A model training or inference method, apparatus and communication device

Abstract

The application provides a model training or reasoning method, device and communication equipment. The model training or reasoning method comprises the following steps: receiving target parameter information for model training or reasoning sent by a second device; determining a target model and a model segmentation point according to the target parameter information; sending model information of the target model and the model segmentation point to the second device; and performing model training or reasoning according to the target model and the model segmentation point.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a model training or inference method and device and communication equipment. BACKGROUND

[0002] With the rapid development of artificial intelligence (AI) technology, in future networks, the architecture of the network will be more and more integrated with AI technology. The existing AI model is trained and / or inferred on the network side device, but the terminal side will obtain more dimensional data than the network side device. In order to obtain more information and better protect the data privacy of users, training is performed on the user equipment (UE) side. In order to reduce the resources consumed by the UE side, the model is divided according to the model split point, part of which is executed on the UE side and part of which is executed on the network side device.

[0003] When the environmental conditions (such as available computing power and communication resources) change, different requirements for the computing power of the UE and the network side device will be generated, and the position of the model split point is particularly important. At present, in the joint training or inference of the model by the terminal and the network side device, the determination method of the model split point is single, which affects the effectiveness of the model training or inference. SUMMARY

[0004] Embodiments of the present application provide a model training or inference method, device and communication equipment to solve the problem of single determination method of the model split point in the joint training or inference of the model by the terminal and the network side device, which affects the effectiveness of the model training or inference.

[0005] To solve the above technical problems, the present application is implemented as follows:

[0006] In a first aspect, embodiments of the present application provide a model training or inference method for a first device, the method comprising:

[0007] receiving target parameter information for model training or inference sent by a second device;

[0008] determining a target model and a model split point according to the target parameter information;

[0009] sending model information of the target model and the model split point to the second device;

[0010] performing model training or inference according to the target model and the model split point.

[0011] In a second aspect, embodiments of the present application provide a model training or inference method for a second device, the method comprising:

[0012] sending target parameter information for model training or inference to a first device;

[0013] receiving model information and a model split point of a target model sent by the first device;

[0014] performing model training or inference according to the target model and the model split point.

[0015] In a third aspect, an embodiment of the present application provides a model training or inference apparatus, and the method comprises:

[0016] a receiving module configured to receive target parameter information for model training or inference sent by a second device;

[0017] a determining module configured to determine a target model and a model split point according to the target parameter information;

[0018] a sending module configured to send model information of the target model and the model split point to the second device;

[0019] a training or inference module configured to perform model training or inference according to the target model and the model split point.

[0020] In a fourth aspect, an embodiment of the present application provides a model training or inference apparatus, and the method comprises:

[0021] a sending module configured to send target parameter information for model training or inference to a first device;

[0022] a receiving module configured to receive model information and a model split point of a target model sent by the first device;

[0023] a training or inference module configured to perform model training or inference according to the target model and the model split point.

[0024] In a fifth aspect, an embodiment of the present application provides a communication device, comprising a first device and a second device, wherein the first device is configured to implement the steps in the method of the first aspect, and the second device is configured to implement the steps in the method of the second aspect.

[0025] In a sixth aspect, an embodiment of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps in the model training or inference method of the first aspect or the second aspect.

[0026] In the embodiment of the present application, the first device receives target parameter information for model training or inference sent by the second device; the first device determines a target model and a model split point according to the target parameter information; the first device sends model information of the target model and the model split point to the second device; and the first device performs model training or inference according to the target model and the model split point. In the above, the first device can dynamically determine the target model and the model split point, and the way of determining the target model and the model split point is more flexible. Since the model split point can be flexibly set according to actual conditions, the model training or inference failure caused by limited computing power of the terminal or the network side device can be avoided, and the effectiveness of the model training or inference can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the description of the embodiments of the present application will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0028] Figure 1 is a model splitting method provided by the embodiment of the present application;

[0029] Figure 2a is one of flowcharts of a model training or inference method provided by the embodiment of the present application;

[0030] Figure 2b is another of flowcharts of a model training or inference method provided by the embodiment of the present application;

[0031] Figure 3a is one of interaction flowcharts of a network side device and a terminal provided by the embodiment of the present application;

[0032] Figure 3b is another of interaction flowcharts of a network side device and a terminal provided by the embodiment of the present application;

[0033] Figure 4 is a third of flowcharts of a model training or inference method provided by the embodiment of the present application;

[0034] Figure 5 is a fourth of flowcharts of a model training or inference method provided by the embodiment of the present application;

[0035] Figure 6a is a third of interaction flowcharts of a network side device and a terminal provided by the embodiment of the present application;

[0036] Figure 6bFigure 4 is a flowchart of an interaction process between a network-side device and a terminal according to an embodiment of the present application;

[0037] Figure 7 Figure 1 is a structural diagram of a model training or inference device according to an embodiment of the present application;

[0038] Figure 8 Figure 2 is another structural diagram of a model training or inference device according to an embodiment of the present application;

[0039] Figure 9 Figure 3 is a third structural diagram of a model training or inference device according to an embodiment of the present application;

[0040] Figure 10 Figure 4 is a fourth structural diagram of a model training or inference device according to an embodiment of the present application;

[0041] Figure 11 Figure 5 is a structural diagram of a communication device according to an embodiment of the present application;

[0042] Figure 12 Figure 6 is a structural diagram of a terminal according to an embodiment of the present application;

[0043] Figure 13 Figure 7 is a structural diagram of a network-side device according to an embodiment of the present application. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0045] With the rapid development of AI technology, in future networks, the architecture of the network will be increasingly integrated with AI technology. Existing AI models are trained and / or inferred on the network side, but the terminal side can obtain more dimensional data than the network side. In order to obtain more information and better protect the data privacy of users, training is performed on the UE side. In order to reduce the resources consumed by the UE side, the model is split, with part of the model executed on the UE side and part of the model executed on the network side. When the environmental conditions (such as available computing power and communication resources) change, different requirements for computing power of the UE and the network side will be generated. In order to ensure that the upper limit resources provided by the UE and the network side are higher than the required resources, the position of the model split point needs to be changed.

[0046] The present application splits the network structure of the model to reduce the processing capability of the terminal and the privacy risk of data, for example, by using a model parallel manner. The model splitting manner is as followsFigure 1 As shown, the AI model is stored in the terminal side and the network side respectively, and in the model training or inference, the UE or the network side will evaluate the current UE, network side and channel state, select a suitable model and its split point.

[0047] The embodiment of the present application provides a model training or inference method, which is used for a first device, and the method comprises the following steps:

[0048] receiving target parameter information for model training or inference sent by a second device;

[0049] determining a target model and a model split point according to the target parameter information;

[0050] sending model information of the target model and the model split point to the second device;

[0051] performing model training or inference according to the target model and the model split point.

[0052] When the terminal and the network side device jointly train or infer the model, the model is split according to the model split point, and the part on one side of the model split point is executed on the terminal side, and the part on the other side of the model split point is executed on the network side device. The model and the model split point can be determined by the terminal or the network side device. That is, in the above description, the first device is the network side device, and the second device is the terminal, or the first device is the terminal, and the second device is the network side device. The following describes that the model and the model split point are determined by the terminal or the network side device.

[0053] Figure 2a FIG. 1 is a flowchart of a model training or inference method provided by an embodiment of the present application, which is applied to a network side device, and the method comprises the following steps:

[0054] Step 101, receiving first parameter information (in this embodiment, the target parameter information is referred to as the first parameter information) for model training or inference sent by a terminal.

[0055] The terminal can periodically send the first parameter information, or send the first parameter information according to a preset condition, for example, the terminal sends the first parameter information after receiving a model assistance request sent by the network side device, and the model assistance request is used for requesting the terminal and the network side device to jointly train or infer the model.

[0056] The first parameter information comprises at least one of the following:

[0057] (1) Task accuracy, which is a kind of index for measuring prediction effect (which can also be referred to as inference effect), and is a ratio of an accurate result predicted by the model to a total number of samples. This parameter can be used for comparison with model accuracy, and is used for model selection.

[0058] (2) Task latency, the maximum latency that the current task can accept, this parameter can be used for comparison with model latency for model selection;

[0059] (3) Power consumption of the terminal, since the power consumption of the terminal (hereinafter also referred to as UE) depends on the battery power, the power consumption provided by the UE is limited. For example, when the following conditions are met, the parameter is reported, when the battery power decreases, the power consumption provided by the UE will decrease, the UE reports this parameter, when the network side receives it, when selecting the model split point, the location selection of the split point will be more inclined to the side close to the input layer of the neural network, when the model is trained or inferred, the UE side will have less calculation amount, thereby reducing the power consumption of the UE;

[0060] (4) Memory of the terminal, UE memory is used to measure the parameter amount that can run the model, for example, when the UE memory becomes less, the UE reports this parameter, at this time the location selection of the model split point will be more inclined to the side close to the input layer of the neural network, and the parameter amount stored by the UE will be smaller;

[0061] (5) Calculation amount of the terminal, used to measure the complexity of the algorithm and / or model that the UE can calculate, the UE reports this parameter in the following ways: periodic reporting; reporting under certain conditions, for example, when the calculation amount of the UE decreases, at this time the location selection of the model split point will be more inclined to the side close to the input layer of the neural network, and the UE side will have less calculation amount, thereby ensuring that the UE can complete the calculation.

[0062] Step 102, determining the model information of the target model and the model split point according to the first parameter information.

[0063] The network side device comprehensively evaluates the task accuracy, task latency, UE power consumption, UE memory, UE calculation amount, data amount, channel state and the like to determine the target model and the model split point. The network side device will have multiple model structures for a training task, and can evaluate and select the model through the indexes of task accuracy, task latency, UE memory and UE calculation amount. For example, the model accuracy and the task latency are compared with the model accuracy and the model latency provided by each model structure, the network side device will select the model whose indexes are closest to the conditions, that is, the model whose model accuracy is higher than the task accuracy and whose model latency is lower than the task latency; if the UE memory and the UE calculation amount are lower than the required resources of the model, the demand for a large model cannot be met, and then a model structure with low complexity is selected.

