WO2023197950A1 - 一种通信方法及相关装置 - Google Patents

一种通信方法及相关装置 Download PDF

Info

Publication number
WO2023197950A1
WO2023197950A1 PCT/CN2023/086876 CN2023086876W WO2023197950A1 WO 2023197950 A1 WO2023197950 A1 WO 2023197950A1 CN 2023086876 W CN2023086876 W CN 2023086876W WO 2023197950 A1 WO2023197950 A1 WO 2023197950A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
network model
communication
model
information
Prior art date
Application number
PCT/CN2023/086876
Other languages
English (en)
French (fr)
Inventor
王坚
徐晨
张公正
李榕
颜敏
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2023197950A1 publication Critical patent/WO2023197950A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/02—Arrangements for optimising operational condition
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W28/00—Network traffic management; Network resource management
    • H04W28/02—Traffic management, e.g. flow control or congestion control
    • H04W28/04—Error control
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W28/00—Network traffic management; Network resource management
    • H04W28/16—Central resource management; Negotiation of resources or communication parameters, e.g. negotiating bandwidth or QoS [Quality of Service]
    • H04W28/18—Negotiating wireless communication parameters
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W74/00—Wireless channel access
    • H04W74/08—Non-scheduled access, e.g. ALOHA
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W76/00—Connection management
    • H04W76/20—Manipulation of established connections
    • H04W76/28—Discontinuous transmission [DTX]; Discontinuous reception [DRX]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W74/00—Wireless channel access
    • H04W74/08—Non-scheduled access, e.g. ALOHA
    • H04W74/0833—Random access procedures, e.g. with 4-step access

Definitions

  • the present application relates to the field of communication, and in particular, to a communication method and related devices.
  • AI technology can also be applied to the network layer and physical layer.
  • AI technology can be used to implement network optimization, mobility management, resource allocation, etc. in the network layer, and can also be used to implement channel coding and decoding, channel prediction, receivers, etc. in the physical layer.
  • AI models can be used to implement communication transceivers to achieve communication between the sending and receiving ends.
  • the transmitting end can use a neural network model to process signals to be sent and send signals
  • the receiving end can use a neural network model to process received signals.
  • Current AI transceivers can only adapt to specific communication scenarios. When the communication scenario changes, the corresponding model needs to be retrained, making it difficult to deploy in a general communication system. How to determine or deploy the AI model used in communication between the sending and receiving ends is an urgent problem that needs to be solved.
  • Embodiments of the present application provide a communication method and related devices, which can determine a neural network model for communication between the sending and receiving ends with low complexity and low signaling overhead.
  • embodiments of the present application provide a communication method, which method includes: determining a first neural network model from one or more pre-trained neural network models; adjusting the first neural network model according to the first information, wherein: One information includes communication resource information and/or channel state information for communication between the first device and the second device; sending the adjusted first neural network model, or sending a sub-model in the adjusted first neural network model ; The adjusted first neural network model is used for communication between the first device and the second device.
  • the neural network model used for communication between the first device and the second device is adjusted according to the trained pre-trained neural network model, and the obtained neural network model is adapted to the first device and the second device. Communication scenario of two devices. Compared with the method in which the first device and the second device retrain the neural network model for communication, the complexity of determining the neural network model can be reduced, and the interaction between the first device and the second device for determining the neural network model can be reduced. signaling overhead.
  • the first neural network model is determined from one or more pre-trained neural network models based on the communication system parameters of the first device and/or the communication system parameters of the second device; the first The communication system parameters of the device include the system bandwidth and frame structure supported by the first device, and the communication system parameters of the second device include the system bandwidth and frame structure supported by the second device.
  • This implementation can enable the determined first neural network model to be adapted to the communication system of the first device and the second device, which is beneficial to improving the application of the adjusted first neural network model to the communication between the first device and the second device. performance, thereby improving the quality of communication between the first device and the second device.
  • the first neural network model is determined based on a third neural network model selected from one or more second neural network models; the one or more second neural network models are determined from one or more second neural network models. or determined in multiple pre-trained neural network models; wherein the input dimension of each second neural network model in one or more second neural network models is the same as its corresponding first input dimension, and/or, each second neural network model
  • the output dimensions of the two neural network models are the same as their corresponding first output dimensions; the first input dimension and/or the first output dimension corresponding to each second neural network model is based on the resource slice applicable to the second neural network model ( resource patch, RP) size, and the communication system parameters of the first device and/or the communication system parameters of the second device.
  • This implementation can make the input dimension of the first neural network model the same as its corresponding first input dimension, and/or the output dimension of the first neural network model and its corresponding first output dimension, thereby making the first neural network
  • the model is adapted to the communication scenario of the first device and the second device.
  • the RP size is determined based on the time domain length of the RP and the frequency domain width of the RP.
  • the second neural network model supports a first service type; the first service type is a service type required by the first device and the second device.
  • This implementation selects one or more second neural network models from pre-trained neural networks based on business types, so that the first neural network model determined from the one or more neural network models supports the first device and the second device requirements. business type.
  • the third neural network model is the second neural network model with the largest number of parameters among the multiple second neural network models.
  • the larger the number of parameters of the neural network model the better the performance.
  • This method can make the determined third neural network model be the second neural network model with the best performance among multiple second neural network models.
  • the third neural network model is the second neural network model with the smallest absolute difference between the calculation amount and the first calculation amount among the plurality of second neural network models.
  • the first calculation amount is based on the sum of the computing power of the first device.
  • the delay requirements, and/or, the computing power and delay requirements of the second device are determined. This method can make the calculation amount of the third neural network model determined from the plurality of second neural network models closest to the maximum range supported by the computing capabilities of the first device and the second device.
  • the third neural network model is the second neural network model with the smallest absolute difference between the parameter quantity and the first parameter quantity among the plurality of second neural network models.
  • the first parameter quantity is based on the sum of the computing power of the first device.
  • the storage space, and/or, is determined by the computing power and storage space of the second device. This method can make the parameter amount of the third neural network model determined from the plurality of second neural network models be closest to the maximum range supported by the storage capabilities of the first device and the second device.
  • the first neural network model is obtained by distilling the third neural network model, and the operation amount of the third neural network model is greater than the first operation amount, and/or, the third neural network model
  • the parameter amount is greater than the first parameter amount; wherein the first operation amount is determined based on the computing power and delay requirements of the first device, and/or the computing power and delay requirements of the second device; the first parameter amount is Determined based on the computing power and storage space of the first device, and/or the computing power and storage space of the first device.
  • the first neural network model can be obtained by reducing the parameter amount and/or operation amount of the third neural network model. In this way, the operation amount of the first neural network model is within the limit supported by the computing capabilities of the first device and the second device. Within the range, the parameter amount is within the range supported by the storage capabilities of the first device and the second device.
  • the first neural network model is a third neural network model; the operation amount of the third neural network model is less than or equal to the first operation amount, and/or the parameter amount of the third neural network model is less than Or equal to the first parameter quantity; the first calculation quantity is determined according to the computing power and delay requirements of the first device, and/or, the computing power and delay requirements of the second device; the first parameter quantity is determined according to the computing power and delay requirements of the first device
  • the computing power and storage space of the second device are determined by the computing power and storage space of the second device. It can be seen that in this implementation, if the calculation amount and parameter amount of the third neural network model are within the range supported by the computing capabilities and storage capabilities of the first device and the second device respectively, the third neural network model can be directly used as The first neural network model.
  • the first neural network model is determined by the model server based on the received model request information; the model request information includes the identification of the third neural network model, as well as the first calculation amount and/or the first parameter. quantity; wherein, the first calculation quantity is determined based on the computing power and delay requirements of the first device, and/or the computing power and delay requirements of the second device, and the first parameter quantity is determined based on the computing power and delay requirements of the first device. and storage space, and/or, determined by the computing power and storage space of the second device. It can be seen that this embodiment can be applied to the situation where one or more pre-trained neural network models are stored in the model server, and the first neural network model can be obtained from the model server.
  • the information of each pre-trained neural network model in one or more pre-trained neural network models is predefined or obtained from the model server; the information of each pre-trained neural network model includes one or more of the following: Each: identification, business type, RP size, input dimension, output dimension, parameter amount and operation amount.
  • the adjusted first neural network model is obtained by training the fourth neural network model based on the channel state information in the first information; the fourth neural network model is obtained based on the first neural network
  • the RP size applicable to the model and the communication resource information in the first information are obtained by adjusting the input dimension and/or output dimension of the first neural network model.
  • the first neural network model is adjusted according to the communication resource information and channel state information communicated between the first device and the second device, so that the adjusted first neural network model can be adapted to the first device and the second device. Communication scenario of two devices.
  • embodiments of the present application provide a communication method, which method includes: receiving an adjusted first neural network model; or, receiving a sub-model in the adjusted first neural network model; wherein, the adjusted first neural network model A neural network model is used for communication between the first device and the second device; the adjusted first neural network model is obtained by adjusting the first neural network model based on the first information, the first information includes the first device and the second device. Communication resource information and/or channel state information for communication between the second device and the first neural network device are determined from one or more pre-trained neural network models. Communication is performed based on the adjusted first neural network model; or communication is performed based on a sub-model in the adjusted first neural network model.
  • the neural network model used for communication between the first device and the second device is adjusted based on the trained pre-trained neural network model, and the determined neural network model is adapted to the first device and the second device. Communication scenario of two devices. Compared with the manner in which the first device and the second device retrain the neural network model for communication, the complexity of determining the neural network model can be reduced, and the signaling overhead of interaction between devices for determining the neural network model can also be reduced.
  • the first neural network model is determined from one or more pre-trained neural network models based on the communication system parameters of the first device and/or the communication system parameters of the second device; wherein, The communication system parameters of the first device include the system bandwidth and frame structure supported by the first device, and the communication system parameters of the second device include the system bandwidth and frame structure supported by the second device.
  • This implementation can adapt the determined first neural network model to the communication scenario of the first device and the second device, which is beneficial to improving the application of the adjusted first neural network model to the communication between the first device and the second device. performance, thereby improving the quality of communication between the first device and the second device.
  • the method also includes: sending communication system parameters.
  • the first neural network model is determined based on a third neural network model selected from one or more second neural network models; the one or more second neural network models are determined from one or more second neural network models. Determined among multiple pre-trained neural network models; wherein the input dimension of each second neural network model in one or more second neural network models is the same as its corresponding first input dimension, and/or, each second The output dimension of the neural network model is the same as its corresponding first output dimension; the first input dimension and/or the first output dimension corresponding to each second neural network model is based on the RP size of the resource slice applicable to the second neural network model. , and the communication system parameters of the first device and/or the communication system parameters of the second device are determined.
  • This implementation can make the input dimension of the first neural network model the same as its corresponding first input dimension, and/or, The output dimension of the first neural network model is the same as its corresponding first output dimension, thereby enabling the first neural network model to be adapted to the communication scenario of the first device and the second device.
  • the RP size is determined based on the time domain length of the RP and the frequency domain width of the RP.
  • the second neural network model supports a first service type; the first service type is a service type required by the first device and the second device.
  • This implementation selects one or more second neural network models from pre-trained neural networks based on business types, so that the first neural network model determined from the one or more neural network models supports the first device and the second device requirements. business type.
  • the method also includes: sending the required service type.
  • the third neural network model is the neural network model with the largest number of parameters among the multiple second neural network models. Generally speaking, the larger the number of parameters of the neural network model, the better the performance. This method can make the determined third neural network model be the second neural network model with the best performance among multiple second neural network models.
  • the third neural network model is the neural network model with the smallest absolute difference between the operation amount and the first operation amount among the plurality of second neural network models.
  • the first operation amount is based on the computing power and delay of the first device. requirements, and/or, the computing power and latency requirements of the second device are determined.
  • This method can make the calculation amount of the third neural network model determined from the plurality of second neural network models closest to the maximum range supported by the computing capabilities of the first device and the second device.
  • the third neural network model is the neural network model with the smallest absolute difference between the parameter quantity and the first parameter quantity among the plurality of second neural network models.
  • the first parameter quantity is based on the computing power and storage space of the first device. , and/or, determined by the computing power and storage space of the second device. This method can make the parameter amount of the third neural network model determined from the plurality of second neural network models be closest to the maximum range supported by the storage capabilities of the first device and the second device.
  • the method also includes: sending computing power and delay requirements, or sending computing power and storage space, or sending computing power, delay requirements, and storage space.
  • the first neural network model is obtained by distilling the third neural network model; the operation amount of the third neural network model is greater than the first operation amount, and/or the third neural network model has The parameter amount is greater than the first parameter amount; the first calculation amount is determined based on the computing power and delay requirements of the first device, and/or, the computing power and delay requirements of the second device; the first parameter amount is determined based on the first device Determined by the computing power and storage space of the device, and/or, the computing power and storage space of the second device.
  • the first neural network model can be obtained by reducing the parameter amount and/or operation amount of the third neural network model. In this way, the parameter amount and operation amount of the first neural network model are within the computing capabilities of the first device and the second device. and within the range supported by storage capabilities.
  • the first neural network model is a third neural network model; the operation amount of the third neural network model is less than or equal to the first operation amount, and/or the parameter amount of the third neural network model is less than Or equal to the first parameter quantity; the first calculation quantity is determined according to the computing power and delay requirements of the first device, and/or, the computing power and delay requirements of the second device; the first parameter quantity is determined according to the computing power and delay requirements of the first device
  • the computing power and storage space of the second device are determined by the computing power and storage space of the second device.
  • the third neural network model can be directly used as the third neural network model.
  • a neural network model if the parameter amount and calculation amount of the third neural network model are within the range supported by the computing capabilities and storage capabilities of the first device and the second device, the third neural network model can be directly used as the third neural network model.
  • the first neural network model is determined by the model server based on the received model request information; the model request information includes the identification of the third neural network model, as well as the first calculation amount and/or the first parameter. quantity; wherein the first calculation quantity is determined according to the computing power and delay requirements of the first device, and/or the computing power and delay requirements of the second device, and the first parameter quantity is determined according to the computing power and delay requirements of the first device.
  • the computing power and storage space, and/or, the computing power and storage space of the second device are determined.
  • the information of each pre-trained neural network model in one or more pre-trained neural network models is predefined or obtained from the model server; the information of each pre-trained neural network model includes one or more of the following: Number: logo, industry Service type, RP size, input dimension, output dimension, parameter amount and operation amount.
