG. 4 .

[0284] Based on the methods shown in Figures 4 to 7 above, the third node can switch the model learning mode of each first node participating in model training (such as switching to the target model learning mode) according to the training status information of the first node. The specific steps are shown in Figure 8 below:

[0285] S801: A first node sends training status information to a third node; correspondingly, the third node receives the training status information from the first node.

[0286] The training status information includes one or more of the following: training loss, number of training rounds, data set status, or second indication information.

[0287] The second indication information is used to indicate whether the first node continues to participate in training.

[0288] Exemplarily, the second indication information may be 1-bit information, where a bit value of 1 indicates that the first node continues to participate in training; a bit value of 0 indicates that the first node does not continue to participate in training. Alternatively, a bit value of 0 indicates that the first node continues to participate in training; a bit value of 1 indicates that the first node does not continue to participate in training.

[0289] Optionally, the third node may instruct the first node to periodically send training status information to the third node through third configuration information.

[0290] The third configuration information is used to indicate a period for the first node to report the training status.

[0291] S802: When the training state information satisfies a first preset condition, the third node sends third indication information to the first node; correspondingly, the first node receives the third indication information from the third node.

[0292] Among them, the third indication information is used to instruct the first node to adopt the target model learning method to perform model training.

[0293] Optionally, the target model learning method can be a distributed learning method, a centralized learning method, a split learning method, or a federated learning method, without restriction.

[0294] Optionally, the first preset condition may be one or more of the following: the training loss is greater than or equal to an eighth preset threshold, the number of training rounds does not match the data set, the error of the data in the data set is greater than or equal to a ninth preset threshold, the amount of data in the data set is less than a tenth preset threshold, or the second indication information is used to instruct the first node not to continue participating in training.

[0295] Among them, the eighth preset threshold, the ninth preset threshold, and the tenth preset threshold can be determined according to actual communication scenarios or actual communication requirements.

[0296] For example, the mismatch between the number of training rounds and the data set may be caused by too few training rounds and too large a data set size, which affects the accuracy of the training results; or too many training rounds and too small a data set size, which causes resource loss.

[0297] S803. The first node switches the model learning mode to the target model learning mode according to the third instruction information.

[0298] Optionally, the first node can switch the learning mode from distributed learning mode to centralized learning mode, from centralized learning mode to distributed learning mode, from federated learning mode to split learning mode, or from split learning mode to federated learning mode, without restriction.

[0299] Based on the method shown in Figure 8 above, when the model training process requires the participation of multiple first nodes, each first node can switch to the target model learning mode with reference to the method shown in Figure 8 above, and the third node instructs each first node to switch to the target model learning mode based on the method shown in Figure 8 above.

[0300] Exemplarily, as shown in FIG8 , taking the case where multiple first nodes include a first node 1 and a first node 2, both the first node 1 and the first node 2 can switch to the target model learning mode with reference to the method shown in FIG8 .

[0301] In an exemplary embodiment, taking online training 1 as split learning as an example, the second indication information in the first training status information 1 instructs the first node 1 to continue to participate in the training, and the second indication information in the first training status 2 information instructs the first node 2 not to continue to participate in the training. Then, the third node sends the third indication information and the training model associated with the first node 2 to the first node, instructing the first node 1 to switch to the centralized training learning mode and continue the model training; at the same time, the third node sends the third indication information to the first node 2, instructing the first node 2 to exit the training.

[0302] Another exemplary example is to take online learning 1 as centralized learning and the first node 1 as an example of individual training. The first node sends first training status information 1 to the third node. The amount of data in the data set in the first training status information 1 is greater than the tenth preset threshold. The third node sends third indication information to the first node 1. The third indication information indicates that the target model learning method is a federated learning method. At the same time, the first node 2 is scheduled so that the first node 2 uses the local data set to perform model training based on the federated learning method.

[0303] Based on the method shown in Figures 4 to 8 above, the third node can also schedule the second node so that the second node replaces the first node for model training. The method includes: the third node obtains multiple first test results; when the multiple first test results all meet the preset test conditions, the third node sends the second configuration information associated with the first node to the second node.

[0304] Different first test results are associated with different first models, and different first models are associated with different training modes.