[0064] The model split point is determined by the power consumption of the UE, the UE memory, the UE computing amount, and the data amount index. The model split point can be selected to split after each layer of the neural network. The network side tends to select a position with less data transmission amount as the split point. When the power consumption of the UE or the UE memory or the UE computing amount is lower than the required resources of the model, the model split point will move to the UE side. When the channel state quality deteriorates, the model split point tends to select a split point position with less data amount.

[0065] The model split point is used to split the target model, and the two parts of the target model obtained can be trained or inferred by the terminal and the network side device respectively. For example, the target model has five layers of neural networks, and the model split point is after the second layer of neural network. Based on the model split point, the target model can be divided into two layers and three layers. The first two layers of the target model are trained or inferred on the terminal, and the last three layers of the target model are trained or inferred on the network side device.

[0066] In step 103, the model information of the target model and the model split point are sent to the terminal.

[0067] The model information can include a model identifier of the target model and model configuration information. The model identifier is used to identify the selected AI model, and the model configuration information is used to configure the target model. For example, the model configuration information includes at least one of model structure configuration, model parameters, model hyperparameters, and loss function.

[0068] In step 104, model training or inference is performed according to the target model and the model split point.

[0069] The network side device sends the model information and the model split point to the terminal. The terminal can perform model training or inference based on the model information and the model split point. The network side device also performs model training or inference according to the target model and the model split point. In the process of jointly training and inferring the target model, the terminal and the network side device also need to interact with intermediate parameters to realize joint training or inference of the target model.

[0070] In this embodiment, the target model and the model split point are determined by the network side device.

[0071] In the embodiment, the first parameter information for model training or inference sent by the terminal is received; the target model and the model segmentation point are determined according to the first parameter information; the model information of the target model and the model segmentation point are sent to the terminal; and the model training or inference is performed according to the target model and the model segmentation point. In the above, the network side device can dynamically determine the target model and the model segmentation point, and the way of determining the target model and the model segmentation point is more flexible. Since the model segmentation point can be flexibly set according to the actual situation, the model training or inference failure caused by the limited computing power of the terminal or the network side device can be avoided, and the effectiveness of the model training or inference can be improved.

[0072] In an embodiment of the present application, the first parameter information for model training or inference sent by the terminal includes:

[0073] The model assistance request is sent to the terminal, and the model assistance request includes the task type of the model and the model action information;

[0074] The first parameter information sent by the terminal is received;

[0075] Alternatively, the first parameter information for model training or inference sent by the terminal includes:

[0076] The first parameter information for model training or inference sent by the terminal is received, and the task type of the model and the model action information are also received.

[0077] In the above, the terminal can send the first parameter information based on the model assistance request sent by the network side device, or the terminal can actively send the first parameter information, for example, periodically. In addition, when the terminal actively sends the first parameter information, the task type of the model and the model action information are also sent. The model task type is used to represent the function of the model, for example, the model task type is a model for requesting traffic prediction. The model action information is used to indicate that the target model is used for training or inference, for example, 0 can be selected as the training identifier, and 1 can be selected as the inference identifier.

[0078] In an embodiment of the present application, the model training or inference is performed according to the target model and the model segmentation point, including:

[0079] The first intermediate parameter sent by the terminal is received;

[0080] If it is determined according to the model action information that the target model is used for inference, then the inference is performed according to the first intermediate parameter and the target model, and the inference result is obtained;

[0081] The inference result is sent to the terminal.

[0082] Specifically, if it is determined according to the model action information that the target model is used for inference, the network side device performs inference according to the first intermediate parameter and the target model, completes the prediction of the data, and the inference process is completed. For example, the network side device can send the inference result (which can also be referred to as a prediction result) to the UE through RRCReconfiguration signaling, and the UE can confirm the prediction result through RRCReconfigurationComplete signaling and transmit a prediction process completion notification to the network side device.

[0083] In an embodiment of the present application, the model training or inference according to the target model and the model split point comprises:

[0084] receiving the first intermediate parameter sent by the terminal;

[0085] If it is determined according to the model action information that the target model is used for training, the target model is trained according to the first intermediate parameter to obtain a second intermediate parameter;

[0086] sending the second intermediate parameter to the terminal;

[0087] receiving a new intermediate parameter sent by the terminal;

[0088] determining the new intermediate parameter as the first intermediate parameter, and performing the step of training the target model according to the first intermediate parameter to obtain a second intermediate parameter until a maximum training round number is reached or a neural network of the target model converges.

[0089] The network side device and the terminal realize joint training of the target model by interacting intermediate parameters. Specifically, the network side device receives the first intermediate parameter sent by the terminal, trains the target model based on the first intermediate parameter to obtain a second intermediate parameter, and sends the second intermediate parameter to the terminal. The terminal trains the target model based on the second intermediate parameter to obtain a new intermediate parameter, and sends the new intermediate parameter to the network side device. The network side device trains the target model according to the new intermediate parameter, and repeatedly performs the above steps until the terminal reaches a maximum training round number or reaches a convergence threshold. In this case, the terminal can notify the network side device to stop training, or the network side device notifies the terminal to stop training.

[0090] In an embodiment of the present application, the receiving of the new intermediate parameter sent by the terminal comprises:

[0091] If the first state information sent by the terminal is received, a new model split point of the target model is determined according to the first state information;

[0092] sending the new model split point to the terminal;

[0093] receiving a new intermediate parameter sent by the terminal based on the new model split point.

[0094] Specifically, if the state of the terminal changes, for example, the power consumption, memory, and calculation amount of the UE change, during the training of the target model, the terminal sends first state information to the network side device, and the first state information can include the current state of the terminal, for example, the power consumption, memory, and calculation amount of the UE, so as to facilitate the network side device to determine the model split point again based on the first state information.

[0095] The network side device sends the new model split point to the terminal, and the terminal trains according to the new model split point and sends the obtained new intermediate parameter to the network side device.

[0096] Through the above process, the network side device can dynamically determine the model split point based on the state of the terminal, guarantee the smooth progress of the terminal model training, and improve the effectiveness of the model training.

[0097] Figure 2b is a flowchart of a model training or inference method provided by an embodiment of the application, applied to a terminal which can interact with a network side device in the embodiment shown in the figure, and the method comprises the following steps: Figure 2a

[0098] Step 201, sending first parameter information (in this embodiment, target parameter information is referred to as first parameter information) for model training or inference to the network side device.

[0099] The terminal can send the first parameter information periodically or according to a preset condition, for example, the terminal sends after receiving the model assistance request sent by the network side device, without limitation, and the model assistance request is used to request the terminal and the network side device to jointly train or infer.

[0100] The first parameter information comprises at least one of the following:

[0101] (1) task accuracy, a kind of index for measuring prediction effect (also known as inference effect), indicating the ratio of accurate results of model prediction to total number of samples, which can be used for comparison with model accuracy and model selection;

[0102] (2) task delay, the maximum delay that can be accepted by the current task, which can be used for comparison with model delay and model selection;

[0103] ​(3) the power consumption of the terminal, since the power consumption of the UE depends on the battery power, the power consumption provided by the UE is limited. For example, when the following conditions are met, the parameter is reported, when the battery power decreases, the power consumption provided by the UE will decrease, the UE reports this parameter, when the network side receives it, when selecting the model segmentation point, the position selection of the segmentation point will be more inclined to the side close to the input layer of the neural network, when the model is trained or inferred, the UE side will have less calculation amount, thereby reducing the power consumption of the UE;

[0104] (4) the memory of the terminal, the UE memory is used to measure the parameter amount of the model that can be run, for example, when the UE memory becomes smaller, the UE reports this parameter, at this time, the position selection of the model segmentation point will be more inclined to the side close to the input layer of the neural network, and the parameter amount stored by the UE will be smaller;

[0105] (5) the calculation amount of the terminal, which is used to measure the complexity of the algorithm and / or model that can be calculated by the UE. The UE can report this parameter in the following ways: periodic reporting; reporting under certain conditions, for example, when the calculation amount of the UE decreases, at this time, the position selection of the model segmentation point will be more inclined to the side close to the input layer of the neural network, and the UE side will have less calculation amount, thereby ensuring that the UE can complete the calculation.

[0106] Step 202, receiving the model information and the model segmentation point of the target model sent by the network side device.

[0107] The target model is a model determined by the network side device for joint training or inference. The model information can include a model identifier of the target model and model configuration information. The model identifier is used to identify the selected AI model, and the model configuration information is used to configure the target model, for example, the model configuration information includes at least one of model structure configuration, model parameters, model hyperparameters, and loss function.

[0108] The model segmentation point is used to segment the target model, and the two parts of the obtained target model can be trained or inferred by the terminal and the network side device respectively. For example, the target model has five layers of neural networks, and the model segmentation point is behind the second layer of neural networks. Based on the model segmentation point, the target model can be divided into two layers and three layers. The first two layers of the target model are trained or inferred on the terminal, and the last three layers of the target model are trained or inferred on the network side device.

[0109] Step 203, performing model training or inference according to the target model and the model segmentation point.

[0110] The terminal can perform model training or inference based on the model information and the model split point, and the network side device also performs model training or inference based on the target model and the model split point. In the process of jointly training and inferring the target model, the terminal and the network side device also need to interact intermediate parameters to achieve joint training or inference of the target model.

[0111] In the above, the target model and the model split point are determined by the network side device.

[0112] In the embodiment, the first parameter information for model training or inference is sent to the network side device, the model information of the target model and the model split point sent by the network side device are received, and model training or inference is performed according to the target model and the model split point. In the above, the network side device can dynamically determine the target model and the model split point, and the determination of the target model and the model split point is more flexible. Since the model split point can be flexibly set according to actual conditions, the model training or inference failure caused by limited computing power of the terminal or the network side device can be avoided, and the effectiveness of model training or inference can be improved.

[0113] In an embodiment of the present application, the first parameter information for model training or inference is sent to the network side device, including:

[0114] The first parameter information for model training or inference is sent to the network side device, and a model assistance request is sent, the model assistance request including model task type and model action information.