  • the adjusted first neural network model is obtained by training the fourth neural network model based on the channel state information in the first information; the fourth neural network model is obtained based on the first neural network
  • the RP size applicable to the model and the communication resource information in the first information are obtained by adjusting the input dimension and/or output dimension of the first neural network model.
  • the first neural network model can be adjusted according to the communication resource information and/or channel state information communicated between the two devices, so that the adjusted first neural network model can be adapted to the first device and the second device. Device communication scenarios.
  • this application also provides a communication device.
  • the communication device has the function of realizing part or all of the embodiments described in the first aspect, or has the function of realizing part or all of the functional embodiments of the second aspect.
  • the functions described can be implemented by hardware, or can be implemented by hardware executing corresponding software.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the structure of the communication device may include a processing unit and a communication unit, and the processing unit is configured to support the communication device to perform corresponding functions in the above method.
  • the communication unit is used to support communication between the communication device and other communication devices.
  • the communication device may further include a storage unit coupled to the processing unit and the communication unit, which stores necessary program instructions and data for the communication device.
  • the communication device includes: a processing unit and a communication unit, and the processing unit is used to control the communication unit to send and receive data/signaling.
  • the processing unit is used to determine a first neural network model from one or more pre-trained neural network models; the processing unit is also used to adjust the first neural network model according to the first information; the first information includes the first device and the second device. communication resource information and/or channel state information communicated between; the communication unit is used to send the adjusted first neural network model, or send a sub-model in the adjusted first neural network model; the adjusted first neural network model A network model is used for communication between the first device and the second device.
  • the communication device includes: a processing unit and a communication unit, and the processing unit is used to control the communication unit to send and receive data/signaling.
  • the communication unit is used to receive the adjusted first neural network model, or to receive a sub-model in the adjusted first neural network model; the adjusted first neural network model is used to conduct communication between the first device and the second device. Communication; the adjusted first neural network model is obtained by adjusting the first neural network model based on the first information, and the first information includes communication resource information and/or channel state information for communication between the first device and the second device. ; The communication unit is also used to communicate based on the adjusted first neural network model, or to communicate based on a sub-model in the adjusted first neural network model.
  • the communication unit may be a transceiver or a communication interface
  • the storage unit may be a memory
  • the processing unit may be a processor
  • the communication device includes: a processor and a transceiver.
  • the processor is used to determine a first neural network model from one or more pre-trained neural network models; the processor is also used to adjust the first neural network model according to the first information; the first information includes the first device and the second device communication resource information and/or channel status information communicated between; the transceiver is used to send the adjusted first neural network model, or send a sub-model in the adjusted first neural network model; the adjusted first neural network model A network model is used for communication between the first device and the second device.
  • the communication device includes: a transceiver.
  • the transceiver is used to receive the adjusted first neural network model, or to receive a sub-model in the adjusted first neural network model; the adjusted first neural network model is used between the first device and the second device.
  • Information the first neural network device is determined from one or more pre-trained neural network models; the transceiver is further configured to communicate based on the adjusted first neural network model, or, based on the adjusted first neural network model Communicate with the submodels in the .
  • the communication device is a chip or a chip system.
  • the processing unit can also be embodied as a processing circuit or a logic circuit; the transceiver unit can be an input/output interface, interface circuit, output circuit, input circuit, pin or related circuit on the chip or chip system.
  • the processor may be used to perform, for example, but not limited to, baseband related processing
  • the transceiver may be used to perform, for example, but not limited to, radio frequency transceiver.
  • the above-mentioned devices may be arranged on separate chips, or at least part or all of them may be arranged on the same chip.
  • processors can be further divided into analog baseband processors and digital baseband processors.
  • the analog baseband processor can be integrated with the transceiver on the same chip, and the digital baseband processor can be set on an independent chip. With the continuous development of integrated circuit technology, more and more devices can be integrated on the same chip.
  • the digital baseband processor can be integrated with a variety of application processors (such as but not limited to graphics processors, multimedia processors, etc.) on the same chip.
  • application processors such as but not limited to graphics processors, multimedia processors, etc.
  • SoC System on a Chip
  • the embodiments of this application do not limit the implementation form of the above devices.
  • this application also provides a processor for executing the various methods mentioned above.
  • the process of sending the above information and receiving the above information in the above method can be understood as the process of the processor outputting the above information, and the process of the processor inputting the above information.
  • the processor When outputting the above information, the processor outputs the above information to the transceiver for transmission by the transceiver. After the above information is output by the processor, it may also need to undergo other processing before reaching the transceiver.
  • the processor receives the above information input, the transceiver receives the above information and inputs it into the processor. Furthermore, after the transceiver receives the above information, the above information may need to undergo other processing before being input to the processor.
  • processor output and reception, input operations rather than the transmitting and receiving operations performed directly by RF circuits and antennas.
  • the above-mentioned processor may be a processor specifically designed to perform these methods, or may be a processor that executes computer instructions in a memory to perform these methods, such as a general-purpose processor.
  • the above-mentioned memory can be a non-transitory memory, such as a read-only memory (Read Only Memory, ROM), which can be integrated on the same chip with the processor, or can be separately provided on different chips.
  • ROM Read Only Memory
  • the present application also provides a communication system, which includes at least one first device and at least one second device according to the above aspect.
  • the system may also include other devices that interact with the first device and the second device in the solution provided by this application.
  • the present application provides a computer-readable storage medium for storing instructions.
  • the instructions are calculated When the machine is running, the method described in any one of the above first aspect or the second aspect is implemented.
  • the present application also provides a computer program product including instructions that, when run on a computer, implement the method described in any one of the above first or second aspects.
  • the application provides a chip system.
  • the chip system includes a processor and an interface.
  • the interface is used to obtain a program or instructions.
  • the processor is used to call the program or instructions to implement the first aspect. function, or used to call the program or instruction to realize the function involved in the second aspect.
  • the chip system further includes a memory, and the memory is used to store necessary program instructions and data for the terminal.
  • the chip system may be composed of chips, or may include chips and other discrete devices.
  • Figure 1 is a schematic structural diagram of a communication system provided by an embodiment of the present application.
  • Figure 2 is a schematic structural diagram of another communication system provided by an embodiment of the present application.
  • Figure 3a is a schematic structural diagram of a fully connected neural network provided by an embodiment of the present application.
  • Figure 3b is a schematic diagram of a neural network training method provided by an embodiment of the present application.
  • Figure 3c is a schematic diagram of a gradient reverse transmission provided by an embodiment of the present application.
  • Figure 4 is an interactive schematic diagram of a communication method provided by an embodiment of the present application.
  • Figure 5 is a schematic diagram of resource division provided by an embodiment of the present application.
  • Figure 6a is a schematic diagram of a model adjustment provided by an embodiment of the present application.
  • Figure 6b is a schematic diagram of another model adjustment provided by the embodiment of the present application.
  • Figure 6c is a schematic diagram of another model adjustment provided by the embodiment of the present application.
  • Figure 7 is an interactive schematic diagram of another communication method provided by an embodiment of the present application.
  • Figure 8a is a schematic diagram of uplink communication or downlink communication provided by an embodiment of the present application.
  • Figure 8b is a schematic diagram of D2D communication provided by an embodiment of the present application.
  • Figure 9 is a schematic diagram of a protocol stack provided by an embodiment of the present application.
  • Figure 10 is a schematic structural diagram of a communication device provided by an embodiment of the present application.
  • Figure 11 is a schematic structural diagram of another communication device provided by an embodiment of the present application.
  • Figure 12 is a schematic structural diagram of a chip provided by an embodiment of the present application.
  • Embodiments of the present application can be applied to fourth generation (4G) communication systems such as long term evolution (LTE) systems, and fifth generation (5th generation, 5G) communications such as new radio (NR) systems.
  • 4G fourth generation
  • 5th generation, 5G fifth generation
  • NR new radio
  • the system can also be applied to short-distance communication systems such as wireless fidelity (WiFi) systems, communication systems that support the integration of multiple wireless technologies, or sixth generation (6th generation, 6G) communication systems that have evolved after 5G. Communication Systems.
  • WiFi wireless fidelity
  • 6G sixth generation
  • wireless communication systems include but are not limited to: narrow band-internet of things (NB-IoT), LTE and three major application scenarios of 5G mobile communication systems: enhanced mobile broadband (enhanced mobile broadband) , eMBB), ultra-reliable low latency communication (URLLC) and massive machine type of communication (mMTC), etc.
  • NB-IoT narrow band-internet of things
  • LTE Long Term Evolution
  • eMBB enhanced mobile broadband
  • URLLC ultra-reliable low latency communication
  • mMTC massive machine type of communication
  • a wireless communication system may include one or more network devices and one or more terminal devices. Among them, network equipment and terminal equipment can communicate with each other, and different terminal equipment can also communicate with each other.
  • Figure 1 takes a wireless communication system including one network device and two terminal devices as an example.
  • Figure 2 is a schematic structural diagram of another communication system provided by an embodiment of the present application.
  • the communication system includes but is not limited to a third device and a second device.
  • the number and shape of the devices shown in Figure 2 are only examples and do not constitute a limitation on the embodiments of the present application.
  • two or more third devices, two or more second devices, and other devices may be included.
  • a first device may be included.
  • the third device may be a network device or a terminal device; the second device may be a network device or a terminal device, and the first device may be a network device or a terminal device.
  • the network device is a device with wireless transceiver functions, which can be an evolved base station (evolved Node B, eNB or eNodeB) in LTE, or a base station in a 5G network or a future evolved public land mobile network ( Base stations in public land mobile network (PLMN), broadband network gateway (BNG), aggregation switches or non-3rd generation partnership project (3rd generation partnership project, 3GPP) access equipment, etc.
  • the network equipment in the embodiments of this application may include various forms of base stations, such as: macro base stations, micro base stations (also called small stations), relay stations, access points, and base stations implemented in communication systems evolved after 5G.
  • Functional equipment access nodes in WiFi systems, transmitting and receiving points (TRP), transmitting points (TP), mobile switching centers and device-to-device (D2D), Equipment that performs base station functions in vehicle-to-everything (V2X) and machine-to-machine (M2M) communications, etc.
  • C-RAN cloud radio access network
  • CU centralized unit
  • DU distributed unit
  • NTN non-terrestrial network
  • Terminal devices may include various handheld devices with wireless communication capabilities, vehicle-mounted devices, wearable devices, computing devices, or other processing devices connected to wireless modems.
  • Terminal equipment can also refer to user equipment (UE), access terminal, customer-premises equipment (CPE), subscriber unit (subscriber unit), user agent, cellular phone, smartphone ( smart phone), wireless data card, personal digital assistant (PDA) computer, tablet computer, wireless modem (modem), handheld device (handset), laptop computer (laptop computer), machine type communication (machine) type communication (MTC) terminals, communication equipment carried on high-altitude aircraft, wearable devices, drones, robots, smart point of sale (POS) machines, terminals in D2D, terminals in V2X, virtual reality ( virtual reality (VR) terminal equipment, augmented reality (AR) terminal equipment, wireless terminals in industrial control (industrial control), wireless terminals in self-driving (self driving), remote medical (remote medical) Wireless terminals, wireless terminals in smart grid, wireless terminals in transportation safety, wireless terminals in smart city, wireless terminal
  • Fully connected neural network can also be called multilayer perceptron (MLP).
  • MLP contains an input layer, an output layer and multiple hidden layers, and each layer contains several nodes, which can be called neurons. Among them, neurons in two adjacent layers are connected in pairs.
  • the MLP shown in Figure 3a includes an input layer, an output layer, and two hidden layers. Among them, the input layer includes 4 neurons, each hidden layer includes 8 neurons, and the output layer includes 6 neurons.
  • w is the weight matrix
  • b is the bias vector
  • f() is the activation function
  • the neural network can represent the mapping relationship from the input data set to the output data set.
  • the initialization of neural networks is usually random and needs to be trained before being put into use.
  • the training of neural network refers to the process of using existing data to determine the mapping relationship from random w and b.
  • the specific method of training the neural network includes: using a loss function (loss function) to evaluate the output results of the neural network, back-propagating the error, and iteratively optimizing w and b through the gradient descent method until the loss function reaches minimum value.
  • the gradient descent process can be expressed as:
  • ⁇ is the parameters to be optimized (such as w and b), L is the loss function; eta is the learning efficiency, which is used to control the step size of gradient descent.
  • the gradient of the previous layer parameters can be calculated recursively from the gradient of the subsequent layer parameters.
  • the gradient of the weight w ij between neuron j and neuron i in Figure 3c can be expressed as:
  • s i is the weighted sum of inputs on neuron i.
  • s i is the weighted sum of inputs on neuron i.
  • It can be called the gradient of the middle layer, that is, s i is regarded as the middle layer.
  • a large model refers to a neural network model with a huge number of parameters that is pre-trained with a large amount of data, such as a fully connected neural network model with a huge number of parameters.
  • Large models have strong information extraction and expression capabilities and can be used to complete a variety of tasks. For example, a large model in the field of natural language processing: Generative Pre-trained Transformer 3 (GPT-3), which contains 175 billion parameters. Using GPT-3 can complete translation, article writing, search, etc. Various natural language processing related tasks.
  • Embodiments of the present application provide a communication method that can determine a first neural network model from one or more pre-trained neural network models, and determine the first neural network model based on the communication resource information for communication between the first device and the second device. and/or the channel state information adjusts the first neural network model, and the adjusted first neural network model can be used for communication between the first device and the second device.
  • the method is capable of determining a neural network model for communication between the first device and the second device with low complexity and low signaling overhead.
  • Figure 4 is an interactive schematic diagram of a communication method provided by an embodiment of the present application.
  • the communication method starts from Chapter 4.
  • the interaction between the third device and the second device is explained.
  • the communication method includes the following steps:
  • the third device determines a first neural network model from one or more pre-trained neural network models.
  • one or more pre-trained neural network models can be one or more large models that have been trained.
  • the first neural network model is determined from one or more pre-trained neural network models based on the communication system parameters of the first device and/or the communication system parameters of the second device; wherein , the communication system parameters of the first device include the system bandwidth and frame structure supported by the first device, and the communication system parameters of the second device include the system bandwidth and frame structure supported by the second device.
  • the communication system parameters may also include carrier frequency, system parameters (numerology), antenna configuration, reference signal location, etc.
  • the system parameters may include subcarrier spacing, the length of each time slot (slot), the number of symbols (symbols) included in each time slot, the length of the cyclic prefix (CP) in each symbol, etc.
