Communication method and communication device

Abstract

The invention provides a communication method and a communication device, relates to the field of communication, and helps to improve the matching performance between associated models when the associated models are set in different devices. The method comprises the following steps: receiving first information from second equipment; wherein the first information is used for indicating the first equipment to be used for training a part of a first model structure of the first model, and the part of the first model structure comprises a model structure of an output layer; or the first information is used for indicating the dimension of the output data of the first model or the dimension of the input data of the second model; a first model structure is determined based on the first information.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the field of communication, and in particular to a communication method and a communication apparatus. BACKGROUND

[0002] At present, artificial intelligence (AI) models are widely used in the field of communication. For example, in some application scenarios, associated models can be set in different electronic devices to process information transmitted between electronic devices, thereby improving the communication efficiency or communication quality between electronic devices. The associated models can be understood as the output data of one model being the input data of another model. For example, model 1 can be set in one electronic device, and model 2 can be set in another electronic device, the output data of model 1 being the input data of model 2, and so on.

[0003] However, the associated models can not be matched, so that the associated models cannot be jointly used to process the information transmitted between devices, thereby affecting the communication efficiency or communication quality between devices. SUMMARY

[0004] The present application provides a communication method and a communication apparatus. The second device can indicate to the first device the part of the first model structure or the dimension of the output data of the first model used by the first device to train the first model, so that the output layer of the first model in the first device matches the input layer of the second model in the second device, or the dimension of the first model output data in the first device is the same as the dimension of the second model input data in the second device, which is beneficial to improve the matching of the associated models.

[0005] In a first aspect, the present application provides a communication method, which comprises: receiving first information from a second device; wherein the first information is used to indicate part of a first model structure used by a first device to train a first model, the part of the first model structure comprising a model structure of an output layer; or the first information is used to indicate a dimension of first model output data or a dimension of second model input data; determining a first model structure based on the first information.

[0006] In one possible implementation, the method is performed by a first communication apparatus. The first communication apparatus can be a first device or a chip or circuit applicable to the first device. The first device can be a terminal device or a network device, for example.

[0007] In one case, the first device is a terminal device, and the second device is a network device. In another case, the first device is a network device, and the second device is a terminal device.

[0008] The communication method of the present application, the second device indicates the model structure of the first model output layer to the first device, so that the model structure of the output layer adopted by the first model trained by the first device can be the same as the model structure of the input layer adopted by the second model trained by the second device; or, the second device indicates the dimension of the first model output data or the dimension of the second model input data to the first device, so that the dimension of the first model output data trained by the first device can be the same as the dimension of the second model input data trained by the second device. In this way, the output layer of the first model trained by the first device matches the input layer of the second model trained by the second device, so that the first model and the second model are matched. Further, the second device can process the output data of the first model based on the second model. It helps to make the communication between the first device and the second device more efficient or the communication quality better.

[0009] In combination with the first aspect, in some embodiments of the first aspect, the part of the first model structure is the model structure of the last n network layers in the first model, and n is a positive integer.

[0010] Since the output data of the first model is the input data of the second model, the effect of the second device processing the data based on the second model is related to the structure and form of the output data of the first model, and the structure and form of the first model output data are usually related to the last n network layers. Therefore, the second device indicates the model structure of the last n network layers to the first device, which helps to match the first model and the second model, so that the effect of the second device processing the output data of the first model based on the second model can be better.

[0011] In combination with the first aspect, in some embodiments of the first aspect, the model structure of the last n network layers includes the type of each network layer in the last n network layers and / or the number of neurons included in each network layer in the last n network layers.

[0012] In this way, the type of the second model input layer can be the same as the type of the first model output output layer, and / or the number of neurons included in the second model input layer can be the same as the number of neurons included in the first model output layer. So that the matching of the first model and the second model is higher.

[0013] In combination with the first aspect, in some embodiments of the first aspect, the method further includes: receiving second information from the second device; wherein the second information is used to indicate all or part of the model structure of the second model, and all or part of the model structure of the second model is used to train the first model; or, the second information is used to indicate all or part of the model structure of the second model and the model parameters of the second model, and all or part of the model structure of the second model and the model parameters of the second model are used to train the first model.

[0014] In this way, the first device can determine the second model based on all or part of the model structure of the second model, or the first device can determine the second model based on all or part of the model structure of the second model and the model parameters of the second model. Thus, the first device can train the first model based on the second model.

[0015] With reference to the first aspect, in some embodiments of the first aspect, the method further includes receiving information from the second device, the information being used to indicate that the first device performs model training alone.

[0016] In this way, the first device can perform model training based on the first model structure by itself.

[0017] With reference to the first aspect, in some embodiments of the first aspect, the method further includes sending third information to the second device, the third information being used to indicate the model structure supported by the first device and / or the dimension of the output data; wherein the part of the first model structure is determined by the second device based on the third information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the second device based on the third information.

[0018] In this way, the second device can determine the first information based on the third information, so that the part of the first model structure indicated by the second device to the first device is the model structure supported by the first device, or the dimension of the first model output data or the dimension of the second model input data indicated by the second device to the first device is the dimension of the output data supported by the first device, so that the first device has a higher efficiency in training the first model based on the first model structure.

[0019] With reference to the first aspect, in some embodiments of the first aspect, the method further includes sending fourth information to the second device, the fourth information being used to indicate that the first device supports performing model training, or the fourth information being used to indicate that the first device does not support performing model training.

[0020] In this way, the second device can determine whether the first device can perform model training.

[0021] With reference to the first aspect, in some embodiments of the first aspect, the fourth information is used to indicate that the first device does not support performing model training, and the fourth information is further used to indicate a reason and / or a first time length for which the first device does not support performing model training.

[0022] In this way, the second device can determine the reason for which the first device does not support model training, and / or the second device can determine the time length for which the first device does not support performing model training.

[0023] With reference to the first aspect, in some embodiments of the first aspect, the method further includes training the first model based on the first model structure.

[0024] In this way, the model structure of the output layer of the first model trained by the first device can be the same as the model structure of the input layer of the second model, or the dimension of the output data of the first model trained by the first device can be the same as the dimension of the input data of the second model. This makes the matching of the first model and the second model higher.

[0025] With reference to the first aspect, in some embodiments of the first aspect, the first model is a model for encoding, and the second model is a model for decoding.

[0026] In this way, the second device can better decode the encoded data output by the first model based on the second model.

[0027] In a second aspect, the present application provides another communication method, which includes determining first information, wherein the first information is used to indicate part of a first model structure of a first model used by a first device to train the first model, the part of the first model structure including a model structure of an output layer; or the first information is used to indicate a dimension of output data of the first model or a dimension of input data of a second model; and sending the first information to the first device.

[0028] In a possible implementation, the method is performed by a second communication apparatus. The second communication apparatus can be the second device or a chip or circuit applicable to the second device.

[0029] With reference to the second aspect, in some embodiments of the second aspect, the part of the first model structure is a model structure of n last network layers in the first model, n being a positive integer.

[0030] With reference to the second aspect, in some embodiments of the second aspect, the model structure of the n last network layers includes a type of each network layer in the n last network layers and / or a number of neurons included in each network layer in the n last network layers.

[0031] With reference to the second aspect, in some embodiments of the second aspect, the method further includes sending second information to the first device, wherein the second information is used to indicate all or part of a model structure of the second model, the all or part of the model structure of the second model being used to train the first model; or the second information is used to indicate the all or part of the model structure of the second model and a model parameter of the second model, the all or part of the model structure of the second model and the model parameter of the second model being used to train the first model.

[0032] With reference to the second aspect, in some embodiments of the second aspect, the method further includes sending information used to indicate that the first device performs model training alone to the first device.

[0033] With reference to the second aspect, in some embodiments of the second aspect, the method further includes: receiving third information from the first device, the third information being used to indicate a model structure supported by the first device and / or a dimension of output data of the first device; wherein the partial first model structure is determined by the second device based on the third information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the second device based on the third information.

[0034] With reference to the second aspect, in some embodiments of the second aspect, the method further includes: receiving fourth information from the first device, the fourth information being used to indicate that the first device supports model training, or the fourth information being used to indicate that the first device does not support model training.

[0035] With reference to the second aspect, in some embodiments of the second aspect, the fourth information is used to indicate that the first device does not support model training, and the fourth information is further used to indicate a reason and / or a first time length for which the first device does not support model training.

[0036] With reference to the second aspect, in some embodiments of the second aspect, the first model is a model for encoding, and the second model is a model for decoding.

[0037] The third aspect provides yet another communication method, which includes: determining fifth information; wherein the fifth information is used to indicate a partial second model structure of a second model used by the second device to train the second model, the partial second model structure including a model structure of an input layer; or the fifth information is used to indicate a dimension of first model output data or a dimension of second model input data; and sending the fifth information to the second device.

[0038] In a possible implementation, the method is performed by a first communication apparatus. The first communication apparatus can be the first device or can be a chip or circuitry applied to the first device, etc.

[0039] With reference to the third aspect, in some embodiments of the third aspect, the partial second model structure is a model structure of a first m-layer network layer in the second model, m being a positive integer.

[0040] With reference to the third aspect, in some embodiments of the third aspect, the model structure of the first m-layer network layer includes: a type of each network layer in the first m-layer network layer and / or a number of neurons included in each network layer in the first m-layer network layer.

[0041] With reference to the third aspect, in some embodiments of the third aspect, the method further includes: sending, to the second device, sixth information; wherein the sixth information is used to indicate all or part of the model structure of the first model, and the all or part of the model structure of the first model is used to train the second model; or the sixth information is used to indicate all or part of the model structure of the first model and the model parameters of the first model, and the all or part of the model structure of the first model and the model parameters of the first model are used to train the second model.

[0042] With reference to the third aspect, in some embodiments of the third aspect, the method further includes: sending, to the second device, information used to instruct the second device to perform model training alone.

[0043] With reference to the third aspect, in some embodiments of the third aspect, the method further includes: receiving seventh information from the second device, the seventh information being used to indicate the model structure supported by the second device and / or the dimension of the input data; wherein the part of the second model structure is determined by the first device based on the seventh information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the first device based on the seventh information.

[0044] With reference to the third aspect, in some embodiments of the third aspect, the method further includes: receiving eighth information from the second device, the eighth information being used to indicate that the second device supports performing model training, or the eighth information being used to indicate that the second device does not support performing model training.

[0045] With reference to the third aspect, in some embodiments of the third aspect, the eighth information is used to indicate that the second device does not support performing model training, and the eighth information is further used to indicate one or more of the following: the reason why the second device does not support performing model training, the first time point, the second time length or the second time point, the first time point being the time point at which the first device requests the second device to perform model training again, the second time length being the duration during which the first device is prohibited from requesting the second device to perform model training, and the second time point being the time point at which the second device can start training the second model.

[0046] With reference to the third aspect, in some embodiments of the third aspect, the first model is a model used for encoding, and the second model is a model used for decoding.

[0047] In a fourth aspect, the present application provides another communication method, which includes: receiving fifth information from a first device; wherein the fifth information is used to indicate part of a second model structure of a second device used to train the second model, and the part of the second model structure includes the model structure of an input layer; or the fifth information is used to indicate the dimension of first model output data or the dimension of second model input data; and determining the second model structure based on the fifth information.

[0048] In a possible implementation, the method is performed by the second communication apparatus. The second communication apparatus can be a second device or a chip or circuit applied to a second device.

[0049] With reference to the fourth aspect, in some embodiments of the fourth aspect, the partial second model structure is a model structure of the first m network layers in the second model, m being a positive integer.

[0050] With reference to the fourth aspect, in some embodiments of the fourth aspect, the model structure of the first m network layers comprises: a type of each of the first m network layers and / or a number of neurons included in each of the first m network layers.

[0051] With reference to the fourth aspect, in some embodiments of the fourth aspect, the method further comprises: receiving sixth information from the first device, wherein the sixth information is used to indicate the whole or partial model structure of the first model, and the whole or partial model structure of the first model is used to train the second model; or the sixth information is used to indicate the whole or partial model structure of the first model and the model parameters of the first model, and the whole or partial model structure of the first model and the model parameters of the first model are used to train the second model.

[0052] With reference to the fourth aspect, in some embodiments of the fourth aspect, the method further comprises: receiving information from the first device, wherein the information is used to indicate that the second device performs model training alone.

[0053] With reference to the fourth aspect, in some embodiments of the fourth aspect, the method further comprises: sending seventh information to the first device, wherein the seventh information is used to indicate the model structure supported by the second device and / or the dimension of the input data; and wherein the partial second model structure is determined by the first device based on the seventh information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the first device based on the seventh information.

[0054] With reference to the fourth aspect, in some embodiments of the fourth aspect, the method further comprises: sending eighth information to the first device, wherein the eighth information is used to indicate that the second device supports model training, or the eighth information is used to indicate that the second device does not support model training.

[0055] With reference to the fourth aspect, in some embodiments of the fourth aspect, the eighth information is used to indicate that the second device does not support model training, and the eighth information is further used to indicate one or more of the following: a reason why the second device does not support model training, a first time point, a second time length or a second time point, the first time point being a time point at which the first device requests the second device to perform model training again, the second time length being a continuous time length during which the first device is prohibited from requesting the second device to perform model training, and the second time point being a time point at which the second device can start training the second model.

[0056] In some embodiments of the fourth aspect, the method further comprises training the second model based on the second model structure.

[0057] In some embodiments of the fourth aspect, the first model is a model for encoding and the second model is a model for decoding.

[0058] In some embodiments of the fourth aspect, the first model is a model for encoding and the second model is a model for decoding.

[0059] In a fifth aspect, a communication apparatus is provided, which is configured to execute the method in any possible implementation of the first aspect, the second aspect, the third aspect or the fourth aspect. Specifically, the communication apparatus comprises modules for executing the method in any possible implementation of the first aspect, the second aspect, the third aspect or the fourth aspect.

[0060] In a sixth aspect, another communication apparatus is provided, which comprises a processor and a memory coupled to the processor. The processor is configured to execute instructions in the memory to implement the method in any possible implementation of the first aspect, the second aspect, the third aspect or the fourth aspect. Optionally, the apparatus further comprises the memory. Optionally, the apparatus further comprises a communication interface, and the processor is coupled to the communication interface.

[0061] In a seventh aspect, a processor is provided, which comprises an input circuit, an output circuit and a processing circuit. The processing circuit is configured to receive a signal through the input circuit and transmit a signal through the output circuit, so that the processor executes the method in any possible implementation of the first aspect, the second aspect, the third aspect or the fourth aspect.

[0062] In a specific implementation process, the processor can be a chip, the input circuit can be an input pin, the output circuit can be an output pin, and the processing circuit can be a transistor, a gate circuit, a flip-flop and various logic circuits, etc. The input signal received by the input circuit can be received and input by, for example but not limited to, a receiver, the signal output by the output circuit can be output to and transmitted by, for example but not limited to, a transmitter, and the input circuit and the output circuit can be the same circuit which is used as the input circuit and the output circuit at different times. The embodiments of the present application do not limit the specific implementation of the processor and various circuits.

[0063] Optionally, the processor is one or more, and the memory is one or more.

[0064] Optionally, the memory can be integrated with the processor, or the memory and the processor are separately arranged.

[0065] In the implementation process, the memory can be a non-transitory memory, for example, a read only memory (ROM), which can be integrated with the processor on the same chip, or arranged separately on different chips. The embodiments of the present application do not limit the type of memory and the arrangement of the memory and the processor.

[0066] It should be understood that the related data interaction process, such as sending indication information, can be the process of outputting indication information from the processor, and receiving capability information can be the process of receiving input capability information by the processor. Specifically, the processed output data can be output to the transmitter, and the input data received by the processor can come from the receiver. Wherein, the transmitter and the receiver can be collectively referred to as the transceiver.

[0067] The processing device in the eighth aspect described above can be a chip, and the processor can be implemented by hardware or software. When implemented by hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented by software, the processor can be a general-purpose processor, which realizes by reading software code stored in the memory. The memory can be integrated in the processor, or located outside the processor and exist independently.

[0068] In the ninth aspect, a computer program product is provided, which includes a computer program (also referred to as code or instruction), which, when executed, causes a computer to execute the method in any possible implementation manner of the first aspect, the second aspect, the third aspect or the fourth aspect.

[0069] In the tenth aspect, a computer readable storage medium is provided, which stores a computer program (also referred to as code or instruction), which, when executed on a computer, causes the computer to execute the method in any possible implementation manner of the first aspect, the second aspect, the third aspect or the fourth aspect. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 is a schematic diagram of a neural network;

[0071] Figure 2 is a schematic diagram of a communication system suitable for the embodiments of the present application;

[0072] Figure 3 is a schematic diagram of a process in which a terminal device reports CSI;

[0073] Figure 4A flowchart of a communication method provided by an embodiment of the present application;

[0074] Figure 5 A process diagram of a model training method provided by an embodiment of the present application;

[0075] Figure 6 A flowchart of another communication method provided by an embodiment of the present application;

[0076] Figure 7 A diagram of another communication system applicable to an embodiment of the present application;

[0077] Figure 8 A flowchart of yet another communication method provided by an embodiment of the present application;

[0078] Figure 9 A flowchart of still another communication method provided by an embodiment of the present application;

[0079] Figure 10 A schematic block diagram of a communication device provided by an embodiment of the present application;

[0080] Figure 11 A schematic block diagram of another communication device provided by an embodiment of the present application;

[0081] Figure 12 A schematic block diagram of a network element function division and protocol layer structure diagram of an O-RAN device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical solutions of the present application more clear and intuitive, the communication method and communication device of the embodiments of the present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described in the present application are only used to explain the present application and should not be used to limit the present application.

[0083] Before introducing the method and device provided by the embodiments of the present application, the following points are explained.

[0084] First, in the embodiments shown below, each term and English abbreviation, such as reference data or differential data, is an exemplary example given for convenience of description, and should not constitute any limitation on the present application. The present application does not exclude the possibility of defining other terms capable of achieving the same or similar functions in existing or future protocols.

[0085] Second, in the embodiments shown below, the first, second and various numerical numbers are only for differentiation for convenience of description, and do not limit the scope of the embodiments of the present application.

[0086] Third, "at least one" means one or more, "multiple" means two or more. "And / or" describes the association relationship of the associated objects, which means that there can be three kinds of 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. The character " / " generally represents that the associated objects before and after are in an "or" relationship. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b and c can mean: a, or b, or c, or a and b, or a and c, or b and c, or a, b and c, where a, b and c can be single or multiple.

[0087] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant national and regional laws, regulations and standards, and provide corresponding operation portal for user to choose authorization or refusal.