[0115] The terminal can actively send the first parameter information, for example, periodically, and also send the model task type and the model action information. The model task type is used to represent the function of the model, for example, the model task type is a model requesting traffic prediction. The model action information is used to indicate that the target model is used for training or inference, for example, 0 can be selected as the training identifier and 1 as the inference identifier.

[0116] In an embodiment of the present application, the terminal can also send the first parameter information based on the model assistance request sent by the network side device, that is, before the first parameter information for model training or inference is sent to the network side device, the method further includes:

[0117] The model assistance request sent by the network side device is received, and the model assistance request includes model task type and model action information.

[0118] In an embodiment of the present application, the model training or inference according to the target model and the model split point includes:

[0119] If it is determined according to the model action information that the target model is used for inference, model inference is performed according to the target model and the model split point, and a first intermediate parameter is obtained;

[0120] The first intermediate parameter is sent to the network side device;

[0121] The inference result sent by the network side device is received.

[0122] Specifically, if it is determined according to the model action information that the target model is used for inference, the terminal sends the first intermediate parameter to the network side device, the network side device performs inference according to the first intermediate parameter and the target model, completes the prediction of data, and the inference process is completed. For example, the network side device can send the inference result (which can also be referred to as a prediction result) to the UE through RRCReconfiguration signaling, and the UE can confirm the prediction result through RRCReconfigurationComplete signaling and transmit a prediction process completion notification to the network side device.

[0123] In an embodiment of the present application, the model training or inference according to the target model and the model split point comprises:

[0124] If it is determined according to the model action information that the target model is used for training, model training is performed according to the target model and the model split point, and a first intermediate parameter is obtained;

[0125] The first intermediate parameter is sent to the network side device;

[0126] The second intermediate parameter sent by the network side device is received.

[0127] The target model is trained according to the second intermediate parameter until a maximum training round number is reached or a neural network of the target model converges.

[0128] The network side device and the terminal implement joint training of the target model by interacting intermediate parameters. Specifically, the terminal sends a first intermediate parameter to the network side device, the network side device receives the first intermediate parameter sent by the terminal, trains the target model based on the first intermediate parameter, obtains a second intermediate parameter, and sends the second intermediate parameter to the terminal. The terminal trains the target model based on the second intermediate parameter, obtains a new intermediate parameter, and sends the new intermediate parameter to the network side device. The network side device trains the target model according to the new intermediate parameter, and repeatedly performs the above steps until the terminal reaches a maximum training round number or a convergence threshold. In this case, the terminal can notify the network side device to stop training, or the network side device notifies the terminal to stop training.

[0129] In an embodiment of the present application, the training of the target model according to the second intermediate parameter is performed until a maximum number of training rounds is reached or the neural network of the target model converges, comprising:

[0130] During the training of the target model according to the second intermediate parameter, if the state of the terminal changes, first state information is obtained;

[0131] The first state information is sent to the network side device;

[0132] The new model split point of the target model sent by the network side device is received;

[0133] The target model is trained based on the new model split point to obtain new intermediate parameters;

[0134] The new intermediate parameters are determined as the first intermediate parameters, and the step of sending the first intermediate parameters to the network side device is performed until a maximum number of training rounds is reached or the neural network of the target model converges.

[0135] Specifically, if the state of the terminal changes, such as changes in UE power consumption, UE memory, and UE computing capacity, during the training of the target model, the terminal sends first state information to the network side device. The first state information can include the current state of the terminal, such as UE power consumption, UE memory, and UE computing capacity, to facilitate the network side device to determine the model split point based on the first state information.

[0136] The network side device sends the new model split point to the terminal, and the terminal trains according to the new model split point and sends the obtained new intermediate parameters to the network side device.

[0137] Through the above process, the network side device can dynamically determine the model split point based on the state of the terminal, ensure the smooth progress of the terminal model training, and improve the effectiveness of the model training.

[0138] As shown in Figure 3a and Figure 3b is a schematic diagram of the interaction between the network side device and the terminal, Figure 3a is a schematic diagram of the model training process, Figure 3b is a schematic diagram of the model inference process;

[0139] As shown in Figure 3a , the model training method comprises:

[0140] Step 311, the network (the network refers to the network side device, the same below) sends a model assistance request to the UE, which can be sent through a model request. The request carries the model task name, the identification of assistance training and / or inference. The model task is, for example, a request for a traffic prediction model; 0 can be selected as the training identification, and 1 as the inference identification. Step 311 is an optional step, which can be executed or not executed.

[0141] Step 312, if step 311 is executed, the UE sends the reported parameters of the UE to the network, otherwise the UE sends the model assistance request, the reported parameters of the UE, and the identification of training and / or inference to the network side.

[0142] Step 313, the network determines which model and model partition point to select according to the current UE state, network side state, and channel state (such as model accuracy, task delay, task accuracy, UE power consumption, UE memory, UE computing amount, network side memory, network side computing amount, model memory, model delay, model computing amount, channel state, data amount, etc.). The network confirms the received model task and the reported parameters of the UE, and notifies the UE of the selected model configuration and model partition point. The model configuration can be performed through signaling such as RRC setup, and the model configuration information includes at least one of: model structure configuration, model parameter, model hyperparameter, etc.

[0143] Step 314, the UE confirms the received parameters.

[0144] Step 315, the UE and the network transmit intermediate parameters for model training or inference.

[0145] The training or inference identification shows that the current state is a training state, and the following steps 316-318 are executed.

[0146] Step 316, when the state of the UE changes, such as UE power consumption, UE memory, UE computing amount, the terminal reports the current parameters through signaling such as Resource reporting information.

[0147] Step 317, the network determines which model partition point to select according to the current UE state, network side state, and channel state. The network side retransmits the partition position of the model to the UE through signaling such as RRCReconfiguration.

[0148] Steps 316-317 are optional steps, which are executed only when the state of the UE changes, otherwise this step is skipped.

[0149] Step 318, the UE confirms the received parameters, for example, through RRCReconfigurationComplete.

[0150] Steps 314-317 are repeated until the UE reaches the maximum number of training rounds or reaches the convergence threshold.

[0151] Step 319, the UE informs the network to stop model training, or the network informs the UE to stop model training, for example, through the Training termination notice.

[0152] Step 3110, the network informs the UE that the training is complete, or the UE informs the network that the training is complete, such as through the Training completed notice.

[0153] As shown in Figure 3b The model inference method includes:

[0154] Step 321, the network (the network refers to the network side device, the same below) sends a model assistance request to the UE, which can be sent through the model request. The request carries the model task name, the identification of assistance training and / or inference. The model task is, for example, the request traffic prediction model; 0 can be selected as the training identification, and 1 as the inference identification. Step 311 is an optional step, which can be executed or not executed.

[0155] Step 322, if step 311 is executed, the UE sends the UE's reported parameters to the network, otherwise the UE sends the model assistance request, the UE's reported parameters, and the identification of training and / or inference to the network side.

[0156] Step 323, the network determines which model and model partition point to select through the current UE state, network side state, and channel state (such as model accuracy, task delay, task accuracy, UE power consumption, UE memory, UE calculation amount, network side memory, network side calculation amount, model memory, model delay, model calculation amount, channel state, data amount, etc.). The network confirms the received model task and the UE's reported parameters, and notifies the UE of the selected model configuration and model partition point. The model configuration can be performed through signaling such as RRC setup, and the model configuration information includes at least one of: model structure configuration, model parameter, model hyperparameter, etc.

[0157] Step 324, the UE confirms the received parameters.

[0158] Step 325, the UE and the network transmit intermediate parameters for model training or inference.

[0159] If the training or inference identification shows that the current state is inference, after the UE transmits the intermediate parameters to the network, the network side will transmit the inference result to the UE, complete the data prediction, and complete the inference process;

[0160] Step 326, the network sends the inference result to the UE, for example, the model prediction result can be transmitted by the network to the UE through RRCReconfiguration;

[0161] Step 327, the UE sends a prediction process completion notification to the network, for example, through RRCReconfigurationComplete, the prediction result is confirmed and the prediction process completion notification is transmitted by the UE to the network.

[0162] When the state of the UE changes, the model partition point between the UE and the network side can be changed in time, and the model training or inference is completed by the way of delivering the position of the model partition point and the training parameters through the air interface. The model partitioning method can effectively reduce the parameters of the terminal, reduce the storage space, and reduce the requirement for the computing power of the terminal. By changing the model partition point, the condition (such as available computing power and communication resources) can be better adapted to the change, and the ability of the AI prediction can be reasonably utilized.

[0163] Figure 4 is a flowchart of a model training or inference method provided by an embodiment of the application, applied to a terminal, in this embodiment, the target model and the model partition point are determined by the terminal, and the method comprises the following steps:

[0164] Step 401, receiving second parameter information (in this embodiment, the target parameter information is referred to as second parameter information) for model training or inference sent by a network side device.

[0165] The network side device can actively send the second parameter information to the terminal, or can send it according to a preset condition, for example, the network side device sends it after receiving the model assistance request sent by the terminal, which is not limited herein, and the model assistance request is used to request the terminal and the network side device to jointly train or infer.

[0166] The second parameter information comprises at least one of the following:

[0167] (1) Model accuracy: a kind of index for measuring prediction effect, indicating the ratio of accurate results of model prediction to total number of samples, for the same task, the network side device has different model sizes of models, this parameter is compared with the task accuracy, and is used to judge the selected model.

[0168] (2) Model delay: the time required for model inference, the network side device reports this parameter to the UE side, and informs the UE of the time required for the whole model prediction, this parameter is compared with the task delay, and is used to judge the selected model.

[0169] (3) Model memory: used to measure the data size of a model. Model memory includes: model memory size of neural network before model split point and model memory size of neural network after model split point. According to this parameter, it can be evaluated whether the UE and the network side device have enough memory to run the model.

[0170] (4) Model calculation amount: used to measure the calculation amount required by the model. Model calculation amount includes: model calculation amount size of neural network before model split point and model calculation amount size of neural network after model split point. According to this parameter, it can be evaluated whether the UE and the network side device have enough calculation amount to run the model.

[0171] (5) Network side device memory: used to measure the parameter size that can run the model, for example, when the network side device memory becomes smaller, the network side device reports this parameter, at this time the position selection of the model split point will be more inclined to the output side of the neural network, and the parameter size stored by the network side device will be smaller.