  • the antenna configuration may include the number of antenna ports
  • the reference signal location may include the time domain resource location, frequency domain resource location, air domain resource location, etc. where the reference signal is located. That is to say, the communication system parameters of the first device represent the resources that the first device can work on, and the communication system parameters of the first device are relevant parameters of the resources on which the first device can work.
  • the communication system parameters of the second device represent the resources that the second device can work on, and the communication system parameters of the second device are relevant parameters of the resources on which the second device can work.
  • the first neural network model determined in step S101 is used to determine the neural network model used for communication between the first device and the second device.
  • the first device may be a third device, or may be another device different from the second device and the third device.
  • the first device may be other network equipment or terminal equipment.
  • the communication system parameters of the first device and/or the communication system parameters of the second device may be communication system parameters determined by the third device from a plurality of predefined groups of communication system parameters, for a group determined by the first device.
  • the communication system parameters are applicable to the first device, and a set of communication system parameters determined for the second device are applicable to the second device.
  • the communication system parameters of the second device may be sent by the second device to the third device. If the first device is another device different from the second device and the third device, the communication system parameters of the first device may be sent by the first device to the third device.
  • the first neural network model is determined based on a third neural network model selected from one or more second neural network models; the one or more second neural network models are determined from one or more pre-trained neural network models. determined in the neural network model.
  • the input dimension of each second neural network model in one or more second neural network models is the same as its corresponding first input dimension
  • the output dimension of each second neural network model is the same as its corresponding first input dimension.
  • the output dimensions are the same.
  • the first input dimension and/or the first output dimension corresponding to each second neural network model is based on the resource patch (RP) size applicable to the second neural network model, as well as the communication system parameters of the first device and /or determined by the communication system parameters of the second device.
  • RP resource patch
  • This implementation can enable the determined first neural network model to be adapted to the communication system of the first device and the second device, which is beneficial to improving the application of the adjusted first neural network model to the communication between the first device and the second device. performance, thereby improving the quality of communication between the first device and the second device.
  • RP may also be called a resource package.
  • the first input dimension and/or the first input dimension corresponding to each pre-trained neural network model in the one or more pre-trained neural network models may be determined based on the RP size applicable to the pre-trained neural network model and the communication system parameter with the smallest value among the communication system parameters of the first device and the communication system parameters of the second device.
  • the communication system parameters include system bandwidth, frame structure, and number of antenna ports. The frame lengths corresponding to the frame structures supported by the first device and the second device respectively have the same value, the system bandwidth has different values, and the number of antenna ports has the same value.
  • the first input dimension and sum corresponding to the pre-trained neural network model can be determined based on the frame length with the same value, the system bandwidth with the smallest value, the number of antenna ports with the smallest value, and the RP size applicable to the pre-trained neural network model. /or the first output dimension.
  • the first device is device 1 and the second device is device 2.
  • the frame structure supported by device 1 has the same frame length as the frame structure supported by device 2, both of which are frame length 1; the system bandwidth supported by device 1 is 1 is less than the system bandwidth 2 supported by device 2, and the number of antenna ports 1 supported by device 1 is greater than the number 2 of antenna ports supported by device 2, then the RP size, frame length 1, system bandwidth 1 and The number of antenna ports is 2, which determines the first input dimension and/or the first output dimension corresponding to the pre-trained neural network model.
  • the RP size applicable to each pre-trained neural network model can be determined based on the time domain length and frequency domain width of the RP applicable to the pre-trained neural network model; or, it can also be determined based on the pre-trained neural network model.
  • the time domain length of RP, the frequency domain width of RP, and the spatial domain width of RP that the network model is applicable to are determined.
  • the RP sizes applicable to different pre-trained neural network models may be different.
  • the unit of the time domain length of the RP is, for example, milliseconds (ms)
  • the unit of the frequency domain width of the RP is, for example, kilohertz (kHz)
  • the spatial domain width of the RP can be expressed as the number of spatial streams occupied by the RP, where The unit is, for example, stream.
  • B sys is the system bandwidth with the smallest median value between the system bandwidth supported by the first device and the system bandwidth supported by the second device
  • S sys is the smallest median frame length corresponding to the frame structure supported by the first device and the second device respectively.
  • the frame length, n is the smallest number of antenna ports among the number of antenna ports supported by the first device and the number of antenna ports supported by the second device.
  • B sys is the value of the same system bandwidth; if the frame length corresponding to the frame structure supported by the first device is the same as that supported by the second device, The frame length corresponding to the frame structure is the same, then S sys is the value of the same frame length; if the number of antenna ports supported by the first device is the same as the number of antenna ports supported by the second device, then n is the same number of antenna ports. value.
  • B rp is the frequency domain width of the RP applicable to the pre-trained neural network model
  • S rp is the time domain length of the RP applicable to the pre-trained neural network model
  • ceil() is an upward rounding function.
  • the resource grid shown in Figure 5 is a resource network on one antenna port.
  • the number of antenna ports n 1, it means that the resource grid for the first device and the second device to work on this antenna port can be divided into 2 RPs.
  • This RP is an RP suitable for a certain pre-trained neural network model, then, The first input dimension or the first output dimension corresponding to the pre-trained neural network model can be determined to be 2; if the number n of antenna ports is 2, it means that the resource network on these two antenna ports can provide work for the first device and the second device.
  • the grid can be divided into 4 RPs in total, then the first input dimension or the first output dimension corresponding to the pre-trained neural network model can be determined to be 4.
  • the value of the communication system parameter of the network device is greater than the value of the communication system parameter of the terminal device.
  • the system bandwidth supported by the network device is greater than the system bandwidth supported by the terminal device, and the frame structure corresponding to the frame structure supported by the network device The length is greater than the frame length corresponding to the frame structure supported by the terminal device.
  • B sys in the above formula is the system bandwidth supported by the terminal equipment in the first device and the second device
  • S sys is the frame length corresponding to the frame structure supported by the terminal equipment in the first device and the second device
  • n is the number of antenna ports supported by the terminal device in the first device and the second device.
  • the second neural network model supports a first service type; wherein the first service type is a service type required by the first device and the second device.
  • This method is beneficial to making the determined first neural network model support the service types required by the first device and the second device.
  • the first service type may be based on the service information of the first device and the service of the second device. The information is certain.
  • the service information of any one of the first device and the second device may include one of the service types such as eMBB, URLLC, and mMTC required by the device; or the service information of any device may include the device's response to the neural network model.
  • the requirements for performance indicators such as communication delay and throughput reflect the performance requirements of the device.
  • the communication delay of the neural network model includes the time for the sender to process the signal to be sent based on the neural network model, the time to transmit the processed signal, and the time for the receiver to process the received signal based on the neural network model. Its unit is, for example, ms. ;
  • the throughput of the neural network model refers to the amount of data sent and/or received based on the neural network model within a certain period of time, and its unit is, for example, million bits per second (Mbps).
  • the first service type may be determined by the third device based on its own service information and service information sent by the second device.
  • the first service type may be the service information sent by the third device to the third device according to the service information sent by the first device to the third device and the service sent by the second device to the third device. The information is certain.
  • the service types supported by each pre-trained neural network model and the expression form of the first service type include those described in Embodiment 1.1 and Embodiment 1.2.
  • the service type supported by each pre-trained neural network model is one of the service types such as eMBB, URLLC, and mMTC
  • the first service type is one of the service types such as eMBB, URLLC, and mMTC.
  • the service type supported by the pre-trained neural network model is the same as the first service type, it means that the pre-trained neural network model supports the first service type, that is, the service type that supports the requirements of the first device and the second device.
  • the service information of each of the first device and the second device includes one of the service types such as eMBB, URLLC, and mMTC
  • the service information of the first device includes the service type and the service type of the second device.
  • the information includes the same service type, and the same service type is the first service type.
  • the same service type included in the service information of the first device and the service information of the second device may be pre-agreed between the first device and the second device.
  • the third device can use the service information of the first device to and the second device's service information, determine the first service type from eMBB, URLLC, mMTC and other service types, and the determined first service type can meet the performance requirements of the first device and the second device.
  • the third device may determine whether the service type included in the service information of the first device can meet the performance requirements of the second device; if so, use the service type included in the service information of the first device as the first service type.
  • the service information of the first device includes the first device's requirements for performance indicators such as communication delay and throughput of the neural network model
  • the service information of the second device includes the situation of one of the service types such as eMBB, URLLC, and mMTC. are similar and will not be described again.
  • the service type supported by each pre-trained neural network model is expressed in the form of one or more performance indicators
  • the first service type is expressed in the value range of each performance indicator in the one or more performance indicators.
  • each performance indicator represented by the first service type The value range of is included in the intersection of the performance requirements of the first device and the performance requirements of the second device.
  • the service information of the first device includes: the communication delay of the neural network model is less than value #1, and the throughput is greater than value #2; the service information of the second device includes: the communication delay of the neural network model is less than value #3, and the throughput is greater than value #2. Amount greater than value #4.
  • the first service type can be expressed as: the communication delay of the neural network model is less than value #5, and the throughput is greater than value #6, where value #5 is less than or equal to the smallest one of value #1 and value #3, Value #6 is greater than or equal to the greater of value #2 and value #4.
  • the service type included in the service information of the first device is different from the service information of the second device.
  • the business types included in are the same.
  • the value range of each performance indicator represented by the first service type is included in the value range of each performance indicator supported by the same service type.
  • the service information of the first device and the service information of the second device both include service type 1, and the communication delay of service type 1 supporting the neural network model is less than the value #1; then, the first service type can be expressed as: neural network model The communication delay is less than value #3, where value #3 is less than or equal to value #1.
  • the service information of the first device includes one of the service types such as eMBB, URLLC, and mMTC
  • the service information of the second device includes the second device's requirements for performance indicators such as communication delay and throughput of the neural network model
  • the value range of each performance indicator represented by the first service type is included in the intersection of the value range of each performance indicator supported by the service type included in the service information of the first device and the performance requirements of the second device.
  • the service information of the first device includes service type 1, and the communication delay of the service type 1 supporting the neural network model is less than value #1.
  • the service information of the second device includes: the communication delay of the neural network model is less than value #2.
  • the first service type can be expressed as: the communication delay of the neural network model is less than value #3, where value #3 is less than or equal to the smallest one of value #1 and value #2.
  • the service information of the first device includes the first device's requirements for performance indicators such as communication delay and throughput of the neural network model
  • the service information of the second device includes the situation of one of the service types such as eMBB, URLLC, and mMTC. are similar and will not be described again.
  • the communication method may further include: the second device sending service information and communication system parameters to a third device;
  • the service information determines the first service type, and then determines one or more pre-trained neural network models from one or more pre-trained neural network models based on the first service type and the communication system parameters of the first device and/or the communication system parameters of the second device.
  • the second neural network model determines a third neural network model from one or more second neural network models.
  • the third device determines one or more second neural networks from one or more pre-trained neural network models according to the first service type and the communication system parameters of the first device and/or the communication system parameters of the second device.
  • the model may include: the third device determines a pre-trained neural network model that supports the first service type from one or more pre-trained neural network models, and then based on the communication system parameters of the first device and/or the communication system of the second device Parameters: determine one or more second neural network models whose input dimensions are equal to their corresponding first input dimensions and/or whose output dimensions are equal to their corresponding first output dimensions from the pre-trained neural network models that support the first business type.
  • the third device determines from one or more pre-trained neural network models that the input dimension is equal to its corresponding first input dimension sum according to the communication system parameter of the first device and/or the communication system parameter of the second device. /or output a pre-trained neural network model with a dimension equal to its corresponding first output dimension, and then determine one or more second neural network models that support the first service type from the determined pre-trained neural network models.
  • the third device determines a pre-trained neural network model that supports the first service type from one or more pre-trained neural network models; at the same time, the third device determines a pre-trained neural network model that supports the first service type according to the communication system parameters of the first device and/or the third device.
  • Communication system parameters of the two devices determine from one or more pre-trained neural network models a pre-trained neural network model whose input dimension is equal to its corresponding first input dimension and/or whose output dimension is equal to its corresponding first output dimension; then , the third device then uses the same pre-trained neural network model obtained in the above two operations as the second neural network model.
  • the third device determines the pre-trained neural network model 1 and the pre-trained neural network model 2 that support the first service type; at the same time, the third device determines that the input dimension is equal to its corresponding first input dimension and/or the output dimension is equal to The corresponding pre-trained neural network model 2 and pre-trained neural network model 3 of the first output dimension; then, use the pre-trained neural network model 2 as the second neural network network model.
  • the third neural network model is the second neural network model. If there are multiple second neural network models, the third neural network model is the second neural network model with the largest number of parameters among the multiple second neural network models; this method can make the determined third neural network model be the second neural network model of the multiple second neural network models. The second best performing neural network model among neural network models.
  • the third neural network model is the second neural network model with the smallest absolute difference between the operation amount and the first operation amount among the plurality of second neural network models, wherein the first operation amount may be based on the first device
  • the computing power and delay requirements, and/or the computing power and delay requirements of the second device are determined; this method can make the determined third neural network model be: among the multiple second neural network models, the calculation amount is closest to The maximum range of second neural network models supported by the computing capabilities of the first device and the second device.
  • the third neural network model is the second neural network model with the smallest absolute difference between the parameter quantity and the first parameter quantity among the plurality of second neural network models, wherein the first parameter quantity may be based on the first device
  • the computing power and storage space, and/or the computing power and storage space of the second device are determined; this method can make the determined third neural network model be: among the multiple second neural network models, the parameter amount is closest to the third neural network model.
  • the computing power of any one of the first device and the second device can be expressed as the number of times the device can operate floating point per second, and its unit is, for example, floating point operation per second. , flop/s).
  • the delay requirement of any one of the first device and the second device refers to the calculation delay requirement of the neural network model by the device.
  • the calculation delay of the neural network model refers to the time required for the neural network model to calculate a certain floating point number.
  • its unit is, for example, second (second, s).
  • the storage space of any one of the first device and the second device refers to the storage space that the device can use when communicating based on the neural network model, and its unit may be a byte.
  • the unit of the first operation quantity is, for example, flop, and the unit of the first parameter quantity is, for example, byte.
  • the computing power, latency requirements and storage space of the second device may be sent by the second device to the third device. If the first device is another device different from the second device and the third device, the computing power, delay requirements and storage space of the first device may be sent by the first device to the third device.
  • the above related descriptions of the first operation quantity and the first parameter quantity can be applied to any position where the first operation quantity and the first parameter quantity are mentioned in the embodiments of the present application, and will not be described again below.