[0088] The technical solutions of the embodiments of the present application can be applied to various communication systems, for example: long term evolution (long term evolution, LTE) system, LTE frequency division duplex (frequency division duplex, FDD) system, LTE time division duplex (time division duplex, TDD), universal mobile communication system (universal mobile telecommunication system, UMTS), 5th generation (5th generation, 5G) system or new radio (new radio, NR), or other future communication systems, etc.

[0089] The terminal device in the embodiments of the present application can also be referred to as: user equipment (user equipment, UE), mobile station (mobile station, MS), mobile terminal (mobile terminal, MT), access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent or user device, etc.

[0090] The terminal device can be a device that provides voice / data connectivity to a user, for example, a handheld device with wireless connection function, a vehicle-mounted device, etc. At present, some examples of terminals are: mobile phones, tablet computers, notebook computers, palm computers, mobile internet devices (MIDs), wearable devices, virtual reality (VR) devices, augmented reality (AR) devices, wireless terminals in industrial control, wireless terminals in self driving, wireless terminals in remote medical surgery, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, wireless terminals in smart homes, cellular phones, cordless phones, session initiation protocol (SIP) phones, wireless local loop (WLL) stations, personal digital assistants (PDAs), handheld devices with wireless communication function, computing devices or other processing devices connected to wireless modems, vehicle-mounted devices, wearable devices, terminal devices in 5G networks, or terminal devices in future evolved public land mobile networks (PLMNs), etc. The embodiments of the present application are not limited thereto.

[0091] By way of example and not limitation, in the embodiments of the present application, the terminal device can also be a wearable device. The wearable device can also be referred to as a wearable smart device, which is a general term for devices that are designed and developed by applying wearable technology to daily wear, such as glasses, gloves, watches, clothing, and shoes. The wearable device is a portable device that is directly worn on the body or integrated into the user's clothes or accessories. The wearable device is not only a hardware device, but also a device that realizes powerful functions through software support and data interaction and cloud interaction. The general wearable smart device includes devices with full functions, large sizes, and the ability to realize complete or partial functions without relying on smart phones, such as smart watches or smart glasses, and devices that focus on a certain type of application function and need to be used in conjunction with other devices such as smart phones, such as various smart wristbands and smart jewelry for monitoring vital signs.

[0092] In addition, in the embodiments of the present application, the terminal device can also be a terminal device in an internet of things (IoT) system. The IoT is an important part of future information technology development, and its main technical feature is to connect objects through communication technology and network, so as to realize the intelligent network of man-machine interconnection and object-object interconnection. The terminal device of the present application can also be a vehicle-mounted unit, a vehicle-mounted module, a vehicle-mounted component, a vehicle-mounted chip or a vehicle-mounted unit built in as one or more components or units in a vehicle. The vehicle can implement the method of the present application through the built-in vehicle-mounted unit, vehicle-mounted module, vehicle-mounted component, vehicle-mounted chip or vehicle-mounted unit. Therefore, the embodiments of the present application can be applied to the Internet of Vehicles, such as vehicle to everything (V2X), long term evolution-vehicle (LTE-V), vehicle-to-vehicle (V2V), etc.

[0093] In addition, the access network device in the embodiments of the present application can also be referred to as a wireless access network device, which can be a transmission reception point (TRP), and can also be an evolved NodeB (eNB or eNodeB) in an LTE system, and can also be a home base station (such as a home evolved NodeB or home NodeB, HNB), a baseband unit (BBU), and can also be a wireless controller in a cloud radio access network (CRAN) scenario, or the access network device can be a relay station, an access point, a vehicle-mounted device, a wearable device, an access network device in a 5G network, or an access network device in a future evolved PLMN network, etc., which can be an access point (AP) in a WLAN, can be a gNB in a new radio (NR) system, can be a satellite base station in a satellite communication system, etc. The embodiments of the present application are not limited.

[0094] The access network device in the Ben embodiment can include a centralized unit (CU) node, or a distributed unit (DU) node, or an access network device including a CU node and a DU node, or an access network device of a control plane CU node (CU-CP node) and a user plane CU node (CU-UP node) and a DU node. The access network device including the CU node and the DU node can split the protocol layers of the access network device, and the functions of part of the protocol layers are controlled by the CU, and the functions of the remaining part or all of the protocol layers are distributed in the DU and controlled by the CU.

[0095] As an implementation manner, the CU deploys a radio resource control (RRC) layer, a packet data convergence protocol (PDCP) layer, and a service data adaptation protocol (SDAP) layer in the protocol stack. The DU deploys a radio link control (RLC) layer, a media access control (MAC) layer, and a physical layer (PHY) layer in the protocol stack. Therefore, the CU has the processing capability of RRC, PDCP, and SDAP. The DU has the processing capability of RLC, MAC, and PHY.

[0096] The above-mentioned function splitting is only an example and does not constitute a limitation on the CU and the DU. That is, there can be other function splitting manners between the CU and the DU, which are not described herein. The functions of the CU can be implemented by one entity or by different entities. For example, the functions of the CU can be further split, for example, the control plane (CP) and the user plane (UP) are separated, that is, the control plane of the CU (CU-CP) and the user plane of the CU (CU-UP). For example, the CU-CP and the CU-UP can be implemented by different functional entities, and the CU-CP and the CU-UP can be coupled with the DU to jointly complete the functions of the access network device.

[0097] In a possible mode, the CU-CP is responsible for the control plane function, mainly including the RRC and the PDCP-C, where the PDCP-C is mainly responsible for the encryption and decryption of the control plane data, the integrity protection, the data transmission, and the like. The CU-UP is responsible for the user plane function, mainly including the SDAP and the PDCP-U, where the SDAP is mainly responsible for processing the data of the core network device and mapping the data flow to a bearer. The PDCP-U is mainly responsible for the encryption and decryption of the data plane, the integrity protection, the header compression, the sequence number maintenance, the data transmission, and the like. The CU-CP and the CU-UP are connected through an E1 interface. The CU-CP represents the access network device connected through the interface between the core network device and the access network device and the core network device. The CU-CP is connected with the DU through an F1-C (control plane). The CU-UP is connected with the DU through an F1-U (user plane). In addition, there is also a possible implementation that the PDCP-C is also in the CU-UP, which is not limited in the application.

[0098] The core network device in the embodiments of the application refers to a device in a core network (CN) that provides service support for a terminal device. At present, the core network device can be an access and mobility management function (AMF) entity, a session management function (SMF) entity, a user plane function (UPF) entity, and the like, which are not listed one by one here. Among them, the AMF entity can be responsible for access management and mobility management of the terminal device; the SMF entity can be responsible for session management, such as session establishment of a user; and the UPF entity can be a functional entity of the user plane, mainly responsible for connecting external networks. It should be noted that the entity in the application can also be referred to as a network element or a functional entity, for example, the AMF entity can also be referred to as an AMF network element or an AMF functional entity, and for example, the SMF entity can also be referred to as an SMF network element or an SMF functional entity, and the like, which are not limited in the application.

[0099] First, some technical terms and symbols involved in the application are introduced.

[0100] 1. Artificial intelligence (AI) is a technology proposed in the 1950s to perform complex calculations by simulating the human brain. Through artificial intelligence technology, a machine can have human intelligence. For example, a machine can use the software and hardware of a computer to simulate certain intelligent behavior of a human being. In order to realize artificial intelligence, a machine learning method or other methods can be adopted, which are not limited.

[0101] 2. A neural network (NN), also known as a model, intelligent model, or neural network model, is a specific implementation of machine learning. A neural network is a computational model that simulates the structure and function of the human brain, composed of multiple layers of nodes (or "neurons"). These nodes are connected by weights and can process and analyze complex data. For example, according to the universal approximation theorem, a neural network can theoretically approximate any continuous function, thus enabling it to learn arbitrary mappings. Therefore, neural networks can accurately abstract and model complex, high-dimensional problems.

[0102] Neural networks have a wide range of applications in artificial intelligence, such as image recognition, speech recognition, natural language processing, and communication systems. Common types of neural networks include feedforward neural networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), generative adversarial networks (GANs), Transformer models, and long short-term memory (LSTM) models.

[0103] Neural networks can run on electronic devices, efficiently enabling artificial intelligence functions through hardware acceleration (such as GPUs or dedicated AI accelerator chips) and optimized software algorithms.

[0104] 3. Network Layers: Neural networks generally consist of multiple layers, which can be referred to as multiple network layers. A network layer can be understood as a basic building block in a neural network. Each network layer may contain one or more neurons. These neurons are connected to neurons in the preceding and following layers via weights. The main function of a network layer is to receive input data, perform specific calculations or transformations, and pass the results to the next layer. The following are some common types of network layers: input layer, hidden layer, and output layer.

[0105] It should be noted that, in the embodiments of the present application, two continuous network layers (or two adjacent network layers) can be understood as that the two network layers are connected by neurons and there is no other network layer between the two neurons. In addition, the front and back order of the network layers is described in the order from input to output. That is, the input layer is the first network layer, and the output layer is the last network layer. For example, in two network layers, the former layer represents a network layer closer to the input layer, and the latter layer represents a network layer closer to the output layer. In addition, for the s-th network layer, the smaller the s is, the closer the network layer is to the input layer, for example, the first layer is the input layer, and the second layer is the layer after the input layer; the larger the s is, the farther the network layer is from the input layer, for example, in the third network layer and the second network layer, the second network layer is closer to the input layer, and s is a positive integer. For the sake of brevity, this will not be described again hereinafter.

[0106] In one implementation, the neural network can include an input layer and an output layer. The input layer of the neural network transmits the result of the neuron processing of the received input to the output layer, and the output layer obtains the output result of the neural network.

[0107] In another implementation, the neural network includes an input layer, a hidden layer and an output layer. The input layer of the neural network transmits the result of the neuron processing of the received input to the intermediate hidden layer, the hidden layer transmits 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. For example, as shown in FIG. 1, the neural network includes an input layer, a hidden layer and an output layer. The hidden layer includes two network layers. Each of the input layer, the hidden layer and the output layer includes a plurality of neurons. Figure 1

[0108] 4. Neuron, the basic computing unit of the neural network. Each neuron performs a weighted sum operation on its input values and generates an output through an activation function. The main components of a neuron include: input weights, that is, each input data has a corresponding weight, indicating its importance in the calculation; bias, that is, an additional parameter for adjusting the value of the output; and activation function, that is, a nonlinear function for introducing nonlinear characteristics, so that the neural network can handle complex patterns.

[0109] ​Activation functions play a crucial role in each neuron of the hidden and output layers. Common activation functions include the Sigmoid function, Tanh function, rectified linear unit (ReLU) function, Softmax function, and leaky ReLU, among others.

[0110] It should be understood that, where the weights can also be referred to as weight matrices, the biases can also be referred to as bias vectors, and the like, which are not specifically limited in the present application.

[0111] 5、The input layer is the first layer of the neural network and is used to receive external input data. The number of neurons in the input layer is usually the same as the number of features of the input data.

[0112] 6、The hidden layer is located between the input layer and the output layer. The neural network can include one or more hidden layers. The neurons in the hidden layer perform a non-linear transformation on the input data through an activation function, thereby capturing complex patterns and features.

[0113] 7、The output layer is the last layer of the neural network and is responsible for outputting the final prediction result.

[0114] Common types of network layers can include, but are not limited to, fully connected layers (dense layers), convolutional layers, pooling layers, and recurrent layers, among others. That is, the input layer, hidden layer, or output layer can be any type of network layer described above.

[0115] 8、The fully connected layer (dense layer), each neuron in the fully connected layer is connected to all neurons in the previous layer.

[0116] 9、The convolutional layer (convolutional layer), commonly used for processing image data, extracts local features through convolution operations.

[0117] 10、The pooling layer (pooling layer), commonly used to reduce data dimensionality and computational complexity while preserving important features.

[0118] 11、The recurrent layer (recurrent layer), commonly used for processing sequence data such as time series or natural language processing tasks, capable of capturing temporal dependencies.

[0119] 12、Loss function, during the training process of a model, a loss function can be defined. The loss function describes the gap or difference between the output value of the model and the ideal target value. The specific form of the loss function is not limited in the present disclosure. The training process of the model is a process of adjusting part or all of the parameters (model parameters) of the model so that the value of the loss function is less than a threshold value or meets the target requirement. For example, the model is a neural network, and one or more of the following parameters can be adjusted during the training process: the number of layers of the neural network, the width of the neural network, the connection relationship between layers, the weight of neurons, the activation function of neurons, or the bias in the activation function, so that the difference between the output of the neural network and the ideal target value is as small as possible.

[0120] 13、Model parameters, are variables learned through optimization algorithms (such as gradient descent) during the training process, which determine the behavior and performance of the model. The specific meaning and form of the model parameters depend on the type and structure of the model. Model parameters may include, but are not limited to, neuron weights, neuron biases, neuron activation functions, or inter-layer connection relationships.

[0121] 14、Encoder, mainly used for encoding the original channel status information (CSI) to reduce data volume and improve transmission efficiency. The encoder can use various encoding techniques (such as linear predictive coding, vector quantization, etc.) to compress the CSI.

[0122] The encoded (compressed) CSI is easier to transmit and has lower bandwidth requirements.

[0123] 15、Quantizer, which can be used to discretize continuous CSI, i.e., map continuous signal values to a finite set of discrete values. The quantization process introduces a certain quantization error, but can reduce the data volume.

[0124] It can be understood that the quantizer is usually designed according to the characteristics of the channel and the requirements of the system to minimize the data volume while ensuring the accuracy of the CSI.

[0125] 16、Decoder, mainly used for restoring the received encoded data to the original data, which can be understood as the inverse process of the encoding of the CSI by the encoder. In the context of CSI, the decoder decodes the encoded CSI into quantized CSI data.

[0126] 17. The dequantizer is mainly used to restore quantized discrete data to a form as close as possible to the original continuous data. In the context of CSI, the dequantizer converts the quantized CSI back to continuous CSI. This process can be understood as the inverse process of discretizing continuous CSI.

[0127] 18. The dimension of the input data, also known as the input dimension, can be used to describe the shape and structure of the data received by the model.

[0128] 19. The dimension of the output data, also known as the output dimension, can be used to describe the shape and structure of the data output by the model.

[0129] 20. Line of sight (LOS) refers to the distance between the transmitter and receiver where the signal propagates in a straight line without any obstructions.

[0130] Under LOS conditions, the signal propagation path is the shortest and the path loss is the least, therefore the signal strength is stronger.

[0131] 21. Non-line of sight (NLOS) refers to a situation where there are obstacles in the propagation path of a signal between the transmitter and receiver. The signal needs to bypass the obstacles through reflection, refraction, scattering, etc., to reach the receiver.

[0132] Under NLOS conditions, the signal propagation path is longer and the path loss is greater, resulting in weaker signal strength.

[0133] It should be understood that LOS can also be called within line of sight, and NLOS can also be called outside line of sight; this application does not make any specific limitation in this regard.

[0134] 22. User Equipment Assistance Information (UAI): Terminal devices can proactively inform network devices of their internal status through UAI messages, which helps network devices to understand the status of terminal devices in a timely manner.

[0135] 23. Minimization Drive Test (MDT): An automated drive test technology that uses network-configured terminal devices to collect, report, and preprocess measurement data. The terminal devices enable the Global Positioning System (GPS) and support MDT functionality, automatically reporting MDT data containing user location information to the network devices. MDT data can be used for big data analytics.

[0136] In the embodiments of the present application, each term and English abbreviation, such as AI, NN, CNN, UAI, and the like, are exemplary examples given for the convenience of description, and should not constitute any limitation on the present application. The present application does not exclude the possibility of defining other terms capable of achieving the same or similar functions in existing or future protocols.

[0137] In order to facilitate the understanding of the embodiments of the present application, the following will be described in conjunction with Figure 2 The communication system applicable to the embodiments of the present application will be described in detail.

[0138] Figure 2 A schematic diagram of the communication system 200 applicable to the embodiments of the present application is shown in Figure 2 The communication system 200 can include at least one network device, such as the network device 210 shown in Figure 2 The communication system 200 can also include at least one terminal device, such as the terminal device 220 shown in Figure 2 The network device 210 and the terminal device 220 can communicate through a wireless link. In one possible case, the network device 210 can act as a transmitting end, and the terminal device 220 can act as a receiving end, and the network device 210 transmits signals to the terminal device 220. In another possible case, the network device 210 can act as a receiving end, and the terminal device 220 can act as a transmitting end, and the terminal device 220 transmits signals to the network device 210.

[0139] Figure 2 One network device 210 and one terminal device 220 are exemplarily shown. Alternatively, the communication system 200 can also include multiple network devices and / or multiple terminal devices. The network device 210 can be a router, a base station, or the like, and the terminal device 220 can be a mobile phone, a tablet computer, a smart bracelet, or the like, and the embodiments of the present application do not limit this.

[0140] Each of the above communication devices, such as the network device 210 or the terminal device 220 in Figure 2 , can be configured with multiple antennas. The multiple antennas can include at least one transmitting antenna for transmitting signals and at least one receiving antenna for receiving signals. In addition, each communication device additionally includes a transmitter chain and a receiver chain, and those skilled in the art can understand that they can include multiple components (such as processors, modulators, multiplexers, demodulators, demultiplexers, or antennas, etc.) related to signal transmission and reception. Therefore, the network device 210 and the terminal device 220 can communicate through multiple antenna technology.

[0141] Optionally, the above communication system 200 can also include a network controller, a mobile management entity, and other network entities, and the embodiments of the present application are not limited thereto.

[0142] It should also be understood that the method provided in the embodiments of this application can be applied to a variety of communication systems, including 5G new radio (NR) systems. Communication system 200 is only an example. This application does not limit the specific architecture of the applicable system, nor does it limit the number and form of various devices contained in each communication system.

[0143] Currently, artificial intelligence (AI) is widely used in the communications field to improve network performance and user experience. For example, AI models can be applied to network communication (NR) systems, intelligently collecting and analyzing data to enhance network performance and user experience. An AI model can be understood as a model used to implement AI functions. AI models can be configured in network devices or terminal devices.

[0144] The following is combined Figure 3 This paper provides a detailed explanation of the application architecture of AI models in NR.

[0145] Figure 3 This is a schematic diagram illustrating an application architecture of an AI model in NR, provided as an embodiment of this application. Figure 3 As shown, the data source module stores input data from gNB, gNB-CU, gNB-DU, UE, or other management entities, serving as a database for AI model training and data analysis inference. The model training module analyzes the training data provided by the data source module to produce the optimal AI model (the trained AI model); the training data can be part or all of the data in the database. Based on the inference data provided by the data source module, the model inference module can use the optimal AI model to make reasonable predictions about network operation and guide network devices to make policy adjustments. Policy adjustments can be uniformly planned by the actor module entities and sent to multiple network entities for execution. Simultaneously, after applying the adjusted policies, the specific performance data of the network entities, such as operational data, will be input back into the database for storage.