[0172] (6) Network side device calculation amount: used to measure the complexity of the algorithm and / or model that can be calculated by the network side device, for example, when the calculation amount of the network side device becomes smaller, the network side device reports this parameter, at this time the position selection of the model split point will be more inclined to the output side of the neural network, and the network side device will have less calculation amount, thereby ensuring that the network side device can complete the calculation.

[0173] (7) Data amount: the size of the parameter exchanged between the network side device and the UE side, used for the UE to judge the selection of the model split point.

[0174] Step 402, determining a target model and a model split point according to the second parameter information.

[0175] The UE comprehensively evaluates the model accuracy, the network side device memory, the model delay, the network side device calculation amount, the model memory, the model calculation amount, and the channel state, and judges whether the required resources for executing the AI model can be achieved, so as to judge the selected AI model. The model identifier of the target model can be used to identify the selected AI model. The selected model split point is evaluated by the indicators of the network side device memory, the network side device calculation amount, the channel state, and the data amount. The position of the model split point can be selected to split after each layer of neural network. The network side device tends to select a position with less data transmission amount as the split point. When the network side device memory and the network side device calculation amount are lower than the required resources of the model, the split point of the model will move to the network side device. When the channel state quality becomes poor, the model split point will tend to select a split point position with less data amount.

[0176] Step 403, sending the model identifier of the target model and the model split point to the network side device.

[0177] The model configuration information of the target model can be configured by the network side device after receiving the model identifier of the target model, and the model configuration information is sent to the terminal. The model configuration information is used to configure the target model, for example, the model configuration information includes at least one of model structure configuration, model parameters, model hyperparameters, and loss function.

[0178] Step 404, model training or inference is performed according to the target model and the model split point.

[0179] In the process of jointly training and inferring the target model by the terminal and the network side device, the intermediate parameters are interacted to realize the joint training or inference of the target model.

[0180] In this embodiment, the second parameter information for model training or inference sent by the network side device is received; the target model and the model split point are determined according to the second parameter information; the model identifier of the target model and the model split point are sent to the network side device; and model training or inference is performed according to the target model and the model split point. In the above, the terminal can dynamically determine the target model and the model split point, and the way of determining the target model and the model split point is more flexible. Since the model split point can be flexibly set according to actual conditions, the failure of model training or inference caused by limited computing power of the terminal or the network side device can be avoided, and the effectiveness of model training or inference can be improved.

[0181] In an embodiment of the present application, before the second parameter information for model training or inference sent by the network side device is received, the method further includes:

[0182] The model assistance request sent by the network side device is received; and a target request is sent to the network side device.

[0183] Alternatively, a target request and the model assistance request are sent to the network side device.

[0184] The target request is a request for the terminal to perform model and split point evaluation, and the model assistance request includes model task type and model action information.

[0185] In the above, before the second parameter information for model training or inference sent by the network side device is received, the terminal can send a target request to the network side device based on the model assistance request sent by the network side device to request the terminal to perform model and split point evaluation; or the terminal actively sends a target request and a model assistance request to the network side device.

[0186] The model task type is used to represent the function of the model, for example, the model task type is a model for requesting traffic prediction. The model action information is used to indicate that the target model is used for training or inference, for example, 0 can be selected as the training identifier and 1 as the inference identifier.

[0187] In an embodiment of the present application, the model training or inference according to the target model and the model split point comprises:

[0188] If it is determined according to the model action information that the target model is used for inference, the model inference is performed according to the target model and the model split point, and a first intermediate parameter is obtained.

[0189] The first intermediate parameter is sent to the network side device.

[0190] The inference result sent by the network side device is received.

[0191] Specifically, if it is determined according to the model action information that the target model is used for inference, the terminal sends the obtained first intermediate parameter to the network side device, the network side device performs inference according to the first intermediate parameter and the target model, completes the prediction of data, and the inference process is completed. For example, the network side device can send the inference result (also referred to as the prediction result) to the UE through RRCReconfiguration signaling, confirm the prediction result through RRCReconfigurationComplete signaling, and transmit a prediction process completion notification from the UE to the network side device.

[0192] In an embodiment of the present application, the model training or inference according to the target model and the model split point comprises:

[0193] If it is determined according to the model action information that the target model is used for training, the model training is performed according to the target model and the model split point, and a first intermediate parameter is obtained.

[0194] The first intermediate parameter is sent to the network side device.

[0195] The second intermediate parameter sent by the network side device is received.

[0196] The target model is trained according to the second intermediate parameter until the maximum number of training rounds is reached or the neural network of the target model converges.

[0197] Specifically, the network side device and the terminal implement joint training on the target model by interacting intermediate parameters. Specifically, the network side device receives the first intermediate parameter sent by the terminal, trains the target model based on the first intermediate parameter to obtain a second intermediate parameter, and sends the second intermediate parameter to the terminal. The terminal trains the target model based on the second intermediate parameter to obtain a new intermediate parameter, and sends the new intermediate parameter to the network side device. The network side device trains the target model according to the new intermediate parameter, and repeats the above steps until the terminal reaches a maximum training round or reaches a convergence threshold. In this case, the terminal can notify the network side device to stop training, or the network side device notifies the terminal to stop training.

[0198] In an embodiment of the present application, the training of the target model according to the second intermediate parameter until reaching a maximum training round or the neural network of the target model converges comprises:

[0199] During the process of training the target model according to the second intermediate parameter, if the second state information sent by the network side device is received, a new model partition point of the target model is determined according to the second state information;

[0200] The new model partition point is sent to the network side device.

[0201] The new intermediate parameter sent by the network side device is received, and the new intermediate parameter is taken as the second intermediate parameter to perform the step of training the target model according to the second intermediate parameter until reaching a maximum training round or the neural network of the target model converges.

[0202] Specifically, if the state of the network side device changes during the process of training the target model, for example, the memory of the network side device and the UE calculation amount change, the network side device sends second state information to the terminal. The second state information can include the current state of the network side device, for example, the power consumption of the network side device, the memory of the network side device, the calculation amount of the network side device, and the like, so as to facilitate the terminal to determine the model partition point again based on the second state information.

[0203] The terminal sends the new model partition point to the network side device, the network side device trains according to the new model partition point, and sends the obtained new intermediate parameter to the terminal, to implement joint training of the target model by the terminal and the network side device.

[0204] Through the above process, the terminal can dynamically determine the model partition point based on the state of the network side device, guaranteeing the smooth progress of the model training of the terminal and the network side device, and improving the effectiveness of the model training.

[0205] Figure 5is a flowchart of a model training or inference method provided by an embodiment of the present application, applied to a network side device which can interact with Figure 4 the terminal in the illustrated embodiment to implement joint training or inference of the model, and the method includes the following steps:

[0206] Step 501: sending second parameter information (in this embodiment, the target parameter information is referred to as second parameter information) for model training or inference to the terminal.

[0207] The network side device can actively send the second parameter information to the terminal, or can send it according to a preset condition, for example, the network side device sends it after receiving a model assistance request sent by the terminal, which is used to request joint training or inference of the terminal and the network side device, without limitation.

[0208] The second parameter information includes at least one of the following:

[0209] (1) Model accuracy: a kind of index for measuring prediction effect, indicating the ratio of accurate results predicted by the model to the total number of samples. For the same task, the network side device will have models of different sizes, and this parameter is compared with the task accuracy to determine the selected model.

[0210] (2) Model latency: the time required for model inference. The network side device reports this parameter to the UE side to inform the UE of the time required for the entire model prediction. This parameter is compared with the task latency to determine the selected model.

[0211] (3) Model memory: used to measure the data size of a model. The model memory includes the model memory size of the neural network before the model split point and the model memory size of the neural network after the model split point. According to this parameter, it can be evaluated whether the UE and the network side device have sufficient memory to run the model.

[0212] (4) Model computation: used to measure the computation required by the model. The model computation includes the model computation size of the neural network before the model split point and the model computation size of the neural network after the model split point. According to this parameter, it can be evaluated whether the UE and the network side device have sufficient computation to run the model.

[0213] (5) Network side device memory: the network side memory is used to measure the parameter size that can run the model. For example, when the network side device memory becomes smaller, the network side device reports this parameter, and at this time, the position of the model split point will be more inclined to the output side of the neural network, and the parameter size stored by the network side device will become smaller

[0214] (6) Network side device computation: used to measure the complexity of the algorithm and / or model that the network side device can calculate, for example, when the network side device reports this parameter when the network side device computation is reduced, the location of the model segmentation point will be more inclined to be close to the output side of the neural network, the network side device will have less computation, and thus ensure that the network side device can complete the calculation.

[0215] (7) Data volume: the size of the parameter exchanged between the network side device and the UE side, used by the UE to judge the selection of the model segmentation point.

[0216] Step 502, receiving the model identifier of the target model and the model segmentation point sent by the terminal.

[0217] The UE comprehensively evaluates the model accuracy, network side device memory, model delay, network side device computation, model memory, model computation, and channel state, and judges whether the required resources for executing the AI model can be achieved, so as to judge the selected AI model. The model identifier of the target model can be used to identify the selected AI model. The selected model segmentation point is evaluated by the network side device memory, network side device computation, channel state, and data volume. The model segmentation point can be selected to be segmented after each layer of neural network. The network side device tends to select a position with less data transmission as the segmentation point. If the network side device memory and the network side device computation are lower than the required resources of the model, the model segmentation point will move to the network side device. When the channel state quality deteriorates, the model segmentation point will tend to select a position with less data volume as the segmentation point.

[0218] The terminal determines the target model and the model segmentation point, and sends the model identifier of the target model and the model segmentation point to the network side device.

[0219] Step 503, model training or inference according to the target model and the model segmentation point.

[0220] The model configuration information of the target model can be configured by the network side device after receiving the model identifier of the target model, and the model configuration information is sent to the terminal. The model configuration information is used to configure the target model, for example, the model configuration information includes at least one of model structure configuration, model parameter, model hyperparameter, and loss function.

[0221] During the joint training and inference of the target model by the terminal and the network side device, intermediate parameters are exchanged to realize the joint training or inference of the target model.