  • the pre-trained neural network model when there is only one pre-trained neural network model, if the input dimension of the pre-trained neural network model is the same as its corresponding first input dimension, and/or the output dimension of the pre-trained neural network model is the same as its corresponding first input dimension, If the output dimensions are the same and the pre-trained neural network model supports the first service type, the pre-trained neural network model is the second neural network model and the third neural network model.
  • the third device can determine one or more second neural network models from the multiple pre-trained neural network models, and then determine the one or more second neural network models from the one or more second neural network models. Determine the third neural network model.
  • Embodiment 2.1 For relevant explanations about the third device determining the first neural network model based on the third neural network model selected from one or more second neural network models, please refer to the descriptions in Embodiment 2.1 to Embodiment 2.3.
  • the first neural network model is obtained by distillation of the third neural network model; the calculation amount of the third neural network model is greater than the first calculation amount, and/or the parameter amount of the third neural network model is greater than the first parameter amount. Distilling the third neural network model can reduce the amount of calculations and/or parameters of the third neural network model while ensuring that the performance of the third neural network model is not affected as much as possible, so that the obtained first neural network model can operation
  • the quantity is within the range supported by the computing capabilities of the first device and the second device, and the parameter quantity is within the range supported by the storage capabilities of the first device and the second device.
  • the process of distilling the third neural network model to obtain the first neural network model can also be performed in combination with reference data, where the reference data reflects the environmental characteristics and data characteristics of the first device and the second device.
  • This method can make the first neural network model more adaptable to the communication scenario of the first device and the second device, thereby helping to improve the performance of the adjusted first neural network model when it is applied to the communication between the first device and the second device. performance, thereby improving communication quality between the first device and the second device.
  • the reference data may be determined based on the historical data of the first device and the historical data of the second device collected by the third device.
  • reference data related to the channel can be obtained based on the historical channel estimate value of the channel between the first device and the second device; based on the historical service flow between the first device and the second device, the reference data related to the channel between the first device and the second device can be obtained.
  • Reference data related to data services between the second device can be obtained based on the historical channel estimate value of the channel between the first device and the second device; based on the historical service flow between the first device and the second device.
  • the first neural network model is a third neural network model; the calculation amount of the third neural network model is less than or equal to the first calculation amount, and/or, the parameter amount of the third neural network model is less than or equal to the first parameter quantity. It can be seen that the calculation amount and parameter amount of the third neural network model are already within the range supported by the computing capabilities and storage capabilities of the first device and the second device respectively. Then, the first device can directly use the third neural network model as The first neural network model.
  • the first operation quantity and the first parameter quantity please refer to the above-mentioned relevant explanations and will not be described again here.
  • the third neural network model may not be used as the first neural network model. network model, but the first neural network model is obtained by distilling the third neural network model.
  • the communication method may also include: the third device performs the following steps when the calculation amount of the third neural network model is greater than that of the third neural network model: a calculation amount, and/or, when the parameter amount of the third neural network model is greater than the first parameter amount, the third neural network model is distilled to obtain the first neural network model; when the calculation amount of the third neural network model is less than or is equal to the first operation amount, and/or when the parameter amount of the third neural network model is less than or equal to the first parameter amount, the third neural network model is used as the first neural network model.
  • the first neural network model is determined by the model server based on the received model request information; the model request information includes the identification of the third neural network model, as well as the first operation amount and/or the first parameter amount.
  • the model request information is sent by the third device to the model server, and the first neural network model is received by the third device from the model server. It can be seen that this embodiment can be applied to the situation where one or more pre-trained neural network models are stored by the model server, and the third device can obtain the first neural network model from the model server.
  • the first operation quantity and the first parameter quantity please refer to the above-mentioned relevant explanations and will not be described again here.
  • the communication method also includes: the third device sends a model request to the model server Information, the model request information includes the identification of the third neural network model, and the first operation quantity and/or the first parameter quantity.
  • the model server may determine the third neural network model from one or more pre-trained neural network models according to the identification of the third neural network model in the model request information, and then determine the third neural network model according to the third neural network model and the first neural network model in the model request information.
  • the first neural network model is determined based on the calculation amount and/or the first parameter amount
  • the first neural network model is sent to the third device.
  • the third device receives the first neural network model from the model server.
  • the model server determines the first neural network model based on the third neural network model, the first calculation amount and the first parameter amount in the model request information, which may include: the model server determines the first neural network model when the calculation amount of the third neural network model is greater than the first parameter amount.
  • the amount of operation, and/or, when the amount of parameters of the third neural network model is greater than the amount of the first parameters the third neural network model is distilled to obtain the first neural network model; when the amount of operations of the third neural network model is less than or equal to first operand, and/or, third When the parameter amount of the neural network model is less than or equal to the first parameter amount, the third neural network model is used as the first neural network model.
  • the information of the pre-trained neural network model includes one or more of the following: identification (can also be called an identifier (identifier, ID)), business type, RP size, input dimension, output dimension, parameter amount and operation quantity. Among them, different pre-trained neural network models have different identifiers.
  • the information for pre-training the neural network model is predefined.
  • the third device can obtain the information of each pre-trained neural network model defined from standard text, product descriptions.
  • the information for pre-trained neural network models is obtained from the model server.
  • the model server actively sends the information of the pre-trained neural network model to the third device; or the third device requests the model server to obtain the information of the pre-trained neural network model, and then the model server sends the information to the third device.
  • the identification of the third neural network model may be determined by the third device based on the information of each pre-trained neural network model in one or more pre-trained neural network models.
  • the pre-trained neural network model may include sub-models.
  • the pre-trained neural network model may include a transmitter sub-model and a receiver sub-model.
  • the transmitter sub-model is used to process the signal to be sent and send the processed signal
  • the receiver sub-model is used to receive the signal and process the received signal.
  • the use of “transmitter sub-model” and “receiver sub-model” to name the two sub-models here is only an exemplary naming method. "First sub-model” and “second sub-model” can also be used. Naming is not limited here. The following article will take the naming method of "transmitter sub-model” and "receiver sub-model” as an example to explain.
  • the input dimension of the pre-trained neural network model refers to the input dimension of the transmitter sub-model in the pre-trained neural network model
  • the output dimension refers to the output dimension of the receiver sub-model in the pre-trained neural network model
  • the parameter amount refers to the parameter amount of the transmitter sub-model. and the sum of the parameters of the receiver sub-model.
  • the calculation amount refers to the sum of the calculation amount of the transmitter sub-model and the calculation amount of the receiver sub-model.
  • processing such as distillation and adjustment of a certain pre-trained neural network model includes processing both the transmitter sub-module and the receiver sub-model in the pre-trained neural network model.
  • distilling the third neural network model includes distilling the transmitter sub-model and the receiver sub-model in the third neural network model. It can be seen that the first neural network model also includes a transmitter sub-model and a receiver sub-model. Wherein, if the first neural network model is obtained by distilling the third neural network model, and the transmitter sub-model in the first neural network model is obtained by distilling the transmitter sub-model in the third neural network model, the first neural network The receiver sub-model in the model is obtained by distilling the receiver sub-model in the third neural network model.
  • the embodiment of this application also provides Table 1, which exemplarily shows the information of each pre-trained neural network model in one or more pre-trained neural network models.
  • the pre-trained neural network models included in the same table The trained neural network models support the same service types, while the pre-trained neural network models included in different tables support different service types, as shown in Table 2 and Table 3.
  • the pre-trained neural network models in Table 2 all support eMBB, and the pre-trained neural network models in Table 3 all support mMTC.
  • the entries in Table 2 and Table 3 may not contain the service type, or they may contain the service type, which is not limited here.
  • the information of multiple pre-trained neural network models is divided into multiple tables according to the frequency domain width of the RP applicable to the pre-trained neural network model, and so on. There is no limitation here on how the information of each pre-trained neural network model in one or more pre-trained neural network models is embodied.
  • the third device adjusts the first neural network model according to the first information, where the first information includes communication resource information and/or channel state information for communication between the first device and the second device.
  • the third device adjusting the first neural network model according to the first information includes: the third device adjusts the transmitter sub-model and the receiver sub-model in the first neural network model. Then, the transmitter sub-model in the adjusted first neural network model is obtained by adjusting the transmitter sub-model in the first neural network model, and the receiver sub-model in the adjusted first neural network model is the first neural network model. It is obtained by adjusting the receiver sub-model in .
  • the adjusted first neural network model is obtained by training the fourth neural network model based on the channel state information in the first information; the fourth neural network model is obtained based on the first neural network model.
  • the RP size applicable to the network model and the communication resource information in the first information are obtained by adjusting the input dimensions and/or output dimensions of the first neural network model.
  • the communication resource information in the first information includes resources used for communication between the first device and the second device.
  • the resources may include time domain resources, frequency domain resources, air domain resources and other resources.
  • the resources used for communication between the first device and the second device may be different from the resources that the device can work represented by the communication system parameters in the previous article.
  • the resources used for communication between the first device and the second device are the resources that the first device and the second device can work on. Some or all of the resources that the two devices can work together are also the resources that the first device and the second device can actually use in communication based on the neural network model.
  • the resources used for communication between the first device and the second device may be allocated by the third device to the first device and the second device, or may be allocated by another device different from the third device to the first device. If assigned to the second device, there is no limitation here.
  • the RP size applicable to the first neural network model please refer to the above-mentioned relevant explanations and will not be repeated here.
  • the channel state information in the first information may be obtained by channel estimation by the receiving device among the first device and the second device. If the communication between the first device and the second device using the neural network model is two-way communication, that is, the first device acts as both the sender and the receiver in the communication, and the second device also acts as both the sender and the receiver in the communication. , the channel state information in the first information includes information obtained by channel estimation performed by the first device and the second device respectively. The information obtained by the second device through channel estimation may be sent by the second device to the third device. If the first device is another device different from the second device and the third device, the information obtained by the first device performing channel estimation may be sent by the first device to the third device.
  • the communication method may also include: if the second input dimension determined according to the RP size applicable to the first neural network model and the communication resource information in the first information is smaller than the input dimension of the first neural network model, then the third The device adjusts the input dimension of the first neural network model to the second input dimension, that is, the input dimension of the fourth neural network model is the second input dimension; otherwise, the input dimension of the fourth neural network model is equal to the input of the first neural network model Dimension, that is, there is no need to adjust the input dimension of the first neural network model.
  • the communication method may also include: if the second output dimension determined according to the RP size applicable to the first neural network model and the communication resource information in the first information is smaller than the output dimension of the first neural network model, then the third The device adjusts the output dimension of the first neural network model to the second output dimension, that is, the output dimension of the fourth neural network model is the second output dimension; otherwise, the output dimension of the fourth neural network model is equal to the output of the first neural network model. dimension, that is, there is no need to adjust the output dimension of the first neural network model.
  • the input dimension of the first neural network model is 4. According to the applicable RP size of the first neural network model and the communication resources in the first information If the second input dimension determined by the information is 2, then the input dimension of the first neural network model is adjusted to 2, that is, the input dimension of the fourth neural network model is 2.
  • the communication resource information includes information for the first device and the second device.
  • the adjusted first neural network model is obtained by training the fourth neural network model based on the channel state information in the first information; the fourth neural network model is obtained based on the first neural network model.
  • the RP size applicable to the neural network model and the communication resource information in the first information are obtained by adjusting the input value and/or output value of the first neural network model.
  • the communication method may also include: if the second input dimension determined according to the RP size applicable to the first neural network model and the communication resource information in the first information is smaller than the input dimension of the first neural network model, then the third The device sets M1 input values of the first neural network model to zero, that is, there are M1 input values of the fourth neural network model that are zero. Among them, M1 is equal to the value obtained by subtracting the second input dimension from the input dimension of the first network model.
  • the communication method may also include: if the second output dimension determined according to the RP size applicable to the first neural network model and the communication resource information in the first information is smaller than the output dimension of the first neural network model, then the third The device sets M2 output values of the first neural network model to zero, that is, M2 output values of the fourth neural network model are zero.
  • M2 is equal to the value obtained by subtracting the second output dimension from the output dimension of the first network model.
  • the input dimension of the first neural network model is 4.
  • the second input dimension determined by the information is 2, which means that the second input dimension is smaller than the input dimension of the first neural network model, then the two input values of the first neural network model are set to zero, that is, the two inputs of the fourth neural network model The value is 0.
  • the communication method may also include: the third device determines the RP corresponding to each input and/or each output of the first neural network model in the first resource, and the RP corresponding to different inputs and/or outputs is different.
  • the first resource is a resource represented by the communication system parameter of the first device and a resource represented by the communication system parameter of the second device, and the first device and the second device can work together. If the second input dimension is smaller than the input dimension of the first neural network model, then the first resource will be The input values of the inputs corresponding to the resources that are not allocated to the first device and the second device are set to zero, and/or, if the second output dimension is smaller than the output dimension of the first neural network model, the first resource that is not allocated is set to zero.
  • the output values of outputs corresponding to the resources allocated to the first device and the second device are set to zero.
  • the input dimension of the first neural network model is 4, and in the resource (ie, the first resource) where the first device and the second device can work together, the input #1 of the first neural network model Corresponds to RP#1, input #2 corresponds to RP#2, input #3 corresponds to RP#3, and input #4 corresponds to RP#4.
  • the resources used for communication between the first device and the second device include RP#1 and RP #2, excluding RP#3 and RP#4, then set the input value of #3 corresponding to the first neural network model corresponding to RP#3 to zero, and set the input value of the first neural network model corresponding to RP#4
  • the input value of #4 is set to zero.
  • the first information may also include device status information of the first device and/or device status information of the second device.
  • the fourth neural network model in addition to training the fourth neural network model according to the channel state information in the first information, the fourth neural network model is also trained.
  • the fourth neural network model may be trained in combination with device status information of the first device and/or device status information of the second device.
  • the device status information of any one of the first device and the second device may include the load (such as central processing unit (CPU) load, memory, etc.) and power of the device.
  • the fourth neural network model can also be distilled to further reduce the amount of parameters and operations of the fourth neural network model. quantity.
  • the third device sends the adjusted first neural network model; correspondingly, the second device receives the adjusted first neural network model.
  • the third device sends the adjusted sub-model of the first neural network model; correspondingly, the second device receives the adjusted sub-model of the first neural network model.
  • the adjusted first neural network model is used for communication between the first device and the second device.
  • the second device communicates based on the adjusted first neural network model; or, the second device communicates based on a sub-model in the adjusted first neural network model.
  • the following will describe respectively whether the third device sends the adjusted first neural network model or the sub-model in the adjusted first neural network model.