[0146] It should be understood that an AI model can be a complete AI model that can run independently on a terminal device or network device (without restrictions on its format and compilation environment); it can also be a functional module that runs on a terminal device or network device and can be used to obtain specific outputs or inputs; or it can even be other non-AI related functional modules. This application does not impose any specific restrictions on this.

[0147] In the field of communication, different terminal devices can be configured to perform different functions. For example, a terminal device can be configured to perform channel measurement or MDT measurement, etc. In order to implement various different functions, the terminal device can set an AI model for implementing various functions. For example, the terminal device can be configured to perform channel measurement, and the terminal device can be provided with a model for implementing the channel measurement function, etc. The AI model provided in the terminal device can be configured by the network device, or can be determined (or selected) by the terminal device from AI models respectively used to implement various different functions.

[0148] In different application environments, the terminal device can select different AI models. The application environment can include, but is not limited to, one or more of the following: an application related to the macro physical properties of the terminal device, such as the speed, moving direction or geographic location of the terminal device; an application related to the software and hardware properties of the terminal device, such as the effective power, computing power, storage space or supported AI model compilation environment of the terminal device; an application related to the channel environment in which the terminal device is located, such as an urban macro (urban macro) environment, an urban micro (urban micro) environment or an indoor hotspot (indoor hotspot); an application related to the communication configuration of the terminal device and the network device, such as the number of receive antennas of the terminal device or the number of transmit antenna ports of the network device; or an application related to the time-frequency domain resources of air interface communication, such as carrier frequency, subcarrier spacing or bandwidth, etc.

[0149] The specific application of AI model in NR is illustrated below.

[0150] 1. The AI model can be applied to energy saving.

[0151] For example, the network device can obtain information such as the load, energy consumption and energy efficiency of the current cell and the neighboring cell, and can also obtain information such as the trajectory and measurement results of the terminal device, so as to predict the network device load by using the information and the AI model. In addition, the network device can combine the cell purpose and / or key performance indicator (KPI) requirements, etc., and take appropriate energy saving measures in a timely manner without affecting network coverage and user access. For example, the energy saving strategy can include, but is not limited to, one or more of the following: deactivating the cell, turning off the carrier, turning off the channel, turning off the time slot or reducing the transmission power, etc.

[0152] However, when the network coverage is affected or cannot meet the UE access demand and / or service demand, the network device needs to update the current energy saving strategy, or directly restore to the normal working state (i.e. not using the energy saving strategy). In addition, the network device can re-predict the network device load, for example, the network device can update the AI model used for predicting the network device load, so as to re-predict the network device load based on the updated AI model.

[0153] 2. The AI model can be applied in load balancing.

[0154] For example, the network device can collect the load, energy consumption and energy efficiency of the current cell and neighboring cells, and the trajectory and measurement results of the terminal device, and use the information and the AI model to predict the network device load. In addition, the network device can determine and implement a load balancing strategy in combination with the AI model and the cell usage and / or KPI requirements. The load balancing strategy is, for example, to select part of the UEs, i.e. to switch the part of the UEs to the neighboring cells, or to receive the part of the UEs from the neighboring cells. In this way, the load balancing between the network devices in the whole slice network can be achieved, and the situation that part of the network devices are overloaded while part of the network devices are idle can be reduced.

[0155] However, in some cases, for example, the selected part of the UEs is unreasonable or the target cell of the UE switching is unreasonable, resulting in UE cell switching failure or UE service being affected; or the AI model is inaccurate in predicting the network device load, resulting in poor load balancing effect; or temporary load abnormal change occurs, resulting in the original load balancing strategy no longer applicable. At this time, it is necessary to exit the use of the current load balancing strategy or modify the current load balancing strategy, and the network device can re-predict the network device load, for example, the network device can update the AI model used for predicting the network device load, so as to re-predict the network device load based on the updated AI model.

[0156] 3. The AI model can be applied in mobility optimization.

[0157] For example, the network device inputs the historical trajectory information of the UE and the measurement information of the UE into the AI model, and outputs the future trajectory of the UE, so as to realize the prediction of the future trajectory of the UE. Based on the predicted trajectory, the network device can determine whether the UE performs cell switching. In addition, in the case where the UE performs cell switching, the network device can issue a switching configuration in advance, and can instruct the target cell to prepare access resources. In this way, the delay of the UE in the switching process can be reduced, and the probability of switching failure and access failure can also be reduced.

[0158] However, when the UE future trajectory output by the AI model is wrong, it will cause UE handover failure, resulting in UE service interruption. In this case, the network device can retrain the AI model or replace the AI model, so as to improve the accuracy of predicting the future trajectory of the UE.

[0159] 4. The AI model can be applied in positioning accuracy enhancements.

[0160] Exemplarily, the network device can obtain original data from a reference UE set (or controlled) by the operator. The original data can include, but is not limited to, one or more of the following: GPS coordinates, signal strength, time stamp, etc. The original data can come from multiple reference UEs in different locations, so as to cover data of different geographical areas and environmental conditions.

[0161] The location management node (LMF, non-RAN side node) and the network device can respectively perform AI model training based on the original data. Assuming that the AI model trained by the LMF is AI model 1, the LMF can predict the positioning (longitude and latitude, etc.) of the UE based on the AI model 1, and the input data of the AI model 1 can be, for example, the signal strength and / or signal arrival time reported by the UE, etc. Assuming that the AI model trained by the network device is AI model 2, the network device can determine whether the LOS condition or the NLOS condition between the sending end and the receiving end based on the AI model 2, and the input data of the AI model 2 can be, for example, the signal strength reported by the UE, etc.

[0162] 5. The AI model can be applied in channel state information reference signal (CSI-RS) feedback enhancement.

[0163] Exemplarily, as shown in Figure 3 , a trained encoder and quantizer can be set in the terminal device. The encoder and quantizer can be trained by the network device based on UE capability information and network device requirements and instructed to the terminal device.

[0164] The terminal device can perform channel measurement based on the reference signal (such as CSI-RS) from the network device to obtain initial CSI; the terminal device performs discretization on the initial CSI based on the quantizer to obtain discrete CSI; and the terminal device performs encoding (or compression) processing on the discrete CSI based on the encoder to obtain encoded CSI; and the terminal device sends the encoded CSI to the network device.

[0165] Network devices can be equipped with decoders and dequantizers. The network device decodes (or decompresses) the encoded CSI from the terminal device based on the decoder to obtain discrete CSI; and processes the discrete CSI based on the dequantizer to obtain the recovered initial CSI.

[0166] It is understandable that the recovered initial CSI can be very close to the original initial CSI, and there can be a certain degree of error between them. Furthermore, the closer the recovered initial CSI is to the original initial CSI, the better the effect of the terminal device and network device transmitting the CSI in the above manner, which helps to improve the communication quality between the terminal device and the network device.

[0167] The functions of the encoder and / or quantizer, and the functions of the decoder and / or dequantizer, can be implemented using AI models. For example, the terminal device may have a pre-trained AI model for discretizing CSI, which can implement the function of the quantizer described above; and the network device may have a pre-trained AI model for de-discretizing the discrete CSI, which can implement the function of the dequantizer described above. And / or, the terminal device may have a pre-trained AI model for encoding CSI, which can implement the function of the encoder described above; and the network device may have a pre-trained AI model for decoding the encoded CSI, which can implement the function of the decoder described above.

[0168] Among them, such as Figure 3 As shown, the output data of the AI ​​model that implements the encoder function is the input data of the AI ​​model that implements the decoder function. For ease of description, multiple models with this relationship will be referred to as related models in the following text.

[0169] It should be understood that the associated models can be two models, such as an AI model that implements the encoder function and an AI model that implements the decoder function; or, the associated models can be more than one model, such as three AI models, for example, model a, model b, and model c, where the output data of model a and the output data of model b can be the input data of model c. These three models can be understood as associated models. This application does not limit the number of associated models.

[0170] It should also be understood that in related models, the output data of some models is the input data of others. For ease of description, these models will be referred to as the first part models and the others as the second part models. The output data of the first part models is the input data of the second part models. For example, the AI ​​model that implements the encoder function can be understood as the first part model, and the AI ​​model that implements the decoder function can be understood as the second part model.

[0171] It should be noted that the names defined for ease of description in the embodiments of this application do not constitute a limitation on the embodiments of this application. The defined names are all examples, and these names can be replaced with others.

[0172] In some scenarios, associated models set up on different devices may exhibit mismatches. This mismatch can also be understood as a mismatch between the output layer of the first model and the input layer of the second model. This mismatch can mean that the output data of the first model cannot be successfully used as input data for the second model. For example, the dimensions of the output data of the first model may differ from the dimensions of the input data of the second model, resulting in lower accuracy or failure for the second model in processing the input data.

[0173] Below, we will use the first part of the model as an AI model that implements the encoder function and the second part of the model as an AI model that implements the decoder function as an example to illustrate the situation of mismatch between related models.

[0174] For ease of description, the AI ​​model that implements the encoder function will be referred to as the compression model, and the AI ​​model that implements the decoder function will be referred to as the decompression model.

[0175] It is understood that the compression model can be a neural network, or it can also be called an encoder, decoder, or encoding model, etc.; the decompression model can also be a neural network, or it can also be called a decoder, decoder, or decoding model, etc. This application does not specifically limit the type and name of the compression model and decompression model.

[0176] like Figure 3 As shown, the output data of the compression model set in the terminal device is the input data of the decompression model set in the network device. If the output layer of the compression model and the input layer of the decompression model do not match, for example, the type of the output layer of the compression model is different from the type of the input layer of the decompression model, and / or, the dimension of the output data of the compression model is different from the dimension of the input data of the decompression model, it may lead to a low success rate of the network device decoding the encoded CSI based on the decompression model.

[0177] To address the above issues, the protocol can currently define a model structure 1 for the first part of the model and a model structure 2 for the second part of the model, so that the first part of the model trained based on model structure 1 matches the second part of the model trained based on model structure 2.

[0178] Taking the first part model as an AI model for implementing an encoder function and the second part model as an AI model for implementing a decoder function as an example, in a case where the model structure 1 and the model structure 2 agreed by the protocol match, the compression model and the decompression model trained based on the model structure 1 and the model structure 2 match. In this way, the form and structure of the encoded CSI output by the compression model are the same as the form and structure of the identifiable input data of the decompression model, so that the success rate of the network device in decoding the encoded CSI output by the compression model based on the decompression model is relatively high.

[0179] However, in different application scenarios, the model structure of the applicable first part model and / or the model structure of the applicable second part model may change. For example, in the application scenario 1, the model structure of the applicable first part model is the first model structure; in the application scenario 2, the model structure of the applicable first part model is the second model structure, and so on. It is difficult for the protocol to agree on the model structure applicable in all application scenarios. Moreover, with the development of communication technology, the model structure of the first part model and / or the second part model agreed by the protocol may no longer be applicable. In this way, after the first part model is trained based on the model structure of the first part model agreed by the protocol, and the second part model is trained based on the model structure of the second part model agreed by the protocol, the accuracy and / or success rate of the data processed by the first part model and the second part model is relatively low, for example, the difference between the recovered initial CSI obtained by the network device based on the decompression model and the initial CSI is relatively large, resulting in a relatively poor channel measurement effect. In addition, the model structure agreed by the protocol may limit the performance of the first part model and the second part model, so that in some application scenarios, the effect of the data processed by each device based on the first part model or the second part model is relatively poor.

[0180] Therefore, the application provides a communication method. In the case that output data of a first model is input data of a second model, in order to match an output layer of the first model trained by a first device with an input layer of the second model trained by a second device, the second device can indicate a partial model structure of the first model or a dimension of the first model output data to the first device, the partial model structure of the first model including a model structure of the output layer. In this way, the model structure of the input layer adopted by the second model trained by the second model can be the same as the model structure of the output layer adopted by the first model, or the dimension of the second model input data can be the same as the dimension of the first model output data. The first model and the second model are matched. Furthermore, the second device can process the output data of the first model based on the second model, which helps to make the processing effect of the first model output data by the second device based on the second model better. For example, assuming that the second model is a decompression model and the first model is a compression model, the success rate of the decompression model in decoding the encoded CSI is higher, and it helps to make the recovered initial CSI closer to the initial CSI.

[0181] In addition, with the development of communication technology, in different application scenarios, in order to meet the communication needs of the first device and the second device, the partial model structure of the first model or the dimension of the first model output data indicated by the second device to the first device can be different. In addition to the partial model structure of the first model indicated by the second device, the remaining partial model structure of the first model can be determined by the first device according to the needs. As can be seen, the communication method of the application helps to meet the communication needs in different scenarios on the premise of matching the first model and the second model. In addition, the first device also has a certain autonomy, so that the first device can determine the remaining partial model structure of the first model according to its own needs. Such a method has less limitation on the performance of the first model and the second model.

[0182] In one case, the first device can be a terminal device, and the second device can be a network device. The output data of the first model trained by the terminal device is the input data of the second model trained by the network device. The first model is, for example, the compression model in the above, and the second model is, for example, the decompression model in the above, etc.

[0183] In another case, the first device can be a network device, and the second device can be a terminal device. The output data of the first model trained by the network device is the input data of the second model trained by the terminal device. For example, the network device processes the initial downlink information based on the first model to output the processed downlink information, and the terminal device processes the processed downlink information based on the second model to obtain the initial downlink information, etc.

[0184] For the sake of brevity, the device types of the first device and the second device will not be described in detail below.

[0185] In addition to the above-mentioned communication method, similarly, the application also provides another communication method. In order to match the output layer of the first model trained by the first device with the input layer of the second model trained by the second device, the first device can also indicate the partial model structure of the second model or the dimension of the second model input data to the second device, the partial model structure of the second model including the model structure of the input layer. In this way, the model structure of the input layer adopted by the second model trained by the second device can be the same as the model structure of the output layer adopted by the first model trained by the first device, or the dimension of the second model input data can be the same as the dimension of the first model output data. So that the first model and the second model are matched.

[0186] It should be understood that the other communication method is similar to the effect of the above-mentioned communication method, and can refer to the description above, which will not be repeated here.

[0187] The communication method of the application will be described in detail below. Figures 4 to 12 The communication method of the application will be described in detail below.

[0188] It should be understood that the first device can be the first device itself, or a chip, chip system or processor supporting the first device to implement the communication method, or a logic module or software capable of implementing all or part of the first device; the second device can be the second device itself, or a chip, chip system or processor supporting the second device to implement the communication method, or a logic module or software capable of implementing all or part of the second device, which is not limited by the application.

[0189] Figure 4 The flowchart of the communication method 400 provided by the embodiment of the application is shown. The method 400 is applicable to the system 200, and the method 400 includes the following steps:

[0190] S401, the second device determines the second model structure, and / or the second device trains the second model.

[0191] The second model structure is a model structure of the second model. In S401, the second device can determine the second model structure, but the second model is not trained completely; or the second device trains the second model; or the second device determines the second model structure and trains the second model based on the second model structure. For example, the second device determines the second model structure before training the second model, and then trains the second model based on the second model structure. The second model structure can be determined by the second device according to an application scenario, for example, according to the type of a task that the second device needs to process based on the second model.

[0192] In this way, the second device can determine the first information based on the second model structure or the second model. For example, the model structure of the first model output layer is determined based on the model structure of the second model structure or the input layer of the second model, and the model structure of the first model output layer is the model structure of the input layer of the second model structure or the second model; or the dimension of the first model output data is determined based on the dimension of the second model structure or the second model input data, and the dimension of the first model output data is the dimension of the second model structure or the second model input data. So that the first model trained by the first device matches the second model.

[0193] It should be noted that S401 is optional. That is, before the second device sends the first information to the first device, the second device can also not determine the second model structure and not train the second model. Or, S401 can be replaced by the second device determining the dimension of the second model input data, etc.

[0194] It should be understood that in the embodiments of the present application, the model structure can be understood as part other than the model parameters. The model structure may, for example, but not limited to, include the type of each network layer, the positional relationship of the network layers, or the number of neurons included in each network layer, etc. The positional relationship of the network layers can also be understood as the front and back positional relationship of the network layers. For example, the network layer closer to the input layer is closer to the front, and the network layer closer to the output layer is closer to the back, etc. The model structure can also be referred to as a structure, etc. The embodiments of the present application do not make specific limitations thereto.

[0195] In S402, the second device sends the first information to the first device. The first information is used to indicate part of the first model structure used by the first device to train the first model, and the part of the first model structure includes the model structure of the output layer; or the first information is used to indicate the dimension of the first model output data or the dimension of the second model input data. Correspondingly, the first device receives the first information from the second device.

[0196] The input data of the second model is the output data of the first model. The first model and the second model are associated models. The first model or the second model can be a feedforward neural network, a CNN, an RNN, a GAN, a Transformer model, or an LSTM model.

[0197] The first model structure is a model structure of the first model. The partial first model structure represents part of the first model structure. For example, the entire first model structure includes X layers of network layers, where X is an integer greater than or equal to 2. The partial first model structure is a model structure of v layers of network layers in the X layers of network layers, where v is a positive integer less than X. The v layers of network layers include the output layer of the first model. When v is 1, the partial first model structure is a model structure of the output layer of the first model.

[0198] It should be understood that, for the convenience of description, the output layer of the first model is referred to as the first output layer hereinafter. The input layer of the second model is referred to as the first input layer. This will not be described again hereinafter.

[0199] In addition, the v layers of network layers can be consecutive v layers of network layers, such as the i-th layer of network layer, the i+1-th layer of network layer, …, the i+v-1-th layer of network layer, and the like, where i is a positive integer. The v layers of network layers can also be completely discontinuous network layers, that is, any two network layers are not adjacent. Alternatively, the v layers of network layers include partially continuous network layers and partially discontinuous network layers, and the like. For example, the v layers of network layers include the j-th layer of network layer, the j+1-th layer of network layer, the j+2-th layer of network layer, the j+5-th layer of network layer, and the j+7-th layer of network layer, where the j-th layer of network layer, the j+1-th layer of network layer, and the j+2-th layer of network layer are three consecutive layers of network layers, and the j+5-th layer of network layer and the j+7-th layer of network layer are discontinuous network layers, where j is a positive integer.

[0200] It should be understood that, regardless of which of the above cases the v layers of network layers are, the v layers of network layers include the first output layer. For the other network layers in the v layers of network layers other than the first output layer, the present embodiments do not limit the positions and quantities of the other network layers.

[0201] In one possible case, before S402, the second device performs S401. S402 can be implemented, for example, by the following possible implementation.

[0202] In the first possible implementation, the first information is used to indicate the partial first model structure.

[0203] The partial first model structure includes a model structure of the first output layer, that is, includes the dimension of the first model output data. The dimension of the first model output data is the same as the dimension of the second model input data.

