Communication method and apparatus

Abstract

A communication method and apparatus. The method comprises: a first communication apparatus sending a first model feature to a second communication apparatus, wherein the first model feature is used for representing model features of a first model, such that the second communication apparatus can execute a corresponding operation on the basis of the first model feature. For example, the second communication apparatus may aggregate, on the basis of the first model feature, a plurality of models which comprise the first model. There is no restriction on whether the plurality of aggregated models have identical model structures, thereby improving the flexibility of model aggregation.

Description

A communication method and apparatus

[0001] Cross-reference to Related Applications

[0002] This application claims priority to the Chinese Patent Application No. 202410869558.7, filed on June 30, 2024, and entitled “A communication method and apparatus”, the content of which is incorporated herein by reference in its entirety. TECHNICAL FIELD

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

[0004] In a wireless communication network, for example, in a mobile communication network, the services supported by the network are increasingly diverse, and thus the needs to be met are increasingly diverse. For example, the network needs to be able to support ultra-high rates, ultra-low latencies, and / or ultra-large connections. This feature makes network planning, network configuration, and / or resource scheduling increasingly complex. These new needs, new scenarios, and new features bring unprecedented challenges to network planning, operation and maintenance, and efficient operation. In order to meet this challenge, artificial intelligence (AI) technology can be introduced into the wireless communication network, thereby realizing network intelligentization. In the wireless communication network, how to describe the features of a model in multiple dimensions and indicate to the communication node at the opposite end is a research direction. SUMMARY

[0005] Embodiments of the present application provide a communication method and apparatus to describe the features of a model in at least one dimension and indicate to the communication node at the opposite end.

[0006] In a first aspect, a communication method is provided. An execution subject of the method is a first communication apparatus, which can be a first node or a module, unit, or component in the first node (for example, a chip, a chip system, a circuit, a processor, or the like). The method includes determining a first model feature and sending a first indication, where the first indication is used to indicate the first model feature. The first model feature is used to represent a model feature of a first model, and the first model feature includes at least one of an application scenario of the first model, a structure of the first model, a function of the first model, or a format of the first model.

[0007] By the above design, the definition of the model feature is introduced, the model feature includes information of at least one dimension of an application scenario of the model, a structure of the model, a function of the model, a format of the model, or a target performance of the model, characteristics of the model are described from different dimensions, flexibility of model management is enhanced, and the determined model feature can be indicated to the communication node at the other end.

[0008] In a possible implementation, the first model feature further includes a target performance of the first model.

[0009] In a possible implementation, the method further includes: sending the first model.

[0010] By the above design, the first communication device sends the first model and the first model feature to the second communication device, the first model feature is used to represent model features of the first model. In a distributed training scenario, the second communication device can be a center node, and the first communication device can be an edge node. The center node can aggregate models according to model features of the models trained by the edge node. For example, models with consistent model features are aggregated together to form a new model, without limiting whether the model structures of the aggregated models are consistent. Compared with being able to only aggregate models with consistent model structures, the flexibility and diversity of model aggregation can be improved, and the limitation on the models trained by the edge node is reduced. Further, since the model structures of the aggregated models can be different, different model aggregation results can be obtained according to different strategies.

[0011] In a possible implementation, the method further includes: receiving a second indication, the second indication being used to indicate a second model feature; and determining the first model according to the second model feature, wherein the second model feature is an expected model feature of the first model for the second communication device, and the second model feature includes at least one of an expected application scenario of the first model, an expected structure of the first model, an expected function of the first model, or an expected format of the first model.

[0012] By the above design, in a distributed training scenario, the first communication device can be a center node, and the second communication device can be an edge node. The center node can indicate a model feature of an expected first model, i.e., a second model feature, to the edge node. The edge node trains a model according to the second model feature to obtain the first model. Further, the edge node can send the first model to the center node, so that the first model trained by the edge node meets the expectation or requirement of the center node.

[0013] In a possible implementation, the second model feature is different from the first model feature.

[0014] In a possible implementation, the method further includes: receiving the first model, the first model being determined according to the first model feature.

[0015] Through the above design, the first communication device indicates the first model feature to the second communication device, the second communication device trains to obtain the corresponding first model according to the first model feature, and sends the first model to the first communication device, so that the first communication device can obtain the first model meeting the model feature requirement.

[0016] In a possible implementation, the method further includes: determining first inference information according to the first model; and sending the first inference information.

[0017] Through the above design, the scheme of the embodiment of the application can be applied to the scene of distributed inference, the first communication device can be a first inference node, and the second communication device can be a second inference node. In addition to sending the corresponding inference information to the second inference node, the first inference node can also send the model feature of the model (or sub-model) corresponding to the inference information to the second inference node, so that the second inference node selects a model matching the model feature of the model in the first inference node for model inference, ensures that the model features of the models for model inference between different inference nodes are consistent, and improves the accuracy and flexibility of distributed inference.

[0018] In a possible implementation, the first model feature is specifically used to represent the model feature of the updated first model, and the determination of the first model feature includes: updating the first model; and determining the first model feature according to the updated first model.

[0019] Through the above design, the scheme of the embodiment of the application can also be applied to the scene of model updating. The first communication device can be a training node, and the second communication device can be an inference node. When the training node updates the model feature of the first model, the training node can indicate the updated model feature to the inference node, and the inference node performs model inference according to the updated model feature, thereby improving the reliability of inference.

[0020] In a possible implementation, the type of the first model feature includes: a fixed model feature of the first model, and / or a variable model feature of the first model, and the fixed model feature of the first model and / or the variable model feature of the first model are used to determine a second model based on the first model.

[0021] Through the above design, the scheme of the embodiment of the application can also be applied to the scenario of sequential training of a bilateral model. The first communication device can be a first training node, the first training node can be a side node of a first trained model, and the second communication device can be a second training node, which can be a side node of a second trained model. The side node of the first trained model can provide effective guidance information such as fixed model and / or variable model features in a reference model to the side node of the second trained model, so that the side node of the second trained model can continue model training based on the reference model according to the effective guidance information, improve the efficiency of model training, and fully utilize the capability of the training node.

[0022] In a second aspect, a communication method is provided, an execution subject of the method is a second communication device, the second communication device can be a second node or a module, unit, or component (for example, a chip, a chip system, a circuit, a processor, or the like) in the second node, and the method includes: receiving a first indication, the first indication is used to indicate a first model feature; wherein the first model feature is used to represent a model feature of a first model, and the first model feature includes at least one of the following: an application scenario of the first model, a structure of the first model, a function of the first model, or a format of the first model.

[0023] In a possible implementation, the first model feature further includes a target performance of the first model.

[0024] In a possible implementation, the method further includes: receiving the first model; and aggregating the first model according to the first model feature.

[0025] In a possible implementation, the method further includes: sending a second indication, the second indication is used to indicate a second model feature, the second model feature is used to determine the first model, the second model feature is an expected model feature of the first model for the second communication device, and the second model feature includes at least one of the following: an expected application scenario of the first model, an expected structure of the first model, an expected function of the first model, or an expected format of the first model.

[0026] In a possible implementation, the second model feature is different from the first model feature.

[0027] In a possible implementation, the method further includes: determining the first model according to the first model feature; and sending the first model.

[0028] In a possible implementation, the method further includes: determining a second model according to the first model feature; receiving first inference information, the first inference information being determined according to the first model; and determining second inference information according to the second model and the first inference information.

[0029] In a possible implementation, the first model feature is specifically used to represent a model feature of the first model after being updated, and the method further includes: updating a model feature of a first model locally saved by the second communication device according to the first model feature.

[0030] In a possible implementation, the type of the first model feature includes: a fixed model feature of the first model, and / or a variable model feature of the first model, and the method further includes: receiving the first model; and determining a second model according to the first model, and the fixed model feature of the first model and / or the variable model feature of the first model.

[0031] In a third aspect, an apparatus is provided, which can implement the method in the first aspect. For example, the apparatus includes means for performing the corresponding method in the first aspect. The apparatus can be implemented by hardware, software, or by executing corresponding software by hardware.

[0032] In a possible design, the apparatus includes units for performing the method in the first aspect.

[0033] In a possible design, the apparatus includes a processor configured to perform the method in the first aspect. Optionally, the apparatus further includes a memory coupled to the processor. The processor is specifically configured to execute a computer program or instructions stored in the memory, so that the apparatus implements the method in the first aspect.

[0034] In a possible design, the apparatus includes a processor and an interface circuit. The interface circuit is configured to receive a signal from another apparatus outside the apparatus and transmit the signal to the processor, or send a signal from the processor to another apparatus outside the apparatus. The processor is configured to implement the method in the first aspect by means of a logic circuit or by executing code instructions.

[0035] Optionally, the apparatus can be a first apparatus, or a module, unit, or component (for example, a chip, a chip system, a circuit, or a processor) in the first apparatus that is one-to-one corresponding to the method / operation / step / action described in the first aspect, or is capable of being matched with the first apparatus.

[0036] In a fourth aspect, an apparatus is provided, which is capable of implementing the method of the second aspect. For example, the apparatus comprises means for performing the method of the second aspect. The apparatus can be implemented by hardware, software or by executing corresponding software by hardware.

[0037] In a possible design of the apparatus, the apparatus comprises units for performing the method of the second aspect.

[0038] In a possible design of the apparatus, the apparatus comprises a processor configured to perform the method of the second aspect. Optionally, the apparatus further comprises a memory coupled to the processor. The processor is specifically configured to execute computer programs or instructions stored in the memory, so that the apparatus implements the method of the second aspect.

[0039] In a possible design of the apparatus, the apparatus comprises a processor and an interface circuit. The interface circuit is configured to receive signals from other apparatuses outside the apparatus and transmit the signals to the processor, or transmit signals from the processor to other apparatuses outside the apparatus. The processor is configured to implement the method of the second aspect by logic circuit or executing code instructions.

[0040] Optionally, the apparatus can be a second apparatus, or a module, unit or component (for example, a chip, a chip system, a circuit or a processor) in the second apparatus, which is specifically configured to perform the method / operation / step / action described in the second aspect, or is capable of matching the second apparatus.

[0041] In a fifth aspect, a computer readable storage medium is provided, which stores computer programs or instructions. When the computer programs or instructions are run on a computer, the computer is caused to implement the method of the first aspect or the second aspect.

[0042] In a sixth aspect, a computer program product is provided, which comprises computer programs or instructions. When the computer programs or instructions are run on a computer, the method of the first aspect or the second aspect is executed.