  • the third device sends the adjusted first neural network model.
  • the third device sends the adjusted first neural network model to the second device. If the first device is another device different from the second device and the third device, the third device sends the adjusted first neural network model to the first device and the second device respectively.
  • the first device can determine the transmitter sub-model from the adjusted first neural network model, and based on The transmitter submodel handles the signal to be transmitted and sends the signal to the second device.
  • the second device may determine a receiver sub-model from the adjusted first neural network model and receive a signal from the first device based on the receiver sub-model and process the received signal.
  • the situation in which the second device serves as the sender in the communication and the first device serves as the receiver in the communication is similar and will not be described again.
  • the third device sends the adjusted sub-model of the first neural network model.
  • the sub-model in the adjusted first neural network model sent by the third device to the second device is: The transmitter sub-model in the first neural network model; if the second device serves as the receiver in the communication, the adjusted sub-model in the first neural network model sent by the third device to the second device is: the adjusted first neural network model A receiver submodel in a neural network model.
  • the third device sends The adjusted transmitter sub-model in the first neural network model is sent to the second device, and the adjusted receiver sub-model in the first neural network model is sent to the second device. If second The device serves as the sender in the communication and the first device serves as the receiver in the communication. The third device sends the adjusted transmitter sub-model of the first neural network model to the second device and sends the adjusted first neural network model to the first device. Receiver submodel in the network model.
  • the communication between the first device and the second device is two-way communication, that is, the first device serves as both the sender and the receiver, and the second device also serves as both the sender and the receiver, if the two-way communication In the two reverse one-way communications, the minimum value of the system bandwidth supported by the first device and the second device is the same, the minimum value of the frame length corresponding to the supported frame structure is the same, and the minimum value of the system bandwidth supported by the first device and the second device is the same. If the resources for communication between the two devices are of the same size, then the first device and the second device can conduct bidirectional communication based on the determined adjusted first neural network model.
  • the first device when the first device acts as a sender, it uses the adjusted transmitter sub-model of the first neural network model to communicate, and when it acts as a receiver, it uses the adjusted receiver sub-model of the first neural network model to communicate.
  • the second device serves as the sender and the receiver respectively is similar and will not be described again.
  • the third device needs to separately execute the communication method provided by the embodiment of the present application for the two reverse one-way communications to obtain the neural network for communication between the first device and the second device under each one-way communication. Model.
  • the third device sends the adjusted first neural network model, or after sending the sub-model in the adjusted first neural network model, the communication method may further include: the third device Obtain performance indicators when the first device and the second device communicate based on the adjusted first neural network model.
  • the performance indicators include one or more of the following: throughput, bit error rate, packet loss rate, communication delay, etc. If the performance index does not meet the preset conditions, the third device obtains a fifth neural network model; the fifth neural network model is obtained by jointly training the adjusted first neural network model by the first device and the second device.
  • the performance index not meeting the preset conditions can indicate that the communication quality based on the adjusted first neural network model is low, that is, the performance of the adjusted first neural network model when applied to communication between the first device and the second device Low. Then, joint training of the adjusted first neural network model by the first device and the second device can improve the performance of the adjusted first neural network model.
  • the preset condition is that throughput is greater than the first value; if the performance index is bit error rate, the preset condition is that the bit error rate is less than the second value; if the performance index is packet loss rate, the preset condition is The condition is that the packet loss rate is less than the third value; if the performance indicator is communication delay, the preset condition is that the communication delay is less than the fourth value; the preset conditions corresponding to other performance indicators are similar to the above, and will not be elaborated here.
  • the first value may be preset or defined in the protocol.
  • the thresholds in the preset conditions corresponding to other performance indicators are determined in a manner similar to the first value, and will not be described again.
  • the first device and the second device jointly train the adjusted first neural network model, which may include: the first device and the second device as the receiver use a loss function to calculate the adjusted first neural network model gradient values of the parameters of the receiver sub-model in the receiver sub-model, and updates the parameters of the receiver sub-model based on the calculated gradient values; the device as the receiver also sends the gradient values of the parameters of the receiver sub-model to the first device and the second device The device that is the sender. Then, the device as the sender can calculate the gradient values of the parameters of the transmitter sub-model in the adjusted first neural network model based on the received gradient values, and update the parameters of the transmitter sub-model based on the calculated gradient values.
  • the transmitter sub-model in the fifth neural network model is obtained by updating the parameters of the transmitter sub-model in the adjusted first neural network model
  • the receiver sub-model in the fifth neural network model is the adjusted first neural network It is obtained by updating the parameters of the receiver sub-model in the model.
  • the performance indicators during communication based on the adjusted first neural network model may be obtained irregularly or periodically by the third device.
  • the third device can obtain the changing trend of the performance indicators when the first device and the second device communicate based on the adjusted first neural network model, so that the third device can obtain the adjusted first neural network model and apply it to the first device and the second device.
  • This method is conducive to determining the adjusted first god
  • the first device and the second device can jointly train the adjusted first neural network model in time to reduce the impact of the performance decrease of the adjusted first neural network model on the first device and the second device. The impact caused by the communication quality between the two devices.
  • the communication method also includes: the third device stores a third of the one or more pre-trained neural network models according to the fifth neural network model.
  • the neural network model is updated.
  • the communication method may also include: the third device sends a fifth neural network model to the model server ;
  • the model server updates the third neural network model among the one or more pre-trained neural network models according to the fifth neural network model.
  • the operation of the third device or model server to update the third neural network model among the one or more pre-trained neural network models may be implemented using transfer learning.
  • a common air interface can be established in advance between the third device and the second device. If the first device is a device different from the second device and the third device, a common air interface may be pre-established between the third device and the first device, and a common air interface may be pre-established between the first device and the second device. If there is a model server in the communication system, a common air interface can also be established in advance between the third device and the model server.
  • the common air interface may be established based on one or more modules such as channel coding, modulation, resource mapping, and precoding, or may be established based on the neural network in AI technology, which is not limited here.
  • the information interacted between different devices can be transmitted through a common air interface.
  • the information exchanged between different devices may include some or all of the following: communication system parameters, business information, computing power, delay requirements, storage space, channel status information, and device status information sent by the second device to the third device. etc., the information of each pre-trained neural network model or the first neural network model sent by the model server to the third device, etc., the adjusted first neural network model or the adjusted first neural network model sent by the third device to the second device. Sub-models in neural network models, etc.
  • the information exchanged between the different devices may also include: communication system parameters, business information, computing power, etc. sent by the first device to the third device. Delay requirements, storage space, channel status information, device status information, etc., the adjusted first neural network model or sub-models in the adjusted first neural network model sent by the third device to the first device, etc.
  • the first device and the second device can establish an AI air interface based on the adjusted first neural network model.
  • the AI air interface is different from the above-mentioned ordinary air interface.
  • the AI air interface can be used for Transmit signals (such as data, information, control signaling, etc.) for communication between the first device and the second device based on the adjusted first neural network model.
  • the signals communicated by the first device and the second device based on the adjusted first neural network model can also be directly multiplexed over a common air interface for transmission. There are no limitations here.
  • the third device determines the first neural network model from one or more pre-trained neural network models; according to the communication resource information and/or channel state information communicated between the two devices, Adjust the first neural network model and send the adjusted first neural network model, where the adjusted first neural network model is used for communication between the two devices. Then, the first device and the second device can communicate based on the adjusted first neural network model. It can be seen that the neural network model used for communication between the first device and the second device is adjusted based on the trained pre-trained neural network model, and the determined neural network model is adapted to the first device and the second device. Communication scenario of two devices.
  • the complexity of determining the neural network model can be reduced, and the signaling overhead of interaction between devices to determine the neural network model can also be reduced.
  • each module may also include multiple sub-modules
  • this embodiment of the present application also provides another communication method, as shown in Figure 7 .
  • the communication method shown in FIG. 7 is a specific implementation method of the communication method shown in FIG. 4 .
  • the communication method includes the following steps:
  • the second device sends the communication system parameters and service information of the second device to the third device; correspondingly, the third device receives the communication system parameters and service information of the second device.
  • the service information of the second device may include one of the service types such as eMBB, URLLC, and mMTC required by the second device; or the service information of the second device may include the communication delay and throughput of the neural network model by the second device. requirements for performance indicators such as quantity.
  • the third device determines the first service type based on its own service information and the service information of the second device.
  • the service information of the third device may include one of the service types such as eMBB, URLLC, and mMTC required by the third device; or the service information of the third device may include the communication delay and throughput of the neural network model by the third device. requirements for performance indicators such as quantity.
  • the third device determines one or more second neural network models from one or more pre-trained neural network models based on the first service type, its own communication system parameters, and the communication system parameters of the second device.
  • each second neural network model supports the first business type; the input dimension of each second neural network model is the same as its corresponding first input dimension, and/or the output dimension of each second neural network model is the same as its corresponding first input dimension.
  • the corresponding first output dimensions are the same.
  • the first input dimension and/or the first output dimension corresponding to each second neural network model is based on the RP size applicable to the second neural network model, and the communication system parameters of the third device and/or the second device. The communication system parameters are determined.
  • the third device determines a third neural network model from one or more second neural network models.
  • the second neural network model when there is one second neural network model, the second neural network model is used as the third neural network model.
  • the second neural network model with the largest number of parameters among the multiple second neural network models is used as the third neural network model; or, the calculation amount among the multiple second neural network models is the same as that of the third neural network model.
  • the second neural network model with the smallest absolute difference between arithmetic quantities is used as the third neural network model; or, the third neural network model with the smallest absolute difference between the parameter quantity and the first parameter quantity among the plurality of second neural network models is used as the third neural network model.
  • the second neural network model serves as the third neural network model.
  • the third device sends model request information to the model server.
  • the model request information includes the identifier of the third neural network model, the first calculation amount, and the first parameter amount.
  • the first calculation amount may be determined based on the computing power and latency requirements of the third device, and/or the computing power and latency requirements of the second device; the first parameter amount may be determined based on the computing power and latency requirements of the third device. and storage space, and/or, determined by the computing power and storage space of the second device.
  • the model server determines the third neural network model from one or more stored pre-trained neural network models according to the identification of the third neural network model.
  • the model server distills the third neural network model to obtain the first Neural network model.
  • the model server uses the third neural network model as the third neural network model.
  • a neural network model is used as the third neural network model.
  • S207a and S207b are parallel steps, and which step is specifically executed can be determined according to the result of the judged condition.
  • the model server sends the first neural network model to the third device; correspondingly, the third device receives the first neural network model from the model server.
  • the third device adjusts the first neural network model according to the first information; the first information includes communication resource information and/or channel state information for communication between the second device and the third device.
  • the third device sends the adjusted first neural network model to the second device.
  • the second device receives the modulated The integrated first neural network model.
  • the third device and the second device communicate based on the adjusted first neural network model.
  • the third device monitors performance indicators when communicating based on the adjusted first neural network model.
  • the third device and the second device jointly train the adjusted first neural network model to obtain a fifth neural network model.
  • the third device sends the fifth neural network model to the model server; accordingly, the model server receives the fifth neural network model.
  • the model server updates the stored third neural network model according to the fifth neural network model.
  • the communication method provided by the embodiment of the present application can determine a neural network model for uplink communication between the network device and the terminal device, and can also determine a neural network model for the network device and the terminal device. Neural network model for downlink communication between end devices.
  • the communication method may determine a neural network model used for device-to-device communication (D2D) between the two terminal devices.
  • D2D device-to-device communication
  • the communication method provided by the embodiment of the present application can also determine the neural network model used by each group of devices, where each group of devices in the multiple groups of devices includes communication two devices, and at least one of the two devices included in different groups of devices is different.
  • group 1 includes the first device 1 and the second device 1
  • group 2 includes the first device 1 and the second device 2
  • group 3 includes the first device 2 and the second device 3 .
  • the third device may determine the group of devices from one or more pre-trained neural network models for each group of devices in the N1 group of devices.
  • the corresponding third neural network model If there are N2 groups of devices corresponding to the same third neural network model, and the model server stores one or more pre-trained neural network models, the model server can perform an operation of obtaining the third neural network model according to the identifier of the third neural network. , without having to perform the operation of obtaining the third neural network model separately for the N2 groups of devices.
  • the third device determines, for each group of devices in the N1 group of devices, the first neural network model corresponding to the group of devices, and executes steps S102, S103 and S104 in Figure 4.
  • N1 and N2 are both integers greater than 1, and N2 is less than or equal to N1.
  • the first device is a third device
  • the third device is a network device
  • the second device is a terminal device.
  • n groups of devices n is an integer greater than 1
  • each group of devices in the n groups of devices includes a third device and a second device
  • different groups of devices include different second devices
  • each group of devices in the n groups of devices Cellular network uplink communication is performed between the included third device and the second device.
  • the third device determines the same third neural network model (i.e., model A) for the n groups of devices. Then, for each group of devices in the n groups of devices, the model server can use model A and the first corresponding to the group of devices. The calculation amount and/or the first parameter amount determines the first neural network model, and the determined first neural network model is sent to the third device.
  • the third device can obtain the first neural network model corresponding to each group of devices in the n groups of devices, and separately execute for each group of devices: based on the communication resource information and/or channel state information for communication between two devices included in the group of devices.
  • the third device may send the transmitter sub-model in model B1 to the second device 1, send the transmitter sub-model in model B2 to the second device 2, ..., and send the transmitter sub-model in model Bn to the second device n.
  • the third device performs uplink communication with the second device 1
  • the second device 1 processes the signal to be sent and transmits the signal based on the transmitter sub-model in the model B1
  • the third device receives the signal and transmits the signal based on the receiving sub-model in the model B1.
  • the received signal is processed.
  • the situation where the third device and the second device 2, ..., and the second device n respectively perform uplink communication are different from the aforementioned third device and the second device 1.
  • the situation of uplink communication is similar and will not be elaborated here.
  • the third device can determine the adjusted first neural network for the n groups of devices respectively. model, and sending the adjusted receiver sub-model in the first neural network model to the second device in each group of devices.
  • the third device processes the signal to be sent and transmits the signal based on the adjusted transmitter sub-model of the first neural network model, and the second device processes the signal to be sent based on the adjusted first neural network model.
  • the receiver sub-model in the network model receives signals and processes the received signals.
  • the first device is a device different from the second device and the third device
  • the third device is a network device
  • both the first device and the second device are terminal devices.
  • Each group of devices in the m groups of devices includes a first device and a second device, and at least one device is different between different groups of devices.
  • the communication between the first device and the second device included in each group of devices in the m groups of devices is D2D.