[0305] The second configuration information is used to indicate one or more of the following: a training task associated with the first node, a training mode associated with the first node, or a training model associated with the first node.

[0306] Based on the above method of the third node scheduling the second node, this application proposes a possible embodiment, the specific steps of which are shown in Figure 9 below:

[0307] S901. The third node splits the training task.

[0308] The description of S901 may refer to the description of S501 in FIG. 5 , and will not be repeated here.

[0309] S902 : The third node sends the first task 1 to the first node 1 .

[0310] The description of S902 may refer to the description of S502 in FIG. 5 , and will not be repeated here.

[0311] S903 : The third node sends the first task 2 to the first node 2 .

[0312] The description of S903 may refer to the description of S503 in FIG. 5 , and will not be repeated here.

[0313] S904 : The third node sends first configuration information 1 to the first node 1 .

[0314] The description of S904 may refer to the description of S504 in FIG. 5 , and will not be repeated here.

[0315] S905 : The third node sends first configuration information 2 to the first node 2 .

[0316] The description of S905 may refer to the description of S505 in FIG. 5 , and will not be repeated here.

[0317] S906: The first node 1 sends the first model 1 to the third node.

[0318] S907: The first node 2 sends the first model 2 to the third node.

[0319] The description of S906 and S907 may refer to the description of S401 in FIG. 4 , and will not be repeated here.

[0320] S908 . The third node tests the first model 1 and obtains a first test result 1 .

[0321] S909 : The third node tests the first model 2 and obtains a first test result 2 .

[0322] The description of S908 and S909 may refer to the description of S402 in FIG. 4 , and will not be repeated here.

[0323] S910. When the first test result 1 meets the preset test condition, the third node sends first indication information 1 to the first node 1; correspondingly, the first node 1 receives the first indication information 1 from the third node.

[0324] S911. When the first test result 2 meets the preset test condition, the third node sends the first indication information 2 to the first node 2; correspondingly, the first node 2 receives the first indication information 2 from the third node.

[0325] The description of S910 and S911 may refer to the description of S403 in FIG. 4 , and will not be repeated here.

[0326] S912. The first node 1 switches the training mode to the target training mode 1 according to the first instruction information 1.

[0327] Among them, the first node 1 can perform model training according to the target training mode 1 to obtain the first model 1.

[0328] It can be understood that the first node 1 can receive the first indication information 1 multiple times and switch to the target training mode 1 according to the first indication information 1; the first node can perform model training according to the target training mode 1, obtain the first model 1, and send the first model 1 to the third node, that is, the first node 1 can send the first model 1 associated with different training modes to the third node.

[0329] The third node can test different first models 1 to obtain different first test results 1.

[0330] S913 : The first node 2 switches the training mode to the target training mode 2 according to the first instruction information 2 .

[0331] Among them, the first node 2 can perform model training according to the target training mode 2 to obtain the first model 2.

[0332] S914: When the plurality of first test results 1 all meet the preset test condition, the third node sends the second configuration information associated with the first node to the second node.

[0333] For example, taking the case where all three first test results 1 meet the preset test conditions, the first test result 1 in the above S908 is the first one, and the third node can test the first model 1 (the training model obtained by the first node 1 performing model training according to the content of the above S912) to obtain the second first test result 1. If the second first test result 1 still meets the preset test conditions, the third node continues to instruct the first node 1 to switch the training mode; similarly, if the third first test result 1 still meets the preset test conditions, the first node 1 may not be suitable for model training (such as the training model is more complex, and the first node has less memory or less computing power resources), and the third node can schedule the second node to replace the first node 1 for model training.

[0334] For example, the first node 1 exits model training, the second node can replace the first node 1, and the first node 2 performs model training through online training 2.

[0335] Based on the methods shown in Figures 4 to 9 above, as shown in Figure 10 below, the third node may also send fourth indication information to the first node, instructing the first node to perform model training in a certain frequency domain.

[0336] S1001. The third node sends a training task to the first node.

[0337] The description of S1001 may refer to the description of S502 or S503 in FIG. 5 , and will not be repeated here.

[0338] S1002: The third node sends first configuration information to the first node.