[0043] In a seventh aspect, a chip is provided, which comprises a processor configured to implement the method of the first aspect or the second aspect. Optionally, the processor is coupled to a memory, and the processor is configured to execute computer programs or instructions stored in the memory, so that the chip implements the method of the first aspect or the second aspect. Further, optionally, the chip comprises the memory.

[0044] In an eighth aspect, a communication system is provided, which comprises a first communication apparatus and a second communication apparatus. The first communication apparatus is configured to implement the method of the first aspect, and the second communication apparatus is configured to implement the method of the second aspect.

[0045] The beneficial effects of the second aspect to the eighth aspect can be referred to the description of the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0046] FIG. 1 is a schematic diagram of a communication system according to an embodiment of the present application;

[0047] FIG. 2 is a schematic diagram of a neuron according to an embodiment of the present application;

[0048] FIG. 3 is a schematic diagram of a neural network according to an embodiment of the present application;

[0049] FIG. 4 is a schematic diagram of an application architecture of AI according to an embodiment of the present application;

[0050] FIG. 5 is a schematic diagram of distributed training according to an embodiment of the present application;

[0051] FIGS. 6a to 6c are schematic diagrams of application scenarios according to embodiments of the present application;

[0052] FIG. 7 is a flowchart of a communication method according to an embodiment of the present application;

[0053] FIGS. 8 and 9 are schematic diagrams of model feature aggregation according to embodiments of the present application;

[0054] FIGS. 10a to 10c are schematic diagrams of structures of conversion models corresponding to different scenarios according to embodiments of the present application;

[0055] FIGS. 11 and 12 are schematic diagrams of model updating according to embodiments of the present application;

[0056] FIGS. 13 to 15 are schematic diagrams of bilateral model training according to embodiments of the present application;

[0057] FIGS. 16 and 17 are schematic diagrams of structures of devices according to embodiments of the present application. DETAILED DESCRIPTION

[0058] In order to make the purposes, technical solutions and advantages of the present application clearer, the embodiments of the present application are described in detail below with reference to the drawings. The specific operation methods, function descriptions and the like in the method embodiments can also be applied to the device embodiments or system embodiments.

[0059] It can be understood that, in the embodiments of the present application, the number of nouns, unless otherwise specified, represents "a singular noun or a plural noun", that is, "one or more". "At least one" refers to one or more, and "multiple" refers to two or more. "And / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. In the textual description of the present application, the character " / ", generally represents that the associated objects before and after are in an "or" relationship; in the formula of the present application, the character " / ", represents that the associated objects before and after are in a "division" relationship. "Including at least one of A, B and C (or) or the like can represent: including A; including B; including C; including A and B; including A and C; including B and C; including A, B and C, where A, B and C can be singular or plural.

[0060] The various numerical numbers involved in the embodiments of the present application are distinguished for the convenience of description, and are not used to limit the scope of the embodiments of the present application. The size of the serial number of the above processes does not mean the order of execution, and the execution order of the processes should be determined according to its function and inherent logic. The ordinal numbers "first", "second" and the like involved in the embodiments of the present application are used to distinguish a plurality of objects, and do not limit the size, order, time sequence, priority or importance of the plurality of objects.

[0061] Figure 1 shows a possible, non-limiting system diagram. As shown in Figure 1, the communication system 1000 includes a radio access network (RAN) 100 and a core network (CN) 200. Optionally, the communication system 1000 also includes the Internet 300.

[0062] Among them, the RAN 100 includes at least one RAN node (such as 110a and 110b in Figure 1, collectively referred to as 110) and at least one terminal (such as 120a-120j in Figure 1, collectively referred to as 120). The RAN 100 can also include other RAN nodes, such as wireless relay devices and / or wireless backhaul devices (not shown in Figure 1) and the like.

[0063] The terminal 120 is connected to the RAN node 110 in a wireless manner. The RAN node 110 is connected to the core network 200 in a wireless or wired manner. The core network device in the core network 200 and the RAN node 110 in the RAN 100 can be different physical devices respectively, or can be the same physical device integrated with the logical functions of the core network and the logical functions of the radio access network.

[0064] RAN100 can be used for cellular systems related to the 3rd generation partnership project (3GPP), such as 4th generation (4G). th generation, 4G), fifth generation (5 th RAN100 can be a generation (5G) mobile communication system, or a future-oriented evolution system (such as a future communication network). RAN100 can also be an open access network (O-RAN or ORAN), a cloud radio access network (CRAN), or a wireless fidelity (WiFi) system. RAN100 can also be a communication system that integrates two or more of the above systems.

[0065] RAN node 110, forming part of the communication system, is used to help terminals achieve wireless access. Multiple RAN nodes 110 in the communication system 1000 can be of the same type or different types. In some scenarios, the roles of RAN node 110 and terminal 120 are relative. For example, network element 120i in Figure 1 can be a helicopter or drone, which can be configured as a mobile base station. For terminals 120j accessing RAN 100 through network element 120i, network element 120i is a base station; but for base station 110a, network element 120i is a terminal.

[0066] In a possible scenario, the RAN node can be a base station, an evolved Node B (eNodeB), an access point (AP), a transmission reception point (TRP), a next generation NodeB (gNB), a next generation base station in future communication networks, a base station, or an access node in a WiFi system, etc. The RAN node can be a macro base station (such as 110a in FIG. 1), a micro base station or an indoor station (such as 110b in FIG. 1), a relay node or a donor node, or a wireless controller in a CRAN scenario. Optionally, the RAN node can also be a server, a wearable device, a vehicle or a vehicle-mounted device, etc. For example, an access network device in vehicle to everything (V2X) technology can be a road side unit (RSU). All or part of the functions of the RAN node in the embodiments of the present application can be implemented by software functions running on hardware, or by virtualized functions instantiated on a platform (such as a cloud platform). The RAN node in the embodiments of the present application can also be a logical node, a logical module or software that can implement all or part of the functions of the RAN node.

[0067] In another possible scenario, multiple RAN nodes cooperate to assist a terminal to implement wireless access, and different RAN nodes respectively implement part of the functions of a base station. For example, the RAN node can be a central unit (CU), a distributed unit (DU), a CU-control plane (CP), a CU-user plane (UP), or a radio unit (RU), etc. The CU and the DU can be separately arranged, or can be included in the same network element, such as a baseband unit (BBU). The RU can be included in a radio frequency device or a radio frequency unit, such as a remote radio unit (RRU), an active antenna processing unit (AAU), or a remote radio head (RRH).

[0068] It can be understood that the CU (or CU-CP and CU-UP), DU or RU can also have different names in different systems, but those skilled in the art can understand their meanings. For example, in the ORAN system, the CU can also be referred to as an open CU (O-CU), the DU can also be referred to as an open DU (O-DU), the CU-CP can also be referred to as an open CU-CP (O-CU-CP), the CU-UP can also be referred to as an open CU-UP (O-CU-UP), and the RU can also be referred to as an open RU (O-RU). For the convenience of description, the CU, CU-CP, CU-UP, DU and RU are taken as examples for description in the present application. Any one of the CU (or CU-CP, CU-UP), DU and RU in the embodiments of the present application can be implemented by a software module, a hardware module, or a combination of a software module and a hardware module.

[0069] The terminal 120 is a device with wireless transceiving function. The terminal 120 can also be referred to as a terminal device, a user equipment (UE), a mobile station, a mobile terminal, etc. The terminal can be widely applied to various scenarios, such as device-to-device (D2D) communication, vehicle to everything (V2X) communication, machine-type communication (MTC), internet of things (IOT), virtual reality, augmented reality, industrial control, autonomous driving, remote medical treatment, smart power grid, smart furniture, smart office, smart wear, smart transportation, smart city, etc. The terminal can be a mobile phone, a tablet computer, a computer with wireless transceiving function, a wearable device, a vehicle, a drone, a helicopter, an airplane, a ship, a robot, a mechanical arm, a smart home device, etc. The embodiments of the present application do not limit the device form of the terminal.

[0070] The RAN nodes 110 and the terminals 120 can be fixed in position or movable. The RAN nodes 110 and the terminals 120 can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; can also be deployed on the water surface; can also be deployed on aircraft, balloons and artificial satellites in the air. The embodiments of the present application do not limit the application scenarios of the RAN nodes 110 and the terminals 120. The RAN nodes 110 and the terminals 120 can be deployed in the same scenario or different scenarios, for example, the RAN nodes 110 and the terminals 120 are deployed on land at the same time; or the RAN nodes 110 are deployed on land and the terminals 120 are deployed on the water surface, etc., which will not be listed one by one.

[0071] The RAN nodes 110 and the terminals 120 can communicate over licensed spectrum, over unlicensed spectrum, or over both licensed and unlicensed spectrum; for example, the RAN nodes 110 and the terminals 120 can communicate over sub-6 gigahertz (GHz) spectrum, over 6 GHz spectrum, or over both sub-6 GHz spectrum and 6 GHz spectrum. Embodiments of the present application do not limit the spectrum resources used for wireless communication.

[0072] The RAN nodes 110 and the terminals 120 can be referred to as communication apparatuses; for example, the network elements 110a and 110b in FIG. 1 can be understood as communication apparatuses having base station functionality, and the network elements 120a-120j can be understood as communication apparatuses having terminal functionality.

[0073] It can be understood that the scheme of embodiments of the present application can be applied to the communication system 1000 shown in FIG. 1, which can correspond to a terrestrial network (TN). Alternatively, the scheme of embodiments of the present application can also be applied to a non terrestrial network (NTN). In the communication system corresponding to the NTN, the RAN nodes 110 in FIG. 1 can be replaced by satellites and ground stations. The satellites are deployed in space, and the ground stations are deployed on the ground, which can be understood as base stations deployed on the ground, and the ground stations can also be referred to as gateways (GWs). The link between the satellite and the terminal is referred to as a user link, and the link between the satellite and the ground station is referred to as a feeder link. The satellites can communicate through inter-satellite links. The working modes of the satellites include transparent and regenerative.

[0074] When the satellite works in the transparent mode, the satellite has the function of signal forwarding, and the ground station has all or part of the functions of the base station, and the ground station can be regarded as a base station. It can be understood that the ground station can be one device (for example, a macro base station, or a micro base station, etc.), or the ground station can implement the corresponding functions by multiple RAN nodes (for example, CUs and DUs, etc.), for details, refer to the foregoing description. Alternatively,

[0075] When the satellite works in the regenerative mode, the satellite has the ability to process digital signals, and the satellite has all or part of the functions of the base station, and the satellite can be regarded as a base station. Further, for the regenerative mode, it can be subdivided into: all functions of the base station are deployed on the satellite, which is referred to as base station full function (for example, CU and DU) on satellite, or part of the functions of the base station are deployed on the satellite, which is referred to as base station partial function (for example, DU) on satellite, and the remaining functions of the base station (for example, CU) are implemented on the ground station.