  • the third device determines the same third neural network model (i.e., model C) for the m group of devices. Then, for each group of devices in the m group of devices, the model server can use the model C and the first model corresponding to the group of devices. The calculation amount and/or the first parameter amount determines the first neural network model, and the determined first neural network model is sent to the third device.
  • the third device can obtain the first neural network model corresponding to each group of devices in the m groups of devices, and separately execute for each group of devices: according to the communication resource information and/or the communication resources for communication between the first device and the second device included in the group of devices.
  • the third device can send the transmitter sub-model in the model D1 to the first device 1 as the sender, and send the transmitter sub-model in the model D2 to the second device 1 as the receiver; ...; to the third device as the sender.
  • a device m sends the transmitter sub-model in the model Dm, and sends the receiver sub-model in the model Dm to the second device m as the receiver.
  • the first device 1 can process the signal to be sent and send the signal based on the transmitter sub-model in the model D1, and the second device 1 can receive the signal and process the received signal based on the receiving sub-model in the model D1.
  • the D2D between the first device and the second device included in other groups of devices in the m group of devices is similar to the aforementioned D2D between the first device 1 and the second device 1 , and will not be elaborated here.
  • the related operations of the third device in the communication method provided by the embodiments of the present application can also be implemented by designing a protocol stack.
  • the "model acquisition and distribution" functional module can be added to the media access control (medio access control, MAC) layer of the protocol stack, as shown in Figure 9.
  • the functional module has the functions of the third device in the above method embodiment, that is, the functional module can have functions such as model selection, model distillation, model adjustment, model delivery, and model update.
  • a "model acquisition and distribution" functional module is added to the controller of the MAC entity.
  • a new protocol layer can be added to the protocol stack, and the protocol layer includes a "model acquisition and distribution" functional module. In the protocol stack, the newly added protocol layer can be located at the same level as the MAC layer.
  • the model selection function of the "model acquisition and distribution" functional module may include: determining a third neural network model from one or more pre-trained neural network models.
  • the model distillation function may include: when the calculation amount of the third neural network model is greater than the first calculation amount, and/or the parameter amount of the third neural network model is greater than the first parameter amount, distilling the third neural network model to obtain the first neural network model; otherwise, the third neural network model is used as the first neural network model.
  • the model adjustment function may include: adjusting the first neural network model according to first information; the first information includes communication resource information and/or channel state information for communication between the first device and the second device.
  • the model delivery function may include: sending the adjusted first neural network model, or sending the adjusted first neural network model A sub-model in the network model; the adjusted first neural network model is used for communication between the first device and the second device.
  • the model update function includes: obtaining the performance indicators when the first device and the second device communicate based on the adjusted first neural network model; when the performance indicators do not meet the preset conditions, obtaining the adjusted pair between the first device and the second device.
  • the first neural network model is jointly trained to obtain a fifth neural network model, and the third neural network model among the one or more pre-trained neural network models is updated according to the fifth neural network model.
  • the third device or the second device may include a hardware structure and/or a software module to implement the above-mentioned functions in the form of a hardware structure, a software module, or a hardware structure plus a software module.
  • a hardware structure a hardware structure plus a software module.
  • Each function Whether one of the above functions is performed as a hardware structure, a software module, or a hardware structure plus a software module depends on the specific application and design constraints of the technical solution.
  • an embodiment of the present application provides a communication device 1000.
  • the communication device 1000 may be a component of a third device (eg, integrated circuit, chip, etc.) or a component of a second device (eg, integrated circuit, chip, etc.).
  • the communication device 1000 may also be other communication units, used to implement the methods in the method embodiments of the present application.
  • the communication device 1000 may include: a communication unit 1001 and a processing unit 1002. Among them, the processing unit 1002 is used to control the communication unit 1001 to send and receive data/signaling.
  • the communication device 1000 may also include a storage unit 1003.
  • the processing unit 1002 is configured to determine a first neural network model from one or more pre-trained neural network models.
  • the processing unit 1002 is also configured to adjust the first neural network model according to the first information; the first information includes communication resource information and/or channel state information for communication between the first device and the second device.
  • the communication unit 1001 is used to send the adjusted first neural network model, or send a sub-model in the adjusted first neural network model; the adjusted first neural network model is used between the first device and the second device. communicate between.
  • the first neural network model is determined from one or more pre-trained neural network models based on the communication system parameters of the first device and/or the communication system parameters of the second device.
  • the communication system parameters of the first device include the system bandwidth and frame structure supported by the first device
  • the communication system parameters of the second device include the system bandwidth and frame structure supported by the second device
  • the first neural network model is determined based on a third neural network model selected from one or more second neural network models; the one or more second neural network models are determined from one or more second neural network models. Determined from multiple pre-trained neural network models.
  • each second neural network model in the one or more second neural network models is the same as its corresponding first input dimension
  • the output dimension of each second neural network model is the same as its corresponding first output dimension.
  • the first input dimension and/or the first output dimension corresponding to each second neural network model is based on the RP size applicable to the second neural network model, as well as the communication system parameters of the first device and/or the second device The communication system parameters are determined.
  • the RP size is determined based on the time domain length of the RP and the frequency domain width of the RP.
  • the second neural network model supports a first service type; the first service type is a service type required by the first device and the second device.
  • the third neural network model is the second neural network model with the largest number of parameters among the multiple second neural network models.
  • the third neural network model is the second neural network model with the smallest absolute difference between the calculation amount and the first calculation amount among the plurality of second neural network models.
  • the first calculation amount is based on the sum of the computing power of the first device. latency requirements, and/or, secondary The computing power and latency requirements of the equipment are determined.
  • the third neural network model is the second neural network model with the smallest absolute difference between the parameter quantity and the first parameter quantity among the plurality of second neural network models.
  • the first parameter quantity is based on the sum of the computing power of the first device.
  • the storage space, and/or, is determined by the computing power and storage space of the second device.
  • the first neural network model is obtained by distilling the third neural network model; the operation amount of the third neural network model is greater than the first operation amount, and/or the third neural network model has The parameter amount is greater than the first parameter amount.
  • the first calculation amount is determined based on the computing power and latency requirements of the first device, and/or the computing power and latency requirements of the second device; the first parameter amount is determined based on the computing power and storage space of the first device, And/or, the computing power and storage space of the second device are determined.
  • the first neural network model is a third neural network model; the operation amount of the third neural network model is less than or equal to the first operation amount, and/or the parameter amount of the third neural network model is less than Or equal to the first parameter quantity.
  • the first calculation amount is determined based on the computing power and latency requirements of the first device, and/or the computing power and latency requirements of the second device; the first parameter amount is determined based on the computing power and storage space of the first device, And/or, the computing power and storage space of the second device are determined.
  • the first neural network model is determined by the model server based on the received model request information; the model request information includes the identification of the third neural network model, and the first operation amount and/or the first parameter amount. .
  • the first calculation amount is determined based on the computing power and latency requirements of the first device, and/or the computing power and latency requirements of the second device, and the first parameter amount is determined based on the computing power and storage space of the first device, And/or, the computing power and storage space of the second device are determined.
  • each pre-trained neural network model in one or more pre-trained neural network models is predefined or obtained from a model server.
  • the information of each pre-trained neural network model includes one or more of the following: identification, business type, RP size, input dimension, output dimension, parameter amount, and calculation amount.
  • the adjusted first neural network model is obtained by training the fourth neural network model based on the channel state information in the first information.
  • the fourth neural network model is obtained by adjusting the input dimension and/or output dimension of the first neural network model according to the RP size applicable to the first neural network model and the communication resource information in the first information.
  • the communication unit 1001 is used to receive the adjusted first neural network model, or to receive the sub-models in the adjusted first neural network model; the adjusted first neural network model is Communication is performed between the first device and the second device.
  • the adjusted first neural network model is obtained by adjusting the first neural network model based on the first information, and the first information includes communication resource information and/or channel state information for communication between the first device and the second device;
  • a neural network device is determined from one or more pretrained neural network models;
  • the communication unit 1001 is also configured to communicate based on the adjusted first neural network model, or to communicate based on a sub-model in the adjusted first neural network model.
  • the first neural network model is determined from one or more pre-trained neural network models based on the communication system parameters of the first device and/or the communication system parameters of the second device.
  • the communication system parameters of the first device include the system bandwidth and frame structure supported by the second device
  • the communication system parameters of the second device include the system bandwidth and frame structure supported by the second device
  • the communication unit 1001 is also used to send communication system parameters.
  • the first neural network model is determined based on a third neural network model selected from one or more second neural network models; the one or more second neural network models are determined from one or more second neural network models. Determined from multiple pre-trained neural network models.
  • the input dimension of each second neural network model in the one or more second neural network models is the same as its corresponding first input dimension, and/or, the output dimension of each second neural network model is the same as its corresponding first output dimension. same.
  • the first input dimension and/or the first output dimension corresponding to each second neural network model is the RP size applicable to the second neural network model, as well as the communication system parameters of the first device and/or the communication of the second device. The system parameters are determined.
  • the RP size is determined based on the time domain length of the RP and the frequency domain width of the RP.
  • the second neural network model supports a first service type; the first service type is a service type required by the first device and the second device.
  • the communication unit 1001 is also used to send the required service type.
  • the third neural network model is the second neural network model with the largest number of parameters among the multiple second neural network models.
  • the third neural network model is the second neural network model with the smallest absolute difference between the calculation amount and the first calculation amount among the plurality of second neural network models.
  • the first calculation amount is based on the sum of the computing power of the first device.
  • the delay requirements, and/or, the computing power and delay requirements of the second device are determined.
  • the third neural network model is the second neural network model with the smallest absolute difference between the parameter quantity and the first parameter quantity among the plurality of second neural network models.
  • the first parameter quantity is based on the sum of the computing power of the first device.
  • the storage space, and/or, is determined by the computing power and storage space of the second device.
  • the communication unit 1001 is also used to send computing power and delay requirements, or to send computing power and storage space, or to send computing power, delay requirements and storage space.
  • the first neural network model is obtained by distilling the third neural network model; the operation amount of the third neural network model is greater than the first operation amount, and/or the third neural network model has The parameter amount is greater than the first parameter amount.
  • the first calculation amount is determined based on the computing power and latency requirements of the first device, and/or the computing power and latency requirements of the second device; the first parameter amount is determined based on the computing power and storage space of the first device, And/or, the computing power and storage space of the second device are determined.
  • the first neural network model is a third neural network model; the operation amount of the third neural network model is less than or equal to the first operation amount, and/or the parameter amount of the third neural network model is less than Or equal to the first parameter quantity.
  • the first calculation amount is determined based on the computing power and latency requirements of the first device, and/or the computing power and latency requirements of the second device; the first parameter amount is determined based on the computing power and storage space of the first device, And/or, the computing power and storage space of the second device are determined.
  • the first neural network model is determined by the model server based on the received model request information; the model request information includes the identification of the third neural network model, and the first operation amount and/or the first parameter amount. .
  • the first calculation amount is determined based on the computing power and latency requirements of the first device, and/or the computing power and latency requirements of the second device, and the first parameter amount is determined based on the computing power and storage space of the first device, And/or, the computing power and storage space of the second device are determined.
  • the information of each neural network model in one or more pre-trained neural network models is predefined, or Obtained from the model server; the information of each pre-trained neural network model includes one or more of the following: identification, business type, RP size, input dimension, output dimension, parameter amount, and calculation amount.
  • the adjusted first neural network model is obtained by training the fourth neural network model based on the channel state information in the first information.
  • the fourth neural network model is obtained by adjusting the input dimension and/or output dimension of the first neural network model according to the RP size applicable to the first neural network model and the communication resource information in the first information.
  • An embodiment of the present application also provides a communication device 1100, as shown in Figure 11.
  • the communication device 1100 may be a third device or a second device, or may be a chip, chip system, or processor that supports the third device to implement the above method, or may be a chip, chip system, or the like that supports the second device to implement the above method. or processor etc.
  • the device can be used to implement the method described in the above method embodiment. For details, please refer to the description in the above method embodiment.
  • the communication device 1100 may include one or more processors 1101 .
  • the processor 1101 may be a general-purpose processor or a special-purpose processor. For example, it can be a baseband processor, a digital signal processor, an application-specific integrated circuit, a field programmable gate array or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components or a central processing unit (Central Processing Unit, CPU).
  • the baseband processor can be used to process communication protocols and communication data
  • the central processor can be used to process communication devices (such as base stations, baseband chips, terminals, terminal chips, distributed units (DU) or centralized units (centralized units)). unit, CU), etc.) to control, execute software programs, and process data of software programs.
  • DU distributed units
  • centralized units centralized units
  • the communication device 1100 may include one or more memories 1102, on which instructions 1104 may be stored, and the instructions may be executed on the processor 1101, so that the communication device 1100 executes the method described in the above method embodiment.
  • the memory 1102 may also store data.
  • the processor 1101 and the memory 1102 can be provided separately or integrated together.
  • the memory 1102 may include, but is not limited to, non-volatile memories such as hard disk drive (HDD) or solid-state drive (SSD), random access memory (RAM), erasable and programmable memory.
  • non-volatile memories such as hard disk drive (HDD) or solid-state drive (SSD), random access memory (RAM), erasable and programmable memory.
  • HDD hard disk drive
  • SSD solid-state drive
  • RAM random access memory
  • erasable and programmable memory erasable and programmable memory
  • Read-only memory erasable programmable ROM, EPROM
  • ROM compact disc read-only memory
  • CD-ROM compact disc read-only memory
  • the communication device 1100 may also include a transceiver 1105 and an antenna 1106.
  • the transceiver 1105 may be called a transceiver unit, a transceiver, a transceiver circuit, etc., and is used to implement transceiver functions.
  • the transceiver 1105 may include a receiver and a transmitter.
  • the receiver may be called a receiver or a receiving circuit, etc., used to implement the receiving function;
  • the transmitter may be called a transmitter, a transmitting circuit, etc., used to implement the transmitting function.
  • the communication device 1100 is a third device: the transceiver 1105 is used to perform S103 in the communication method shown in Figure 4, and is used to perform S201, S205, S208, S210, S211, S213 and S214 in the communication method shown in Figure 9 .
  • the processor 1101 is configured to execute S101 and S102 in the communication method shown in FIG. 4, and to execute S202 to S204, S209, S212, and S213 in the communication method shown in FIG. 9.
  • the communication device 1100 is the second device: the transceiver 1105 is used to perform S103 in the communication method shown in FIG. 4, and is used to perform S201, S210, S211, and S213 in the communication method shown in FIG. 9.