[0339] The description of S1002 may refer to the description of S504 or S505 in FIG. 5 , and will not be repeated here.

[0340] Optionally, the first node may perform model training (such as online training 1) according to the training task and the first configuration information.

[0341] S1003. The third node sends fourth indication information to the first node; correspondingly, the first node receives the fourth indication information from the third node.

[0342] The fourth indication information is used to indicate the frequency domain information when the first node performs model training.

[0343] Optionally, the third node may configure a channel measurement signal for the first node, detect the uplink signal-to-noise ratio (or interference) of the first node, and determine whether to send the fourth indication information.

[0344] Furthermore, the first node may determine the frequency domain information based on the fourth indication information, perform model training on the frequency domain (as shown in FIG10 , the third node may switch from frequency band 1 to frequency band 2 for model training), and obtain the first model.

[0345] Optionally, the frequency domain information may be a specific frequency band or a frequency domain offset, without limitation.

[0346] In an exemplary embodiment, during the model training process, when the first node is at the edge of the cell, the third node detects that the uplink signal-to-noise ratio of the first node is low. The third node can send a fourth indication message to the first node, and the fourth indication message is used to instruct the first node to switch to a low frequency band for model training (such as the first node can switch from the time division duplex (TDD) band to the frequency division duplex (FDD) band); or, the fourth indication message can indicate a supplementary uplink (SUL), that is, the downlink channel of the first node remains in the TDD band, and the uplink channel uses the FDD band for supplementation to improve the uplink signal-to-noise ratio.

[0347] In another exemplary embodiment, the third node may send fourth indication information to the first node, where the fourth indication information may be a predefined frequency domain offset. The first node may perform model training on the current frequency domain after the predefined frequency domain offset.

[0348] It should be noted that the various embodiments of this application can be implemented independently or in combination, without limitation. Unless otherwise specified or there is a logical conflict, the terms and / or descriptions of the different embodiments provided in this application are consistent and can be referenced by each other. The technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0349] It is understood that in the embodiments of the present application, the execution subject may perform some or all of the steps in the embodiments of the present application. These steps or operations are merely examples, and the embodiments of the present application may also perform other operations or variations of various operations. In addition, the various steps may be performed in a different order than those presented in the embodiments of the present application, and it is possible that not all operations in the embodiments of the present application need to be performed.

[0350] The above mainly introduces the solutions provided by this application from the perspective of interaction between various devices. Accordingly, this application also provides a communication device, which is used to implement the various methods described above. The communication device can be the third node in the above method embodiments, or a device that includes the above third node, or a component that can be used for the third node; alternatively, the communication device can be the first node involved in the above method embodiments, or a device that includes the first node, or a component that can be used for the first node.

[0351] It is understandable that, in order to realize the above functions, the communication device includes hardware structures and / or software modules corresponding to the execution of each function. It should be easily appreciated by those skilled in the art that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0352] The embodiment of the present application can divide the functional modules of the communication device according to the above method embodiment. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiment of the present application is schematic and is only a logical functional division. In actual implementation, there may be other division methods.

[0353] In one implementation scenario, taking the communication device as the third node in the above method embodiment as an example, FIG11 shows a schematic structural diagram of a third node 100. The third node 110 includes a processing module 1101 and a transceiver module 1102.

[0354] In some embodiments, the third node 110 may further include a storage module (not shown in FIG. 11 ) for storing program instructions and data.

[0355] In some embodiments, the transceiver module 1102, which may also be referred to as a transceiver unit, is configured to implement a transmitting and / or receiving function. The transceiver module 1102 may be composed of a transceiver circuit, a transceiver, a transceiver, or a communication interface.

[0356] In some embodiments, the transceiver module 1102 may include a receiving module and a sending module, which are respectively used to execute the receiving and sending steps performed by the third node in the above-mentioned method embodiment, and / or used to support other processes of the technology described in this document; the processing module 1101 may be used to execute the processing steps (such as determination, generation, etc.) performed by the terminal in the above-mentioned method embodiment, and / or used to support other processes of the technology described in this document.