[0076] Satellite and ground station can be referred to as communication devices, for example, satellite can be understood as a communication device with satellite function, and ground station can be understood as a communication device with ground station function.

[0077] It can be understood that in the communication system corresponding to TN, the RAN node is used to help the terminal to realize wireless access, and it can also have other different descriptions, such as RAN entity, access node, access network device, etc.; in the communication system corresponding to NTN, the satellite and the ground station help the terminal to realize wireless access. In the subsequent description of the present application, if there is no special description, the node or device that helps the terminal to realize wireless access is described as "access network device".

[0078] In the embodiments of the present application, the artificial intelligence (AI) model is a specific method for implementing AI function, and the AI model represents the mapping relationship between the input and the output of the model. The AI model can be a neural network or other machine learning model. Among them, the AI model can be referred to as a model. AI-related operations can include at least one of data collection, model training, model information publishing, model inference (model reasoning), or reasoning result publishing, etc.

[0079] Taking a neural network as an example, the neural network is a specific implementation form of machine learning technology. According to the universal approximation theorem, the neural network can theoretically approximate any continuous function, so that the neural network has the ability to learn any mapping. The traditional communication system needs to rely on rich expert knowledge to design the communication module, while the deep learning communication system based on neural network can automatically discover the implicit pattern structure from a large amount of data set, establish the mapping relationship between the data, and obtain better performance than the traditional modeling method.

[0080] The idea of neural network comes from the neuron structure of brain organization. Each neuron performs weighted summation operation on its input value, and generates output through an activation function. As shown in FIG. 2, it is a schematic diagram of neuron structure. Assuming that the input of the neuron is x = [x0, x1, …, x n ], the weight corresponding to each input is w = [w0, w1, …, w n ], and the bias of weighted summation is b. The form of the activation function can be diversified. Assuming that the activation function of a neuron is y = f(z) = max(0, z), the output of the neuron is: For example, the activation function of a neuron is y = f(z) = z, and the output of the neuron is: b can be various possible values such as a decimal, an integer (including 0, a positive integer, or a negative integer), or a complex number, etc. The activation functions of different neurons in a neural network can be the same or different.

[0081] A neural network generally includes a multi-layer structure, and each layer can include one or more neurons. Increasing the depth and / or width of a neural network can improve the expressiveness of the neural network, providing a more powerful information extraction and abstract modeling capability for complex systems. The depth of a neural network can refer to the number of layers included in the neural network, and the number of neurons included in each layer can be referred to as the width of the layer. As shown in FIG. 3, an illustration of the layer relationship of a neural network is shown. In one implementation, a neural network includes an input layer and an output layer. The input layer of the neural network processes the received input through neurons and passes the result to the output layer, and the output layer obtains the output result of the neural network. In another implementation, a neural network includes an input layer, a hidden layer, and an output layer. The input layer of the neural network processes the received input through neurons and passes the result to the intermediate hidden layer, and the hidden layer passes the calculation result to the output layer or the adjacent hidden layer, and finally the output layer obtains the output result of the neural network. A neural network can include one or more layers of sequentially connected hidden layers, without limitation. During the training process of a neural network, a loss function can be defined. The loss function describes the gap or difference between the output value of the neural network and the ideal target value, and the specific form of the loss function is not limited by the embodiments of the present application. The training process of a neural network is a process of adjusting the parameters of the neural network, such as the number of layers, the width of the neural network, the weights of the neurons, and / or the parameters in the activation function of the neurons, so that the value of the loss function is less than a threshold value or meets the target requirements.

[0082] As shown in FIG. 4, an illustration of an application framework of AI is shown. A data source is used to store training data and inference data. A model training node obtains an AI model by analyzing or training the training data provided by the data source, and deploys the AI model in a model inference node. Optionally, the model training node can also update the AI model that has been deployed in the model inference node. The model inference node can also feed back relevant information of the deployed model to the model training node, so that the model training node optimizes or updates the deployed AI model, etc.

[0083] The AI model represents a mapping relationship between the input and the output of the model. The AI model is learned by the model training node, which is equivalent to learning the mapping relationship between the input and the output of the model by the model training node using the training data. The model inference node uses the AI model to perform inference based on inference data provided by the data source to obtain an inference result. The method can also be described as follows: The model inference node inputs the inference data into the AI model, and obtains the output of the AI model, which is the inference result. The inference result can indicate the configuration parameter used (executed) by the execution object and / or the operation executed by the execution object. The inference result can be uniformly planned by an actor entity and sent to one or more execution objects (for example, network entities) for execution.

[0084] The model training method involved in the embodiments of the present application is described below. It can be understood that the description is only for understanding the scheme of the embodiments of the present application, and does not limit the embodiments of the present application. The training of the AI model can be divided into two categories: one is centralized training; and the other is distributed training.

[0085] 1. Centralized training

[0086] The centralized training, also known as non-distributed training, is a training method in which a node centrally trains an AI model, and sends the trained AI model to other nodes for AI inference according to the AI model. This training method usually has high requirements for the node. When the parameter quantity of the AI model is large and the structure of the AI model is complex, only a cloud server or an access network device with strong capability can implement AI training, and a terminal or other device with weak capability usually cannot implement AI model training alone. Alternatively, the cloud server can be located in the Internet 300 in the communication system 1000 shown in FIG. 1.

[0087] 2. Distributed training

[0088] In the distributed training, a center node (master node) and multiple edge nodes (local nodes) are provided. Each edge node can train an AI model using locally collected data, and send the trained AI model to the center node. The center node aggregates the AI models sent by the edge nodes. This training method can fully utilize the capabilities of the edge nodes and protect the privacy of the local data of the edge nodes. In a possible implementation, the center node can be a cloud server or an access network device, and the edge node can be a terminal.

[0089] For example, federated learning is a typical application of distributed training. Taking the number of edge nodes participating in training as 3, which are respectively called edge node 1 to edge node 3, as an example, as shown in FIG. 5, the training process of federated learning includes:

[0090] Step 1, broadcast global model: in the first round of training process, the global model can be called the initialized AI model, which is referred to as the initial AI model. The center node broadcasts the initial AI model to each edge node. Then start the iterative training, and each iteration training process is:

[0091] Step 2, local training: each edge node trains the initial AI model using local data to obtain the gradient of the trained AI model. Optionally, the gradient can be understood as: the change amount of the parameters (such as weights) in the neural network of the trained AI model relative to the parameters in the neural network of the AI model before training.

[0092] Step 3, report gradient: each edge node reports the gradient trained by itself to the center node.

[0093] Step 4, global aggregation: after the center node receives the gradient reported by each edge node, the gradient is aggregated, and the parameters of the AI model are updated according to the aggregated gradient.

[0094] After that: the center node calculates the loss function of the AI model after updating the parameters; if the loss function meets the condition, the model training is terminated; if the loss function does not meet the condition, steps 1-4 are repeated. The difference is that in the training process of other rounds except the first round: the center node sends the aggregated gradient to the edge participating in the training, and the edge node updates the parameters and gradient of the AI model trained locally according to the aggregated gradient sent by the center node, and continues to train the updated AI model using local data. This training method requires that the AI models of all edge nodes have the same structure, so that the center node can effectively aggregate the gradients sent by each edge node.

[0095] For another example, ring learning is another typical application of distributed training. In ring learning, multiple center nodes are included, and each center node is associated with multiple edge nodes. Each center node aggregates the AI models sent by the multiple edge nodes associated therewith and sends them to the next center node until the loss function of the trained AI model meets the condition. This training method also requires that the AI models of all edge nodes under each center node have the same structure.

[0096] Therefore, in the distributed training, when the structures of the multiple AI models are the same, the central node can effectively aggregate the gradients of the multiple AI models or the AI model. If the structures of the AI models are inconsistent, the central node cannot aggregate, which limits the flexibility of AI model aggregation, and the fixed model can only provide one aggregation result.

[0097] In view of the above, the embodiments of the present application provide a communication method, in which: a first communication device sends a first model feature to a second communication device, the first model feature is used to represent the model feature of the first model, so that the second communication device can perform corresponding operations according to the first model feature. For example, the second communication device can aggregate multiple models including the first model according to the first model feature. The model structures of the aggregated multiple models are not limited, thereby improving the flexibility of AI model aggregation. Further, due to the different model structures of the multiple models, multiple aggregation results may be generated. It can be understood that the scheme of the embodiments of the present application is not limited to be applied in the distributed training scene. For example, in addition to the distributed training scene, the scheme of the embodiments of the present application can also be applied in the scenes of non-distributed training, distributed inference, model updating, or training of double-sided models, etc., which will be described in detail below.

[0098] In the following description, the first communication device and the first communication device are executed as the main line, and the scheme of the embodiments of the present application is described. The application scenarios of the scheme of the embodiments of the present application and the first communication device and the first communication device are exemplarily described below.

[0099] As shown in FIG. 6a, the scheme of the embodiments of the present application can be applied in the scene of communication between a terminal and a network (NW). For example, the first communication device and the first communication device can be a terminal and a network device respectively, and the network device can be an access network device (such as a base station), or a core network device, etc., which is not limited; or the first communication device can be a module, unit or component (chip, chip system, circuit, processor or other, etc.) in the terminal, and the second communication device can be a module, unit or component, etc. in the network device. When the network device is an access network device, the terminal and the access network device can communicate through a wireless link. Or, when the network device is a core network device, the communication between the terminal and the core network device is forwarded through the access network device.

[0100] As shown in FIG. 6b, the scheme of the embodiments of the present application can be applied in the scene of communication between a terminal and a terminal. For example, the first communication device and the second communication device can be a first terminal and a second terminal respectively, and the first terminal and the second terminal can communicate through a sidelink (SL). Or, the first communication device can be a module, unit or component, etc. in the first terminal, and the second communication device can be a module, unit or component, etc. in the second terminal.

[0101] As shown in FIG. 6c, the scheme of the embodiments of the present application can be applied to the scenario of communication between network devices. For example, the first communication apparatus and the second communication apparatus can be first network devices or second network devices respectively. Alternatively, the first communication apparatus can be a module, unit or component in the first network device, and the second communication apparatus can be a module, unit or component in the second network device. The first network device and the second network device can communicate through a backhaul link. The backhaul link can be a wired backhaul link, such as an optical fiber, a copper cable, or a wireless backhaul link, such as a microwave.