[0222] In the embodiment, the second parameter information for model training or inference is sent to the terminal, the model identifier of the target model and the model segmentation point sent by the terminal are received, and model training or inference is performed according to the target model and the model segmentation point. In the above, the terminal can dynamically determine the target model and the model segmentation point, and the way of determining the target model and the model segmentation point is more flexible. Since the model segmentation point can be flexibly set according to the actual situation, the model training or inference failure caused by the limited computing power of the terminal or the network side device can be avoided, and the effectiveness of the model training or inference can be improved.

[0223] In an embodiment of the present application, before the second parameter information for model training or inference is sent to the terminal, the method further comprises:

[0224] sending a model assistance request to the terminal, and receiving the target request and the model assistance request sent by the terminal;

[0225] Alternatively, the target request and the model assistance request sent by the terminal are received.

[0226] The target request is a request for the terminal to evaluate the model and the segmentation point, and the model assistance request includes the task type of the model and the model action information.

[0227] In the above, before the second parameter information for model training or inference is sent to the terminal, the network side device can actively send a model assistance request to the network side device, and the terminal can send a target request to the network side device based on the model assistance request sent by the network side device to request the terminal to evaluate the model and the segmentation point. Alternatively, the network side device receives the target request and the model assistance request sent by the terminal.

[0228] The model task type is used to represent the function of the model, for example, the model task type is a model that requests traffic prediction. The model action information is used to indicate that the target model is used for training or inference, for example, 0 can be selected as the training identifier and 1 as the inference identifier.

[0229] In an embodiment of the present application, the model training or inference according to the target model and the model segmentation point comprises:

[0230] receiving the first intermediate parameter sent by the terminal;

[0231] If it is determined according to the model action information that the target model is used for inference, inference is performed according to the first intermediate parameter and the target model to obtain an inference result.

[0232] The inference result is sent to the terminal.

[0233] Specifically, if it is determined according to the model action information that the target model is used for inference, the terminal sends the obtained first intermediate parameter to the network side device, the network side device performs inference according to the first intermediate parameter and the target model, completes the prediction of the data, and the inference process is completed. For example, the network side device can send the inference result (also referred to as the prediction result) to the UE through RRCReconfiguration signaling, confirm the prediction result through RRCReconfigurationComplete signaling, and transmit a prediction process completion notification from the UE to the network side device.

[0234] In an embodiment of the present application, the model training or inference according to the target model and the model split point comprises:

[0235] receiving the first intermediate parameter sent by the terminal;

[0236] If it is determined according to the model action information that the target model is used for training, the target model is trained according to the first intermediate parameter to obtain a second intermediate parameter;

[0237] sending the second intermediate parameter to the terminal;

[0238] receiving the new intermediate parameter sent by the terminal, determining the new intermediate parameter as the first intermediate parameter, and performing the step of training the target model according to the first intermediate parameter to obtain a second intermediate parameter until a maximum training round number is reached or a neural network of the target model converges.

[0239] Specifically, the network side device and the terminal realize joint training of the target model by interacting intermediate parameters. Specifically, the network side device receives the first intermediate parameter sent by the terminal, trains the target model based on the first intermediate parameter to obtain a second intermediate parameter, and sends the second intermediate parameter to the terminal. The terminal trains the target model based on the second intermediate parameter to obtain a new intermediate parameter, and sends the new intermediate parameter to the network side device. The network side device trains the target model according to the new intermediate parameter, and repeatedly performs the above steps until the terminal reaches a maximum training round number or reaches a convergence threshold. In this case, the terminal can notify the network side device to stop training, or the network side device notifies the terminal to stop training.

[0240] In an embodiment of the present application, the training of the target model according to the first intermediate parameter to obtain a second intermediate parameter comprises:

[0241] If the state of the network side device changes, a second state information is obtained;

[0242] sending the second state information to the terminal;

[0243] receiving the new model split point of the target model sent by the terminal;

[0244] training the target model based on the new model split point, to obtain the second intermediate parameter.

[0245] Specifically, if the state of the network side device changes during the training of the target model, for example, the memory of the network side device, the UE calculation amount changes, the network side device sends the second state information to the terminal. The second state information can include the current state of the network side device, for example, the power consumption of the network side device, the memory of the network side device, the calculation amount of the network side device, etc., so as to facilitate the terminal to determine the model split point based on the second state information.

[0246] The terminal sends the new model split point to the network side device, and the network side device trains according to the new model split point and sends the obtained new intermediate parameter to the terminal, realizing the joint training of the target model by the terminal and the network side device.

[0247] Through the above process, the terminal can dynamically determine the model split point based on the state of the network side device, guarantee the smooth progress of the model training of the terminal and the network side device, and improve the effectiveness of the model training.

[0248] As shown in Figure 6a and Figure 6b is an interaction schematic diagram of the network side device and the terminal, Figure 6a is a flowchart of model training, Figure 6b is a flowchart of model inference.

[0249] As shown in Figure 6a , the model training method comprises:

[0250] Step 611, the network (the network refers to the network side device, the same below) sends a model assistance training request to the UE, and the identification of assistance training and / or inference, for example, through model request, the message carries the model task name, such as requesting traffic prediction model, this step is optional.

[0251] Step 612, the UE sends a UE evaluation model and split point request to the network, if step 611 exists, the received data is confirmed, otherwise the UE sends a model assistance training request to the network side, and the identification of assistance training and / or inference.

[0252] Step 613, the network confirms the received model task, UE evaluation model and split point request, and the network reports the index of the candidate model on the network side according to the model task, for example, through Resource reporting information.

[0253] The indicators include at least one of the following: model accuracy, network side memory, model latency, network side computation, model memory, model computation, data volume, etc.

[0254] Step 614: The UE determines which model and model split point to select through the current UE, network side, and channel state (such as task accuracy, task latency, model accuracy, UE power consumption, UE memory, UE computation, network side memory, network side computation, model memory, model latency, model computation, channel state, data volume, etc.). The UE notifies the network side of the selected model and model split point through signaling such as model mode selection.

[0255] Step 615: The network side confirms the received parameters through signaling such as model mode selection acknowledge, and transmits the model configuration required by the UE according to the model selected by the UE side, including one or more of the following: model structure configuration, model parameter, model hyperparameter, loss function, etc.

[0256] Step 616: The UE and the network side transmit intermediate parameters for model training or inference.

[0257] If it is a training state, the following steps 617-6111 are performed.

[0258] Steps 617-618 are optional steps, which are executed only when the network state changes, otherwise this step is skipped.

[0259] Step 617: When the network side state changes, the network side reports the existing state through signaling such as Resource reporting information, including one or more of the following: network side memory, network side computation.

[0260] Step 618: The UE determines which model and model split point to select through the current UE state, network side state, and channel state. The UE retransmits the model split point information through signaling such as model split point selection.

[0261] Step 619: The UE confirms the received parameters through signaling such as model split point selection acknowledge.

[0262] The above steps 615-618 are repeated until the UE reaches the maximum training number of rounds or reaches the convergence threshold.

[0263] Step 6110, the UE informs the network to stop model training through signaling such as Training termination notice, or the network informs the UE to stop model training through signaling such as Training termination notice.

[0264] Step 6111, the network informs the UE that the training is completed through signaling such as Training completed, or the UE informs the network that the training is completed through signaling such as Training completed.

[0265] As shown in Figure 6b The model training method includes:

[0266] Step 621, the network (the network refers to the network side device, the same below) sends a model assisted training request to the UE, and the identification of assisted training and / or inference is sent, for example, through model request, and the message carries a model task name, such as requesting a traffic prediction model, and this step is optional.

[0267] Step 622, the UE sends a request for UE evaluation model and segmentation point to the network, and if step 621 exists, the received data is confirmed, otherwise the UE sends a model assisted training request to the network side, and the identification of assisted training and / or inference is sent.

[0268] Step 623, the network confirms the received model task and UE evaluation model and segmentation point request, and the network reports the indicators of the candidate model on the network side according to the model task, for example, through Resource reporting information.

[0269] The indicators include at least one of the following: model accuracy, network side memory, model latency, network side calculation amount, model memory, model calculation amount, data amount, etc.

[0270] Step 624, the UE determines which model and model segmentation point to select through the current UE, network side, and channel state (such as task accuracy, task latency, model accuracy, UE power consumption, UE memory, UE calculation amount, network side memory, network side calculation amount, model memory, model latency, model calculation amount, channel state, data amount, etc.). The UE informs the network side of the selected model and model segmentation position through signaling such as model mode selection.

[0271] Step 625: The network side acknowledges the received parameters via signaling such as model mode selection acknowledge, and transmits the model configuration required by the UE according to the model selected by the UE, including one or more of the following: model structure configuration, model parameters, model hyperparameters, loss function, etc.

[0272] Step 626: The UE transmits intermediate parameters to the network side for model training or inference.

[0273] If the current state is in inference, the UE will transmit the parameters to the network side, and the network side will transmit the inference results to the UE to complete the data prediction and the inference process will be completed.

[0274] Step 627: The network sends the inference results to the UE. For example, it can use RRCReconfiguration to confirm the received information and transmit the model prediction results to the UE.

[0275] Step 628: The UE sends a prediction process completion notification to the network, for example via RRCReconfigurationComplete, to confirm the prediction result and transmit the prediction process completion notification to the network.

[0276] The model training or inference method provided in this application, which enables joint training or inference by the UE and the network, can save UE computing power, protect UE data privacy, and make model segmentation more suitable for network structures. When the state of the channel, UE, or network changes, the model segmentation points used for training or inference by the UE or network will change. By reporting parameters through the UE or network, the UE or network can adaptively select a suitable model and model segmentation point position based on these parameters, making AI more adaptable to changing communication environments and preventing model training interruptions due to insufficient training resources.

[0277] like Figure 7 The image shows a model training or inference apparatus provided in an embodiment of this application. The first model training or inference apparatus 700 includes:

[0278] The first receiving module 701 is used to receive first parameter information sent by the terminal for model training or inference;

[0279] The determining module 702 is used to determine the model information and model segmentation points of the target model based on the first parameter information;

[0280] The first sending module 703 is used to send the model information and the model segmentation points to the terminal;

[0281] The training or inference module 704 is configured to perform model training or inference according to the target model and the model split point.