[0357] Illustratively, transceiver module 1102 is configured to receive a first model from a first node, where the first model is a trained model of the first node. Processing module 1101 is configured to test the first model and obtain a first test result. Transceiver module 1102 is further configured to send first indication information to the first node when the first test result meets a preset test condition, where the first indication information is configured to instruct the first node to switch to a target training mode.

[0358] In one possible implementation, the first test result satisfies the preset test conditions, including one or more of the following situations: the test accuracy of the first test result is less than or equal to a first preset threshold; or, the first test result indicates that the first model is overfitting; or, the test accuracy of the first test result remains unchanged or decreases within multiple test cycles; or, the difference in network traffic associated with the first test result within multiple test cycles is greater than or equal to a second preset threshold; wherein the network traffic is used to indicate the amount of data used to train the first model; or, the difference in the signal-to-noise ratio associated with the first test result within multiple test cycles is greater than or equal to a third preset threshold; or, the difference in the interference associated with the first test result within multiple test cycles is greater than or equal to a fourth preset threshold.

[0359] In one possible implementation, the transceiver module 1102 is also used to send first configuration information to the first node; wherein the first configuration information includes one or more of the following: hyperparameter information, model test cycle information, and model maximum retained version number information; the model test cycle information is used to indicate the number of model training rounds experienced by the first node when reporting the first model, and the model maximum retained version number information is used to indicate the maximum number of first models saved within the first preset time, and different first models are obtained through different training rounds.

[0360] In one possible implementation, the transceiver module 1102 is further configured to receive a switching request from the first node, wherein the switching request is used to request switching to the training mode. The processing module 1101 is further configured to test the first model according to the switching request to obtain a first test result.

[0361] In a possible implementation, the transceiver module 1102 is further configured to receive a handover request from the first node. The transceiver module 1102 is further configured to send first indication information to the first node according to the handover request.

[0362] In one possible implementation, the switching request is used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease of the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease of the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold.

[0363] In one possible implementation, the transceiver module 1102 is further configured to receive training status information from the first node; wherein the training status information includes one or more of the following: training loss, number of training rounds, data set status, or second indication information; the second indication information is used to indicate whether the first node should continue to participate in training. The transceiver module 1102 is further configured to send third indication information to the first node when the training status information satisfies a first preset condition; wherein the first preset condition is one or more of the following: the training loss is greater than or equal to an eighth preset threshold, the number of training rounds does not match the data set, the error of the data in the data set is greater than or equal to a ninth preset threshold, the amount of data in the data set is less than a tenth preset threshold, or the second indication information is used to indicate that the first node should not continue to participate in training; the third indication information is used to indicate that the first node should use a target model learning method for model training.

[0364] In one possible implementation, the transceiver module 1102 is further used to send fourth indication information to the first node; wherein the fourth indication information is used to indicate frequency domain information when the first node performs model training.

[0365] In one possible implementation, transceiver module 1102 is further configured to obtain multiple first test results, wherein different first test results are associated with different first models, and different first models are associated with different training modes. Transceiver module 1102 is further configured to, when all of the multiple first test results meet a preset test condition, send second configuration information associated with the first node to the second node, wherein the second configuration information is configured to indicate one or more of the following: a training task associated with the first node, a training mode associated with the first node, or a training model associated with the first node.

[0366] In this application, the third node 110 is presented in the form of functional modules divided in an integrated manner. The "module" here can refer to an ASIC, a circuit, a processor and memory that executes one or more software or firmware programs, an integrated logic circuit, and / or other devices that can provide the above functions.

[0367] In some embodiments, in terms of hardware implementation, those skilled in the art may conceive that the third node 110 may be in the form of the communication device 30 shown in FIG. 3 .

[0368] As an example, the functions / implementation process of the processing module 1101 in FIG11 can be implemented by the processor 301 in the communication device 30 shown in FIG3 calling the computer-executable instructions stored in the memory 303. The functions / implementation process of the transceiver module 1102 in FIG11 can be implemented by the communication interface 304 in the communication device 30 shown in FIG3.

[0369] In some embodiments, when the third node 110 in Figure 11 is a chip or a chip system, the function / implementation process of the transceiver module 1102 can be implemented through the input and output interface (or communication interface) of the chip or chip system, and the function / implementation process of the processing module 1101 can be implemented through the processor (or processing circuit) of the chip or chip system.