[0102] It can be understood that when the execution subject is a terminal, a module, unit or component in a network device, the receiving / sending can be understood as input / output, i.e., the module communicates with other modules or other components of the terminal or the network device. At this time, the processing performed by a single execution subject can also be divided into processing performed by multiple execution subjects, which can be logically and / or physically separated. For example, the processing performed by the access network device can be divided into processing performed by at least one of the CU, the DU, the RU, etc.

[0103] FIG. 7 is a schematic interaction diagram of a communication method 700 provided by the embodiments of the present application. It can be understood that steps 710 to 730 are only for describing the process of the communication method 700, and should not constitute a limitation on the method 700. Steps 710 to 730 can be divided into more steps, or combined into fewer steps, and the order of steps 710 to 730 is not limited.

[0104] Step 710: The first communication apparatus determines a first model feature.

[0105] The first model feature is used to represent the model feature of the first model. Optionally, the first model can also be referred to as an AI / machine learning (ML) model. The AI / ML model is a specific method for implementing AI functions, and the AI model describes the mapping relationship between the input and the output of the model.

[0106] Step 720: The first communication apparatus sends a first indication, and the second communication apparatus receives the first indication.

[0107] The first indication is used to indicate the first model feature. The first model feature includes at least one of the following:

[0108] 1. Application scenario of the first model:

[0109] The application scenario of the first model refers to a scenario in which the first model is applied. For example, the application scenario of the first model includes an indoor factory (InF), an indoor office, a rural macro cell (RMA), an urban macro cell (UMA), or an urban micro cell (UMI), and the like. For example, when the application scenario of the first model is an InF, the first model is applied in the scenario of the InF.

[0110] 2. Structure of the first model

[0111] The structure of the first model can be a fully connected network, an embedding layer, a convolutional layer, or a pooling layer, and the like. For example, the first model is a transformer model. The transformer model is a deep learning model mainly used for natural language processing. A typical application of the transformer model is to give a corresponding reply to a user's question. For example, the user inputs a corresponding question into the transformer model, and the transformer model processes the question input by the user and outputs a corresponding reply. The transformer model includes an encoder and a decoder. The encoder includes one or more encoder blocks, each of which includes a multi-head attention module, and each multi-head attention module includes one or more attention modules. The model structure of the transformer model can include at least one of the following: the number of encoder blocks included in the encoder, the number of attention modules included in each encoder block, the number of decoder blocks included in the decoder, the number of attention modules included in each decoder block, and the like.

[0112] 3. Function of the first model

[0113] The function of the first model refers to a function implemented by the first model. For example, the function of the first model includes modulation and coding scheme (MCS) prediction, channel state information (CSI) prediction, beam prediction, or positioning, and the like.

[0114] 4. Format of the first model

[0115] For example, the format of the first model comprises: a quantization index of the first model, and an input / output dimension of the first model. The quantization index of the first model can comprise a quantization manner of the first model and a corresponding quantization codebook. For example, when the quantization manner of the first model is scalar quantization, the corresponding quantization codebook is a codebook for scalar quantization. Or, when the quantization manner of the first model is vector quantization, the corresponding quantization codebook is a codebook for vector quantization. It can be understood that model quantization refers to the process of converting floating-point numbers in the neural network corresponding to the model into fixed-point numbers. For example, after the training of a model is completed, the parameters in the neural network corresponding to the model are represented by floating-point numbers. When the model is sent to the corresponding application node for model inference, the floating-point numbers need to be converted into fixed-point numbers.

[0116] 5. The target performance of the first model:

[0117] For example, the target performance of the first model comprises at least one of the following: accuracy, system performance (such as throughput, or link performance (such as signal to interference plus noise ratio, SINR, block error rate, BLER), etc. The accuracy refers to the accuracy of the first model when implementing the corresponding function. For example, the first model is used for MCS prediction, and in the prediction result of the first model: the MCS obtained by 8 times of prediction is the same as the actual MCS, and the MCS obtained by 2 times of prediction is different from the actual MCS, then the accuracy of the MCS prediction of the first model can be considered as 80%. It can be understood that the actual MCS refers to the MCS that matches the current wireless channel.

[0118] Optionally, in step 730, the second communication device performs corresponding processing according to the first model feature.

[0119] Through the above design, the definition of the model feature is introduced, and the model feature comprises information of at least one dimension of the following: application scenario of the model, structure of the model, function of the model, format of the model, or target performance of the model, etc. The characteristics of the model are described from different dimensions, and the flexibility of model management is enhanced.

[0120] The scheme of the flow in FIG. 7 can be applied to scenarios such as distributed training, non-distributed training, distributed inference, model updating, or training of a bilateral model. In different scenarios, when the second communication device receives the first indication for indicating the first model feature, the second communication device can perform corresponding processing. The scheme of the embodiment of the present application is described below in combination with specific application scenarios:

[0121] The scheme of the embodiments of the present application can be applied to a distributed training scenario: in the distributed training scenario, an edge node can perform model training using local data, and send the trained model to a center node, and the center node aggregates the models sent by the plurality of edge nodes. The second communication device can be a center node, or a module, unit or component applied to a center node. For example, the second communication device can aggregate the first model according to the first model feature. For example, the second communication device obtains a plurality of models, and the plurality of models include the first model. The second communication device can aggregate according to the model features corresponding to the plurality of models. For example, the models with consistent model features are aggregated together to generate a new model. The following three examples are used to illustrate the process of the present application:

[0122] The first communication device can be a terminal, or a module, unit or component applied to a terminal. The second communication device can be a network device, or a module, unit or component applied to a network device. In the following description, the first communication device is taken as a terminal, and the second communication device is taken as a network device:

[0123] Example 1: In the distributed training scenario: the terminal acts as an edge node, and the terminal performs model training using local data to obtain a corresponding model, and the corresponding model includes at least a first model. The terminal obtains a model feature of the first model, which can be referred to as a first model feature, and generates a first indication according to the first model feature, and the first indication is used to indicate the first model feature. For example, the first indication can explicitly or implicitly indicate the first model feature. For example, when the first indication includes information of the first model feature, the first indication can explicitly indicate the first model feature. Or, when the first indication includes other information related to the first model feature, the other information has a corresponding relationship or an association relationship with the first model feature, for example, the other information is an index or a number of the first model feature, and the first indication can implicitly indicate the first model feature. The terminal sends the first model and the first indication to the center node, i.e., the network device, and the network device aggregates the models with consistent model features. This way can make full use of the training capability of each edge node and reduce the limitation on the edge node. As can be seen, in this example 1, the terminal sends the first model to the network device in addition to the first indication. Correspondingly, the network device receives the first model in addition to the first indication.

[0124] Example 1.1: The network device aggregates the models with consistent application scenario features: each terminal performs model training using local data to obtain a corresponding model. Each terminal sends the trained model and the indication information of the application scenario feature of the model to the network device. The network device aggregates the models according to the application scenario features of the models reported by the terminals.

[0125] As shown in FIG. 8, terminal 1 and terminal 2 train using local data to obtain model 1 and model 2. The application scenario features of model 1 and model 2 are both INF, and model 1 can be referred to as INF model 1 and model 2 can be referred to as INF model 2. Terminal 1 sends model 1 and indication information of the application scenario feature INF of model 1 to the network device, and terminal 2 sends model 2 and indication information of the application scenario feature INF of model 2 to the network device. Since the application scenario features of model 1 and model 2 are consistent and both are INF, the network device aggregates model 1 and model 2 to obtain an aggregated model, which can be referred to as an INF aggregation (AG) model. For example, the application scenario feature of model 1 reported by terminal 3 is UMA, and model 1 can be referred to as UMA model 1. The application scenario feature of model 2 reported by terminal 4 is UMA, and model 2 can be referred to as UMA model 2. Since the application scenario features of model 1 reported by terminal 3 and model 2 reported by terminal 4 are consistent and both are UMA, the network device can aggregate model 3 and model 4 to obtain an aggregated model, which can be referred to as a UMA AG model.

[0126] Example 1.1.2: The network device aggregates models with consistent functional features: Each terminal trains a model using local data. Each terminal sends the trained model and indication information of the functional feature of the model to the network device. The network device aggregates the models according to the functional features of the models reported by the terminals.

[0127] As shown in FIG. 9, terminal 1 trains using local data to obtain two sub-models, referred to as sub-model 1 and sub-model 2. The function of sub-model 1 is MCS prediction, and the function of sub-model 2 is beam prediction. The terminal sends the two trained sub-models and indication information of the functional features of the corresponding sub-models to the network device. Terminal 2 trains using local data to obtain two sub-models, referred to as sub-model 1 and sub-model 2. The function of sub-model 1 is beam prediction, and the function of sub-model 2 is positioning. The terminal sends the two trained sub-models and indication information of the functional features of the corresponding sub-models to the network device. Since the functions of sub-model 2 reported by terminal 1 and sub-model 1 reported by terminal 2 are consistent and both are beam prediction, the network device aggregates sub-model 2 reported by terminal 1 and sub-model 1 reported by terminal 2 to obtain an aggregated model, and the function of the aggregated model is MCS prediction. The aggregated model can be referred to as a beam prediction AG model.

[0128] It can be understood that, in the embodiments of the present application, when the network device aggregates the models according to the model features of the models, the network device can aggregate the models with consistent model features together to generate a new model. The structural parameters of the multiple models with consistent model features can be the same or different. When the structural parameters of the multiple models are inconsistent, the network device can adjust at least one of the multiple models to make the structural parameters of the multiple models consistent, and then aggregate the multiple models with consistent structural parameters. For example, the structural parameters of model 1, model 2 and model 3 are inconsistent, and the network device can select one model from the three models, and adjust the structural parameters of the remaining two models to be consistent with the structural parameters of the model. Alternatively, the network device can obtain a target structural parameter, and adjust the structural parameters of at least one of model 1, model 2 and model 3 to make the structural parameters of the three models consistent with the target structural parameter.

[0129] Example 2: In the scenario of distributed training, the network device as a center node sends a second indication to a terminal as an edge node, and the second indication is used to indicate the expected model feature of the network device for the first model, which can be referred to as a second model feature. It can be understood that the dimension of the information included in the second model feature can be the same as the dimension of the information included in the first model feature. For example, the second model feature can include information of at least one of the following dimensions: the application scenario of the first model (for ease of distinction, it can be referred to as the application scenario of the second communication device expected or the expected first model), the structure of the first model (for ease of distinction, it can be referred to as the structure of the second communication device expected or the expected first model), the function of the first model (for ease of distinction, it can be referred to as the function of the second communication device expected or the expected first model), the format of the first model (for ease of distinction, it can be referred to as the format of the second communication device expected or the expected first model), or the target performance of the first model (for ease of distinction, it can be referred to as the target performance of the second communication device expected or the expected first model).