[0282] Further, the first receiving module 701 comprises:

[0283] The first sending sub-module is configured to send a model assistance request to the terminal, wherein the model assistance request comprises task type and model action information of the model.

[0284] The first receiving sub-module is configured to receive the first parameter information sent by the terminal.

[0285] Alternatively, the first receiving module 701 comprises:

[0286] The second receiving sub-module is configured to receive first parameter information for model training or inference sent by the terminal, and task type and model action information of the model.

[0287] Further, the training or inference module 704 comprises:

[0288] The second receiving sub-module is configured to receive the first intermediate parameter sent by the terminal.

[0289] The obtaining module is configured to, if it is determined according to the model action information that the target model is used for inference, perform inference according to the first intermediate parameter and the target model to obtain an inference result.

[0290] The second sending sub-module is configured to send the inference result to the terminal.

[0291] Further, the training or inference module 704 comprises:

[0292] The third receiving sub-module is configured to receive the first intermediate parameter sent by the terminal.

[0293] The training sub-module is configured to, if it is determined according to the model action information that the target model is used for training, perform training on the target model according to the first intermediate parameter to obtain a second intermediate parameter.

[0294] The third sending sub-module is configured to send the second intermediate parameter to the terminal.

[0295] The fourth receiving sub-module is configured to receive a new intermediate parameter sent by the terminal.

[0296] The determining sub-module is configured to determine the new intermediate parameter as the first intermediate parameter, and perform the step of performing training on the target model according to the first intermediate parameter to obtain a second intermediate parameter until a maximum number of training rounds is reached or a neural network of the target model converges.

[0297] Furthermore, the fourth receiving submodule includes:

[0298] The determining unit is configured to determine a new model segmentation point of the target model based on the first state information if it receives the first state information sent by the terminal.

[0299] A sending unit is used to send the new model segmentation point to the terminal;

[0300] The receiving unit is used to receive new intermediate parameters sent by the terminal based on the new model segmentation point.

[0301] Furthermore, the first parameter information includes at least one of the following:

[0302] Task accuracy;

[0303] Task delay;

[0304] The power consumption of the terminal;

[0305] The terminal's memory;

[0306] The computational load of the terminal.

[0307] Furthermore, the model information includes model configuration information, which includes at least one of the following:

[0308] Model structure configuration;

[0309] Model parameters;

[0310] Model hyperparameters;

[0311] Loss function.

[0312] The first model training or inference device 700 provided in this application embodiment can achieve... Figure 2a The various processes implemented in the method embodiments achieve the same technical effect, and will not be described again here to avoid repetition.

[0313] like Figure 8 The image shows a model training or inference apparatus provided in an embodiment of this application. The second model training or inference apparatus 800 includes:

[0314] The first sending module 801 is used to send first parameter information for model training or inference to the network-side device;

[0315] The first receiving module 802 is used to receive the model information and model segmentation points of the target model sent by the network-side device;

[0316] The training or inference module 803 is used to perform model training or inference based on the target model and the model segmentation points.

[0317] Further, the first sending module 801 is configured to send first parameter information for model training or inference and a model assistance request to the network side device, where the model assistance request comprises a task type of the model and model action information.

[0318] Further, the second model training or inference apparatus 800 further comprises a second receiving module configured to receive a model assistance request sent by the network side device, where the model assistance request comprises a task type of the model and model action information.

[0319] Further, the training or inference module 803 comprises:

[0320] an inference sub-module configured to, if it is determined according to the model action information that the target model is used for inference, perform model inference according to the target model and the model split point to obtain first intermediate parameters;

[0321] a first sending sub-module configured to send the first intermediate parameters to the network side device;

[0322] a first receiving sub-module configured to receive inference results sent by the network side device.

[0323] Further, the training or inference module 803 comprises:

[0324] a first training sub-module configured to, if it is determined according to the model action information that the target model is used for training, perform model training according to the target model and the model split point to obtain first intermediate parameters;

[0325] a second sending sub-module configured to send the first intermediate parameters to the network side device;

[0326] a second receiving sub-module configured to receive second intermediate parameters sent by the network side device;

[0327] a second training sub-module configured to train the target model according to the second intermediate parameters until a maximum training round number is reached or a neural network of the target model converges.

[0328] Further, the second training sub-module comprises:

[0329] a first obtaining unit configured to, if a state of the terminal changes during the process of training the target model according to the second intermediate parameters, obtain first state information;

[0330] a sending unit configured to send the first state information to the network side device;

[0331] A receiving unit is configured to receive a new model split point of the target model sent by the network-side device.

[0332] A second obtaining unit is configured to train the target model based on the new model split point to obtain a new intermediate parameter.

[0333] A determining unit is configured to determine the new intermediate parameter as the first intermediate parameter, and perform the step of sending the first intermediate parameter to the network-side device until a maximum training round number is reached or a neural network of the target model converges.

[0334] Further, the first parameter information includes at least one of the following:

[0335] Task accuracy;

[0336] Task latency;

[0337] Power consumption of the terminal;

[0338] Memory of the terminal;

[0339] Computational load of the terminal.

[0340] Further, the model information includes model configuration information, and the model configuration information includes at least one of the following:

[0341] Model structure configuration;

[0342] Model parameter;

[0343] Model hyperparameter;

[0344] Loss function.

[0345] The second model training or inference apparatus 800 provided by the embodiments of the present application can implement each process of the method embodiment and achieve the same technical effects, and thus details are not repeated here. Figure 2b

[0346] As shown in Figure 9 A model training or inference apparatus is provided by the embodiments of the present application, and the third model training or inference apparatus 900 includes:

[0347] A first receiving module 901 is configured to receive second parameter information for model training or inference sent by a network-side device.

[0348] A determining module 902 is configured to determine a target model and a model split point according to the second parameter information.

[0349] A first sending module 903 is configured to send a model identifier of the target model and the model split point to the network-side device. ​

[0350] a training or inference module 904, configured to perform model training or inference according to the target model and the model split point.

[0351] Further, the third model training or inference apparatus 900 further comprises:

[0352] a transceiving module, configured to receive a model assistance request sent by the network-side device, and send a target request to the network-side device.

[0353] or, send the target request and the model assistance request to the network-side device.

[0354] The target request is a request for the terminal to perform model and split point evaluation, and the model assistance request comprises a task type of the model and model action information.

[0355] Further, the training or inference module 904 comprises:

[0356] an inference sub-module, configured to, if it is determined according to the model action information that the target model is used for inference, perform model inference according to the target model and the model split point, to obtain a first intermediate parameter.

[0357] a first sending sub-module, configured to send the first intermediate parameter to the network-side device.

[0358] a first receiving sub-module, configured to receive an inference result sent by the network-side device.

[0359] Further, the training or inference module 904 comprises:

[0360] a first training sub-module, configured to, if it is determined according to the model action information that the target model is used for training, perform model training according to the target model and the model split point, to obtain a first intermediate parameter.

[0361] a second sending sub-module, configured to send the first intermediate parameter to the network-side device.

[0362] a second receiving sub-module, configured to receive a second intermediate parameter sent by the network-side device.

[0363] a second training sub-module, configured to perform training on the target model according to the second intermediate parameter, until a maximum training round number is reached or a neural network of the target model converges.

[0364] Further, the second training sub-module comprises:

[0365] The determining unit is configured to, in the process of training the target model according to the second intermediate parameter, if the second state information sent by the network side device is received, determine a new model split point of the target model according to the second state information.

[0366] The sending unit is configured to send the new model split point to the network side device.

[0367] The receiving unit is configured to receive a new intermediate parameter sent by the network side device, and execute the step of training the target model according to the second intermediate parameter until a maximum training round number or a neural network of the target model converges, by taking the new intermediate parameter as a second intermediate parameter.

[0368] Further, the second parameter information includes at least one of the following:

[0369] Model precision;

[0370] Model latency;

[0371] Model memory;

[0372] Model calculation amount;

[0373] Memory of the network side device;

[0374] Calculation amount of the network side device;

[0375] Data amount.

[0376] Further, the model configuration information of the target model is configured by the network side device, and the model configuration information includes at least one of the following:

[0377] Model structure configuration;

[0378] Model parameter;

[0379] Model hyperparameter;

[0380] Loss function.

[0381] The third model training or reasoning apparatus 900 provided by the embodiment of the application can realize each process of the method embodiment and achieve the same technical effect, and thus details are not repeated here. Figure 4

[0382] As shown in Figure 10 , a model training or reasoning apparatus is provided in the embodiment of the application, and the fourth model training or reasoning apparatus 1000 includes:

[0383] The sending module 1001 is configured to send second parameter information for model training or reasoning to a terminal. ​

[0384] The receiving module 1002 is configured to receive the model identifier of the target model and the model split point sent by the terminal.

[0385] The training or inference module 1003 is configured to perform model training or inference according to the target model and the model split point.

[0386] Further, the fourth model training or inference apparatus 1000 further comprises:

[0387] The transceiving module is configured to send a model assistance request to the terminal, and receive a target request sent by the terminal.

[0388] Alternatively, the target request and the model assistance request sent by the terminal are received.

[0389] The target request is a request for the terminal to perform model and split point evaluation, and the model assistance request comprises a task type of the model and model action information.

[0390] Further, the training or inference module 1003 comprises:

[0391] The first receiving submodule is configured to receive the first intermediate parameter sent by the terminal.

[0392] The first obtaining submodule is configured to, if it is determined according to the model action information that the target model is used for inference, perform inference according to the first intermediate parameter and the target model to obtain an inference result.

[0393] The first sending submodule is configured to send the inference result to the terminal.

[0394] Further, the training or inference module 1003 comprises:

[0395] The second receiving submodule is configured to receive the first intermediate parameter sent by the terminal.

[0396] The second obtaining submodule is configured to, if it is determined according to the model action information that the target model is used for training, perform training on the target model according to the first intermediate parameter to obtain a second intermediate parameter.

[0397] The sending submodule is configured to send the second intermediate parameter to the terminal.