  • the processor 1101 is configured to execute S104 in the communication method shown in FIG. 4, and to execute S213 in the communication method shown in FIG. 9.
  • the processor 1101 may include a transceiver for implementing receiving and transmitting functions.
  • the transceiver may be a transceiver circuit, an interface, or an interface circuit.
  • the transceiver circuits, interfaces or interface circuits used to implement the receiving and transmitting functions can be separate or integrated together.
  • the above-mentioned transceiver circuit, interface or interface circuit It can be used for reading and writing code/data, or the above-mentioned transceiver circuit, interface or interface circuit can be used for signal transmission or transfer.
  • the processor 1101 can store instructions 1103, and the instructions 1103 are run on the processor 1101, which can cause the communication device 1100 to execute the method described in the above method embodiment.
  • the instructions 1103 may be fixed in the processor 1101, in which case the processor 1101 may be implemented by hardware.
  • the communication device 1100 may include a circuit, and the circuit may implement the sending or receiving or communication functions in the foregoing method embodiments.
  • the processor and transceiver described in the embodiments of this application can be implemented in integrated circuits (ICs), analog ICs, radio frequency integrated circuits (RFICs), mixed signal ICs, application specific integrated circuits (application specific integrated circuits). circuit (ASIC), printed circuit board (PCB), electronic equipment, etc.
  • the processor and transceiver can also be manufactured using various IC process technologies, such as complementary metal oxide semiconductor (CMOS), n-type metal oxide-semiconductor (NMOS), P-type Metal oxide semiconductor (positive channel metal oxide semiconductor, PMOS), bipolar junction transistor (BJT), bipolar CMOS (BiCMOS), silicon germanium (SiGe), gallium arsenide (GaAs), etc.
  • CMOS complementary metal oxide semiconductor
  • NMOS n-type metal oxide-semiconductor
  • PMOS P-type Metal oxide semiconductor
  • BJT bipolar junction transistor
  • BiCMOS bipolar CMOS
  • SiGe silicon germanium
  • GaAs gallium arsenide
  • the communication device described in the above embodiments may be the first device or the second device, but the scope of the communication device described in the embodiments of the present application is not limited thereto, and the structure of the communication device may not be limited by FIG. 11 .
  • the communication device may be a stand-alone device or may be part of a larger device.
  • the communication device may be:
  • the IC collection may also include a storage component for storing data and instructions;
  • ASIC such as modem (modulator)
  • the communication device may be a chip or a chip system
  • the chip 1200 shown in FIG. 12 includes a processor 1201 and an interface 1202.
  • the number of processors 1201 may be one or more, and the number of interfaces 1202 may be multiple.
  • the processor 1201 may be a logic circuit, and the interface 1202 may be an input-output interface, an input interface or an output interface.
  • the chip 1200 may also include memory 1203 .
  • Processor 1201 configured to determine a first neural network model from one or more pre-trained neural network models.
  • the processor 1201 is also configured to adjust the first neural network model according to the first information; the first information includes communication resource information and/or channel state information for communication between the first device and the second device.
  • Interface 1202 used to send the adjusted first neural network model, or send a sub-model in the adjusted first neural network model; the adjusted first neural network model is used between the first device and the second device communicate.
  • Interface 1202 used to receive the adjusted first neural network model, or receive a sub-model in the adjusted first neural network model; the adjusted first neural network model is used between the first device and the second device communicate.
  • the adjusted first neural network model is obtained by adjusting the first neural network model based on the first information, and the first information includes communication resource information and/or channel state information for communication between the first device and the second device;
  • a neural network device is determined from one or more pretrained neural network models.
  • the interface 1202 is also used to communicate based on the adjusted first neural network model, or to communicate based on a sub-model in the adjusted first neural network model.
  • the communication device 1100 and the chip 1200 can also perform the implementation described above for the communication device 1000.
  • the various illustrative logical blocks and steps listed in the embodiments of this application can be implemented by electronic hardware, computer software, or a combination of both. Whether such functionality is implemented in hardware or software depends on the specific application and overall system design requirements. Those skilled in the art can use various methods to implement the described functions for each specific application, but such implementation should not be understood as exceeding the protection scope of the embodiments of the present application.
  • This application also provides a computer-readable storage medium for storing computer software instructions. When the instructions are executed by a communication device, the functions of any of the above method embodiments are implemented.
  • This application also provides a computer program product for storing computer software instructions. When the instructions are executed by a communication device, the functions of any of the above method embodiments are implemented.
  • This application also provides a computer program that, when run on a computer, implements the functions of any of the above method embodiments.
  • This application also provides a communication system, which includes at least one first device and at least one second device according to the above aspect.
  • the system further includes at least one model server of the above aspects.
  • the system may also include other devices that interact with the first device and the second device in the solution provided by this application.
  • the above embodiments may be implemented in whole or in part by software, hardware, firmware, or any combination thereof.
  • software When implemented using software, it may be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on the computer, the processes or functions described in the embodiments of the present application are generated in whole or in part.
  • the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.
  • the computer instructions may be stored in or transmitted from one computer-readable storage medium to another, e.g., the computer instructions may be transferred from a website, computer, server, or data center Transmission to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) means.
  • the computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains one or more available media integrated therein.
  • the available media may be magnetic media (eg, floppy disk, hard disk, magnetic tape), optical media (eg, high-density digital video disc (DVD)), or semiconductor media (eg, SSD), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Quality & Reliability (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请提供了一种通信方法及相关装置。该方法可从一个或多个预训练神经网络模型中确定第一神经网络模型；根据第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息调整第一神经网络模型，并发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型，该调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。可见，第一设备和第二设备之间进行通信所采用的神经网络模型，是根据已训练好的预训练神经网络模型进行调整得到的，与第一设备和第二设备重新训练神经网络模型以用于通信的方式相比，可降低确定神经网络模型的复杂度。

Description

一种通信方法及相关装置
本申请要求于2022年4月11日提交中国国家知识产权局、申请号为202210375312.5、申请名称为“一种通信方法及相关装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及通信领域，尤其涉及一种通信方法及相关装置。
背景技术
随着人工智能(artificial intelligence,AI)技术的不断发展，AI技术还可应用于网络层和物理层。例如，AI技术可用于实现网络层中的网络优化、移动性管理、资源分配等，还可用于实现物理层中的信道编译码、信道预测、接收机等。
通信场景中，可以采用AI模型实现通信收发机来实现收发两端之间的通信。例如，发送端可以采用神经网络模型以用于对待发送的信号进行处理以及发送信号，接收端可以采用神经网络模型以用于对接收的信号进行处理。当前的AI收发机只能适配特定的通信场景，当通信场景发送变化后需要重新训练相应的模型，难以在通用的通信系统中部署。如何确定或部署收发两端之间通信时采用的AI模型是一个亟待解决的问题。
发明内容
本申请实施例提供了一种通信方法及相关装置，能够以低复杂度和低信令开销，确定用于收发两端之间通信的神经网络模型。
第一方面，本申请实施例提供一种通信方法，该方法包括：从一个或多个预训练神经网络模型中确定第一神经网络模型；根据第一信息调整第一神经网络模型，其中，第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型；该调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
可见，第一设备和第二设备之间进行通信所采用的神经网络模型，是根据已训练好的预训练神经网络模型进行调整得到的，且得到的神经网络模型适配于第一设备和第二设备的通信场景。与第一设备和第二设备重新训练神经网络模型以用于通信的方式相比，可降低确定神经网络模型的复杂度，还减少了第一设备和第二设备间为确定神经网络模型而交互的信令开销。
一种可选的实施方式中，第一神经网络模型是根据第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定的；第一设备的通信系统参数包括第一设备支持的系统带宽和帧结构，第二设备的通信系统参数包括第二设备支持的系统带宽和帧结构。该实施方式可使得确定的第一神经网络模型能够适配于第一设备和第二设备的通信系统，有利于提高调整后的第一神经网络模型应用于第一设备和第二设备之间通信时的性能，进而可提高第一设备和第二设备之间进行通信的质量。
一种可选的实施方式中，第一神经网络模型是基于从一个或多个第二神经网络模型中选择的第三神经网络模型确定的；该一个或多个第二神经网络模型是从一个或多个预训练神经网络模型中确定的；其中，一个或多个第二神经网络模型中每个第二神经网络模型的输入维度与其对应的第一输入维度相同，和/或，每个第二神经网络模型的输出维度与其对应的第一输出维度相同；每个第二神经网络模型对应的第一输入维度和/或第一输出维度，是根据该第二神经网络模型适用的资源片(resource patch，RP)大小，以及第一设备的通信系统参数和/或第二设备的通信系统参数确定的。
该实施方式可使得第一神经网络模型的输入维度与其对应的第一输入维度相同，和/或，第一神经网络模型的输出维度与其对应的第一输出维度相同，进而可使得第一神经网络模型适配于第一设备和第二设备的通信场景。
可选的，RP大小是根据RP的时域长度和RP的频域宽度确定的。
一种可选的实施方式中，第二神经网络模型支持第一业务类型；第一业务类型是第一设备和第二设备需求的业务类型。该实施方式基于业务类型从预训练的神经网络中选择一个或多个第二神经网络模型，使得从一个或多个神经网络模型中确定的第一神经网络模型支持第一设备和第二设备需求的业务类型。