[0130] Optionally, the second indication can explicitly or implicitly indicate the second model feature. The terminal generates the first model according to the second model feature. For example, the terminal can train the first model using the local data and the second model feature expected by the network device, or the terminal can select a model matching the second model feature from the trained models, referred to as the first model. The terminal obtains the actual model feature of the first model, referred to as the first model feature. The terminal sends the first model and the first indication to the network device, where the first indication is used to indicate the first model feature. The network device performs model aggregation on a plurality of models including the first model according to the first model feature. It can be understood that the actual model feature of the first model (the first model feature) can be the same as or different from the model feature of the first model expected by the network device (the second model feature). In a possible implementation, when the second model feature is different from the first model feature, the terminal sends the first indication to the network device to indicate the first model feature.

[0131] For example, the network device indicates a certain model feature to the terminal, which can be the model feature expected by the network device (referred to as the second model feature). The terminal reports a model matching the model feature expected by the network device to the network device, and the network device aggregates the models with the same model feature. In this way, the terminal can independently train a plurality of models based on the local data, and each model has a different model feature. When the terminal receives the expected model feature indicated by the network device, the terminal can select a model matching the expected model feature indicated by the network device from the plurality of models and report the model to the network device. Alternatively, the terminal can train a model according to the expected model feature indicated by the network device. The terminal reports the trained model to the network device. The network device aggregates the models with the same model feature. Optionally, the terminal can report the actual model feature (the first model feature) of the trained model to the network device. In a possible implementation, the expected model feature indicated by the network device can be different from the actual model feature of the model trained by the terminal. The network device can determine whether to perform model aggregation according to the actual model feature of the model reported by the terminal.

[0132] For example, taking the first model as a conversion model, the model structure of the conversion model is as follows:

[0133] As shown in FIG. 10a, the model structure of the conversion model can correspond to a large-scale model, in which: the conversion model comprises an encoder and a decoder; the encoder comprises a plurality of encoding blocks, and the decoder comprises a plurality of decoding blocks. For example, as shown in FIG. 10a, the encoder comprises four encoding blocks, and the indexes of the four encoding blocks are 0 to 3 in sequence; the decoder comprises four decoding blocks, and the indexes of the four decoding blocks are 0 to 3 in sequence. It can be understood that, when the conversion model is used for natural language processing, the encoding block 0 processes the question input by the user, and then sends the output result to the encoding block 1, the encoding block 1 sends the output result to the encoding block 2 in sequence, the encoding block 2 sends the corresponding output result to the encoding block 3, and the encoding block 3 sends the final output result to the decoding blocks 0 to 3, and the decoding block 3 outputs the final processing result. Further, each encoding block or decoding block comprises a plurality of attention modules.

[0134] As shown in FIG. 10b, the model structure of the conversion model can correspond to a medium-scale model, which is different from the model structure shown in FIG. 10a in that: the encoder comprises only one encoding block, and the decoder comprises only one decoding block, and further, each encoding block or decoding block comprises a plurality of attention modules. It can be understood that, in FIG. 10b, the number of attention modules included in each encoding block or decoding block is not limited, which is exemplified by taking two attention modules included in each encoding block or decoding block as an example. In one description, each encoding block or decoding block can comprise a multi-head attention module, and the multi-head attention module comprises a plurality of attention modules.

[0135] As shown in FIG. 10c, the model structure of the conversion model can correspond to a small-scale model, which is different from the model structure shown in FIG. 10b in that: each encoding block and decoding block comprises only one attention module.

[0136] The network device indicates the expected model structure features (i.e., the second model features) to the terminal, for example, the number of encoding blocks and decoding blocks, and / or the number of attention modules, etc. The terminal trains the corresponding model according to the model structure features indicated by the network device, and reports the trained model to the network device, and the network device aggregates the models with consistent model structure features, etc.

[0137] Example 3: In the distributed training scenario, or other training scenarios (for example, a non-distributed training scenario): the terminal sends the indication information of the expected model features (i.e., the first model features) to the network device, the network device obtains the model (i.e., the first model) meeting the first model features expected by the terminal, and the network device sends the first model to the terminal. For example:

[0138] The terminal sends a first indication to the network device, where the first indication is used to indicate a first model feature. The network device determines a first model according to the first model feature, where the first model is determined according to the first model feature. For example, the network device performs model training according to the first model feature to obtain a first model that meets the first model feature. Alternatively, the network device determines a model that meets the first model feature from the trained models, which is referred to as the first model; and the network device sends the first model that meets the expected first model feature to the terminal. Optionally, in a distributed training scenario, the terminal can continue to train the first model by using local data, and send the trained model to the center node, i.e., the network device. Further, optionally, the network device can aggregate the models according to the model features of the models.

[0139] Through the above design, the center node aggregates the models trained by the edge nodes according to the model features, for example, the models with consistent model features are aggregated together to form a new model, without limiting whether the model structures of the aggregated models are consistent. Compared with the case where only the models with consistent model structures can be aggregated, the flexibility and diversity of model aggregation are improved, and the limitation on the models trained by the edge nodes is reduced. Further, since the model structures of the aggregated models can be different, different model aggregation results can be obtained according to different strategies.

[0140] It can be understood that the scheme of the embodiments of the present application can also be applied to a non-distributed training scenario. In the non-distributed training scenario, a node (which can be referred to as a training node) performs model training, and sends the trained model to another node (which can be referred to as an inference node), and the inference node performs model inference according to the model to obtain a corresponding inference result. Optionally, the training node can be the network device, and the inference node can be the terminal, or vice versa, without limitation.

[0141] When the scheme of the embodiments of the present application is applied in the non-distributed training scenario, the first communication device sends the model feature of the first model, i.e., the first model feature, to the second communication device in addition to sending the trained first model to the second communication device. The second communication device processes the first model according to the first model feature. For example, the second communication device can perform inference on the first model according to the first model feature.

[0142] The first communication device can be a training node, or a module, unit or component located in the training node, and the second communication device can be an inference node, or a module, unit or component located in the inference node. Taking the first communication device as a training node and the second communication device as an inference node as an example, for example:

[0143] In the distributed training scenario, when the training node sends the first model trained to the inference node, the training node can also send the model feature of the first model (i.e., the first model feature) to the inference node; the inference node performs model inference on the first model according to the first model feature. For example, the first model feature includes the application scenario feature of the first model, and the inference node can perform model inference on the first model in the corresponding scenario. Or, the first model feature includes the model structure feature of the first model, and the inference node can determine whether the model structure feature of the first model matches the computing resources of the inference node locally; if so, the inference node performs model inference on the first model; or if not, the inference node does not perform model inference on the first model. For example, the model feature of the first model sent by the training node to the inference node (i.e., the first model feature) can correspond to a large-scale model (see the description in FIG. 10a), and if the computing resources of the inference node locally are limited and insufficient to support large-scale model inference, the inference node can not perform model inference on the first model.

[0144] For another example, in the distributed training scenario, the inference node can send indication information of the expected model feature (i.e., the first model feature) to the training node. The training node can train the corresponding first model according to the first model feature sent by the inference node, and send the first model to the inference node. Optionally, the training node can also send the actual model feature of the first model trained by the training node (which can be referred to as the third model feature) to the inference node. The model feature of the first model trained by the training node can not be completely consistent with the model feature indicated or expected by the inference node.

[0145] The scheme of the embodiments of the present application can also be applied to the distributed inference scenario: distributed inference refers to model inference performed on multiple nodes, and each node corresponds to a sub-model. For example, the network device is the first inference node, and the terminal is the second inference node. The network device performs model inference on a sub-model to obtain inference information, and sends the inference information as an intermediate result to the terminal. The terminal takes the inference information as input, continues to perform model inference on the sub-model, and obtains the final inference result. In the distributed inference scenario, the last inference node sends the indication information of the model feature of the sub-model used for model inference in the last inference node to the next inference node, so that the next inference node can select a sub-model consistent with the model feature of the last inference node to continue model inference, and ensure that the model features of the sub-models used for model inference between different inference nodes are aligned. It can be understood that at least one sub-model is deployed in each inference node, and the sub-models with consistent model features can be paired for distributed inference.

[0146] When the scheme of the embodiments of the present application is applied to the scenario of distributed reasoning: the first communication device determines first reasoning information according to the first model. For example, the first communication device can perform model reasoning by using the first model to obtain the first reasoning information. The first communication device sends the first reasoning information to the second communication device. Further, the first communication device can obtain the model feature of the first model, referred to as the first model feature. The first communication device sends the first indication to the second communication device, and the first indication is used to indicate the first model feature. The second communication device determines the second model according to the first model feature. For example, the second communication device determines the model that meets the first model feature from a plurality of models, and the model is referred to as the second model. The second communication device determines the second reasoning information according to the received first reasoning information and the second model. For example, the second communication device inputs the first reasoning information into the second model, and the output of the second model is the second reasoning information. It can be understood that the second reasoning information can be a final reasoning result, or can be an intermediate reasoning result, and at this time the second communication device can further send the second reasoning information and the model feature of the second model to the communication device corresponding to another reasoning node to continue model reasoning, which is not limited.

[0147] The first communication device can be a first reasoning node, or a module, unit or component applied to the first reasoning node; and the second communication device can be a second reasoning node, or a module, unit or component applied to the second reasoning node. For example, the first reasoning node can be a network device, and the second reasoning node can be a terminal, or vice versa. Alternatively, the first reasoning node and the second reasoning node are both terminals, or both are network devices.

[0148] It can be understood that in the scenario of distributed reasoning, the models of a plurality of reasoning nodes cooperate with each other to perform joint reasoning to determine a final reasoning result. The model corresponding to each node for reasoning can also be referred to as a sub-model. For example, the first model and the second model can be referred to as the first sub-model and the second sub-model respectively.

[0149] For example, taking the first reasoning node as a network device and the second reasoning node as a terminal as an example, the scheme provided by the embodiments of the present application is described.

[0150] For example, at least one target performance is predefined, and each target performance can include information of three dimensions of accuracy, system performance and link performance. Each target performance corresponds to an identity (ID). The identity of the target performance is used to uniquely identify a target performance, and the identity is alternatively described as an index or a number. For example, as shown in Table 1, four target performances can be predefined, and the identities of the four target performances are 0 to 3 in turn. The specific parameters of each target performance in the four target performances can be referred to Table 1.