[0370] Since the third node 110 provided in this embodiment can execute the above method, the technical effects that can be obtained can refer to the above method embodiments and will not be repeated here.

[0371] In another implementation scenario, taking the communication device as the first node in the above method embodiment as an example, FIG12 shows a schematic structural diagram of a first node 120. The first node 120 includes a processing module 1201 and a transceiver module 1202.

[0372] In some embodiments, the first node 120 may further include a storage module (not shown in FIG. 12 ) for storing program instructions and data.

[0373] In some embodiments, the transceiver module 1202, which may also be referred to as a transceiver unit, is configured to implement a transmitting and / or receiving function. The transceiver module 1202 may be composed of a transceiver circuit, a transceiver, a transceiver, or a communication interface.

[0374] In some embodiments, the transceiver module 1202 may include a receiving module and a sending module, which are respectively used to execute the receiving and sending steps performed by the first node in the above method embodiment, and / or used to support other processes of the technology described in this document; the processing module 1201 may be used to execute the processing steps (such as determination, generation, etc.) performed by the first node in the above method embodiment, and / or used to support other processes of the technology described in this document.

[0375] Illustratively, transceiver module 1202 is configured to send a first model to a third node, where the first model is a model trained by the first node. Transceiver module 1202 is further configured to receive first indication information from the third node, where the first indication information is configured to instruct the first node to switch to a target training mode. Processing module 1201 is configured to switch the training mode to the target training mode based on the first indication information.

[0376] In one possible implementation, the transceiver module 1202 is further configured to receive first configuration information from a third node; wherein the first configuration information includes one or more of the following: hyperparameter information, model test cycle information, and information about the maximum number of retained versions of the model; the model test cycle information is used to indicate the number of model training rounds experienced by the first node when reporting the first model, and the maximum number of retained versions of the model is used to indicate the maximum number of first models saved within a first preset time period. Different first models are obtained through different training rounds, and the performance of the saved first model is better than that of the unsaved first model. The processing module 1201 is further configured to train the model according to the first configuration information to obtain the first model.

[0377] In one possible implementation, the processing module 1201 is further configured to measure one or more of the following to obtain a first measurement result: a channel or a training gradient. The transceiver module 1202 is further configured to send a handover request to the third node when the first measurement result satisfies a second preset condition; wherein the second preset condition includes one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within a second preset time, the value of the increase or decrease in the uplink signal-to-noise ratio within a third preset time is greater than or equal to a fifth preset threshold, the value of the increase or decrease in the downlink signal-to-noise ratio within the third preset time is greater than or equal to a fifth preset threshold, the rate of increase or decrease in the uplink signal-to-noise ratio within a fourth preset time is greater than or equal to a sixth preset threshold, the rate of increase or decrease in the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to a sixth preset threshold, or the gradient variance of the first model is greater than or equal to a seventh preset threshold.

[0378] In one possible implementation, the switching request is used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease of the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease of the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold.

[0379] In one possible implementation, the transceiver module 1202 is also used to send training status information to the third node; wherein the training status information includes one or more of the following: training loss, number of training rounds, data set status, or second indication information; the second indication information is used to indicate whether the first node continues to participate in training.

[0380] In one possible implementation, the transceiver module 1202 is further configured to receive third instruction information from a third node; wherein the third instruction information is configured to instruct the first node to perform model training using a target model learning method;

[0381] In one possible implementation, the transceiver module 1202 is further configured to receive fourth indication information from the third node, wherein the fourth indication information is used to indicate frequency domain information for the first node during model training. The processing module 1201 is further configured to determine the frequency domain information for model training based on the fourth indication information.

[0382] In this application, the first node 120 is presented in the form of various functional modules divided in an integrated manner. The "module" here can refer to an ASIC, a circuit, a processor and memory that executes one or more software or firmware programs, an integrated logic circuit, and / or other devices that can provide the above functions.

[0383] In some embodiments, in terms of hardware implementation, those skilled in the art may conceive that the first node 120 may take the form of the communication device 30 shown in FIG. 3 .