[0204] In this possible implementation, the partial first model structure can be determined by the second device according to the second model structure or the second model. For example, the second device can determine the model structure of the first output layer based on the model structure of the first input layer, so as to match the first input layer with the first output layer. Alternatively, since the dimension of the first model output data is the same as the dimension of the second model input data, the second model structure includes the dimension of the second model input data, the second device can determine the dimension of the first model output data based on the dimension of the second model input data, and then determine the partial first model structure based on the dimension of the first model output data. The partial first model structure includes the dimension of the first model output data which is the same as the dimension of the second model input data included in the second model structure.

[0205] It can be understood that the partial first model structure includes the model structure of the first output layer, and the second model structure includes the model structure of the first input layer. For the model structure of the first output layer and the model structure of the first input layer, the dimension of the output data included in the model structure of the first output layer (i.e., the dimension of the first model output data) is the same as the dimension of the input data included in the model structure of the first input layer (i.e., the dimension of the second model input data). The content in the model structure of the first output layer other than the dimension of the output data can be the same as or different from the content in the model structure of the first input layer other than the dimension of the input data, or can be partially the same. The embodiments of the present application do not make specific limitations on this.

[0206] In the second possible implementation, the first information is used to indicate the dimension of the first model output data or the dimension of the second model input data.

[0207] The dimension of the first model output data is the same as the dimension of the second model input data. Therefore, when the first information is used to indicate the dimension of the first model output data, the first device can determine the dimension of the first model output data. Alternatively, when the first information is used to indicate the dimension of the second model input data, since the dimension of the first model output data is the same as the dimension of the second model input data, the first device can determine the dimension of the first model output data based on the dimension of the second model input data.

[0208] In this possible implementation, the first information can be determined by the second device based on the dimension of the second model input data, and the second model structure includes the dimension of the second model input data. For example, the second device can determine the dimension of the output data of the first model based on the dimension of the second model input data, etc.

[0209] In another possible case, S401 is optional content, i.e., before S402, the second device does not perform S401. Then S402 can be implemented, for example, through the following possible implementation.

[0210] In a third possible implementation, the first information is used to indicate a partial first model structure.

[0211] In this possible implementation, the partial first model structure can be determined by the second device according to an application scenario. For example, the second device can determine the partial first model structure based on a load condition of the second device and / or a compression rate required by a service, etc.

[0212] Alternatively, in some possible implementations, S401 can also be replaced by the second device determining a dimension of the second model input data.

[0213] Since the dimension of the second model input data is the same as that of the first model output data, the second device can determine the dimension of the first model output data based on the dimension of the second model input data. For the remaining content of the partial first model other than the dimension of the first model output data, the second device can determine according to an application scenario. For example, the second device can determine the remaining content based on a load condition of the second device and / or a compression rate required by a service, etc.

[0214] And, optionally, S401 can be executed after S402, and then the second model structure determined by the second device is determined based on the partial first model structure. Since the dimension of the second model input data is the same as that of the first model output data, the second device can determine the dimension of the second model input data. Further, the second model structure is determined based on the dimension of the second model input data. Then, the second device can train the second model based on the second model structure.

[0215] Optionally, the model structure of the first input layer can be the same as that of the first output layer. In this way, the first input layer and the first output layer are completely matched, so that the second device can successfully process the output data of the first model based on the second model. Alternatively, the dimension of the second model input data included in the second model structure is the same as that of the first model output data. The content of the second model structure other than the dimension of the second model input data can be determined by the second device based on a load condition of the second device and / or a compression rate required by a service, etc.

[0216] In a fourth possible implementation, the first information is used to indicate a dimension of the first model output data.

[0217] In this possible implementation, the dimension of the first model output data can be determined by the second device according to an application scenario. For example, the second device can determine the dimension of the first model output data based on a load condition of the second device and / or a compression rate required by a service, etc.

[0218] Alternatively, in some possible implementation, S401 can also be replaced by determining, by the second device, the dimension of the second model input data.

[0219] In this way, the second device can indicate the dimension of the second model structure input data to the first device through the first information; or the second device determines the dimension of the first model output data based on the dimension of the second model structure input data, and indicates the dimension of the first model output data to the first device through the first information.

[0220] The dimension of the first model output data is the same as the dimension of the second model input data. Therefore, S401 can be performed after S402, for example, the second device trains the second model after S402, and the dimension of the second model input data is the same as the dimension of the first model output data.

[0221] S403, the first device determines the first model structure based on the first information.

[0222] The first model structure can be understood as the entire model structure of the first model. That is, the first device determines the entire first model structure based on the partial first model structure. Alternatively, the first device determines the entire first model structure based on the dimension of the first model output data. In this way, the first output layer can match the first input layer, or the dimension of the output data of the first model structure is the same as the dimension of the input data of the second model structure.

[0223] Subsequently, the method 400 further comprises S404, the first device trains the first model based on the first model structure.

[0224] Since the first model is trained based on the first model structure, the model structure of the first model trained by the first device is the first model structure.

[0225] The communication method of the present application, the second device indicates the model structure of the output layer of the first model to the first device, so that the model structure of the output layer of the first model trained by the first device can be the same as the model structure of the input layer of the second model trained by the second device; or, the second device indicates the dimension of the first model output data or the dimension of the second model input data to the first device, so that the dimension of the first model output data trained by the first device can be the same as the dimension of the second model input data trained by the second device. In this way, the output layer of the first model trained by the first device matches the input layer of the second model trained by the second device, so that the first model and the second model are matched. Further, the second device can process the first model output data based on the second model. The first device and the second device can jointly process the information transmitted between the first device and the second device based on the first model and the second model respectively, so that the communication efficiency between the first device and the second device is higher or the communication quality is better.

[0226] In addition, part of the first model structure or the dimension of the first model output data (or replaced by the dimension of the second model input data) can be determined by the second device according to the application scenario, so that the first model and the second model are suitable for the current application scenario; for example, the second device determines part of the first model structure or the dimension of the first model output data (or replaced by the dimension of the second model input data) based on the load condition of the second device and / or the compression rate of the service requirement, so that the first model trained by the first device and the second model trained by the second device can meet the compression rate of the service requirement and can adapt to the current load condition of the second device. In addition, with the development of communication technology, in order to meet different needs, the second device can indicate different part of the first model structure or the dimension of the first model output data (or replaced by the dimension of the second model input data).

[0227] In addition, since the second device does not indicate all the first model structure to the first device, the first device can also determine the remaining part of the first model structure autonomously. In this way, the first model structure constructed by the first device is suitable for the application scenario where the first device is located, which facilitates to meet the needs of the first device.

[0228] Next, part of the first model structure is described in detail.

[0229] As an optional embodiment, part of the first model structure is the model structure of the last n network layers in the first model, and n is a positive integer.

[0230] Among them, the last n network layers include the output layer of the first model, that is, the first output layer, and the last n network layers are the n network layers closest to the first output layer.

[0231] In an example, n can be 1, and the n last network layers can be the first output layer. In another example, n can be 2, and the n last network layers can be the first output layer and a hidden layer closest to the first output layer.

[0232] Since the output data of the first model is the input data of the second model, the effect of the second device processing the data based on the second model is related to the structure and form of the output data of the first model, and the structure and form of the output data of the first model are usually related to the n last network layers. Therefore, the second device helps to match the first model and the second model by indicating the model structure of the n last network layers to the first device, so that the effect of the second device processing the output data of the first model based on the second model can be better. For example, the first model is a compression model, and the second model is a decompression model. The higher the matching of the first model and the second model, the closer the initial CSI recovered by the second device based on the decompression model to the initial CSI determined by the first device.

[0233] On the basis of the above-mentioned embodiments, optionally, the model structure of the n last network layers comprises: a type of each network layer in the n last network layers and / or a number of neurons included in each network layer in the n last network layers.

[0234] The type of the network layer may, for example, be one of a fully connected layer, a convolutional layer, a pooling layer, or a recurrent layer. The number of neurons can be any positive integer. The neurons are related to the type of the first model, for example, the first model can be a CNN, and each network layer in the n last network layers includes neurons in the CNN; the first model can be a Transformer model, and each network layer in the n last network layers includes neurons in the Transformer model; or, the first model can be an LSTM model, and each network layer in the n last network layers includes neurons in the LSTM model, and the like.

[0235] In an example, n can be 1, and the n last network layers can be the first output layer. The model structure of the first output layer (part of the model structure of the first model) comprises: a type of the first output layer and / or a number of neurons included in the first output layer. The type of the first output layer may, for example, be a fully connected layer, and the number of neurons included in the first output layer may, for example, be ten neurons, and the like. The dimension of the output data of the first model may, for example, be 1x10 or 10x1, and the like.

[0236] In order to make the input layer (the first input layer) of the second model more matched with the first output layer, the model structure of the first output layer can comprise the type of the first output layer and the number of neurons included in the first output layer. In this way, the type of the first input layer can be the same as the type of the first output layer, and the number of neurons included in the first input layer can be the same as the number of neurons included in the first output layer.

[0237] In another example, n can be 2. The subsequent n network layers consist of a first output layer and a hidden layer closest to the first output layer. For ease of description, this hidden layer is referred to as the first hidden layer. The model structure of the subsequent n network layers (partial first model structure) includes: the type of the first output layer and / or the number of neurons in the first output layer, as well as the type of the first hidden layer and the number of neurons in the first hidden layer. The type of the first output layer can be, for example, a fully connected layer, and the number of neurons in the first output layer can be, for example, ten neurons. The dimension of the first model output data is, for example, 1×10 or 10×1. The first hidden layer can, for example, include 15 neurons.

[0238] To better match the first input layer with the first output layer, when n is greater than 2, the subsequent n network layers include the first output layer and other network layers. The model structure of the subsequent n network layers (partial first model structure) may include: the type of the first output layer and the number of neurons included in that output layer, as well as the type of each network layer in the other network layers and / or the number of neurons included in each network layer in the other network layers.

[0239] It is understandable that the value of n can be larger, the type of each network layer can be other types, and the number of neurons included in each network layer can also be other numbers. For the sake of brevity, they will not be listed here.

[0240] The following is a detailed explanation of S404, which is the method by which the first device trains the first model.

[0241] Since the first model and the second model are related models, S404 can be implemented in the following way: the first device trains the first model based on the first model structure and the second model structure.

[0242] For example, such as Figure 5 As shown, the model structure of the first model is the first model structure determined by the first device, and the model structure of the second model is the second model structure. The first device inputs initial data into the first model, and the first model outputs processed data; the first device inputs the processed data into the second model, and the second model outputs the restored initial data.

[0243] The closer the recovered initial data is to the original data, the more ideal the data processing effect of the first and second models is. For example, if the first model is a compression model and the second model is a decompression model, the initial data can be the initial CSI, the processed data is the encoded CSI, and the recovered initial data is the decompressed CSI of the second model, which is the recovered initial CSI of the second model.

[0244] It should be understood that the initial data can be data in a dataset. The dataset is a dataset used for training the first model and / or the second model. The dataset can be sent by the second device to the first device, so that the first device training the first model and the second device training the second model can use the same dataset.

[0245] For example, the method 400 further includes that the second device sends information 1 to the first device, the information 1 being used to indicate the dataset. Correspondingly, the first device receives the information 1 from the second device.

[0246] It should be understood that the information 1 and the first information can be carried in the same signaling or in different signaling. Moreover, when the information 1 and the first information are carried in the same signaling, the information 1 and the first information can be carried in the same or different fields of the signaling, which is not limited in the present application.

[0247] Alternatively, the dataset can also be acquired by the first device in other manners. For brevity, they are not listed one by one here.

[0248] Then, the first device can calculate a loss factor based on the initial data and the recovered initial data. The loss factor can be used to represent the training effect of the first model. When the loss factor meets a preset condition, the first device can determine that the first model is trained. When the loss factor does not meet the preset condition, the first device can update the model parameters of the first model based on the loss factor to obtain an updated first model. Then, the first device can continue to input the initial data into the updated first model and repeatedly execute the above operations until it is determined that the loss factor meets the preset condition.

[0249] It should be noted that for the second model structure used in the process of training the first model by the first device, the second model structure can be a model structure constructed by the first device itself. For example, the first device determines the model structure of the first input layer based on the part of the first model structure indicated by the first information, that is, the model structure of the first input layer is the same as the model structure of the first output layer included in the part of the first model structure. Or, the first device determines the dimension of the second model input data based on the dimension of the first model output data or the dimension of the second model input data indicated by the first information.

[0250] Alternatively, the second model structure can be indicated by the second device through signaling, which is as follows.

[0251] As an optional embodiment, the method 400 further includes: the second device sending second information to the first device. The second information is used to indicate all or part of the model structure of the second model, and the all or part of the model structure of the second model is used to train the first model; or the second information is used to indicate all or part of the model structure of the second model and the model parameters of the second model, and the all or part of the model structure of the second model and the model parameters of the second model are used to train the first model. Correspondingly, the first device receives the second information from the second device.

[0252] In this way, when the second information indicates all the model structure of the second model, i.e., all the second model structure, the first device can obtain all the second model structure. When the second information indicates part of the model structure of the second model, i.e., part of the second model structure, the first device can obtain part of the second model structure, and for the remaining part of the second model structure, the first device can construct it by itself according to the application scenario. For example, the second model structure includes Y layers of network layers, Y is an integer greater than or equal to 2, and the second information can indicate h layers of network layers in the Y layers of network layers, h is a positive integer less than Y. The first device can construct the remaining Y-h layers of network layers in the Y layers of network layers by itself.

[0253] In addition, when the second information indicates all or part of the model structure of the second model, but does not indicate the model parameters of the second model, the model parameters of the second model can be determined by the first device. In this way, the data amount of the second information is small, so that the signaling overhead is small.

[0254] When the second information indicates all or part of the model structure of the second model, and indicates the model parameters of the second model, Figure 5 The model parameters of the second model indicated by the second information are the model parameters of the second model indicated by the second information, and the model structure of the second model is constructed based on all or part of the model structure indicated by the second information. For example, the second device can indicate all or part of the model structure of the second model and the model parameters of the second model to the first device through the second information after training the second model. In this way, the first device can determine the second model based on all or part of the model structure of the second model and the model parameters of the second model indicated by the second information, and the second model determined by the first device is close to the second model trained by the second device. The first device can train the first model based on the second model, which helps to make the trained first model and the second model trained by the second device better matched.

[0255] It should be noted that the second information and the first information can be carried in the same signaling or in different signaling. Moreover, when the second information and the first information are carried in the same signaling, the second information and the first information can be carried in the same field or different fields of the signaling, which is not limited in the present application.

[0256] Alternatively, in some possible implementation, the second information can also be understood as the first information. That is, the first information in S402 can also be replaced by the second information. In this case, the model structure of the second model indicated by the second information includes the model structure of the first input layer, and the model structure of the first input layer is the same as the model structure of the first output layer. Therefore, the second information can also be understood as the model structure of the first output layer indicated by the second device to the first device.

[0257] On the basis of the above embodiment, the method 400 further includes that the second device sends information for indicating that the first device performs model training alone to the first device. Correspondingly, the first device receives the information for indicating that the first device performs model training alone from the second device.

[0258] The model training alone can also be understood as that the first device can construct the first model structure and the second model structure by itself, and perform model training based on the first model structure and the second model structure. The second model structure can be constructed by the first device based on the first information, or can be constructed by the first device based on the second information. The process of model training performed by the first device can refer to the process shown in Figure 5 , which is not described herein again.

[0259] It should be understood that the information for indicating that the first device performs model training alone and the second information can be carried in the same or different signaling, and the information for indicating that the first device performs model training alone and the first information can also be carried in the same or different signaling. The present application does not make a specific limitation in this regard.

[0260] On the basis of the above embodiment, the second device can also indicate the model parameters of the first model to the first device, so that the first device can train the first model based on the model parameters of the first model.

[0261] Exemplarily, the method 400 further includes that the second device sends information 2 for indicating the model parameters of the first model to the first device. Correspondingly, the first device receives the information 2 from the second device.

[0262] It should be understood that the information 2 and the first information can be carried in the same signaling or in different signaling. Moreover, when the information 2 and the first information are carried in the same signaling, the information 2 and the first information can be carried in the same field or different fields of the signaling, which is not limited in the present application.

[0263] It should be noted that the above are all described by taking the first device training the first model as an example. In some possible implementation, the first model can be trained by other devices, for example, the third device. And the third device can send the trained first model to the first device after training the first model. In this case, the third device trains the first model in a manner similar to the process shown in the above, and the description can be referred to the description above, which will not be repeated here. Figure 5

[0264] And in this case, the first device can send the first information and / or the second information to the third device, so that the third device can construct the first model structure and / or the second model structure based on the first information and / or the second information. The present application does not make specific limitation on this.

[0265] Wherein, the third device can be a server, or other types of devices. The data set used by the third device to train the first model can be sent by the first device to the third device, or can be obtained by the third device in other ways. The present application does not make specific limitation on this.

[0266] In addition, the capability of the first device can affect whether the first device can train the first model. Therefore, the first device can also improve the efficiency of model training in the following manner.

[0267] As an optional embodiment, the method 400 further includes: the first device sends third information to the second device, the third information being used to indicate the model structure supported by the first device and / or the dimension of the output data; wherein, the part of the first model structure is determined by the second device based on the third information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the second device based on the third information. Correspondingly, the second device receives the third information from the first device.

[0268] Wherein, the model structure supported by the first device includes at least one model structure, and the dimension of the output data supported by the first device includes at least one dimension.

[0269] In the case that the third information is used to indicate the model structure supported by the first device, the second device can determine the part of the first model structure based on the model structure supported by the first device.

[0270] ​The model structure supported by the first device can include at least one model structure supported by the first device. Each of the at least one model structure can be a model structure of part of the network layers in the first model structure. For example, when the first model structure includes X network layers, one of the at least one model structure includes a model structure of q network layers, and the q network layers belong to part of the X network layers. The model structure of the q network layers can be the same as or different from the number of network layers included in the part of the first model structure indicated by the first information. For example, the part of the first model structure includes n network layers. The q and the n can be the same or different. The X is an integer greater than or equal to 2, and the q and the n are positive integers less than the X.

[0271] The part of the first model structure indicated by the first information can be determined according to the at least one model structure. For example, the part of the first model structure indicated by the first information can be determined according to one of the at least one model structure.

[0272] Suppose that the part of the first model structure indicated by the first information is determined according to the first one of the at least one model structure, the part of the first model structure can be the same as the first one of the at least one model structure. In this way, the part of the first model structure determined by the second device is the model structure supported by the first device.

[0273] Alternatively, the part of the first model structure indicated by the first information includes the first one of the at least one model structure. For example, the part of the first model structure includes a model structure of n network layers, and the first one of the at least one model structure can include a model structure of q network layers, and the n network layers can include the q network layers. In this way, the second device can autonomously determine a model structure of other network layers (n-q network layers) on the basis of the model structure supported by the first device, so that the part of the first model structure indicated by the second device can also meet the requirements of the second device.