[0151] Table 1

[0152] The network device is deployed with sub-model 1-1 and sub-model 1-2. Among them, the target performance characteristics corresponding to sub-model 1-1 and sub-model 1-2 are different. For example, the IDs of the target performance corresponding to sub-model 1-1 and sub-model 1-2 are 1 and 2 respectively. The terminal side is deployed with sub-model 2-1 and sub-model 2-2. Among them, the target performance corresponding to sub-model 2-1 and sub-model 2-2 is different. For example, the IDs of the target performance corresponding to sub-model 2-1 and sub-model 2-2 are also 1 and 2 respectively.

[0153] The network device selects sub-model 1-1 for model inference to obtain first inference information, which is used as an intermediate output result. The network device sends the first inference information and the target performance corresponding to sub-model 1-1 to the terminal. It can be understood that the ID of the target performance corresponding to sub-model 1-1 is 1, and the network device can send the target performance ID 1 to the terminal.

[0154] The terminal selects a model that meets the target performance from the multiple sub-models deployed in the terminal according to the target performance sent by the network device.

[0155] For example, the target performance sent by the network device is target performance ID 1, and among the sub-model 2-1 and sub-model 2-2 deployed in the terminal, sub-model 2-1 meets the condition, and the ID of the target performance corresponding to it is also ID 1, so the terminal selects sub-model 2-1. Further, the terminal can input the first inference information into sub-model 2-1 as input to continue model inference to obtain second inference information. It can be understood that the second inference information can be the final inference result or an intermediate inference result, which is not limited.

[0156] It can be understood that since the ID of the target performance of sub-model 1-1 is 1, the target performance IDs of the sub-model 2-1 and sub-model 2-2 deployed in the terminal are 1 and 2 respectively. The target performance ID of sub-model 2-1 matches the target performance ID of sub-model 1-1, both of which are 1. Therefore, the terminal selects sub-model 2-1 to continue model inference, thereby ensuring that the target performance of the model for model inference between each inference node is consistent.

[0157] Through the above design, in the distributed inference scenario, the last inference node sends the model characteristics of the model (or sub-model) corresponding to the inference information to the next inference node in addition to sending the corresponding inference information to the next inference node, so that the next inference node selects a model that matches the model characteristics of the model for model inference in the last inference node to continue model inference, thereby ensuring that the model characteristics of the model for model inference between different inference nodes are consistent, and improving the accuracy and flexibility of distributed inference.

[0158] The scheme of the embodiments of the present application can also be applied to a model updating scenario: the training node sends the model features of the updated model to the inference node when performing model updating. The inference node updates according to the model features of the updated model. Further, the inference node performs model inference using the updated model features.

[0159] When the scheme of the embodiments of the present application is applied to a model updating scenario: the first communication device updates the first model; the first communication device determines the first model features according to the updated first model, and the first model features are specifically used to represent the model features of the updated first model; the first communication device sends the first indication to the second communication device, and the first indication is used to indicate the first model features; further, the second communication device can update the model features of the first model saved locally by the second communication device according to the first model features.

[0160] It can be understood that the first communication device can be a training node, or a module, unit or component located in the training node, etc.; the second communication device can be an inference node, or a module, unit or component located in the inference node, etc. In a possible implementation manner, the training node can be a network device, and the inference node can be a terminal, or vice versa. Alternatively, the training node and the inference node are both network devices. Alternatively, the training node and the inference node are both terminals, etc., without limitation.

[0161] It can be understood that when the training node performs model updating, the model features of the model can be updated. For example, the structure of the model, the format of the model, etc. are updated. Therefore, in the embodiments of the present application, when the training node updates the model, if the update causes the corresponding update of the model features of the model, the training node can send the updated model features (i.e. the first model features) to the inference node.

[0162] In the following description, the training node is taken as a network device, and the inference node is taken as a terminal, as an example, to illustrate the scheme of the embodiments of the present application:

[0163] 1. The network device trains the first model and indicates the first model to the terminal.

[0164] 2. The network device updates the first model, and the network device obtains the model features of the updated first model, referred to as the first model features; the network device sends the first indication to the terminal, and the first indication is used to indicate the first model features.

[0165] For example, the network device updates the quantization manner in the model format in the model feature of the first model when updating the first model: for example, the quantization manner of the first model is updated from scalar quantization to vector quantization, and the quantization codebook is updated from a scalar quantization codebook to a vector quantization codebook. As shown in FIG. 11, the quantization manner of the first model before updating is scalar quantization, and the corresponding quantization codebook is codebook 1, which can be understood as a code type corresponding to scalar quantization. The network device updates the first model, and the updated first model is referred to as a second model. The quantization manner of the second model is updated to vector quantization, and the corresponding quantization codebook is codebook 2, which can be understood as a codebook of vector quantization.

[0166] For another example, the network device updates the application scenario in the model feature of the first model when updating the first model. For example, referring to FIG. 12, the application scenario of the first model before updating is INF, and the application scenario of the second model after updating is UMA.

[0167] 3. The terminal updates the model feature of the first model according to the first indication, and further performs model inference according to the updated model feature of the first model.

[0168] For example, the model feature of the first model includes at least one of the following: an application scenario of the first model, a structure of the first model, a function of the first model, a format of the first model, or a target performance of the first model. It can be understood that if the application scenario of the first model is updated, the terminal performs model inference using the first model under the updated application scenario. Or if the structure of the first model is updated, the terminal updates the structure of the first model. Or if the format of the first model is updated, the format of the first model is updated, for example, the quantization manner of the first model is updated, and / or the input / output dimension is updated. Or if the target performance of the first model is updated, the target performance of the first model is updated. Further, the terminal performs model inference using the first model under the condition of meeting the target performance.

[0169] Through the above design, when the training node updates the model feature of the first model, the training node can indicate the updated model feature to the inference node, and the inference node performs model inference according to the updated model feature, thereby improving the reliability of inference.

[0170] The scheme of the present application can also be applied to the scenario of a bilateral model. In the scenario of a bilateral model, one model is deployed in one node, the one node performs model inference using the model to obtain a corresponding output result, and sends the corresponding output result to another node. Another model is deployed in another node, and the other model takes the received output result as input and inputs it into another model to obtain a corresponding result.

[0171] For example, the bilateral model discussed in 3GPP Release 18 / 19, a specific use case is CSI compression, which is used to reduce the overhead of CSI transmission. The terminal measures the reference signal and obtains the CSI. The terminal compresses the CSI through the model to obtain the compressed CSI. The terminal sends the compressed CSI to the network device. When the network device receives the compressed CSI, it decompresses the CSI through the model to obtain the complete CSI. The process of CSI compression involves two models, among which the model deployed on the terminal side is called the CSI generation part model, and the model deployed on the network device side is called the CSI reconstruction part model. Among them, the training method of the bilateral model includes three kinds:

[0172] The first kind: the network device or the terminal trains the CSI generation part model and the CSI reconstruction part model, and sends the trained CSI generation part model or CSI reconstruction part model to the opposite side. The opposite side can directly use the received model. Or, the opposite side takes the received model as a reference model and further trains the reference model. For example, the network device trains the CSI generation part model and the CSI reconstruction part model. The network device sends the CSI generation part model to the terminal, and the terminal can directly use the CSI generation part model or further train the received CSI generation part model as a reference model.

[0173] The second kind: the network device or the terminal trains the CSI generation part model and the CSI reconstruction part model, and sends the trained CSI generation part model or CSI reconstruction part model and the corresponding data set to the opposite side. The opposite side further trains the model based on the received data set. For example, the network device trains the CSI generation part model and the CSI reconstruction part model, and sends the CSI generation part model and the corresponding data set to the terminal. The terminal further trains the received CSI generation part model based on the received data set to obtain the final CSI generation part model.

[0174] The third kind: the network device and the terminal train the CSI generation part model and the CSI reconstruction part model together. In each round of training process, the output of the interaction model and the updated gradient are exchanged.

[0175] Among them, the first and second training methods can be considered as sequential training, that is, one side first trains, and the other side further trains based on the reference model sent by the first training side to obtain the updated model. Alternatively, the first training side can also send the target performance to the other side as a guide, and the other side further trains the reference model according to the target performance to ensure that the updated model and the target performance of the model on the opposite side can be aligned.

[0176] In the sequential training, one side provides a reference model, the other side updates based on the reference model, and the side trained first can also send a target performance, hoping that the other side will train according to the target performance, so that the updated model is aligned with the target performance of the model on the other side. However, since the other side lacks training guidance information, it can only update the received reference model according to its own ability, so the updated model may not be able to achieve the target performance. Moreover, this way of providing a target performance limits the ability of the terminal to retrain, and does not fully utilize the processing capabilities of different terminals. Terminals with strong capabilities can support more complex models in structure or function.

[0177] In the embodiments of the present application, the side trained first sends the reference model and the fixed model features and / or the variable model features in the reference model to the other side. The other side does not change the fixed model features of the reference model, and optimizes the variable model features in the reference model through training. Since the side trained first provides training guidance information to the other side, indicating which features in the reference model are fixed and cannot be changed, and which features in the reference model can be changed, the other side can train the reference model according to the above indication, so that the model obtained by training can meet the requirements of the side trained first. Further, this way does not limit the training ability of the other side, and fully utilizes the processing capabilities of terminals.

[0178] Specifically, when the scheme of the embodiments of the present application is applied to the training scenario of a bilateral model: the first communication device can send a first model and a first indication to the second communication device, the first indication being used to indicate first model features, the first model features being used to represent model features of the first model, and the type of the first model features specifically including: fixed model features of the first model, and / or variable model features of the first model. It can be understood that the fixed model features can refer to fixed and unchanged features in the first model, and the variable model features can refer to changeable features in the first model. The first model as a reference model is used to determine a second model. The second communication device can determine the second model according to the first model, and the fixed model features of the first model and / or the variable model features of the first model. Optionally, the second model can be a model deployed in the second communication device.

[0179] It can be understood that the first communication device can be a first training node, or a module, unit or component located in the first training node, etc. The second communication device can be a second training node, or a module, unit or component located in the second training node, etc. It can be understood that the first training node can be understood as a side node that performs model training first, and the second training node can be understood as another side node that performs model training later. The first training node can be a network device, and the second training node can be a terminal, or vice versa. Alternatively, the first training node and the second training node can both be network devices, or terminals, etc.