[0384] As an example, the functions / implementation process of the processing module 1201 in FIG12 can be implemented by the processor 301 in the communication device 30 shown in FIG3 calling the computer-executable instructions stored in the memory 303. The functions / implementation process of the transceiver module 1202 in FIG12 can be implemented by the communication interface 304 in the communication device 30 shown in FIG3.

[0385] In some embodiments, when the first node 120 in Figure 12 is a chip or a chip system, the function / implementation process of the transceiver module 1202 can be implemented through the input and output interface (or communication interface) of the chip or chip system, and the function / implementation process of the processing module 1201 can be implemented through the processor (or processing circuit) of the chip or chip system.

[0386] Since the first node 120 provided in this embodiment can execute the above method, the technical effects that can be obtained can refer to the above method embodiments and will not be repeated here.

[0387] As a possible product form, the terminal device or first node described in the embodiment of the present application can also be implemented using the following: one or more field programmable gate arrays (FPGA), programmable logic devices (PLD), controllers, state machines, gate logic, discrete hardware components, any other suitable circuits, or any combination of circuits that can perform the various functions described throughout this application.

[0388] As another possible product form, the third node or the first node described in the embodiment of the present application can be implemented by a general bus architecture. For ease of explanation, refer to Figure 13, which is a structural diagram of a communication device 130 provided in an embodiment of the present application. The communication device 130 includes a processor 1301 and a transceiver 1302. The communication device 130 can be a third node, or a chip or module therein; alternatively, the communication device 130 can be a first node, or a chip or module therein. Figure 13 only shows the main components of the communication device 130. In addition to the processor 1301 and the transceiver 1302, the communication device may further include a memory 1303.

[0389] Optionally, processor 1301 is primarily used to process communication protocols and communication data, control the entire communication device, execute software programs, and process software program data. Memory 1303 is primarily used to store software programs and data. Transceiver 1102 may include a radio frequency circuit and an antenna. The radio frequency circuit is primarily used to convert baseband signals into radio frequency signals and process radio frequency signals. The antenna is primarily used to transmit and receive radio frequency signals in the form of electromagnetic waves.

[0390] Optionally, the processor 1301 , the transceiver 1302 , and the memory 1303 may be connected via a communication bus.

[0391] When the communication device is powered on, the processor 1301 can read the software program in the memory 1303, interpret and execute the instructions of the software program, and process the data of the software program. When data needs to be sent wirelessly, the processor 1301 performs baseband processing on the data to be sent and outputs the baseband signal to the radio frequency circuit. The radio frequency circuit performs radio frequency processing on the baseband signal and then transmits the radio frequency signal to the outside in the form of electromagnetic waves through the antenna. When data is sent to the communication device, the radio frequency circuit receives the radio frequency signal through the antenna, converts the radio frequency signal into a baseband signal, and outputs the baseband signal to the processor 1301. The processor 1301 converts the baseband signal into data and processes the data.

[0392] In another implementation, the RF circuit and antenna may be provided independently of the processor performing baseband processing. For example, in a distributed scenario, the RF circuit and antenna may be remotely arranged independent of the communication device.

[0393] In some embodiments, embodiments of the present application further provide a communication device, the communication device including a processor, configured to implement the method in any of the above method embodiments. The communication device may be the third node or the first node in the above method embodiments.

[0394] As a possible implementation, the communication device further includes a memory. The memory is used to store necessary computer programs and data. The computer program may include instructions, and the processor may invoke the instructions in the computer program stored in the memory to instruct the communication device to perform any of the above-described method embodiments. Of course, the memory may not be located in the communication device.

[0395] As another possible implementation, the communication device also includes an interface circuit, which is a code / data read / write interface circuit, and the interface circuit is used to receive computer execution instructions (computer execution instructions are stored in a memory, may be read directly from the memory, or may pass through other devices) and transmit them to the processor.

[0396] As another possible implementation, the communication device further includes a communication interface, where the communication interface is used to communicate with a module outside the communication device.

[0397] It can be understood that the communication device can be a chip or a chip system. When the communication device is a chip system, it can be composed of chips or include chips and other discrete devices. The embodiments of the present application do not specifically limit this.

[0398] The present application also provides a computer-readable storage medium having a computer program or instruction stored thereon, which implements the functions of any of the above method embodiments when executed by a computer.