[0274] It should be noted that the number of network layers included in any two of the at least one model structure can be the same or different. For example, the model structure supported by the first device includes a first one of the at least one model structure, a second one of the at least one model structure, and a third one of the at least one model structure; the first one of the at least one model structure includes a model structure of 3 network layers, the second one of the at least one model structure includes a model structure of 1 network layer, and the third one of the at least one model structure includes a model structure of 3 network layers. The number of network layers included in each of the at least one model structure is not limited in the present application.

[0275] Optionally, each of the at least one model structure can include a model structure of an output layer.

[0276] In this way, the second device can determine the model structure of the at least one output layer supported by the first device, so that the second device can determine the model structure of the first output layer based on the at least one model structure. The model structure of the first output layer included in the part of the first model structure indicated by the first information is the model structure of the output layer supported by the first device, so that the probability of the first device successfully training the first model is higher.

[0277] It should be further noted that, in order to match the second model input layer with the first model output layer, the model structure of the second model input layer needs to be the same as the model structure of the first model output layer. Therefore, the second device can also determine the part of the first model structure according to the self requirement and / or the model structure supported by the second device, and the model structure supported by the first device.

[0278] For example, assuming that the model structure supported by the first device includes the model structure of at least one output layer, the second device determines the model structure of the first model output layer as one of the model structure of at least one output layer, which is also the model structure of the second model input layer, which can also be the model structure supported by the second device. For example, the model structure of at least one output layer supported by the first device includes: the model structure of output layer 1, the model structure of output layer 2 and the model structure of output layer 3; assuming that the model structure of the input layer supported by the second device includes: the model structure of input layer 1 and the model structure of input layer 2. If the model structure of the input layer 2 is the same as the model structure of the output layer 1, the second device can determine the model structure of the output layer 1 as the model structure of the first output layer, that is, the model structure of the first input layer.

[0279] In this way, the second model structure can be the model structure supported by the second device, so that the probability of the second device successfully training the second model is higher.

[0280] In the case that the third information is used to indicate that the first device supports the dimension of the output data, the second device can determine the dimension of the first model output data or the dimension of the second model input data based on the dimension of the output data supported by the first device.

[0281] It can be understood that the dimension of the first model output data and the dimension of the second model input data are the same, so that the second device determines the dimension of the first model output data based on the dimension of the output data supported by the first device, that is, the second device determines the dimension of the second model input data based on the dimension of the output data supported by the first device.

[0282] For example, the first device makes the second device determine the dimension of the second model input data by sending the third information to the second device before the second device determines the second model structure, so that the second device can determine the dimension of the output data supported by the first device based on the third information. The dimension of the second model input data determined by the second device can belong to the dimension of the output data supported by the first device. And the dimension of the first model output data or the dimension of the second model input data indicated by the second device to the first device belongs to the dimension of the output data supported by the first device.

[0283] In this way, the dimension of the first model output data or the dimension of the second model input data indicated by the second device is the dimension of the output data supported by the first device, so that the probability of the first device successfully training the first model is higher.

[0284] In mode 2, optionally, the second device can also determine the dimension of the first model output data or the dimension of the second model input data based on the dimension of the output data supported by the first device and the dimension of the input data supported by the second device.

[0285] Since the dimension of the first model output data and the dimension of the second model input data are the same, in order to make the second device successfully train the second model, the dimension of the second model input data needs to be the dimension of the input data supported by the second device. Therefore, through the above mode, not only can the first device support the dimension of the first model output data, but also the second device can support the dimension of the second model input data. It helps to make the probability of the first device successfully training the first model higher, and helps to make the probability of the second device successfully training the second model higher.

[0286] Mode 3, in the case that the third information is used to indicate the model structure supported by the first device, the model structure supported by the first device can include the model structure of the output layer, and the model structure of the output layer can include the dimension of the output data of the output layer. Therefore, the second device can also determine the dimension of the first model output data or the dimension of the second model input data based on the model structure supported by the first device.

[0287] Wherein, the model structure supported by the first device can refer to the description in mode 1.

[0288] In the case that the model structure supported by the first device is at least one model structure, the model structure supported by the first device can include at least one dimension of the output data, and the at least one dimension of the output data is the dimension of the output data supported by the first device. The dimension of the first model output data or the dimension of the second model input data can be one of the at least one dimension of the output data.

[0289] For example, the model structure supported by the first device is at least one model structure, including a first model structure, a second model structure, and a third model structure. The first model structure includes a first dimension of output data, the second model structure includes a second dimension of output data, and the third model structure includes a third dimension of output data. The dimension of the first model output data or the dimension of the second model input data can be the second dimension of output data.

[0290] In the third mode, the second device can determine at least one dimension of output data supported by the first device based on the model structure supported by the first device. Optionally, the second device can further determine the dimension of the first model output data or the dimension of the second model input data based on the at least one dimension of output data and the dimension of input data supported by the second device.

[0291] The dimension of the first model output data and the dimension of the second model input data are the same. The dimension of the first model output data or the dimension of the second model input data determined by the second device can be one of the at least one dimension of output data, and the second device supports the dimension of the second model input data. For example, the at least one dimension of output data supported by the first device includes a first dimension of output data, a second dimension of output data, and a third dimension of output data. The dimension of input data supported by the second device is the same as the first dimension of output data. The second device can determine the first dimension of output data (or the dimension of input data supported by the second device) as the dimension of the first model output data or the dimension of the second model input data.

[0292] In the fourth mode, the third information is used to indicate the case where the first device supports the dimension of output data. The second device can determine part of the first model structure based on the dimension of output data supported by the first device. The dimension of output data of the first model structure belongs to the dimension of output data supported by the first device.

[0293] In the fourth mode, in the part of the first model structure, the content other than the dimension of output data of the first model structure can be determined by the second device. For example, the second device can determine the content other than the dimension of output data of the first model structure according to the application scenario. For example, the part of the first model structure includes a model structure of n layers of network layers. The dimension of the first model output data in the model structure of n layers of network layers is determined based on the dimension of output data supported by the first device. The remaining content in the model structure of n layers of network layers is determined by the second device.

[0294] On the basis of the above embodiments, in the case where the first device is a terminal device and the second device is a network device, the third information can be carried in the UE capability information.

[0295] In this way, the second device can determine the model structure supported by the first device and / or the dimension of the output data supported by the first device based on the UE capability information before determining the first information, which helps the second device to indicate the partial first model structure supported by the first device or the dimension of the output data to the first device.

[0296] Optionally, the third information can also be a response to the information 3. For example, the method 400 further includes that the second device sends the information 3 to the first device, the information 3 being used to request the model structure supported by the first device and / or the dimension of the output data supported by the first device. Correspondingly, the first device receives the information 3 from the second device. And the first device sends the third information to the second device in response to the information 3.

[0297] The information 3 can also be understood as information used to request the UE capability information, which is carried in a message such as a capability enqiry message, for example.

[0298] In this way, the second device can send the information 3 to the first device before determining the first information, so that the second device can obtain the partial first model structure supported by the first device and / or the dimension of the output data supported by the first device.

[0299] It should be noted that in the above embodiment, the first information can also be understood as a request sent by the second device to the first device, the request being used to request the first device to perform model training. In order to enable the second device to determine whether the first device is capable of performing model training, the first device can send a response to the second device, which is as follows.

[0300] As an optional embodiment, the method 400 further includes that the first device sends fourth information to the second device, the fourth information being used to indicate that the first device supports performing model training, or the fourth information being used to indicate that the first device does not support performing model training. Correspondingly, the second device receives the fourth information from the first device.

[0301] The fourth information can be understood as a response to the first information. In the case where the fourth information is used to indicate that the first device supports performing model training, the fourth information can be an acknowledgement (ACK) message, for example. In the case where the fourth information is used to indicate that the first device does not support performing model training, the fourth information can be a negative acknowledgement (NACK) message, for example.

[0302] In one case, the first device can send the fourth information to the second device after receiving the first information and before training the first model to completion.

[0303] Exemplarily, the first device can determine whether the first device supports the partial first model structure according to a model structure supported by the first device; for example, if the first device determines not to support the partial first model structure, the first device can send fourth information to the second device, and the fourth information is used to indicate that the first device does not support model training, and the fourth information can be a NACK message. If the first device determines to support the partial first model structure, the first device can send fourth information to the second device, and the fourth information is used to indicate that the first device supports model training, and the fourth information can be an ACK message.

[0304] Alternatively, the first device can determine whether the first device supports the dimension of the first model output data indicated by the first information according to the dimension of the output data supported by the first device (if the first information indicates the dimension of the second model input data, since the dimension of the first model output data is the same as the dimension of the second model input data, the first information can also be understood as indicating the dimension of the first model output data); if the first device determines not to support the dimension of the first model output data, the first device can send fourth information to the second device, and the fourth information is used to indicate that the first device does not support model training, and the fourth information can be a NACK message. If the first device determines to support the dimension of the first model output data, the first device can send fourth information to the second device, and the fourth information is used to indicate that the first device supports model training, and the fourth information can be an ACK message.

[0305] In this case, the first device can train the first model in an online training manner, and can respond to the first information through the fourth information. The online training manner can also be understood as a training manner in which the first device updates the model parameters of the first model in real time. For example, the first device trains and updates the first model immediately after obtaining data, which is sample data used to train the first model.

[0306] In addition to the first device not supporting the partial first model structure and / or the first device not supporting the dimension of the first model output data, the first device can also determine not to support model training based on other manners. Exemplarily, before sending the fourth information to the second device, the first device determines that the power is lower than a first threshold value, etc. The first threshold value can be a preset positive number, and the first threshold value is greater than 0 or less than 1.

[0307] Optionally, in the case where the fourth information is used to indicate that the first device does not support model training, the fourth information is also used to indicate the reason why the first device does not support model training and / or a first duration, the first duration being a duration during which the first device does not support model training.

[0308] The reason for not supporting model training may be, for example but not limited to, one or more of the following: the partial first model structure is not supported, the dimension of the first model output data is not supported, or the power is low (e.g., the power is lower than a first threshold).

[0309] It should be noted that, in a case where the partial first model structure indicated by the first information is determined by the second device based on the model structure supported by the first device (the third information), since the partial first model structure is the model structure supported by the first device, the reason for not supporting model training indicated by the fourth information does not include that the partial first model structure is not supported. Similarly, in a case where the dimension of the first model output data or the dimension of the second model input data indicated by the first information is determined by the second device based on the dimension of the output data supported by the first device (the third information), since the dimension of the first model output data is the dimension of the output data supported by the first device, the reason for not supporting model training indicated by the fourth information does not include that the dimension of the first model output data is not supported.

[0310] The first time length is a time length during which the first device does not support model training, and the first time length may be, for example, 30 minutes, 1 hour, or 3 hours, or the like. Alternatively, the first time length may also be replaced by a time length during which the first device does not support the partial first model structure. That is, during the first time length, the first device cannot train the first model based on the partial first model structure.

[0311] In order to enable the second device to determine the time at which the first device can perform model training, in a case where the fourth information is used to indicate the first time length, the fourth information may also be used to indicate a starting time at which the first device does not support model training, and then the second device may determine that, from the starting time, during the first time length, the first device does not support model training.

[0312] Alternatively, in a case where the first device supports model training, the first device may also send the fourth information to the second device after the first model training is completed. The fourth information may also be understood as information indicating that the first model training is completed.

[0313] In this case, the fourth information may be carried in a UAI, for example.

[0314] In this way, the second device can determine that the first device has trained the first model, and then the second device and the first device can perform communication based on the trained first model and the trained second model.

[0315] The embodiment of the present application also provides another communication method 600, which will be described below in combination with Figure 6 The method 600 will be described in detail.

[0316] Figure 6A flowchart of a communication method 600 provided by an embodiment of the present application is shown. The method 600 is applicable to the system 200, and the method 600 includes the following steps:

[0317] S601, the first device determines the first model structure, and / or the first device trains the first model.

[0318] It can be understood that in S601, the first device can determine the first model structure, but the first model is not trained; or the first device trains the first model; or the first device determines the first model structure and trains the first model based on the first model structure. For example, the first device determines the first model structure before training the first model, and then trains the first model based on the first model structure. The first model structure can be determined by the first device according to the application scenario, for example, according to the type of task that the first device needs to process based on the first model.

[0319] In this way, the first device can determine the fifth information based on the first model structure or the first model. For example, the model structure of the second model input layer is determined based on the model structure of the first model structure or the output layer of the first model, and the model structure of the second model input layer is the model structure of the output layer of the first model structure or the first model. Or, the dimension of the second model input data is determined based on the dimension of the first model structure or the first model output data, and the dimension of the second model input data is the dimension of the first model structure or the first model output data. So that the second model trained by the second device matches the first model.

[0320] It should be noted that S601 is optional. That is, before the first device sends the fifth information to the second device, the first device can also not determine the first model structure and not train the first model. Or, S601 can be replaced by the first device determining the dimension of the first model output data, etc.

[0321] It should be understood that in the method 600, the understanding of the first model structure and the second model structure can refer to the description in the method 400, and the understanding of the first model and the second model can also refer to the description in the method 400, which will not be repeated here.

[0322] S602, the first device sends the fifth information to the second device. Wherein, the fifth information is used to indicate part of the second model structure used by the second device to train the second model, and the part of the second model structure includes the model structure of the input layer; or, the fifth information is used to indicate the dimension of the first model output data or the dimension of the second model input data. Correspondingly, the second device receives the fifth information from the first device.

[0323] The second model structure is a model structure of the second model. The partial second model structure is a partial model structure in the total second model structure. For example, the total second model structure includes Y layers of network layers, Y being an integer greater than or equal to 2. The partial second model structure is a model structure of o layers of network layers in the Y layers of network layers, o being a positive integer less than Y. The o layers of network layers include an input layer of the second model. In a case where o is 1, the partial second model structure is a model structure of the input layer of the second model.

[0324] It should be understood that, for the convenience of description, the output layer of the first model in the method 600 is still referred to as the first output layer. The input layer of the second model is referred to as the first input layer. This will not be described again hereinafter.

[0325] In addition, the o layers of network layers can be consecutive o layers of network layers, for example, the i-th layer of network layers, the (i+1)-th layer of network layers, …, the (i+o-1)-th layer of network layers, i being a positive integer. The o layers of network layers can also be completely discontinuous network layers, that is, any two network layers are not adjacent. Alternatively, the o layers of network layers include partially continuous network layers and partially discontinuous network layers, etc. For example, the o layers of network layers include the j-th layer of network layers, the (j+1)-th layer of network layers, the (j+2)-th layer of network layers, the (j+5)-th layer of network layers, and the (j+7)-th layer of network layers, wherein the j-th layer of network layers, the (j+1)-th layer of network layers, and the (j+2)-th layer of network layers are three consecutive layers of network layers, and the (j+5)-th layer of network layers and the (j+7)-th layer of network layers are discontinuous network layers, j being a positive integer.

[0326] It should be understood that, in any of the above cases, the o layers of network layers include the first input layer. For other network layers in the o layers of network layers other than the first input layer, the present embodiment does not limit the positions and the number of the other network layers.

[0327] In a possible case, before S602, the first device performs S601. S602 can be implemented, for example, by the following possible implementation.

[0328] In the first possible implementation, the fifth information is used to indicate the partial second model structure.

[0329] The partial second model structure includes a model structure of the first input layer, that is, a dimension of the second model input data. The dimension of the first model output data is the same as the dimension of the second model input data.

[0330] In this possible implementation, the partial second model structure can be determined by the first device according to the first model structure or the first model. For example, the first device can determine the model structure of the first input layer based on the model structure of the first output layer, so as to match the first output layer with the first input layer. Alternatively, the first model structure includes the dimension of the first model output data, the first device can determine the dimension of the second model input data based on the dimension of the first model output data, and then determine the partial second model structure based on the dimension of the second model input data. The partial second model structure includes the dimension of the second model input data which is the same as the dimension of the first model output data included in the first model structure.

[0331] It can be understood that the partial second model structure includes the model structure of the first input layer, and the first model structure includes the model structure of the first output layer. For the model structure of the first output layer and the model structure of the first input layer, the dimension of the output data included in the model structure of the first output layer (i.e., the dimension of the first model output data) is the same as the dimension of the input data included in the model structure of the first input layer (i.e., the dimension of the second model input data). The content of the model structure of the first output layer except for the dimension of the output data can be the same as or different from the content of the model structure of the first input layer except for the dimension of the input data, or can be partially the same. The embodiments of the present application do not make specific limitations on this.

[0332] In the second possible implementation, the fifth information is used to indicate the dimension of the second model input data or the dimension of the first model output data.

[0333] The dimension of the second model input data is the same as the dimension of the first model output data. Therefore, when the fifth information is used to indicate the dimension of the second model input data, the second device can determine the dimension of the second model input data. Alternatively, when the fifth information is used to indicate the dimension of the first model output data, the second device can determine the dimension of the second model input data based on the dimension of the first model output data since the dimension of the first model output data is the same as the dimension of the second model input data.

[0334] In this possible implementation, the fifth information can be determined by the first device based on the dimension of the first model output data, and the first model structure includes the dimension of the first model output data. For example, the first device can determine the dimension of the output data of the second model based on the dimension of the first model output data.

[0335] In another possible case, S601 is optional content, that is, the first device does not perform S601 before S602. S602 can be implemented, for example, through the following possible implementation.

[0336] In the third possible implementation, the fifth information is used to indicate the partial second model structure.

[0337] In this possible implementation, the partial second model structure can be determined by the first device according to an application scenario. For example, the first device can determine the partial second model structure based on a load condition of the second device and / or a compression rate required by a service.

[0338] Alternatively, in some possible implementations, S601 can also be replaced by the first device determining a dimension of the first model output data.

[0339] Since the dimension of the second model input data is the same as the dimension of the first model output data, the first device can determine the dimension of the second model input data based on the dimension of the first model output data. For the remaining content of the partial second model other than the dimension of the second model input data, the first device can determine the remaining content according to an application scenario. For example, the first device can determine the remaining content based on a load condition of the second device and / or a compression rate required by a service.

[0340] And, optionally, S601 can be executed after S602, and then the first model structure determined by the first device is determined based on the partial second model structure. Since the dimension of the second model input data is the same as the dimension of the first model output data, the first device can determine the dimension of the first model output data. Then, the first model structure is determined based on the dimension of the first model output data. Then, the first device can train the first model based on the first model structure.

[0341] Optionally, the model structure of the first input layer can be the same as the model structure of the first output layer. In this way, the first input layer and the first output layer are completely matched, so that the second device can successfully process the output data of the first model based on the second model. Alternatively, the dimension of the first model output data included in the first model structure is the same as the dimension of the second model input data. The content of the first model structure other than the dimension of the first model output data can be determined by the first device based on a load condition of the second device and / or a compression rate required by a service.