[0398] The receiving submodule is configured to receive a new intermediate parameter sent by the terminal, determine the new intermediate parameter as the first intermediate parameter, and perform the step of performing training on the target model according to the first intermediate parameter to obtain a second intermediate parameter until a maximum training round number is reached or a neural network of the target model converges.

[0399] Further, the second obtaining sub-module comprises:

[0400] The first obtaining unit is configured to obtain second state information if a state of the network-side device changes.

[0401] The sending unit is configured to send the second state information to the terminal.

[0402] The receiving unit is configured to receive a new model split point of the target model sent by the terminal.

[0403] The second obtaining unit is configured to train the target model based on the new model split point to obtain the second intermediate parameter.

[0404] Further, the second parameter information comprises at least one of the following:

[0405] Model accuracy;

[0406] Model latency;

[0407] Model memory;

[0408] Model calculation amount;

[0409] Memory of the network-side device;

[0410] Calculation amount of the network-side device;

[0411] Data amount.

[0412] Further, model configuration information of the target model is configured by the network-side device, and the model configuration information comprises at least one of the following:

[0413] Model structure configuration;

[0414] Model parameter;

[0415] Model hyperparameter;

[0416] Loss function.

[0417] The fourth model training or inference apparatus 1000 provided by the embodiment of the present application can realize each process of the method embodiment and achieve the same technical effect, and thus details are not repeated here. Figure 5

[0418] ​The apparatus in the embodiments of the present application can be an electronic device, for example, an electronic device with an operating system, or a component in an electronic device, for example, an integrated circuit or a chip. The electronic device can be a terminal or other device than a terminal. The other device can be a server, a network attached storage (NAS), etc., which is not limited in the embodiments of the present application.

[0419] Optionally, as shown in Figure 11 the embodiments of the present application further provide a communication device 1100, which comprises a processor 1101 and a memory 1102, and the memory 1102 stores programs or instructions executable on the processor 1101. For example, when the communication device 1100 is a terminal, the programs or instructions are executed by the processor 1101 to implement the steps of the above-mentioned Figure 2b or Figure 4 method embodiments, and the same technical effects can be achieved. When the communication device 1100 is a network side device, the programs or instructions are executed by the processor 1101 to implement the steps of the above-mentioned Figure 2a 、 Figure 5 method embodiments, and the same technical effects can be achieved. To avoid repetition, details are not described herein.

[0420] Figure 12 A hardware structure diagram of a terminal for implementing the embodiments of the present application is shown.

[0421] The terminal 1200 includes, but is not limited to, at least part of components such as a radio frequency unit 1201, a network module 1202, an audio output unit 1203, an input unit 1204, a sensor 1205, a display unit 1206, a user input unit 1207, an interface unit 1208, a memory 1209, and a processor 1210.

[0422] Those skilled in the art can understand that the terminal 1200 can further include a power supply (such as a battery) for supplying power to each component. The power supply can be logically connected to the processor 1210 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. Figure 12 The terminal structure shown in the above-mentioned embodiments does not constitute a limitation on the terminal, and the terminal can include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements, which are not described herein.

[0423] It should be understood that in the embodiments of the present application, the input unit 1204 can include a graphics processing unit (GPU) 12041 and a microphone 12042. The graphics processor 12041 processes image data of a still picture or a video obtained by an image capture device (such as a camera) in a video capture mode or an image capture mode. The display unit 1206 can include a display panel 12061, which can be configured in the form of a liquid crystal display, an organic light-emitting diode, etc. The user input unit 1207 includes at least one of a touch panel 12071 and other input devices 12072. The touch panel 12071 is also called a touch screen. The touch panel 12071 can include two parts of a touch detection device and a touch controller. The other input devices 12072 can include, but are not limited to, a physical keyboard, function keys (such as volume control keys, on-off keys, etc.), a trackball, a mouse, a joystick, and the like, which will not be described here.

[0424] In the embodiments of the present application, after the radio frequency unit 1201 receives the downlink data from the network side device, it can be transmitted to the processor 1210 for processing. In addition, the radio frequency unit 1201 can send uplink data to the network side device. Generally, the radio frequency unit 1201 includes, but is not limited to, an antenna, an amplifier, a transceiver, a coupler, a low noise amplifier, a duplexer, etc.

[0425] The memory 1209 can be used to store software programs or instructions and various data. The memory 1209 can mainly include a first storage area storing programs or instructions and a second storage area storing data, wherein the first storage area can store an operating system, application programs or instructions required by at least one function (such as a sound playing function, an image playing function, etc.), and the like. In addition, the memory 1209 can include a volatile memory or a non-volatile memory, or the memory 1209 can include both volatile and non-volatile memories. The non-volatile memory can be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a Random Access Memory (RAM), a Static RAM (SRAM), a Dynamic RAM (DRAM), a Synchronous DRAM (SDRAM), a Double Data Rate SDRAM (DDR SDRAM), an Enhanced SDRAM (ESDRAM), a Synch link DRAM (SLDRAM), and a Direct Rambus RAM (DRRAM). The memory 1209 in the embodiments of the present application includes but is not limited to these and any other suitable types of memories.

[0426] The processor 1210 can include one or more processing units; optionally, the processor 1210 integrates an application processor and a modem processor, wherein the application processor mainly processes operations related to an operating system, a user interface, and an application program, and the modem processor mainly processes wireless communication signals, such as a baseband processor. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 1210.

[0427] The radio frequency unit 1201 is configured to send first parameter information for model training or inference to a network side device; receive model information and a model segmentation point of a target model sent by the network side device;

[0428] The processor 1210 is configured to perform model training or inference according to the target model and the model segmentation point.

[0429] Further, the radio frequency unit 1201 is further configured to send, to the network side device, first parameter information used for model training or inference, and a model assistance request, the model assistance request including a task type of the model and model action information.

[0430] Further, the radio frequency unit 1201 is further configured to receive a model assistance request sent by the network side device, the model assistance request including a task type of the model and model action information.

[0431] Further, the processor 1210 is further configured to, if it is determined according to the model action information that the target model is used for inference, perform model inference according to the target model and the model split point, to obtain first intermediate parameters.

[0432] The radio frequency unit 1201 is further configured to send the first intermediate parameters to the network side device, and receive an inference result sent by the network side device.

[0433] Further, the processor 1210 is further configured to, if it is determined according to the model action information that the target model is used for training, perform model training according to the target model and the model split point, to obtain first intermediate parameters.

[0434] The radio frequency unit 1201 is further configured to send the first intermediate parameters to the network side device, and receive second intermediate parameters sent by the network side device.

[0435] The processor 1210 is further configured to train the target model according to the second intermediate parameters until a maximum training round number is reached or a neural network of the target model converges.

[0436] Further, the processor 1210 is further configured to, in the process of training the target model according to the second intermediate parameters, if a state of the terminal changes, obtain first state information.

[0437] The radio frequency unit 1201 is further configured to send the first state information to the network side device, and receive a new model split point of the target model sent by the network side device.

[0438] The processor 1210 is further configured to train the target model based on the new model split point, to obtain new intermediate parameters, determine the new intermediate parameters as the first intermediate parameters, and perform the step of sending the first intermediate parameters to the network side device until a maximum training round number is reached or a neural network of the target model converges.

[0439] Further, the first parameter information includes at least one of the following:

[0440] Task accuracy;

[0441] task latency;

[0442] power consumption of the terminal;

[0443] memory of the terminal;

[0444] computational load of the terminal.

[0445] Further, the model information includes model configuration information, and the model configuration information includes at least one of:

[0446] model structure configuration;

[0447] model parameters;

[0448] model hyperparameters;

[0449] loss function.

[0450] In another embodiment, the radio frequency unit 1201 is configured to receive second parameter information for model training or inference sent by the network side device;

[0451] The processor 1210 is further configured to determine a target model and a model split point according to the second parameter information;

[0452] The radio frequency unit 1201 is configured to send a model identifier of the target model and the model split point to the network side device;

[0453] The processor 1210 is further configured to perform model training or inference according to the target model and the model split point.

[0454] Further, the radio frequency unit 1201 is configured to receive a model assistance request sent by the network side device; and send a target request to the network side device;

[0455] Alternatively, the radio frequency unit 1201 is configured to send a target request and the model assistance request to the network side device;

[0456] The target request is a request for the terminal to perform model and split point evaluation, and the model assistance request includes model task type and model action information.

[0457] Further, the processor 1210 is further configured to, if it is determined according to the model action information that the target model is used for inference, perform model inference according to the target model and the model split point to obtain first intermediate parameters;

[0458] The radio frequency unit 1201 is configured to send the first intermediate parameters to the network side device; and receive an inference result sent by the network side device.

[0459] Further, the processor 1210 is further configured to, if it is determined according to the model action information that the target model is used for training, perform model training according to the target model and the model split point, and obtain a first intermediate parameter.

[0460] The radio frequency unit 1201 is configured to send the first intermediate parameter to the network side device, and receive a second intermediate parameter sent by the network side device.

[0461] The processor 1210 is further configured to train the target model according to the second intermediate parameter until a maximum training round number is reached or a neural network of the target model converges.

[0462] Further, the processor 1210 is further configured to, in the process of training the target model according to the second intermediate parameter, if second state information sent by the network side device is received, determine a new model split point of the target model according to the second state information.

[0463] The radio frequency unit 1201 is configured to send the new model split point to the network side device, and receive a new intermediate parameter sent by the network side device.

[0464] The processor 1210 is further configured to take the new intermediate parameter as a second intermediate parameter, and perform the step of training the target model according to the second intermediate parameter until a maximum training round number is reached or a neural network of the target model converges.

[0465] Further, the second parameter information includes at least one of the following:

[0466] Model precision;

[0467] Model latency;

[0468] Model memory;

[0469] Model calculation amount;

[0470] Memory of the network side device;

[0471] Calculation amount of the network side device;

[0472] Data amount.

[0473] Further, model configuration information of the target model is configured by the network side device, and the model configuration information includes at least one of the following:

[0474] Model structure configuration;

[0475] Model parameter;

[0476] Model hyperparameter;

[0477] Loss function.