[0180] It can be understood that in the embodiments of the present application, the fixed model features or variable model features of the first model can include information of multiple dimensions. For example, the fixed model features of the first model can include information of at least one dimension of the application scenario of the model, the structure of the model, the function of the model, or the format of the model, etc. Similarly, the variable model features of the first model can also include information of at least one dimension. For example, the fixed model features of the first model can be specifically the function of the model, and the variable model features of the first model can be specifically the structure of the model. In the following example, it is specifically illustrated that which features in the model structure are fixed model features, and which features in the model structure are variable model features.

[0181] Taking the first communication device or the first training node as a network device, and the second communication device or the second training node as a terminal as an example, the scheme of the embodiments of the present application is illustrated by combining the following three examples:

[0182] Example 1: The network device trains model X and model Y', and model X is a model deployed in the network device. The network device sends model Y' as a reference model (i.e. the first model) to the terminal. Further, the fixed model features and / or variable model features (i.e. the first model features) in the reference model Y' are indicated to the terminal. The terminal determines model Y according to the reference model Y' and the first model features, and model Y is a model deployed in the terminal. The specific process is as follows:

[0183] Step 1: The network device obtains model X and model Y' through training. Model X is a model deployed in the network device, and model Y' is a reference model.

[0184] Step 2: The network device sends model Y' to the terminal, and the indication information of the fixed model features and / or variable model features in model Y'.

[0185] For example, the fixed model feature of the model Y' can be the structure of the attention module, and the variable model feature of the model Y' can be the number of attention modules, that is, the number of attention modules can be changed. After receiving the model Y', the terminal can continue to train the model Y' based on the training set. During the training process, the structure of the attention module in the model Y' remains unchanged, and through training, the number of attention modules in the model Y' can be increased. For example, referring to FIG. 13, the model Y' is taken as a reference model, and the number of attention modules included in the decoding block of the model Y' is 1. The model Y obtained by the terminal through training includes a plurality of attention modules in the decoding block. In FIG. 12, an example in which the number of attention modules included in the decoding block is 2 is illustrated.

[0186] For another example, the fixed model feature of the model Y' can be the structure of the decoding block, and the variable model feature of the model Y' can be the number of decoding blocks; or the fixed model feature of the model Y' can be the normalization layer, and the variable model feature of the model Y' can be the normalization manner, and the like.

[0187] Step 3: The terminal determines the model Y according to the model Y' and the fixed model feature and / or the variable model feature in the model Y'.

[0188] It can be understood that the fixed model feature in the model Y' remains unchanged, and the variable model feature in the model Y' can be optimized during the training process, and the obtained model Y is taken as the model deployed on the terminal side. Further, the terminal can use the model Y to perform model inference.

[0189] Example 2: The network device trains the model X, and the model X is taken as a reference model. The terminal trains the model Y based on the reference model X, and the specific process is as follows:

[0190] Step 1: The network device trains the model X, and the model X is taken as a reference model deployed on the terminal side.

[0191] Step 2: The network device sends the reference model X and the fixed model feature and / or the variable model feature in the reference model X to the terminal.

[0192] Step 3: The terminal determines the model Y according to the reference model X and the fixed model feature and / or the variable model feature in the reference model X, and the model Y is taken as a model deployed on the terminal.

[0193] For example, the terminal can keep the fixed model feature in the reference model X unchanged, and optimize the variable model feature in the reference model X in the joint training process of the reference model X and the model Y. For example, as shown in FIG. 14, the fixed model feature of the reference model X is the structure of the attention module, and the variable model feature of the reference model X is the number of attention modules, that is, the number of attention modules can be changed. In the joint training process of the reference model X and the model Y, the structure of the attention module in the reference model X is fixed, and the number of attention modules in the reference model X is increased to optimize the model. Alternatively, the attention module of the decoding block of the model Y can also be referred to as a masked attention module.

[0194] It can be understood that in example 2, the model X is a possible implementation of the first model in the process of FIG. 7, and the fixed model feature and / or the variable model feature of the model X are a possible implementation of the first model feature in the process of FIG. 7.

[0195] Example 3: The network device trains the model X and the model Y', and the model X and the model Y' are used as reference models. The terminal trains the model Y based on the reference models, that is, the model X and the model Y'. The specific process is as follows:

[0196] Step 1: The network device trains the model X and the model Y', and the model X is a model deployed in the network device. The model X and the model Y' can be used as reference models on the terminal side.

[0197] Step 2: The network device sends the model X and the model Y' to the terminal, and the fixed model feature and / or the variable model feature in the model X and the fixed model feature and / or the variable model feature in the model Y'.

[0198] Step 3: The terminal determines the model Y according to the model X and the fixed model feature and / or the variable model feature in the model X, and the model Y' and the fixed model feature and / or the variable model feature in the model Y'. The model Y is a model deployed in the terminal.

[0199] For example, the terminal can keep the fixed model features in the model X and the model Y' unchanged, and optimize the variable model features in the model X and the model Y' in the process of jointly training the model X and the model Y'. For example, as shown in FIG. 15, the fixed model features of the model X and the model Y' are the structure of the attention module, and the variable model features of the model X and the model Y' are the number of attention modules, that is, the number of attention modules can be changed. In the process of jointly training the model X and the model Y', the structure of the attention module of the model X and the model Y' is fixed, and the number of attention modules in the model X and the model Y' is increased to optimize the model. It can be understood that the terminal can obtain the model X' and the model Y by jointly training the model X and the model Y'. Among them, the model Y is a model deployed in the terminal, and further the terminal can use the model Y for model inference. In the example of FIG. 15, the attention module of the decoding block of the model Y' and the model Y is described as a masked attention module.

[0200] It can be understood that in example 3, the model X and the model Y' are a possible implementation of the first model in the process of FIG. 7, and the fixed model features and the variable model features of the model X and the fixed model features and the variable model features of the model Y' are a possible implementation of the first model features in the process of FIG. 7.

[0201] Through the above design, in the scenario of sequential training of the bilateral model, the first side node trained can provide the fixed model and / or the variable model features of the reference model and other effective guidance information to the second side node trained, so that the second side node trained can continue model training based on the above effective guidance information, improve the efficiency of model training, and fully utilize the capacity of the training node.

[0202] In the above embodiments of the present application, the method provided by the embodiments of the present application is introduced from the perspective of interaction between the first communication device and the second communication device. In order to realize each function in the method provided by the embodiments of the present application, the first communication device and the second communication device can include hardware structures and / or software modules, and realize the above functions in the form of hardware structures, software modules, or hardware structures and software modules. Whether a certain function in the above functions is executed in the form of hardware structure, software module, or hardware structure and software module depends on the design constraint conditions of the specific application of the technical solution.

[0203] FIG. 16 and FIG. 17 are structural schematic diagrams of possible communication apparatuses provided in the embodiments of the present application. The communication apparatuses can implement one or more of the corresponding functions in the method embodiments described above. For example, the functions implemented by the first communication apparatus or the second communication apparatus, and thus the beneficial effects possessed by the method embodiments described above can be achieved. In the embodiments of the present application, the communication apparatus can be a first node, or a unit, module or component (such as a chip, chip system, circuit or processor, etc.) applied in the first node, or the communication apparatus can be a second node, or a unit, module or component applied in the second node. It can be understood that the first node and the second node can be a terminal and a network device respectively, or vice versa; or the first node and the second node can both be network devices; or the first node and the second node can both be terminals.

[0204] In the following description, the communication apparatus is taken as an example including a processing unit and a transceiver unit. It can be understood that the "unit" can also be replaced by "module" or "component", etc. For example, the processing unit in the following description can also be replaced by a processing module or a processing component. The transceiver unit can also be replaced by a transceiver module or a transceiver component. For example, the transceiver component can refer to a communication module.

[0205] As shown in FIG. 16, the communication apparatus 1600 includes a processing unit 1610 and a transceiver unit 1620. The communication apparatus 1600 is configured to implement the functions of the first communication apparatus or the second communication apparatus in FIG. 7 described above.

[0206] Optionally, the transceiver unit 1620 can also be referred to as an output unit, an interface unit, or a communication unit, etc. In a possible implementation, the transceiver unit 1620 includes at least one of a sending unit or a receiving unit. The sending unit and the receiving unit can be integrated together, or two independent units, etc.

[0207] When the communication apparatus 1600 is configured to implement the functions of the first communication apparatus in FIG. 7, specifically: the processing unit 1610 is configured to determine a first model feature; and the transceiver unit 1620 is configured to send a first indication, where the first indication is used to indicate the first model feature; and the first model feature is used to represent a model feature of a first model, and the first model feature includes at least one of an application scenario of the first model, a structure of the first model, a function of the first model, or a format of the first model.

[0208] In a possible implementation, the first model feature further includes a target performance of the first model.

[0209] In a possible implementation, the transceiver unit 1620 is further configured to send the first model.

[0210] In a possible implementation, the transceiver 1620 is further configured to receive a second indication, where the second indication is used to indicate a second model feature, and the second model feature is an expected model feature of the second communication apparatus for the first model, and the second model feature includes at least one of an expected application scenario of the first model, an expected structure of the first model, an expected function of the first model, or an expected format of the first model; and determine the first model according to the second model feature.

[0211] In a possible implementation, the second model feature is different from the first model feature.

[0212] In a possible implementation, the transceiver 1620 is further configured to receive the first model, where the first model is determined according to the first model feature.

[0213] In a possible implementation, the processing unit 1610 is further configured to determine first inference information according to the first model, and the transceiver 1620 is further configured to send the first inference information.

[0214] In a possible implementation, the first model feature is specifically used to represent a model feature of the updated first model, and the processing unit 1610, when determining the first model feature, is specifically configured to update the first model, and determine the first model feature according to the updated first model.

[0215] In a possible implementation, the type of the first model feature includes a fixed model feature of the first model and / or a variable model feature of the first model, and the fixed model feature of the first model and / or the variable model feature of the first model are used to determine the second model based on the first model.

[0216] When the communication apparatus 1600 is configured to implement the functions of the second communication apparatus in FIG. 7, specifically: the transceiver 1620 is configured to receive a first indication, where the first indication is used to indicate a first model feature, and the first model feature is used to represent a model feature of a first model, and the first model feature includes at least one of an application scenario of the first model, a structure of the first model, a function of the first model, or a format of the first model; and optionally, the processing unit 1610 is configured to perform corresponding processing according to the first model feature.

[0217] In a possible implementation, the first model feature further includes a target performance of the first model.

[0218] In a possible implementation, the transceiver 1620 is further configured to receive the first model; and aggregate the first model according to the first model feature.