[0342] In the fourth possible implementation, the fifth information is used to indicate a dimension of the second model output data.

[0343] In this possible implementation, the dimension of the second model input data can be determined by the first device according to an application scenario. For example, the first device can determine the dimension of the second model input data based on a load condition of the second device and / or a compression rate required by a service.

[0344] Alternatively, in some possible implementations, S601 can also be replaced by the first device determining a dimension of the first model output data.

[0345] In this way, the first device can indicate the dimension of the first model structure output data to the second device through the fifth information; or the first device determines the dimension of the second model input data based on the dimension of the first model structure output data, and indicates the dimension of the second model input data to the second device through the fifth information.

[0346] The dimension of the second model input data is the same as the dimension of the first model output data. Therefore, S601 can be executed after S602, for example, the first device trains the first model after S602, and the dimension of the first model output data is the same as the dimension of the second model input data.

[0347] S603, the second device determines the second model structure based on the fifth information.

[0348] The second model structure can be understood as the entire model structure of the second model. That is, the second device determines the entire second model structure based on part of the second model structure. Or, the second device determines the entire second model structure based on the dimension of the second model output data. In this way, the first output layer can match the first input layer, or the dimension of the first model structure output data is the same as the dimension of the second model structure input data.

[0349] After that, the method 600 further includes S604, the second device trains the second model based on the second model structure.

[0350] Since the second model is trained based on the second model structure, the model structure of the second model trained by the second device is the second model structure.

[0351] The communication method of the present application, the first device indicates the model structure of the second model input layer to the second device, so that the model structure of the output layer used by the first model trained by the first device can be the same as the model structure of the input layer used by the second model trained by the second device; or, the first device indicates the dimension of the second model input data or the dimension of the first model output data to the second device, so that the dimension of the second model input data trained by the second device can be the same as the dimension of the first model output data trained by the first device. In this way, the output layer of the first model trained by the first device matches the input layer of the second model trained by the second device, so that the first model and the second model match. Further, the second device can process the first model output data based on the second model.

[0352] In addition, the dimension of the partial second model structure or the second model input data (or alternatively the dimension of the first model output data) can be determined by the first device according to an application scenario, so that the first model and the second model are suitable for the current application scenario; for example, the first device determines the dimension of the partial second model structure or the second model input data (or alternatively the dimension of the first model output data) based on the load condition of the second device and / or the compression rate of the service requirement, so that the first model trained by the first device and the second model trained by the second device can meet the compression rate of the service requirement and can adapt to the load condition of the second device. In addition, with the development of communication technology, in order to meet different requirements, the first device can indicate different dimensions of the partial second model structure or the second model input data (or alternatively the dimension of the first model output data).

[0353] In addition, since the first device does not indicate the entire second model structure to the second device, the second device can also autonomously determine the remaining partial second model structure. In this way, the second model structure constructed by the second device is suitable for the application scenario in which the second device is located, and the requirements of the second device can be met.

[0354] In the following, the partial second model structure is described in detail.

[0355] As an optional embodiment, the partial second model structure is the model structure of the first m network layers in the second model, and m is a positive integer.

[0356] The first m network layers include the input layer of the second model, i.e., the first input layer, and the first m network layers are the m network layers closest to the first input layer.

[0357] In an example, m can be 1, and the first m network layers are the first input layer. In another example, m can be 2, and the first m network layers can be the first input layer and a hidden layer closest to the first input layer.

[0358] Since the output data of the first model is the input data of the second model, the effect of the second device processing data based on the second model is related to the structure and form of the output data of the first model, and the structure and form of the second model input data are usually related to the first m network layers. Therefore, by indicating the model structure of the first m network layers to the second device, the first device helps to match the first model and the second model, so that the effect of the second device processing the output data of the first model based on the second model can be better. For example, the first model is a compression model, and the second model is a decompression model. The higher the matching of the first model and the second model, the closer the initial CSI recovered by the second device based on the decompression model to the initial CSI determined by the first device.

[0359] On the basis of the above-mentioned embodiments, optionally, the model structure of the first m-layer network layer comprises: a type of each network layer in the first m-layer network layer and / or a number of neurons included in each network layer in the first m-layer network layer.

[0360] The type of the network layer may, for example, be one of a fully connected layer, a convolutional layer, a pooling layer or a recurrent layer. The number of neurons may be any positive integer. The neurons are related to the type of the second model, for example, the second model may be a CNN, and the neurons included in each network layer in the first m-layer network layer are neurons in the CNN; the second model may be a Transformer model, and the neurons included in each network layer in the first m-layer network layer are neurons in the Transformer model; or, the second model may be an LSTM model, and the neurons included in each network layer in the first m-layer network layer are neurons in the LSTM model, and the like.

[0361] In an example, m may be 1, and the first m-layer network layer is the first input layer. The model structure (part of the second model structure) of the first input layer comprises: a type of the first input layer and / or a number of neurons included in the first input layer. The type of the first input layer may, for example, be a fully connected layer, and the number of neurons included in the first input layer may, for example, be ten neurons, and the like. The dimension of the second model input data may, for example, be 1x10 or 10x1, and the like.

[0362] In order to make the first input layer more matched with the first output layer, the model structure of the first input layer may comprise a type of the first input layer and a number of neurons included in the first input layer. In this way, the type of the first input layer may be the same as the type of the first output layer, and the number of neurons included in the first input layer may be the same as the number of neurons included in the first output layer.

[0363] In another example, m may be 2, and the first m-layer network layer is the first input layer and a 1-layer hidden layer closest to the first input layer, which is referred to as the second hidden layer for ease of description. The model structure (part of the second model structure) of the first m-layer network layer comprises: a type of the first input layer and / or a number of neurons included in the first input layer, and a type of the second hidden layer and a number of neurons included in the second hidden layer. The type of the first input layer may, for example, be a fully connected layer, and the number of neurons included in the first input layer may, for example, be ten neurons, and the like. The dimension of the second model input data may, for example, be 1x10 or 10x1, and the like. The second hidden layer may, for example, comprise 15 neurons.

[0364] To make the first input layer and the first output layer more matched, in the case that m is greater than 2, the first m-layer network layer includes the first input layer and other network layers. The model structure (part of the second model structure) of the first m-layer network layer can include: the type of the first input layer and the number of neurons included in the first input layer, and the type of each of the other network layers and / or the number of neurons included in each of the other network layers.

[0365] It can be understood that m can also be greater, the type of each network layer can also be other types, and the number of neurons included in each network layer can also be other numbers. For the sake of brevity, they will not be listed one by one here.

[0366] Next, the way in which the second device trains the second model, S604, is described in detail.

[0367] Since the first model and the second model are associated models, S604 can be implemented in the following way: the second device trains the first model based on the first model structure and the second model structure.

[0368] It should be understood that the way in which the second device trains the second model is similar to the way in which the first device trains the first model, and the way in which the second device trains the second model can also be referred to Figure 5 the process shown. The difference is that the second device updates the model parameters of the second model based on the loss factor. For the sake of brevity, it will not be repeated here.

[0369] It should be understood that the second model can also be trained based on initial data, which can be data in a dataset. The dataset can also be sent by the first device to the second device, so that the first device trains the first model and the second device trains the second model can use the same dataset.

[0370] For example, the method 600 further includes: the first device sends information 4 to the second device, the information 4 being used to indicate the dataset. Correspondingly, the second device receives the information 4 from the first device.

[0371] It should be understood that the information 4 and the fifth information can be carried in the same signaling or in different signaling. And when the information 4 and the fifth information are carried in the same signaling, the information 4 and the fifth information can be carried in the same or different fields of the signaling, which is not limited in the present application.

[0372] Alternatively, the dataset can also be obtained by the second device in other ways. For the sake of brevity, they will not be listed one by one here.

[0373] It should be noted that the first model structure used by the second device in the process of training the second model can be a model structure constructed by the second device itself. For example, the second device determines the model structure of the first output layer based on the part of the second model structure indicated by the fifth information, that is, the model structure of the first output layer is the same as the model structure of the input layer included in the part of the second model structure. Or, the second device determines the dimension of the first model output data based on the dimension of the first model output data or the dimension of the second model input data indicated by the fifth information.

[0374] Alternatively, the first model structure can be indicated by the first device through signaling, specifically as follows.

[0375] As an optional embodiment, the method 600 further includes: the first device sends sixth information to the second device. Wherein, the sixth information is used to indicate all or part of the model structure of the first model, and all or part of the model structure of the first model is used to train the second model; or, the sixth information is used to indicate all or part of the model structure of the first model and the model parameters of the first model, and all or part of the model structure of the first model and the model parameters of the first model are used to train the second model. Correspondingly, the second device receives the sixth information from the first device.

[0376] In this way, when the sixth information indicates all the model structure of the first model, that is, all the first model structure, the second device can obtain all the first model structure. When the sixth information indicates part of the model structure of the first model, that is, part of the first model structure, the second device can obtain part of the first model structure, and for the remaining part of the first model structure, the second device can construct it according to the application scenario. For example, the first model structure includes X layer network layers, X is an integer greater than or equal to 2, and the sixth information can indicate u layer network layers in the X layer network layers, u is a positive integer less than X. The second device can construct the remaining X-u layer network layers in the X layer network layers.

[0377] And when the sixth information indicates all or part of the model structure of the first model, but does not indicate the model parameters of the first model, the model parameters of the first model can be determined by the second device. In this way, the data amount of the sixth information is small, so that the signaling overhead is small.

[0378] When the sixth information indicates all or part of the model structure of the first model, and indicates the model parameters of the first model, Figure 5The model parameters of the first model shown in the above embodiment are the model parameters indicated by the sixth information, and the model structure of the first model is constructed based on all or part of the model structure indicated by the sixth information. For example, after the first device trains the first model, the first device can indicate all or part of the model structure of the first model and the model parameters of the first model to the second device through the sixth information. In this way, the second device can determine the first model based on all or part of the model structure of the first model indicated by the sixth information and the model parameters of the first model, and the first model determined by the second device is close to the first model trained by the first device. The second device can train the second model based on the first model, which helps to make the trained second model and the second model trained by the first device better match.

[0379] It should be noted that the sixth information and the fifth information can be carried in the same signaling or in different signaling. Moreover, when the sixth information and the fifth information are carried in the same signaling, the sixth information and the fifth information can be carried in the same or different fields of the signaling, which is not limited in the present application.

[0380] Alternatively, in some possible implementation, the sixth information can also be understood as the fifth information. That is, the fifth information in S602 can also be replaced by the sixth information. In this case, all or part of the model structure of the first model indicated by the sixth information includes the model structure of the first output layer, which is the same as the model structure of the first input layer. Therefore, the sixth information can also be understood as the model structure of the first input layer (the input layer of the second model) indicated by the first device to the second device.

[0381] On the basis of the above embodiment, optionally, the method 600 further includes that the first device sends information for indicating the second device to separately perform model training to the second device. Correspondingly, the second device receives the information for indicating the second device to separately perform model training from the first device.

[0382] Wherein, separately performing model training can also be understood as that the second device can construct the first model structure and the second model structure by itself, and perform model training based on the first model structure and the second model structure. The first model structure can be constructed by the second device based on the fifth information, or can be constructed by the second device based on the sixth information. The process of model training performed by the second device can refer to the process shown in the above embodiment, which will not be described herein again. Figure 5

[0383] It should be understood that the information for indicating the second device to separately perform model training and the sixth information can be carried in the same or different signaling, and the information for indicating the second device to separately perform model training and the fifth information can also be carried in the same or different signaling. The present application does not make a specific limitation in this regard.​

[0384] On the basis of the above-mentioned embodiments, the first device can further indicate the model parameters of the second model to the second device, so that the second device can train the second model based on the model parameters of the second model.

[0385] Exemplarily, the method 600 further includes that the first device sends information 5 to the second device, the information 5 being used to indicate the model parameters of the second model. Correspondingly, the second device receives the information 5 from the first device.

[0386] It should be understood that the information 5 and the fifth information can be carried in the same signaling or in different signaling. Moreover, when the information 5 and the fifth information are carried in the same signaling, the information 5 and the fifth information can be carried in the same or different fields of the signaling. The present application does not make a specific limitation in this regard.

[0387] It should be noted that the above is described by taking the training of the second model by the second device as an example. In some possible implementation manners, the second model can be trained by other devices, for example, a fourth device. Moreover, after training the second model, the fourth device can send the trained second model to the second device. In this case, the manner in which the fourth device trains the second model is similar to the process shown in the above-mentioned Figure 5 . For details, refer to the description hereinabove, which will not be described herein again.

[0388] Moreover, in this case, the second device can send the fifth information and / or the sixth information to the fourth device, so that the fourth device can construct the first model structure and / or the second model structure based on the fifth information and / or the sixth information. The present application does not make a specific limitation in this regard.

[0389] The fourth device can be a server or other types of devices. The data set used by the fourth device to train the second model can be sent by the second device to the fourth device, or can be acquired by the fourth device in other manners. The present application does not make a specific limitation in this regard.

[0390] In addition, the capability of the second device can affect whether the second device can train the second model. Therefore, the second device can further improve the efficiency of model training in the following manners.

[0391] As an optional embodiment, the method 600 further includes that the second device sends seventh information to the first device, the seventh information being used to indicate the model structures supported by the second device and / or the dimensions of input data. Part of the second model structures is determined by the first device based on the seventh information, and / or the dimensions of the second model input data or the dimensions of the first model output data are determined by the first device based on the seventh information. Correspondingly, the first device receives the seventh information from the second device.

[0392] The model structures supported by the second device include at least one model structure, and the dimensions of the input data supported by the second device include at least one dimension.

[0393] In a first mode, when the seventh information indicates the model structures supported by the second device, the first device can determine the partial second model structure based on the model structures supported by the second device.

[0394] The model structures supported by the second device can include at least one model structure supported by the second device. Each of the at least one model structure can be a model structure of a partial network layer in the second model structure. For example, when the second model structure includes Y network layers, one of the at least one model structure includes a model structure of p network layers, and the p network layers belong to the partial network layers in the Y network layers. The model structure of the p network layers can be the same as or different from the number of network layers included in the partial second model structure indicated by the fifth information. For example, the partial second model structure indicated by the fifth information includes m network layers. The p and m can be the same or different. Y is an integer greater than or equal to 2, and p and m are positive integers less than Y.

[0395] The partial second model structure indicated by the fifth information can be determined according to the at least one model structure. For example, the partial second model structure indicated by the fifth information can be determined according to one of the at least one model structure.

[0396] Suppose the partial second model structure indicated by the fifth information is determined according to a second model structure of the at least one model structure, the partial second model structure can be the same as the second model structure. In this way, the partial second model structure determined by the first device is the model structure supported by the second device.

[0397] Alternatively, the partial second model structure indicated by the fifth information includes the second model structure. For example, the partial second model structure includes a model structure of m network layers, and the second model structure can include a model structure of p network layers, and the m network layers can include the p network layers. In this way, the first device can also autonomously determine the model structure of other partial network layers (m-p network layers) based on the model structures supported by the second device, so that the partial second model structure indicated by the first device can also meet the needs of the first device.

[0398] It should be noted that the number of network layers included in any two of the at least one model structure supported by the second device can be the same or different. For example, the model structure supported by the second device includes: a first model structure, a second model structure, and a third model structure; the first model structure includes a model structure of 3-layer network layers, the second model structure includes a model structure of 1-layer network layers, and the third model structure includes a model structure of 3-layer network layers, etc. The number of network layers included in each model structure is not limited in the present application.

[0399] Optionally, each of the at least one model structure supported by the second device can include a model structure of the first input layer (the input layer of the second model).

[0400] In this way, the first device can determine the model structure of the at least one input layer supported by the second device, so that the first device can determine the model structure of the first input layer based on the model structure of the at least one input layer. The model structure of the first input layer included in the partial second model structure indicated by the fifth information is the model structure of the input layer supported by the second device, so that the probability of successfully training the second model by the second device is higher.

[0401] It should be further noted that in order to match the input layer of the second model with the output layer of the first model, the model structure of the input layer of the second model needs to be the same as the model structure of the output layer of the first model. Therefore, the first device can also determine the partial second model structure according to the requirements of the first device and / or the model structure supported by the first device, and the model structure supported by the second device.

[0402] For example, assuming that the model structure supported by the second device includes at least one model structure of the input layer, the first device determines the model structure of the input layer of the second model to be one of the at least one model structure of the input layer, and the model structure of the input layer is also the model structure of the output layer of the first model, which can also be the model structure supported by the first device. For example, the at least one model structure of the input layer supported by the second device includes: a model structure 1 of the input layer, a model structure 2 of the input layer, and a model structure 3 of the input layer; assuming that the model structure of the output layer supported by the first device includes: a model structure 1 of the output layer and a model structure 2 of the output layer. If the model structure 2 of the output layer is the same as the model structure 1 of the input layer, the first device can determine the model structure 1 of the input layer as the model structure of the first input layer, that is, the model structure of the first output layer.

[0403] In this way, the first model structure can be the model structure supported by the first device, so that the probability of successfully training the first model by the first device is higher.

[0404] In the second mode, when the seventh information is used to indicate that the second device supports the dimension of the input data, the first device can determine the dimension of the first model output data or the dimension of the second model input data based on that the second device supports the dimension of the input data.

[0405] It can be understood that the dimension of the first model output data and the dimension of the second model input data are the same. Therefore, the first device determines the dimension of the second model input data based on that the second device supports the dimension of the input data, that is, the first device determines the dimension of the first model output data based on that the second device supports the dimension of the input data.

[0406] For example, the second device sends the seventh information to the first device before the first device determines the first model structure, so that the first device can determine the dimension of the input data supported by the second device based on the seventh information. Then, the dimension of the first model output data determined by the first device can belong to the dimension of the input data supported by the second device. In addition, the dimension of the second model input data or the dimension of the first model output data indicated by the first device to the second device belongs to the dimension of the input data supported by the second device.

[0407] In this way, the dimension of the second model input data or the dimension of the first model output data indicated by the first device is the dimension of the input data supported by the second device, so that the probability of successfully training the second model by the second device is higher.

[0408] In the second mode, optionally, the first device can also determine the dimension of the first model output data or the dimension of the second model input data based on the dimension of the input data supported by the second device and the dimension of the output data supported by the first device.

[0409] Since the dimension of the first model output data and the dimension of the second model input data are the same, in order to enable the first device to successfully train the first model, the dimension of the first model output data needs to be the dimension of the output data supported by the first device. Therefore, through the above mode, not only can the second device support the dimension of the second model input data, but also the first device can support the dimension of the first model output data. This helps to increase the probability of successfully training the first model by the first device, and helps to increase the probability of successfully training the second model by the second device.