[0399] The present application also provides a computer program product, which implements the functions of any of the above method embodiments when executed by a computer.

[0400] Those skilled in the art will appreciate that, for the sake of convenience and brevity of description, the specific working processes of the above-described systems, devices, and units may refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0401] It is understood that the systems, devices, and methods described in this application may also be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed may be through some interface, indirect coupling or communication connection of devices or units, and may be electrical, mechanical, or other forms.

[0402] The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Components shown as units may or may not be physical units. Some or all of these units may be selected to achieve the objectives of this embodiment as needed.

[0403] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0404] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using a software program, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes (or functions) described in the embodiments of the present application are implemented. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more media that can be integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a DVD), or a semiconductor medium (eg, a solid state drive (SSD)). In the embodiment of the present application, the computer may include the aforementioned device.

[0405] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art can understand and implement other changes to the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple situations. A single processor or other unit can implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0406] Although the present application has been described with reference to specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present application. Accordingly, this specification and the drawings are merely illustrative of the present application as defined by the appended claims and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art may make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, the present application is intended to include such modifications and variations as fall within the scope of the claims of the present application and their equivalents.

Claims

1. A model training method, characterized in that: include: Receiving a first model from a first node; wherein the first model is a model trained by the first node; Testing the first model to obtain a first test result; When the first test result meets a preset test condition, first indication information is sent to the first node; wherein the first indication information is used to instruct the first node to switch to a target training mode.

2. The method according to claim 1, characterized in that: The first test result satisfies the preset test condition including one or more of the following situations: The test accuracy of the first test result is less than or equal to a first preset threshold; or, The first test result indicates that the first model is overfitted; or, The test accuracy of the first test result remains unchanged or decreases over multiple test cycles; or, The difference of the network traffic associated with the first test result in multiple test cycles is greater than or equal to a second preset threshold; wherein the network traffic is used to indicate the amount of data used to train the first model; or, The difference of the signal-to-noise ratio associated with the first test result in multiple test cycles is greater than or equal to a third preset threshold; or, A difference in interference associated with the first test result within a plurality of test cycles is greater than or equal to a fourth preset threshold.

3. The method according to claim 1 or 2, characterized in that: The method further comprises: Sending first configuration information to the first node; wherein the first configuration information includes one or more of the following: hyperparameter information, model test cycle information, and maximum number of retained versions of the model; The test cycle information of the model is used to indicate the number of rounds of model training experienced when the first node reports the first model. The maximum retained version number information of the model is used to indicate the maximum number of first models saved within the first preset time. Different first models are obtained through different training rounds.

4. The method according to any one of claims 1 to 3, characterized in that: The step of testing the first model to obtain a first test result includes: receiving a switching request from the first node; wherein the switching request is used to request switching to a training mode; The first model is tested according to the switching request to obtain the first test result.

5. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: receiving a handover request from the first node; Send the first indication information to the first node according to the switching request.

6. The method according to claim 5, characterized in that The switching request is used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease of the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease of the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold.

7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: Receiving training status information from the first node; wherein the training status information includes one or more of the following: training loss, number of training rounds, data set status, or second indication information; the second indication information is used to indicate whether the first node continues to participate in training; When the training status information meets the first preset condition, third indication information is sent to the first node; wherein the first preset condition is one or more of the following: the training loss is greater than or equal to the eighth preset threshold, the number of training rounds does not match the data set, the error of the data in the data set is greater than or equal to the ninth preset threshold, the amount of data in the data set is less than the tenth preset threshold, or the second indication information is used to instruct the first node not to continue to participate in training; the third indication information is used to instruct the first node to use the target model learning method for model training.

8. The method according to any one of claims 1 to 7, characterized in that: The method further comprises: Send fourth indication information to the first node; wherein the fourth indication information is used to indicate frequency domain information when the first node performs model training.

9. The method according to any one of claims 1 to 8, characterized in that: The method further comprises: Obtain multiple first test results; wherein different first test results are associated with different first models, and different first models are associated with Link different training modes; When the multiple first test results all meet the preset test conditions, the second configuration information associated with the first node is sent to the second node; wherein the second configuration information is used to indicate one or more of the following: the training task associated with the first node, the training mode associated with the first node, or the training model associated with the first node.