[0219] In a possible implementation, the transceiver 1620 is further configured to send a second indication, where the second indication is used to indicate a second model feature, the second model feature is used to determine the first model, and the second model feature is an expected model feature of the second communication device for the first model, and the second model feature includes at least one of an expected application scenario of the first model, an expected structure of the first model, an expected function of the first model, or an expected format of the first model.

[0220] In a possible implementation, the second model feature is different from the first model feature.

[0221] In a possible implementation, the processing unit 1610 is further configured to determine the first model according to the first model feature, and the transceiver 1620 is further configured to send the first model.

[0222] In a possible implementation, the processing unit 1610 is further configured to determine a second model according to the first model feature, and the transceiver 1620 is further configured to receive first inference information determined according to the first model, and the processing unit 1610 is further configured to determine second inference information according to the second model and the first inference information.

[0223] In a possible implementation, the first model feature is specifically used to represent a model feature of the updated first model, and the processing unit 1610 is further configured to update a model feature of the first model locally saved by the second communication device according to the first model feature.

[0224] In a possible implementation, the type of the first model feature includes a fixed model feature of the first model and / or a variable model feature of the first model, the transceiver 1620 is further configured to receive the first model, and the processing unit 1610 is further configured to determine a second model according to the first model and the fixed model feature of the first model and / or the variable model feature of the first model.

[0225] For more details of the processing unit 1610 and the transceiver 1620, refer to the description in FIG. 7 in the method embodiments above, which will not be repeated here.

[0226] It can be understood that the division of the units in the embodiments of the present application is schematic, and is merely logical function division. Actual implementation can have another division manner. In addition, each functional unit in the embodiments of the present application can be integrated in one physical device (for example, in a processor), or each functional unit can be a separate physical device, or two or more units can be integrated in one unit for implementation. The integrated unit can be implemented in the form of hardware, or in the form of a software functional module, etc.

[0227] As shown in FIG. 17, the communication apparatus 1700 includes a processor 1710 and an interface circuit 1720. The processor 1710 and the interface circuit 1720 are coupled to each other. It can be understood that the interface circuit 1720 can be a transceiver or an input / output interface. Optionally, the communication apparatus 1700 can further include a memory 1730, used to store instructions executed by the processor 1710 or to store input data required by the processor 1710 to execute instructions or to store data generated after the processor 1710 executes instructions. Optionally, the processor 1710 and the memory 1730 are integrated together.

[0228] When the communication apparatus 1700 is used to implement the method shown in FIG. 7, the processor 1710 is used to implement the functions of the processing unit 1610, and the interface circuit 1720 is used to implement the functions of the transceiving unit 1620.

[0229] When the above communication apparatus is a chip applied to a terminal, the chip implements the functions of the terminal in the above method embodiments. The chip receives information sent by an access network device to the terminal through other modules (such as a radio frequency module or an antenna) in the terminal; or the chip sends information to other modules (such as a radio frequency module or an antenna) in the terminal, and the information is sent by the terminal to the access network device.

[0230] When the above communication apparatus is a module applied to a network device (for example, an access network device), the module implements the functions of the network device in the above method embodiments. The module receives information from other modules (such as a radio frequency module or an antenna) in the network device, and the information is sent by a terminal to the network device; or the module sends information to other modules (such as a radio frequency module or an antenna) in the network device, and the information is sent by the network device to the terminal. The module of the network device here can be a chip of the network device, or a DU or other module. The DU here can be a DU under the O-RAN architecture.

[0231] The embodiment of the present application further provides a communication device, comprising a processor, wherein the processor is configured to implement the functions of the first communication device or the second communication device in FIG. 7. Optionally, the communication device further comprises a memory, wherein the memory is coupled to the processor. The processor is specifically configured to execute the computer program or the instruction stored in the memory, so that the communication device implements the functions of the first communication device or the second communication device in FIG. 7.

[0232] The embodiment of the present application further provides a communication device, comprising a processor and an interface circuit, wherein the interface circuit is configured to receive a signal from another device outside the device and transmit the signal to the processor or send a signal from the processor to another device outside the device, and the processor is configured to implement the functions of the first communication device or the second communication device in FIG. 7 by means of a logic circuit or an execution code instruction.

[0233] The embodiment of the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores instructions, which can also be referred to as a computer program or a computer program code. The instructions are configured to run on a computer, so that the computer implements the functions of the first communication device or the second communication device in FIG. 7 in the method embodiment.

[0234] The embodiment of the present application further provides a computer program product, comprising a computer program or an instruction, wherein the computer program or the instruction is configured to implement the functions of the first communication device or the second communication device in FIG. 7 when running on a computer.

[0235] The embodiment of the present application further provides a chip, comprising a processor, wherein the processor is configured to execute a computer program or an instruction stored in a memory, so as to implement the functions of the first communication device or the second communication device in FIG. 7. Optionally, the chip further comprises a memory, wherein the memory is coupled to the processor.

[0236] The embodiment of the present application further provides a communication system, comprising a first communication device and a second communication device. The first communication device is configured to implement the functions of the first communication device in FIG. 7, and the second communication device is configured to implement the functions of the second communication device in FIG. 7.

[0237] It is to be understood that the processor in the embodiments of the present application can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor can be a microprocessor or any conventional processor.

[0238] The memory in the embodiments of the present application can be a random access memory (RAM), a flash memory, a read-only memory (ROM), a programmable read-only memory (PROM), an erasable PROM (EPROM), an electrically EPROM (EEPROM), a register, a hard disk, a mobile hard disk, a CD-ROM or any other form of storage medium well known in the art.

[0239] The method steps in the embodiments of the present application can be implemented in hardware or in software instructions executable by a processor. The software instructions can be composed of corresponding software modules, which can be stored in a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, a register, a hard disk, a mobile hard disk, a CD-ROM or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor, so that the processor can read information from and write information to the storage medium. The storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an ASIC.

[0240] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer programs or instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are performed. The computer can be a general purpose computer, a special purpose computer, a computer network, a network device, a user equipment or other programmable apparatus. The computer programs or instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer programs or instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center through wired or wireless manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center and the like integrated with one or more available media. The available medium can be a magnetic medium, for example, a floppy disk, a hard disk, a magnetic tape; or an optical medium, for example, a digital video disc; or a semiconductor medium, for example, a solid state disk. The computer readable storage medium can be a volatile or non-volatile storage medium, or can include both volatile and non-volatile storage media.

[0241] In various embodiments of the present application, the terms and / or descriptions of different embodiments are consistent and can be referred to each other if there is no special description and logical conflict, and the technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationship.

Claims

1. A communication method characterized by comprising: The method is applied to a first communication device, comprising: determining a first model feature; sending a first indication, the first indication being used to indicate the first model feature; wherein the first model feature is used to represent a model feature of a first model, and the first model feature comprises at least one of an application scenario of the first model, a structure of the first model, a function of the first model, or a format of the first model.

2. The method of claim 1, wherein, The first model feature further comprises a target performance of the first model.

3. The method according to claim 1 or 2, c h a r a c t e r i z e d in that, Further comprising: sending the first model.

4. The method of any one of claims 1 to 3, wherein, Further comprising: receiving a second indication, the second indication being used to indicate a second model feature, the second model feature being an expected model feature of the first model for a second communication device, and the second model feature comprising at least one of an expected application scenario of the first model, an expected structure of the first model, an expected function of the first model, or an expected format of the first model; determining the first model according to the second model feature.

5. The method of claim 4, wherein, The second model feature is different from the first model feature.

6. The method of claim 1 or 2, wherein, Further comprising: receiving the first model, the first model being determined according to the first model feature.

7. The method of claim 1 or 2, wherein, Further comprising: determining first inference information according to the first model; sending the first inference information.

8. The method of claim 1 or 2, wherein, The first model feature is specifically used to represent a model feature of an updated first model, and the determining of the first model feature comprises: updating the first model; determining the first model feature according to the updated first model.

9. The method of any one of claims 1 to 3, wherein, The type of the first model feature comprises a fixed model feature of the first model and / or a variable model feature of the first model, and the fixed model feature of the first model and / or the variable model feature of the first model are used to determine a second model based on the first model.

10. A communication method characterized by comprising: The method is applied to a second communication device, comprising: receiving a first indication, the first indication being used to indicate a first model feature; wherein the first model feature is used to represent a model feature of a first model, and the first model feature comprises at least one of an application scenario of the first model, a structure of the first model, a function of the first model, or a format of the first model.

11. The method of claim 10, wherein, The first model feature further comprises a target performance of the first model.

12. The method of claim 10 or 11, wherein, Further comprising: receiving the first model; aggregating the first model according to the first model feature.

13. The method of any one of claims 10 to 12, wherein, Further comprising: sending a second indication, the second indication being used to indicate a second model feature, the second model feature being used to determine the first model, and the second model feature being an expected model feature of the first model for the second communication device, and the second model feature comprising at least one of an expected application scenario of the first model, an expected structure of the first model, an expected function of the first model, or an expected format of the first model.

14. The method of claim 13, wherein, The second model feature is different from the first model feature.

15. The method of claim 10 or 11, wherein, Further comprising: determining the first model according to the first model feature; sending the first model.

16. The method of claim 10 or 11, wherein, Further comprising: determining a second model according to the first model feature; receive first inference information, the first inference information being determined according to the first model; determine second inference information according to the second model and the first inference information.

17. The method of claim 10 or 11, wherein, The first model feature is specifically used to represent the model feature of the updated first model, and further comprises: update the model feature of the first model saved locally by the second communication device according to the first model feature.

18. The method of claim 10 or 11, wherein, The type of the first model feature comprises a fixed model feature of the first model and / or a variable model feature of the first model, and further comprises: receive the first model; determine a second model according to the first model, and the fixed model feature of the first model and / or the variable model feature of the first model.

19. A communications device, characterized by comprise units for implementing the method according to any one of claims 1 to 9; or comprise units for implementing the method according to any one of claims 10 to 18.

20. A communications device, characterized by comprise a processor configured to cause the communication device to perform the method according to any one of claims 1 to 9, or configured to cause the communication device to perform the method according to any one of claims 10 to 18.

21. A computer-readable storage medium, characterized in that, The computer readable storage medium has stored instructions which, when executed, cause the communication device to perform the method according to any one of claims 1 to 9, or the method according to any one of claims 10 to 18.

22. A computer program product, characterised in that, The computer program product comprises instructions which, when executed, cause the communication device to perform the method according to any one of claims 1 to 9, or the method according to any one of claims 10 to 18.

Images