[0410] In the third mode, when the seventh information is used to indicate the model structure supported by the second device, the model structure supported by the second device can include the model structure of the input layer, and the model structure of the input layer can include the dimension of the input data of the input layer. Therefore, the first device can also determine the dimension of the first model output data or the dimension of the second model input data based on the model structure supported by the second device.

[0411] The model structure supported by the second device can refer to the description in the first mode.

[0412] In the case that the model structure supported by the second device is at least one model structure, the model structure supported by the second device can include at least one dimension of input data, and each of the at least one dimension of input data is a dimension of input data supported by the second device. The dimension of the first model output data or the dimension of the second model input data can be one of the at least one dimension of input data.

[0413] For example, the model structure supported by the second device is at least one model structure, including a first model structure, a second model structure and a third model structure. The first model structure includes a dimension of input data 1, the second model structure includes a dimension of input data 2, and the third model structure includes a dimension of input data 3. The dimension of the first model output data or the dimension of the second model input data can be the dimension of input data 2.

[0414] In the third mode, the first device can determine at least one dimension of input data supported by the second device based on the model structure supported by the second device. Optionally, the first device can further determine the dimension of the first model output data or the dimension of the second model input data based on the at least one dimension of input data and the dimension of output data supported by the first device.

[0415] The dimension of the first model output data and the dimension of the second model input data are the same. The dimension of the first model output data or the dimension of the second model input data determined by the first device can be one of the at least one dimension of input data, and the first device supports the dimension of the first model output data. For example, the at least one dimension of input data supported by the second device includes a dimension of input data 1, a dimension of input data 2 and a dimension of input data 3; the dimension of output data supported by the first device is the same as the dimension of input data 1, then the first device can determine the dimension of input data 1 (or the dimension of output data supported by the first device) as the dimension of the second model input data or the dimension of the first model output data.

[0416] In the fourth mode, the seventh information is used to indicate the dimension of input and output data supported by the second device, and the first device can determine part of the second model structure based on the dimension of input data supported by the second device. The dimension of input data of the second model structure belongs to the dimension of input data supported by the second device.

[0417] In the fourth mode, the part of the second model structure, except the dimension of the input data of the second model structure, can be determined by the first device itself. For example, the first device can determine the part of the second model structure according to the application scenario. For example, the part of the second model structure includes a model structure of m layers of network layers, the dimension of the input data of the second model in the model structure of m layers of network layers is determined based on the dimension of the input data supported by the second device, and the rest of the model structure of m layers of network layers is determined by the first device itself.

[0418] Optionally, the seventh information can also be a response to the information 5. For example, the method 600 further includes: the first device sends the information 5 to the second device, the information 5 being used to request the model structure supported by the second device and / or the dimension of the input data. Correspondingly, the second device receives the information 5 from the first device. And the second device sends the seventh information to the first device in response to the information 5.

[0419] In this way, the first device can send the information 5 to the second device before determining the fifth information, so that the first device can obtain the part of the second model structure or the dimension of the input data supported by the second device.

[0420] It should be noted that in the above embodiment, the fifth information can also be understood as a request sent by the first device to the second device, the request being used to request the second device to perform model training. In order to enable the first device to determine whether the second device can perform model training, the second device can send a response to the first device, as follows.

[0421] As an optional embodiment, the method 600 further includes: the second device sends the eighth information to the first device, the eighth information being used to indicate that the second device supports performing model training, or the eighth information being used to indicate that the second device does not support performing model training. Correspondingly, the first device receives the eighth information from the second device.

[0422] The eighth information can be understood as a response to the fifth information. In the case where the eighth information is used to indicate that the second device supports performing model training, the eighth information can be, for example, an ACK message. In the case where the eighth information is used to indicate that the second device does not support performing model training, the eighth information can be, for example, a NACK message.

[0423] In one case, the second device can send the eighth information to the first device after receiving the fifth information and before completing the training of the second model.

[0424] Exemplarily, the second device can determine whether the second device supports the second model structure according to a model structure supported by the second device; for example, if the second device judges that part of the second model structure is not supported, the second device can send eighth information to the first device, and the eighth information is used to indicate that the second device does not support model training, then the eighth information can be a NACK message. If the second device judges that part of the second model structure is supported, the second device can send eighth information to the first device, and the eighth information is used to indicate that the second device supports model training, then the eighth information can be an ACK message.

[0425] Alternatively, the second device can determine whether the second model input data dimension indicated by the fifth information is supported by the second device according to the dimension of the input data supported by the second device (if the fifth information indicates the dimension of the first model output data, since the dimension of the first model output data is the same as the dimension of the second model input data, the fifth information can also be understood as indicating the dimension of the second model input data); for example, if the second device judges that the dimension of the second model input data is not supported, the second device can send eighth information to the first device, and the eighth information is used to indicate that the second device does not support model training, then the eighth information can be a NACK message. If the second device judges that the dimension of the second model input data is supported, the second device can send eighth information to the first device, and the eighth information is used to indicate that the second device supports model training, then the eighth information can be an ACK message.

[0426] Optionally, in the case where the eighth information is used to indicate that the second device does not support model training, the eighth information is also used to indicate one or more of the following: the reason why the second device does not support model training, the first time, the second duration or the second time, the first time is the time when the first device requests the second device to perform model training again, the second duration is the duration during which the first device is prohibited from requesting the second device to perform model training, and the second time is the time when the second device can start training the second model.

[0427] For example, the reason why the second device does not support model training can be one or more of the following: not supporting part of the second model structure, not supporting the dimension of the second model input data, or insufficient model training resources (for example, too many devices simultaneously requesting model training) and the like.

[0428] It should be noted that, in a case where the partial second model structure indicated by the fifth information is determined by the first device based on the model structure supported by the second device (the seventh information), since the partial second model structure is the model structure supported by the second device, the reason for not supporting model training indicated by the eighth information does not include not supporting the partial second model structure. Similarly, in a case where the dimension of the first model output data or the dimension of the second model input data indicated by the fifth information is determined by the first device based on the dimension of the input data supported by the second device (the seventh information), since the dimension of the second model input data is the dimension of the input data supported by the second device, the reason for not supporting model training indicated by the eighth information does not include not supporting the dimension of the second model input data.

[0429] The first time can be a future time. Since the second device does not support model training when receiving the fifth information from the first device, the second device can indicate to the first device, through the eighth information, the first time at which model training can be performed, the first time being later than the time corresponding to the time at which the second device receives the fifth information. In this way, the first device can send the fifth information to the second device again at the first time to request the second device to perform model training again.

[0430] The second duration is a duration during which the second device does not support model training, and the second duration may, for example, be 30 minutes, 1 hour, or 3 hours, etc.

[0431] In order to enable the first device to determine the time at which the second device can perform model training, in a case where the eighth information is used to indicate the second duration, the eighth information can also be used to indicate a starting time at which the second device does not support model training, and the first device can determine that, from the starting time, within the second duration, the second device does not support model training.

[0432] Alternatively, the second duration can also be replaced by a time period during which the second device does not support model training, for example, the eighth information is used to indicate a starting time and an ending time at which the second device does not support model training, and the first device can determine that, from the starting time to the ending time, the second device does not support model training. In this way, the first device can send the fifth information to the second device again after the ending time to request the second device to perform model training again.

[0433] It should be noted that, in a case where the eighth information is used to indicate the first time, the eighth information can not indicate the second duration; or, in a case where the eighth information is used to indicate the second duration, the eighth information can not indicate the first time. That is, the first device can determine the time at which the second device can be requested to perform model training again based on the first time or the second duration, and therefore the second device can indicate one of the first time or the second duration to the first device.

[0434] The second time can be a future time. Since the second device does not support model training when receiving the fifth information from the first device, the second device can indicate to the first device, through the eighth information, a second time at which training of the second model can be started, the second time being later than a time corresponding to the second device receiving the fifth information. In this way, the first device can determine the time at which the second model is to be trained starting from the second time.

[0435] In a case where the eighth information is used to indicate the second time, it can be understood that the eighth information is also used to indicate that the second device has successfully received the fifth information (the request for model training) from the first device, and that the second device is waiting for training of the second model, and that the second device can start training of the second model after the second time.

[0436] Since, in a case where the eighth information is used to indicate the second time, it is indicated that the second device is waiting for training of the second model, and that the second device can start training of the second model at the second time. In this case, the first device can not need to request the second device to perform model training again, so the eighth information can not indicate the first time and the second duration.

[0437] Therefore, the eighth information can be used to indicate any of the following: the reason why the second device does not support model training and / or the first time; the reason why the second device does not support model training and / or the second duration; or, the reason why the second device does not support model training and / or the second time.

[0438] In another case, in a case where the second device supports model training, the second device can also send the eighth information to the first device after training of the second model is completed. The eighth information can also be understood as information indicating that training of the second model is completed.

[0439] In this way, the first device can determine that the second device has completed training of the second model, and thus the second device and the first device can communicate based on the trained first model and the trained second model.

[0440] It can be understood that the communication method provided by the embodiments of the present application, for example, the method 400 and the method 600, can be applied to the communication system 200. One of the first device and the second device performing the method 400 or the method 600 can be understood as a network device. The network device is also an access network device. The access network device can adopt an O-RAN architecture, and the access network device adopting the O-RAN architecture can also be referred to as an O-RAN device.

[0441] In a case where the network device 210 in the communication system 200 is an O-RAN device, the system to which the embodiments of the present application are applicable also includes a communication system 700. As Figure 7As shown, the communication system 700 can include at least one open distributed unit (O-DU), for example Figure 7 an open distributed unit 710 as shown; and further include at least one open radio unit (O-RU), for example Figure 7 an open radio unit 720 as shown; the communication system 700 can further include at least one terminal device, for example Figure 7 a terminal device 730 as shown. The open distributed unit 710 and the open radio unit 720 can be understood as the network device 210 in the communication system 200, for example; the terminal device 730 can be understood as the terminal device 220 in the communication system 200, for example.

[0442] The open distributed unit 710 is provided with a baseband processing function, is provided with a complete protocol layer function, and is mainly responsible for high layer protocol functions such as data encryption and integrity protection. At the same time, it is provided with a physical layer high layer processing function.

[0443] The open radio unit 720 is provided with a physical layer bottom layer signal processing function, and is mainly responsible for transmitting and receiving radio frequency signals.

[0444] The open distributed unit 710 can transmit information to the open radio unit 720 through a front-haul interface; the open radio unit 720 can transmit information to the open distributed unit 710 through the front-haul interface.

[0445] The open radio unit 720 and the terminal device 730 can communicate through a wireless link. In one possible case, the open radio unit 720 can act as a transmitting end, and the terminal device 730 can act as a receiving end, and the open radio unit 720 transmits a signal to the terminal device 730; in another possible case, the open radio unit 720 can act as a receiving end, and the terminal device 730 can act as a transmitting end, and the terminal device 730 transmits a signal to the open radio unit 720.

[0446] In the case where the communication method provided in the embodiments of the present application is applied to the communication system 700, the implementation of the communication method can be, for example, as shown in Figure 8 or Figure 9 as shown.

[0447] First, taking the case where the communication method 400 is applied to the communication system 700 and the second device is an O-RAN device, the way in which the communication method provided in the present application is applied to the O-RAN device is described.

[0448] Figure 8 A flowchart of a communication method 800 provided in an embodiment of the present application is shown. As Figure 8As shown, the method 800 includes the following steps:

[0449] S801, the terminal device sends third information to the O-DU through the O-RU, the third information being used to indicate a model structure supported by the terminal device and / or a dimension of output data; wherein part of the first model structure is determined by the O-DU based on the third information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the O-DU based on the third information. Correspondingly, the O-DU receives the third information from the terminal device.

[0450] Wherein, the terminal device can be understood as the first device in the method 400. In addition, the understanding of the third information, the first model, the second model, the first model structure and the second model structure can refer to the description in the method 400.

[0451] The O-RU to the O-DU can be understood as the second device in the method 400. That is, the O-RU is used to perform the steps of the second device interacting with the first device, for example, the second device sending information to the first device can be understood as the O-RU sending information to the terminal device, and the first device sending information to the second device can be understood as the terminal device sending information to the O-RU.

[0452] And the O-DU is used to perform the steps of the second device processing information. For example, the O-RU can send the third information received from the terminal device to the O-DU through the front-haul interface, so that the O-DU can determine part of the first model structure based on the third information, or determine the dimension of the first model output data or the dimension of the second model input data based on the third information.

[0453] S802, the O-DU determines the second model structure, and / or the O-DU device trains the second model.

[0454] It should be understood that the implementation of S802 is similar to the implementation of S401, that is, the steps of the second device processing information are performed by the O-DU, and the implementation of S802 can refer to the description above, which will not be repeated here.

[0455] S803, the O-DU sends first information to the terminal device through the O-RU. Wherein, the first information is used to indicate part of the first model structure used by the terminal device to train the first model, and the part of the first model structure includes the model structure of the output layer; or, the first information is used to indicate the dimension of the first model output data or the dimension of the second model input data. Correspondingly, the terminal device receives the first information from the O-DU.

[0456] It should be understood that the implementation of S803 is similar to the implementation of S402, that is, the O-DU can send the first information to the terminal device after determining the first information, and the implementation of S803 can refer to the description above, which will not be repeated here.

[0457] S804, the terminal device determines a first model structure based on the first information.

[0458] S805, the terminal device trains a first model based on the first model structure.

[0459] The implementation of S804 and S805 is similar to the implementation of S403 and S404, which can refer to the description above, which will not be repeated here.

[0460] It should be understood that the method 800 is similar to the implementation of the method 400, except that the steps of the second device processing information in the method 400 can be understood as being performed by the O-DU, for example, the second model is trained by the O-DU. In the method 400, the second device sending information to the first device can be understood as the O-DU determining the information and sending the information to the O-RU through the fronthaul interface, and the O-RU sending the information to the terminal device. For example, in the method 400, the second device sending the second information to the first device can be understood as the O-DU determining the second information and sending the second information to the O-RU through the fronthaul interface, and the O-RU sending the second information to the terminal device; in the method 400, the second device sending the first information to the first device can be understood as the O-DU determining the first information and sending the first information to the O-RU through the fronthaul interface, and the O-RU sending the first information to the terminal device. The specific implementation of the method 800 can refer to the description in the method 400, which will not be repeated here.

[0461] In addition, in the method 400, the first device sending information to the second device can also be understood as the terminal device sending the information to the O-RU, and then the O-RU sending the information to the O-DU through the fronthaul interface, so that the O-DU can process the information. For example, in the method 400, the first device sending the third information to the second device can also be understood as the terminal device sending the third information to the O-RU, and then the O-RU sending the third information to the O-DU through the fronthaul interface, so that the O-DU can determine the first information based on the third information.

[0462] It should be noted that for the fourth information, the terminal device can send the fourth information to the O-RU; correspondingly, the O-RU can receive the fourth information from the terminal device. Since the fourth information is a response to the first information, after receiving the fourth information from the terminal device, the O-RU can process the fourth information by itself, or the O-RU can send the fourth information to the O-DU for processing.

[0463] Then, the communication method 600 is applied in the communication system 700, and the second device is an O-RAN device, and the application of the communication method provided by the present application to the O-RAN device is described.

[0464] Figure 9 A flowchart of a communication method 900 provided by an embodiment of the present application is shown. As shown in the figure, the method 900 includes the following steps: Figure 9

[0465] S901, the O-DU sends the seventh information through the O-RU terminal device, the seventh information is used to indicate the model structure supported by the O-DU and / or the dimension of the input data; wherein part of the second model structure is determined by the terminal device based on the seventh information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the terminal device based on the seventh information. Correspondingly, the terminal device receives the seventh information from the O-DU.

[0466] Wherein, the terminal device can be understood as the first device in the method 600. In addition, the understanding of the seventh information, the first model, the second model, the first model structure and the second model structure can refer to the description in the method 600.

[0467] The O-RU to the O-DU can be understood as the second device in the method 600. That is, the O-RU is used to execute the steps of the second device interacting with the first device, for example, the second device sending information to the first device can be understood as the O-RU sending information to the terminal device, and the first device sending information to the second device can be understood as the terminal device sending information to the O-RU. And the O-DU is used to execute the steps of the second device processing information.

[0468] S902, the terminal device determines the first model structure, and / or the terminal device trains the first model.

[0469] It should be understood that the implementation of S902 is similar to that of S601, that is, the steps of the second device processing information are executed by the O-DU, and the implementation of S902 can refer to the description above, which will not be described here.

[0470] S903, the terminal device sends the fifth information to the O-DU through the O-RU. Wherein, the fifth information is used to indicate the part of the second model structure used by the O-DU to train the second model, and the part of the second model structure includes the model structure of the input layer; or, the fifth information is used to indicate the dimension of the first model output data or the dimension of the second model input data. Correspondingly, the O-DU receives the fifth information from the terminal device.

[0471] ​It should be understood that the embodiments of S903 are similar to the embodiments of S602, that is, the O-DU performs the step of training the second model, and therefore, the O-RU sends the fifth information to the O-DU. The embodiments of S903 can refer to the description above, and will not be described here.

[0472] S904, the O-DU determines the second model structure based on the fifth information.

[0473] S905, the O-DU trains the second model based on the second model structure.

[0474] The embodiments of S904 and S905 are similar to the embodiments of S603 and S604, and can refer to the description above, and will not be described here.

[0475] It should be understood that the method 900 is similar to the embodiments of the method 600, except that the steps of the second device processing information in the method 600 can be understood as being performed by the O-DU. The method 600 in which the second device sends information to the first device can be understood as the O-DU sending the information to the O-RU through the front-haul interface, and the O-RU sending the information to the terminal device. That is, the step of determining the information sent to the terminal device can be performed by the O-DU, and then sent to the terminal device through the O-RU. For example, the second device sending the seventh information to the first device in the method 600 can be understood as the O-DU determining the seventh information and sending the seventh information to the O-RU through the front-haul interface, and the O-RU sending the seventh information to the terminal device. In the method 600, the first device sends information to the second device, which can be understood as the terminal device sending the information to the O-RU; the O-RU sends the information to the O-DU through the front-haul interface, so that the O-DU can process the information. For example, the first device sending the fifth information to the second device in the method 600 can be understood as the terminal device sending the fifth information to the O-RU, and the O-RU sending the fifth information to the O-DU through the front-haul interface. The specific embodiments of the method 900 can refer to the description in the method 600, and will not be described here.

[0476] It should also be understood that the steps of each of the above embodiments can also be coupled to each other, and the present application does not limit this. 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 by their functions and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0477] It should be noted that the size of the serial number of the above methods does not mean the order of execution, and the execution order of the processes should be determined by their functions and internal logic.