[0478] The terminal provided by the embodiments of the present application can realize Figure 2b or Figure 4 The method embodiments realize various processes and achieve the same technical effects. To avoid repetition, details are not described herein.

[0479] The embodiments of the present application also provide a network side device. As shown in Figure 13 , the network side device 1300 includes a processor 1301, a network interface 1302, and a memory 1303. The network interface 1302 is, for example, a common public radio interface (CPRI).

[0480] Specifically, the network side device 1300 of the embodiments of the present application further includes instructions or programs stored in the memory 1303 and executable on the processor 1301, and the processor 1301 invokes the instructions or programs in the memory 1303 to execute the method performed by each module shown in Figure 7 or 10, and achieve the same technical effects. To avoid repetition, details are not described herein.

[0481] The embodiments of the present application also provide a readable storage medium, and the readable storage medium stores programs or instructions. When the programs or instructions are executed by a processor, each process of the method embodiments shown in Figure 2a , Figure 2b , Figure 4 or Figure 5 is realized, and the same technical effects can be achieved. To avoid repetition, details are not described herein.

[0482] The processor is the processor in the terminal described in the above embodiments. The readable storage medium includes a computer readable storage medium, such as a computer readable memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0483] The embodiments of the present application also provide a chip. The chip includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is configured to execute programs or instructions to realize each process of the method embodiments shown in Figure 2a , Figure 2b , Figure 4 or Figure 5 , and achieve the same technical effects. To avoid repetition, details are not described herein.

[0484] It should be understood that the chip mentioned in the embodiments of the present application can also be referred to as a system chip, a system chip, a chip system, or a system on chip, etc.

[0485] The embodiment of the present application further provides a computer program / program product stored in a storage medium, which is executed by at least one processor to implement each process of the method embodiments shown in the above Figure 2a , Figure 2b , Figure 4 or Figure 5 The same technical effects can be achieved, and thus details are not described herein again to avoid repetition.

[0486] It should be noted that in this document, the term "comprising" or "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that processes, methods, articles or devices including a series of elements not only include those elements, but also include other elements not explicitly listed, or further include elements inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, but can also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method can be performed in an order different from that described, and various steps can also be added, omitted or combined. In addition, the features described with reference to certain examples can be combined in other examples.

[0487] From the above description of the embodiments, those skilled in the art can clearly understand that the above method embodiments can be realized by means of software and necessary general hardware platforms, of course, they can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a plurality of instructions for making a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) execute the methods described in various embodiments of the present application.

[0488] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above specific embodiments, which are only illustrative and not limiting. Those skilled in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the protection scope of the claims.

Claims

1. A model training or inference method for a first device, characterized in that, The method includes: Receive target parameter information sent by the second device for model training or inference; The target model and model segmentation points are determined based on the target parameter information; The model information of the target model and the model segmentation points are sent to the second device; Model training or inference is performed based on the target model and the model segmentation points; The first device is a network-side device, and the second device is a terminal; The receiving of target parameter information for model training or inference sent by the second device includes: Send a model assistance request to the second device, the model assistance request including the model's task type and model action information; Receive the target parameter information sent by the second device; The target parameter information includes at least one of the following: Task accuracy; The power consumption of the terminal; The terminal's memory; The computational load of the terminal; Alternatively, receiving the target parameter information for model training or inference sent by the second device includes: Receive target parameter information for model training or inference, as well as model task type and model action information, sent by the second device; or, The first device is a terminal, and the second device is a network-side device; Before receiving the target parameter information for model training or inference sent by the second device, the method further includes: Receive a model assistance request sent by the second device; send a target request to the second device; Alternatively, send a target request and a model assistance request to the second device; The target request is a request for the first device to evaluate the model and segmentation points, and the model assistance request includes the model's task type and model action information. The target parameter information includes at least one of the following: Model accuracy; Model memory; Model computational complexity; The memory of the network-side device; The computational load of the network-side devices; Data volume.

2. The method according to claim 1, characterized in that, The step of training or inferring the model based on the target model and the model segmentation points includes: Receive the first intermediate parameter sent by the second device; If the target model is determined to be used for inference based on the model action information, then inference is performed based on the first intermediate parameter and the target model to obtain the inference result; The reasoning result is sent to the second device.

3. The method according to claim 1, characterized in that, The step of training or inferring the model based on the target model and the model segmentation points includes: Receive the first intermediate parameter sent by the second device; If the target model is determined to be used for training based on the model action information, then the target model is trained based on the first intermediate parameter to obtain the second intermediate parameter; Send the second intermediate parameter to the second device; Receive new intermediate parameters sent by the second device; The new intermediate parameter is determined as the first intermediate parameter, and the target model is trained based on the first intermediate parameter to obtain the second intermediate parameter, until the maximum number of training rounds is reached or the neural network of the target model converges.

4. The method according to claim 3, characterized in that, The receipt of new intermediate parameters sent by the second device includes: If the first status information sent by the second device is received, a new model segmentation point of the target model is determined based on the first status information; Send the new model segmentation points to the second device; Receive new intermediate parameters sent by the second device based on the new model segmentation point.

5. The method according to claim 1, characterized in that, The step of training or inferring the model based on the target model and the model segmentation points includes: If the target model is determined to be used for inference based on the model action information, then model inference is performed based on the target model and the model segmentation point to obtain the first intermediate parameter; Send the first intermediate parameter to the second device; Receive the inference results sent by the second device.

6. The method according to claim 1, characterized in that, The step of training or inferring the model based on the target model and the model segmentation points includes: If the target model is determined to be used for training based on the model action information, then the model is trained based on the target model and the model segmentation points to obtain the first intermediate parameters; Send the first intermediate parameter to the second device; Receive the second intermediate parameter sent by the second device; The target model is trained according to the second intermediate parameter until the maximum number of training rounds is reached or the neural network of the target model converges.

7. The method according to claim 6, characterized in that, The step of training the target model based on the second intermediate parameter until the maximum number of training rounds is reached or the neural network of the target model converges includes: During the training of the target model based on the second intermediate parameter, if the second state information sent by the second device is received, a new model segmentation point of the target model is determined based on the second state information. Send the new model segmentation points to the second device; The system receives new intermediate parameters sent by the second device, uses these new intermediate parameters as second intermediate parameters, and then trains the target model based on the second intermediate parameters until the maximum number of training rounds is reached or the neural network of the target model converges.

8. The method according to claim 1, characterized in that, The model information includes model configuration information, which includes at least one of the following: Model structure configuration; Model parameters; Model hyperparameters; Loss function.

9. A model training or inference method for a second device, characterized in that, The method includes: Send target parameter information for model training or inference to the first device; Receive the model information and model segmentation points of the target model sent by the first device; Model training or inference is performed based on the target model and the model segmentation points; The first device is a network-side device, and the second device is a terminal; Sending target parameter information for model training or inference to the first device includes: Send target parameter information for model training or inference, and a model assistance request to the first device, the model assistance request including the model's task type and model action information; The target parameter information includes at least one of the following: Task accuracy; The power consumption of the terminal; The terminal's memory; The computational load of the terminal; or, The first device is a terminal, and the second device is a network-side device; Before sending the target parameter information for model training or inference to the first device, the method further includes: Send a model assistance request to the first device; and receive a target request sent by the first device; Alternatively, receive the target request and model assistance request sent by the first device; The target request is a request for the first device to evaluate the model and segmentation points, and the model assistance request includes the model's task type and model action information; The target parameter information includes at least one of the following: Model accuracy; Model memory; Model computational complexity; The memory of the network-side device; The computational load of the network-side devices; Data volume.

10. The method according to claim 9, characterized in that, The step of training or inferring the model based on the target model and the model segmentation points includes: If the target model is determined to be used for inference based on the model action information, then model inference is performed based on the target model and the model segmentation point to obtain the first intermediate parameter; Send the first intermediate parameter to the first device; Receive the inference results sent by the first device.

11. The method according to claim 9, characterized in that, The step of training or inferring the model based on the target model and the model segmentation points includes: If the target model is determined to be used for training based on the model action information, then the model is trained based on the target model and the model segmentation points to obtain the first intermediate parameters; Send the first intermediate parameter to the first device; Receive the second intermediate parameter sent by the first device; The target model is trained according to the second intermediate parameter until the maximum number of training rounds is reached or the neural network of the target model converges.

12. The method according to claim 11, characterized in that, The step of training the target model based on the second intermediate parameter until the maximum number of training rounds is reached or the neural network of the target model converges includes: During the training of the target model based on the second intermediate parameter, if the state of the second device changes, the first state information is obtained. Send the first status information to the first device; Receive new model segmentation points of the target model sent by the first device; The target model is trained based on the new model segmentation points to obtain new intermediate parameters; The new intermediate parameter is determined as the first intermediate parameter, and the step of sending the first intermediate parameter to the first device is performed until the maximum number of training rounds is reached or the neural network of the target model converges.

13. The method according to claim 9, characterized in that, The step of training or inferring the model based on the target model and the model segmentation points includes: Receive the first intermediate parameter sent by the first device; If the target model is determined to be used for inference based on the model action information, then inference is performed based on the first intermediate parameter and the target model to obtain the inference result; The reasoning result is sent to the first device.

14. The method according to claim 9, characterized in that, The step of training or inferring the model based on the target model and the model segmentation points includes: Receive the first intermediate parameter sent by the first device; If the target model is determined to be used for training based on the model action information, then the target model is trained based on the first intermediate parameter to obtain the second intermediate parameter; Send the second intermediate parameter to the first device; The system receives new intermediate parameters sent by the first device, determines the new intermediate parameters as the first intermediate parameters, and performs the step of training the target model based on the first intermediate parameters to obtain the second intermediate parameters, until the maximum number of training rounds is reached or the neural network of the target model converges.

15. The method according to claim 14, characterized in that, The step of training the target model based on the first intermediate parameters to obtain the second intermediate parameters includes: If the state of the second device changes, then the second state information is obtained; Send the second status information to the first device; Receive new model segmentation points of the target model sent by the first device; The target model is trained based on the new model segmentation points to obtain the second intermediate parameters.

16. A communication device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the model training or inference method as described in any one of claims 1 to 15.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the model training or inference method as described in any one of claims 1 to 15.

Images