10. A model training method, characterized in that: include: Sending a first model to a third node; wherein the first model is a model trained by the first node; Receiving first indication information from the third node; wherein the first indication information is used to instruct the first node to switch to a target training mode; According to the first indication information, the training mode is switched to the target training mode.

11. The method according to claim 10, characterized in that The method further comprises: Receive first configuration information from the third node; wherein the first configuration information includes one or more of the following: hyperparameter information, model test cycle information, and maximum number of retained versions of the model; The test cycle information of the model is used to indicate the number of rounds of model training experienced when the first node reports the first model, and the maximum number of retained versions of the model is used to indicate the maximum value of the number of first models saved within the first preset time, and different first models are obtained through different training rounds; The model is trained according to the first configuration information to obtain the first model.

12. The method according to claim 10 or 11, characterized in that: The method further comprises: Measuring one or more of the following to obtain a first measurement result: a channel, or a training gradient; When the first measurement result meets the second preset condition, sending a switching request to the third node; wherein the second preset condition includes one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the increase or decrease of the uplink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the value of the increase or decrease of the downlink signal-to-noise ratio within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold.

13. The method according to claim 12, characterized in that The switching request is used to indicate one or more of the following: the training accuracy of the first model decreases, the training loss of the first model remains unchanged or increases within the second preset time, the value of the uplink signal-to-noise ratio increased or decreased within the third preset time is greater than or equal to the fifth preset threshold, the value of the downlink signal-to-noise ratio increased or decreased within the third preset time is greater than or equal to the fifth preset threshold, the rate of increase or decrease of the uplink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, the rate of increase or decrease of the downlink signal-to-noise ratio within the fourth preset time is greater than or equal to the sixth preset threshold, or the gradient variance of the first model is greater than or equal to the seventh preset threshold.

14. The method according to any one of claims 10 to 13, characterized in that: The method further comprises: Sending training status information to the third node; wherein the training status information includes one or more of the following: training loss, number of training rounds, data set status, or second indication information; the second indication information is used to indicate whether the first node continues to participate in training.

15. The method according to claim 14, characterized in that Receiving third indication information from the third node; wherein the third indication information is used to instruct the first node to perform model training using a target model learning method; According to the third indication information, the model learning mode is switched to the target model learning mode.

16. The method according to any one of claims 10 to 15, characterized in that: The method further comprises: Receiving fourth indication information from the third node; wherein the fourth indication information is used to indicate frequency domain information when the first node performs model training; According to the fourth indication information, the frequency domain information during model training is determined.

17. A communication device, characterized in that: include: A transceiver module, configured to receive a first model from a first node; wherein the first model is a model trained by the first node; A processing module, used for testing the first model to obtain a first test result; The transceiver module is further used to send first indication information to the first node when the first test result meets a preset test condition; wherein the first indication information is used to instruct the first node to switch to a target training mode.

18. A communication device, characterized in that: include: A transceiver module, used to send a first model to a third node; wherein the first model is a model trained by the first node; The transceiver module is further used to receive first indication information from the third node; wherein the first indication information is used to instruct the first node to switch to a target training mode; A processing module is used to switch the training mode to the target training mode according to the first indication information.

19. A communication device, characterized in that: The communication device includes a processor; the processor is used to run a computer program or instruction, or to use a logic circuit to enable the communication device to execute the model training method as described in any one of claims 1-9, or to enable the communication device to execute the model training method as described in any one of claims 10-16.

20. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions or programs, which, when executed on a computer, enable the model training method as described in any one of claims 1 to 9, or enable the communication device to execute the model training method as described in any one of claims 10 to 16.

21. A computer program product, characterized in that The computer program product includes computer instructions; when part or all of the computer instructions are executed, the model training method as described in any one of claims 1 to 9 is executed, or the model training method as described in any one of claims 10 to 16 is executed.

22. A communication system, characterized in that: The communication system includes a third node and a first node; wherein the third node is used to execute the model training method as described in any one of claims 1-9, and the first node is used to execute the model training method as described in any one of claims 10-16.

Images