[0478] The communication method of the embodiments of the present application is described in detail above. Figures 4 to 9 , the communication method of the embodiments of the present application is described in detail above.Figure 10 and Figure 11 The communication apparatus is used for implementing the embodiments of the present application. The communication apparatus includes modules or units corresponding to each part of the above-mentioned embodiments. The modules or units can be software or hardware, or a combination of software and hardware. The communication apparatus is only briefly described below, and the details of the implementation can refer to the description of the above-mentioned method embodiments.

[0479] Figure 10 A schematic block diagram of a communication apparatus 1000 provided by an embodiment of the present application is shown in FIG. 10. As shown in FIG. 10, the communication apparatus 1000 includes a transceiver module 1001 and a processing module 1002. Figure 10

[0480] In a possible implementation, the communication apparatus 1000 is used to implement the steps corresponding to the first device (or terminal device) in the above-mentioned method 400, method 600, method 800 or method 900.

[0481] The transceiver module 1001 is configured to receive first information from a second device, wherein the first information is used to indicate a partial first model structure of the apparatus 1000 used for training the first model, the partial first model structure including a model structure of an output layer; or the first information is used to indicate a dimension of first model output data or a dimension of second model input data; and the processing module 1002 is configured to determine a first model structure based on the first information.

[0482] Optionally, the partial first model structure is a model structure of the last n network layers in the first model, and n is a positive integer.

[0483] Optionally, the model structure of the last n network layers includes a type of each network layer in the last n network layers and / or a number of neurons included in each network layer in the last n network layers.

[0484] Optionally, the transceiver module 1001 is further configured to receive second information from the second device, wherein the second information is used to indicate all or part of a model structure of a second model, and the all or part of the model structure of the second model is used for training the first model; or the second information is used to indicate all or part of the model structure of the second model and a model parameter of the second model, and the all or part of the model structure of the second model and the model parameter of the second model are used for training the first model.

[0485] Optionally, the transceiver module 1001 is further configured to receive information from the second device, wherein the information is used to indicate that the apparatus 1000 performs model training alone.

[0486] ​Optionally, the transceiver 1001 is further configured to send third information to the second device, the third information being used to indicate a model structure supported by the apparatus 1000 and / or a dimension of output data; wherein the partial first model structure is determined by the second device based on the third information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the second device based on the third information.

[0487] Optionally, the transceiver 1001 is further configured to send fourth information to the second device, the fourth information being used to indicate that the apparatus 1000 supports model training, or the fourth information being used to indicate that the apparatus 1000 does not support model training.

[0488] Optionally, the fourth information is used to indicate that the apparatus 1000 does not support model training, and the fourth information is further used to indicate a reason and / or a first time length that the apparatus 1000 does not support model training, the first time length being a duration that the apparatus 1000 does not support model training.

[0489] Optionally, the processing module 1002 is further configured to train the first model based on the first model structure.

[0490] Optionally, the first model is a model for encoding, and the second model is a model for decoding.

[0491] In a possible implementation, the communication apparatus 1000 is configured to implement the steps corresponding to the second device (or the network device) in the above method 400, method 600, method 800 or method 900.

[0492] It can be understood that, for the method 800 and the method 900, the transceiver 1001 can also be understood as an O-RU, and the processing module 1002 can also be understood as an O-DU.

[0493] The processing module 1002 is configured to determine first information, wherein the first information is used to indicate a partial first model structure used by the first device to train the first model, the partial first model structure including a model structure of an output layer; or the first information is used to indicate a dimension of the first model output data or a dimension of the second model input data; and the transceiver 1001 is configured to send the first information to the first device.

[0494] Optionally, the partial first model structure is a model structure of n last network layers in the first model, n being a positive integer.

[0495] Optionally, the model structure of the n last network layers includes a type of each network layer in the n last network layers and / or a number of neurons included in each network layer in the n last network layers.

[0496] Optionally, the transceiver 1001 is further configured to send, to the first device, second information; wherein the second information is used to indicate all or part of model structure of the second model, and the all or part of model structure of the second model is used to train the first model; or the second information is used to indicate all or part of model structure of the second model and model parameters of the second model, and the all or part of model structure of the second model and the model parameters of the second model are used to train the first model.

[0497] Optionally, the transceiver 1001 is further configured to send, to the first device, information used to indicate that the first device performs model training alone.

[0498] Optionally, the transceiver 1001 is further configured to receive third information from the first device, and the third information is used to indicate model structure and / or dimension of output data supported by the first device; wherein the part of the first model structure is determined by the apparatus 1000 based on the third information, and / or the dimension of the first model output data or the dimension of the second model input data is determined by the apparatus 1000 based on the third information.

[0499] Optionally, the transceiver 1001 is further configured to receive fourth information from the first device, and the fourth information is used to indicate that the first device supports model training, or the fourth information is used to indicate that the first device does not support model training.

[0500] Optionally, the fourth information is used to indicate that the first device does not support model training, and the fourth information is further used to indicate a reason and / or a first time length that the first device does not support model training, and the first time length is a duration that the first device does not support model training.

[0501] Optionally, the first model is a model used for encoding, and the second model is a model used for decoding.

[0502] It should be understood that the communication apparatus 1000 herein is embodied in the form of functional modules. The term “module” herein can refer to an application specific integrated circuit (ASIC), an electronic circuit, a processor (shared, dedicated or group) and memory for executing one or more software or firmware programs, a combinational logic circuit, and / or other suitable components that provide the described functionality. In an optional example, those skilled in the art can understand that the communication apparatus 1000 can be embodied as the terminal device or the network device in the above embodiments, and the communication apparatus 1000 can be used to perform the respective processes and / or steps corresponding to the first device or the second device in the above method embodiments, and thus details are not repeated here.

[0503] The communication apparatus 1000 in the above embodiments has functions of implementing corresponding steps performed by the first device or the second device in the above methods; the functions can be implemented by hardware, or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the functions. In the embodiments of the present application, Figure 10 The communication apparatus 1000 in the above embodiments can also be a chip, for example, a SOC.

[0504] Figure 11 A structure diagram of the communication apparatus 1100 provided by the embodiments of the present application is shown. The communication apparatus 1100 includes a processor 1101, a transceiver 1102 and a memory 1103. The processor 1101, the transceiver 1102 and the memory 1103 communicate with each other through internal connection paths. The memory 1103 is used to store instructions, such as computer degree codes, etc. The processor 1101 is used to execute the instructions stored in the memory 1103 to control the transceiver 1102 to send and / or receive signals.

[0505] It should be understood that the communication apparatus 1100 can be specifically the first device or the second device in the above embodiments, and can be used to execute the respective steps and / or processes corresponding to the first device or the second device in the above method embodiments. Optionally, the memory 1103 can include read-only memory and random access memory, and provide instructions and data to the processor. A part of the memory can also include non-volatile random access memory. For example, the memory can also store device type information. The processor 1101 can be used to execute the instructions stored in the memory, and when the processor 1101 executes the instructions stored in the memory, the processor 1101 is used to execute the respective steps and / or processes of the above method embodiments. The transceiver 1102 can include a transmitter 11021, a receiver 11022 and an antenna 11023. The transmitter 11021 can be used to implement the respective steps and / or processes corresponding to the transmitter for executing sending actions. For example, the transmitter 11021 can be used to send information to another device through the antenna 11023. The receiver 11022 can be used to implement the respective steps and / or processes corresponding to the receiver for executing receiving actions. For example, the receiver 11022 can be used to receive information from another device through the antenna 11023.

[0506] It should be understood that in the embodiments of the present application, the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0507] In the implementation process, each step of the above method can be completed by the integrated logic circuit of hardware in the processor or the instruction in the form of software. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as hardware processor execution completion, or executed by hardware and software modules in the processor. The software module can be located in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, register, etc. The storage medium is located in the memory, and the processor executes the instruction in the memory, and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0508] Figure 12 A network element function division and protocol layer structure diagram of an O-RAN device is shown in the embodiments of the present application.

[0509] In some examples, the CU is a logical node that carries the radio resource control (RRC) layer, the service data adaptation protocol (SDAP) layer, the packet data convergence protocol (PDCP) layer, and other control functions of the access network device. The CU is connected to network nodes such as core network devices through some interfaces, which can be E2 interfaces, etc. Optionally, the CU can have part of the functions of the core network device. The CU (such as the PDCP layer and higher layers) is connected to the DU (such as the RLC layer and lower layers) through some interfaces, which can be F1 interfaces, etc. In some examples, these interfaces (such as F1 interfaces) can provide control plane (C-Plane) and user plane (U-Plane) functions, such as interface management, system information management, UE context management, and RRC message transmission, etc. F1AP is the application protocol of the F1 interface, which defines the signaling procedures of F1 in some examples. The F1 interface supports the control plane F1-C and the user plane F1-U.

[0510] In some examples, the CU can be split into a CU-CP (control unit-control plane) and a CU-UP (control unit-user plane), where the CU-CP is a logical node carrying the RRC layer and the PDCP-C (control plane part of PDCP) layer, used to implement the control plane function of the CU. The CU-CP can interact with a network element in the core network used to implement the control plane function. The network element in the core network used to implement the control plane function can be an access and mobility function network element, such as an access and mobility management function (AMF) in a 5G system. The AMF network element is used to be responsible for mobility management in a mobile network, such as location updating of a terminal device, registration of the terminal device to a network, handover of the terminal device, and the like. The CU-UP is a logical node carrying the SDAP layer and the PDCP-U (user plane part of PDCP) layer, used to implement the user plane function of the CU. The CU-UP can interact with a network element in the core network used to implement the user plane function. The network element in the core network used to implement the user plane function, for example, a user plane function (UPF) in a 5G system, is used to be responsible for forwarding and receiving data in a terminal device. The above configuration of the CU and the DU is merely an example, and the CU and the DU can be configured to have functions as needed. For example, the CU or the DU can be configured to have functions of more protocol layers, or the CU or the DU can be configured to have partial processing functions of the protocol layers. For example, partial functions of the RLC layer and functions of protocol layers above the RLC layer are arranged in the CU, and the remaining functions of the RLC layer and functions of protocol layers below the RLC layer are arranged in the DU. For another example, the functions of the CU or the DU can be divided according to a service type or other system requirements, for example, according to a delay requirement. Functions that require a processing time to meet a relatively low delay requirement are arranged in the DU, and functions that do not require the processing time to meet the delay requirement are arranged in the CU.

[0511] In some examples, a DU is a logical node that hosts radio link control (RLC) layer, medium access control (MAC) layer, higher physical layer (higher PHY) layer, and other functions. In some examples, a DU can control at least one RU. The DU is connected with the RUs through some interfaces, which can be a fronthaul interface. In some examples, the higher PHY layer includes parts of PHY layer processing, such as forward error correction (FEC) encoding and decoding, scrambling, modulation and demodulation, and other processing functions.

[0512] In some examples, an RU is a logical node that hosts lower physical layer (lower PHY) and radio frequency (RF) processing, which can also be referred to as radio frequency chain (RF chain). In some examples, an RU can be a 3GPP transmission reception point (TRP) or a remote radio head (RRH) or other similar functional entity. In some examples, the low-PHY includes parts of PHY processing, such as fast Fourier transform (FFT), inverse fast fourier transformation (IFFT), digital beamforming and filtering, and other processing functions. The RU communicates with one or more UEs through a wireless link.

[0513] The DU and the RU can be co-located or not co-located. The DU and the RU exchange control plane information and user plane information through a lower-layer split CUS-plane (LLS-CUS) interface via a lower-layer split-control, user and synchronization (LLS-CUS) interface. The LLS-CUS can include a LLS-C interface and a LLS-U interface that provide control plane (C-plane) and user plane (U-plane), respectively. In some examples, the control plane (C-plane) refers to real-time control between the DU and the RU. The DU and the RU exchange management information through a LLS-M interface of the lower-layer split link, and the management plane (M-plane) refers to non-real-time management operations between the DU and the RU.

[0514] The DU and the RU can cooperate to jointly implement the functions of the PHY layer. One DU can be connected with one or more RUs. The functions of the DU and the RU can be configured in various manners according to design. For example, the DU is configured to implement baseband functions, and the RU is configured to implement radio frequency functions. For another example, the DU is configured to implement high-layer functions in the PHY layer, and the RU is configured to implement low-layer functions in the PHY layer or implement the low-layer functions and radio frequency functions. The high-layer functions in the PHY layer can include a part of functions of the PHY layer that are closer to the MAC layer, and the low-layer functions in the PHY layer can include another part of functions of the PHY layer that are closer to the radio frequency side.

[0515] The CU (or CU-CP and CU-UP), the DU, or the RU can also have different names in different systems, but those skilled in the art can understand their meanings. For example, in an ORAN system, the CU can also be referred to as an O-CU (open CU), the DU can also be referred to as an O-DU, the CU-CP can also be referred to as an O-CU-CP, the CU-UP can also be referred to as an O-CU-UP, and the RU can also be referred to as an O-RU. For the convenience of description, the CU, the CU-CP, the CU-UP, the DU, and the RU are taken as examples for description in this application.

[0516] The application further provides a computer-readable storage medium for storing a computer program for implementing the method shown in the above method embodiments.

[0517] The application further provides a computer program product including a computer program (also referred to as code or instructions), which, when running on a computer, can execute the method shown in the above method embodiments.

[0518] Those of ordinary skill in the art can realize that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0519] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device, and module can refer to the corresponding processes in the foregoing method embodiments, which will not be described herein.

[0520] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the described device embodiments are merely schematic. For example, the division of the modules is merely logical function division. There can be another division manner for the actual implementation, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or modules, and can be in electrical, mechanical or other forms.

[0521] The modules illustrated as separated components can or can not be physically separated, and the components illustrated as modules can or can not be physical modules, i.e., can be located in one place, or can be distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.

[0522] In addition, the functional modules in each embodiment of the present application can be integrated into a processing module, or each module can be physically present alone, or two or more modules can be integrated into one module.

[0523] If the functions are realized in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product in essence or the part of the technical solutions that make contributions to the prior art, or part of the technical solutions. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0524] The above description is merely a specific implementation of the present application, but the protection scope of the embodiments of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the embodiments of the present application, which should be covered within the protection scope of the embodiments of the present application. Therefore, the protection scope of the embodiments of the present application should be subject to the protection scope of the claims.

Claims

1. A communication method, characterized in that, The method includes: Receive the first information from the second device; Wherein, the first information is used to indicate a portion of the first model structure used by the first device to train the first model, the portion of the first model structure including the model structure of the output layer; or, The first information is used to indicate the dimension of the output data of the first model or the dimension of the input data of the second model; Based on the first information, the first model structure is determined.

2. The method according to claim 1, characterized in that, The first model structure is the model structure of the last n layers of the network in the first model, where n is a positive integer.

3. The method according to claim 2, characterized in that, The model structure of the last n network layers includes: the type of each network layer in the last n network layers and / or the number of neurons included in each network layer in the last n network layers.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Receive second information from the second device; wherein the second information is used to indicate all or part of the model structure of the second model, and the all or part of the model structure of the second model is used to train the first model; or, the second information is used to indicate all or part of the model structure of the second model and the model parameters of the second model, and the all or part of the model structure of the second model and the model parameters of the second model are used to train the first model.

5. The method according to any one of claims 1 to 4, characterized in that, The method further includes: Receive information from the second device instructing the first device to perform model training independently.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Send third information to the second device, the third information being used to indicate the model structure and / or the dimension of the output data supported by the first device; wherein, the portion of the first model structure is determined by the second device based on the third information, and / or, the dimension of the first model output data or the dimension of the second model input data is determined by the second device based on the third information.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Send a fourth message to the second device, the fourth message being used to indicate that the first device supports model training, or the fourth message being used to indicate that the first device does not support model training.

8. The method according to claim 7, characterized in that, The fourth information is used to indicate that the first device does not support model training. The fourth information is also used to indicate the reason why the first device does not support model training and / or the first duration, where the first duration is the duration during which the first device does not support model training.

9. The method according to any one of claims 1 to 8, characterized in that, The method further includes: The first model is trained based on the first model structure.

10. The method according to any one of claims 1 to 9, characterized in that, The first model is a model for encoding, and the second model is a model for decoding.

11. A communication method, characterized in that, The method includes: Determine the first piece of information; Wherein, the first information is used to indicate a portion of the first model structure used by the first device to train the first model, the portion of the first model structure including the model structure of the output layer; or, The first information is used to indicate the dimension of the output data of the first model or the dimension of the input data of the second model; Send the first information to the first device.

12. The method according to claim 11, characterized in that, The first model structure is the model structure of the last n layers of the network in the first model, where n is a positive integer.

13. The method according to claim 12, characterized in that, The model structure of the last n network layers includes: the type of each network layer in the last n network layers and / or the number of neurons included in each network layer in the last n network layers.

14. The method according to any one of claims 11 to 13, characterized in that, The method further includes: Send second information to the first device; wherein the second information is used to indicate all or part of the model structure of the second model, and the all or part of the model structure of the second model is used to train the first model; or, the second information is used to indicate all or part of the model structure of the second model and the model parameters of the second model, and the all or part of the model structure of the second model and the model parameters of the second model are used to train the first model.

15. The method according to claim 14, characterized in that, The method further includes: Send information to the first device instructing the first device to perform model training independently.

16. The method according to any one of claims 11 to 15, characterized in that, The method further includes: The system receives third information from the first device, the third information indicating the model structure and / or the dimension of the output data supported by the first device; wherein, the portion of the first model structure is determined by the second device based on the third information, and / or, the dimension of the first model output data or the dimension of the second model input data is determined by the second device based on the third information.

17. The method according to any one of claims 11 to 16, characterized in that, The method further includes: The device receives fourth information from the first device, the fourth information indicating that the first device supports model training, or the fourth information indicating that the first device does not support model training.

18. The method according to claim 17, characterized in that, The fourth information is used to indicate that the first device does not support model training. The fourth information is also used to indicate the reason why the first device does not support model training and / or the first duration, where the first duration is the duration during which the first device does not support model training.

19. The method according to any one of claims 11 to 18, characterized in that, The first model is a model for encoding, and the second model is a model for decoding.

20. A communication device, characterized in that, include: A processor coupled to a memory storing computer-executable instructions, the processor executing the computer-executable instructions stored in the memory, causing the processor to perform the method as claimed in any one of claims 1 to 10, or to perform the method as claimed in any one of claims 11 to 19.

21. A computer-readable storage medium, characterized in that, Used to store a computer program, the computer program including instructions for implementing the method as claimed in any one of claims 1 to 10, or instructions for performing the method as claimed in any one of claims 11 to 19.

22. A computer program product, said computer program product comprising computer program code, characterized in that, When the computer program code is run on a computer, it causes the computer to implement the method as described in any one of claims 1 to 10, or to perform the method as described in any one of claims 11 to 19.

Images