一种可选的实施方式中，如果第二神经网络模型为多个，第三神经网络模型是多个第二神经网络模型中参数量最大的第二神经网络模型。一般来说，神经网络模型的参数量越大，性能越好，该方式可使得确定的第三神经网络模型是多个第二神经网络模型中性能最好的第二神经网络模型。
或者，第三神经网络模型是多个第二神经网络模型中运算量与第一运算量之间的绝对差值最小的第二神经网络模型，第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的。该方式可使得从多个第二神经网络模型中确定的第三神经网络模型的运算量最接近于第一设备和第二设备的运算能力所支持的最大范围。
或者，第三神经网络模型是多个第二神经网络模型中参数量与第一参数量之间的绝对差值最小的第二神经网络模型，第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。该方式可使得从多个第二神经网络模型中确定的第三神经网络模型的参数量最接近于第一设备和第二设备的存储能力所支持的最大范围。
一种可选的实施方式中，第一神经网络模型是对第三神经网络模型进行蒸馏得到的，该第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量；其中，第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量是根据第一设备的算力和存储空间，和/或，第一设备的算力和存储空间确定的。该实施方式可通过减少第三神经网络模型的参数量和/或运算量获得第一神经网络模型，这样，第一神经网络模型的运算量在第一设备和第二设备的运算能力所支持的范围内，参数量在第一设备和第二设备的存储能力所支持的范围内。
一种可选的实施方式中，第一神经网络模型为第三神经网络模型；第三神经网络模型的运算量小于或等于第一运算量，和/或，第三神经网络模型的参数量小于或等于第一参数量；第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。可见，该实施方式中，如果第三神经网络模型的运算量和参数量已分别在第一设备和第二设备的运算能力和存储能力所支持的范围内，可直接将第三神经网络模型作为第一神经网络模型。
一种可选的实施方式中，第一神经网络模型是模型服务器根据接收的模型请求信息确定的；该模型请求信息包括第三神经网络模型的标识，以及第一运算量和/或第一参数量；其中，第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的，第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。可见，该实施方式可应用于由模型服务器存储一个或多个预训练神经网络模型的情况下，可从模型服务器处获取第一神经网络模型。
可选的，一个或多个预训练神经网络模型中每个预训练神经网络模型的信息是预定义的，或是从模型服务器获取的；每个预训练神经网络模型的信息包括以下一个或多个：标识、业务类型、RP大小、输入维度、输出维度、参数量和运算量。
一种可选的实施方式中，调整后的第一神经网络模型是根据第一信息中的信道状态信息，对第四神经网络模型进行训练得到的；第四神经网络模型是根据第一神经网络模型适用的RP大小，以及第一信息中的通信资源信息，对第一神经网络模型的输入维度和/或输出维度进行调整得到的。该实施方式中，根据第一设备和第二设备之间进行通信的通信资源信息和信道状态信息调整第一神经网络模型，可使得调整后的第一神经网络模型适配于第一设备和第二设备的通信场景。
第二方面，本申请实施例提供一种通信方法，该方法包括：接收调整后的第一神经网络模型；或者，接收调整后的第一神经网络模型中的子模型；其中，调整后的第一神经网络模型用于第一设备和第二设备之间进行通信；调整后的第一神经网络模型是基于第一信息对第一神经网络模型调整得到的，该第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息，第一神经网络设备是从一个或多个预训练神经网络模型中确定的。基于调整后的第一神经网络模型进行通信；或者，基于调整后的第一神经网络模型中的子模型进行通信。
可见，第一设备与第二设备之间进行通信所采用的神经网络模型，是根据已训练好的预训练神经网络模型进行调整得到的，且确定的神经网络模型适配于第一设备和第二设备的通信场景。与第一设备和第二设备重新训练神经网络模型以用于通信的方式相比，可降低确定神经网络模型的复杂度，还减少了设备间为确定神经网络模型而交互的信令开销。
一种可选的实施方式中，第一神经网络模型是根据第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定的；其中，第一设备的通信系统参数包括第一设备支持的系统带宽和帧结构，第二设备的通信系统参数包括第二设备支持的系统带宽和帧结构。该实施方式可使得确定的第一神经网络模型适配于第一设备和第二设备的通信场景，有利于提高调整后的第一神经网络模型应用于第一设备和第二设备之间通信时的性能，进而可提高第一设备和第二设备进行通信的质量。
可选的，该方法还包括：发送通信系统参数。
一种可选的实施方式中，第一神经网络模型是基于从一个或多个第二神经网络模型中选择的第三神经网络模型确定的；一个或多个第二神经网络模型是从一个或多个预训练神经网络模型中确定的；其中，一个或多个第二神经网络模型中每个第二神经网络模型的输入维度与其对应的第一输入维度相同，和/或，每个第二神经网络模型的输出维度与其对应的第一输出维度相同；每个第二神经网络模型对应的第一输入维度和/或第一输出维度，是根据该第二神经网络模型适用的资源片RP大小，以及第一设备的通信系统参数和/或第二设备的通信系统参数确定的。
该实施方式可使得第一神经网络模型的输入维度与其对应的第一输入维度相同，和/或， 第一神经网络模型的输出维度与其对应的第一输出维度相同，进而可使得第一神经网络模型适配于第一设备和第二设备的通信场景。
可选的，RP大小是根据RP的时域长度和RP的频域宽度确定的。
一种可选的实施方式中，第二神经网络模型支持第一业务类型；第一业务类型是第一设备和第二设备需求的业务类型。该实施方式基于业务类型从预训练的神经网络中选择一个或多个第二神经网络模型，使得从一个或多个神经网络模型中确定的第一神经网络模型支持第一设备和第二设备需求的业务类型。
可选的，该方法还包括：发送需求的业务类型。
一种可选的实施方式中，如果第二神经网络模型为多个，第三神经网络模型是多个第二神经网络模型中参数量最大的神经网络模型。一般来说，神经网络模型的参数量越大，性能越好，该方式可使得确定的第三神经网络模型是多个第二神经网络模型中性能最好的第二神经网络模型。
或者，第三神经网络模型是多个第二神经网络模型中运算量与第一运算量之间的绝对差值最小的神经网络模型，第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的。该方式可使得从多个第二神经网络模型中确定的第三神经网络模型的运算量最接近于第一设备和第二设备的运算能力所支持的最大范围。
或者，第三神经网络模型是多个第二神经网络模型中参数量与第一参数量之间的绝对差值最小的神经网络模型，第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。该方式可使得从多个第二神经网络模型中确定的第三神经网络模型的参数量最接近于第一设备和第二设备的存储能力所支持的最大范围。
可选的，该方法还包括：发送算力和时延要求，或者发送算力和存储空间，或者发送算力、时延要求和存储空间。
一种可选的实施方式中，第一神经网络模型是对第三神经网络模型进行蒸馏得到的；第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量；第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。该实施方式可通过减少第三神经网络模型的参数量和/或运算量获得第一神经网络模型，这样，第一神经网络模型的参数量和运算量在第一设备和第二设备的运算能力和存储能力所支持的范围内。
一种可选的实施方式中，第一神经网络模型为第三神经网络模型；第三神经网络模型的运算量小于或等于第一运算量，和/或，第三神经网络模型的参数量小于或等于第一参数量；第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。可见，该实施方式中，如果第三神经网络模型的参数量和运算量已在第一设备和第二设备的运算能力和存储能力所支持的范围内，可直接将第三神经网络模型作为第一神经网络模型。
一种可选的实施方式中，第一神经网络模型是模型服务器根据接收的模型请求信息确定的；该模型请求信息包括第三神经网络模型的标识，以及第一运算量和/或第一参数量；其中，第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的，第一参数量是根据所述第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。可选的，一个或多个预训练神经网络模型中每个预训练神经网络模型的信息是预定义的，或是从模型服务器获取的；每个预训练神经网络模型的信息包括以下一个或多个：标识、业 务类型、RP大小、输入维度、输出维度、参数量和运算量。
一种可选的实施方式中，调整后的第一神经网络模型是根据第一信息中的信道状态信息，对第四神经网络模型进行训练得到的；第四神经网络模型是根据第一神经网络模型适用的RP大小，以及第一信息中的通信资源信息，对第一神经网络模型的输入维度和/或输出维度进行调整得到的。该实施方式中，可根据两个设备之间进行通信的通信资源信息和/或信道状态信息调整第一神经网络模型，可使得调整后的第一神经网络模型适配于第一设备和第二设备的通信场景。
第三方面，本申请还提供一种通信装置。该通信装置具有实现上述第一方面所述的部分或全部实施方式的功能，或者具有实现上述第二方面所述的部分或全部功能实施方式的功能。所述功能可以通过硬件实现，也可以通过硬件执行相应的软件实现。所述硬件或软件包括一个或多个与上述功能相对应的单元或模块。
在一种可能的设计中，该通信装置的结构中可包括处理单元和通信单元，所述处理单元被配置为支持通信装置执行上述方法中相应的功能。所述通信单元用于支持该通信装置与其他通信装置之间的通信。所述通信装置还可以包括存储单元，所述存储单元用于与处理单元和通信单元耦合，其保存通信装置必要的程序指令和数据。
一种实施方式中，所述通信装置包括：处理单元和通信单元，处理单元用于控制通信单元进行数据/信令收发。
处理单元用于从一个或多个预训练神经网络模型中确定第一神经网络模型；处理单元还用于根据第一信息调整第一神经网络模型；第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；通信单元用于发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型；调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
另外，该方面中，通信装置其他可选的实施方式可参见上述第一方面的相关内容，此处不再详述。
另一种实施方式中，所述通信装置包括：处理单元和通信单元，处理单元用于控制通信单元进行数据/信令收发。
通信单元用于接收调整后的第一神经网络模型，或者，接收调整后的第一神经网络模型中的子模型；调整后的第一神经网络模型用于第一设备和第二设备之间进行通信；调整后的第一神经网络模型是基于第一信息对第一神经网络模型调整得到的，第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；通信单元还用于基于调整后的第一神经网络模型进行通信，或者，基于调整后的第一神经网络模型中的子模型进行通信。
另外，该方面中，通信装置其他可选的实施方式可参见上述第二方面的相关内容，此处不再详述。
作为示例，通信单元可以为收发器或通信接口，存储单元可以为存储器，处理单元可以为处理器。
一种实施方式中，所述通信装置包括：处理器和收发器。其中，处理器用于从一个或多个预训练神经网络模型中确定第一神经网络模型；处理器还用于根据第一信息调整第一神经网络模型；第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；收发器用于发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型；调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
另外，该方面中，通信装置其他可选的实施方式可参见上述第一方面的相关内容，此处 不再详述。
另一种实施方式中，所述通信装置包括：收发器。其中，收发器用于接收调整后的第一神经网络模型，或者，接收调整后的第一神经网络模型中的子模型；调整后的第一神经网络模型用于第一设备和第二设备之间进行通信；调整后的第一神经网络模型是基于第一信息对第一神经网络模型调整得到的，第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；第一神经网络设备是从一个或多个预训练神经网络模型中确定的；收发器还用于基于调整后的第一神经网络模型进行通信，或者，基于调整后的第一神经网络模型中的子模型进行通信。
另外，该方面中，通信装置其他可选的实施方式可参见上述第二方面的相关内容，此处不再详述。
另一种实施方式中，该通信装置为芯片或芯片系统。所述处理单元也可以体现为处理电路或逻辑电路；所述收发单元可以是该芯片或芯片系统上的输入/输出接口、接口电路、输出电路、输入电路、管脚或相关电路等。
在实现过程中，处理器可用于进行，例如但不限于，基带相关处理，收发器可用于进行，例如但不限于，射频收发。上述器件可以分别设置在彼此独立的芯片上，也可以至少部分的或者全部的设置在同一块芯片上。例如，处理器可以进一步划分为模拟基带处理器和数字基带处理器。其中，模拟基带处理器可以与收发器集成在同一块芯片上，数字基带处理器可以设置在独立的芯片上。随着集成电路技术的不断发展，可以在同一块芯片上集成的器件越来越多。例如，数字基带处理器可以与多种应用处理器(例如但不限于图形处理器，多媒体处理器等)集成在同一块芯片之上。这样的芯片可以称为系统芯片(System on a Chip,SoC)。将各个器件独立设置在不同的芯片上，还是整合设置在一个或者多个芯片上，往往取决于产品设计的需要。本申请实施例对上述器件的实现形式不做限定。
第四方面，本申请还提供一种处理器，用于执行上述各种方法。在执行这些方法的过程中，上述方法中有关发送上述信息和接收上述信息的过程，可以理解为由处理器输出上述信息的过程，以及处理器输入的上述信息的过程。在输出上述信息时，处理器将该上述信息输出给收发器，以便由收发器进行发射。该上述信息在由处理器输出之后，还可能需要进行其他的处理，然后才到达收发器。类似的，处理器接收输入的上述信息时，收发器接收该上述信息，并将其输入处理器。更进一步的，在收发器收到该上述信息之后，该上述信息可能需要进行其他的处理，然后才输入处理器。
对于处理器所涉及的发送和接收等操作，如果没有特殊说明，或者，如果未与其在相关描述中的实际作用或者内在逻辑相抵触，则均可以更加一般性的理解为处理器输出和接收、输入等操作，而不是直接由射频电路和天线所进行的发送和接收操作。
在实现过程中，上述处理器可以是专门用于执行这些方法的处理器，也可以是执行存储器中的计算机指令来执行这些方法的处理器，例如通用处理器。上述存储器可以为非瞬时性(non-transitory)存储器，例如只读存储器(Read Only Memory，ROM)，其可以与处理器集成在同一块芯片上，也可以分别设置在不同的芯片上，本申请实施例对存储器的类型以及存储器与处理器的设置方式不做限定。
第五方面，本申请还提供了一种通信系统，该系统包括上述方面的至少一个第一设备至少一个第二设备。在另一种可能的设计中，该系统还可以包括本申请提供的方案中与第一设备、第二设备进行交互的其他设备。
第六方面，本申请提供了一种计算机可读存储介质，用于储存指令，当所述指令被计算 机运行时，实现上述第一方面或第二方面任一项所述的方法。
第七方面，本申请还提供了一种包括指令的计算机程序产品，当其在计算机上运行时，实现上述第一方面或第二方面任一项所述的方法。
第八方面，本申请提供了一种芯片系统，该芯片系统包括处理器和接口，所述接口用于获取程序或指令，所述处理器用于调用所述程序或指令以实现第一方面所涉及的功能，或者用于调用所述程序或指令以实现第二方面所涉及的功能。在一种可能的设计中，所述芯片系统还包括存储器，所述存储器，用于保存终端必要的程序指令和数据。该芯片系统，可以由芯片构成，也可以包括芯片和其他分立器件。
附图说明
图1是本申请实施例提供的一种通信系统的结构示意图；
图2是本申请实施例提供的另一种通信系统的结构示意图；
图3a是本申请实施例提供的一种全连接神经网络的结构示意图；
图3b是本申请实施例提供的一种神经网络的训练方式的示意图；
图3c是本申请实施例提供的一种梯度反向传输的示意图；
图4是本申请实施例提供的一种通信方法的交互示意图；
图5是本申请实施例提供的一种资源划分的示意图；
图6a是本申请实施例提供的一种模型调整的示意图；
图6b是本申请实施例提供的另一种模型调整的示意图；
图6c是本申请实施例提供的又一种模型调整的示意图；
图7是本申请实施例提供的另一种通信方法的交互示意图；
图8a是本申请实施例提供的一种上行通信或下行通信的示意图；
图8b是本申请实施例提供的一种D2D通信的示意图；
图9是本申请实施例提供的一种协议栈的示意图；
图10是本申请实施例提供的一种通信装置的结构示意图；
图11是本申请实施例提供的另一种通信装置的结构示意图；
图12是本申请实施例提供的一种芯片的结构示意图。
具体实施方式
下面结合本申请实施例中的附图对本申请实施例中的技术方案进行清楚、完整的描述。
为了更好的理解本申请实施例公开的通信方法，对本申请实施例适用的通信系统进行描述。
本申请实施例可应用于长期演进(long term evolution，LTE)系统等第四代(4th generation，4G)通信系统，新无线(new radio，NR)系统等第五代(5th generation，5G)通信系统，还可以应用于无线保真(wireless fidelity，WiFi)系统等短距通信系统，支持多种无线技术融合的通信系统，或者是第六代(6th generation，6G)通信系统等5G之后演进的通信系统。本申请实施例中，无线通信系统包括但不限于：窄带物联网系统(narrow band-internet of things，NB-IoT)、LTE以及5G移动通信系统的三大应用场景：增强移动宽带(enhanced mobile broadband，eMBB)、超可靠低时延通信(ultra-reliable low latency communication，URLLC)和海量机器类通信(massive machine type of communication，mMTC)等。
一种无线通信系统的架构如图1所示。无线通信系统可包括一个或多个网络设备以及一个或多个终端设备。其中，网络设备和终端设备之间可进行通信，不同终端设备之间也可互相通信。图1以无线通信系统包括一个网络设备以及两个终端设备为例。
请参阅图2，图2是本申请实施例提供的另一种通信系统的结构示意图，该通信系统包括但不限于一个第三设备和一个第二设备。图2所示的设备数量和形态用于举例并不构成对本申请实施例的限定，实际应用中可以包括两个或两个以上的第三设备，两个或两个以上的第二设备，还可以包括第一设备。其中，第三设备可以是网络设备，也可以是终端设备；第二设备也可以是网络设备，或是终端设备，第一设备可以是网络设备，或终端设备。
本申请实施例中，网络设备是具有无线收发功能的设备，其可以是LTE中的演进型基站(evolved Node B，eNB或eNodeB)，或者5G网络中的基站或者未来演进的公共陆地移动网络(public land mobile network，PLMN)中的基站，宽带网络业务网关(broadband network gateway，BNG)，汇聚交换机或者非第三代合作伙伴项目(3rd generation partnership project，3GPP)接入设备等。可选的，本申请实施例中的网络设备可以包括各种形式的基站，例如：宏基站、微基站(也称为小站)、中继站、接入点、5G之后演进的通信系统中实现基站功能的设备、WiFi系统中的接入节点，传输接收点(transmitting and receiving point，TRP)、发射点(transmitting point，TP)、移动交换中心以及设备到设备(device-to-device，D2D)、车辆外联(vehicle-to-everything，V2X)、机器到机器(machine-to-machine，M2M)通信中承担基站功能的设备等，还可以包括云接入网(cloud radio access network，C-RAN)系统中的集中式单元(centralized unit，CU)和分布式单元(distributed unit，DU、非陆地通信网络(non-terrestrial network，NTN)通信系统中的网络设备，即可以部署于高空平台或者卫星，本申请实施例对此不作具体限定。
终端设备可以包括各种具有无线通信功能的手持设备、车载设备、可穿戴设备、计算设备或连接到无线调制解调器的其它处理设备。终端设备也可以指用户设备(user equipment，UE)、接入终端、客户终端设备(customer-premises equipment，CPE)、用户单元(subscriber unit)、用户代理、蜂窝电话(cellular phone)、智能手机(smart phone)、无线数据卡、个人数字助理(personal digital assistant，PDA)电脑、平板型电脑、无线调制解调器(modem)、手持设备(handset)、膝上型电脑(laptop computer)、机器类型通信(machine type communication，MTC)终端、高空飞机上搭载的通信设备、可穿戴设备、无人机、机器人、智能销售点(point of sale，POS)机、D2D中的终端、V2X中的终端、虚拟现实(virtual reality，VR)终端设备、增强现实(augmented reality，AR)终端设备、工业控制(industrial control)中的无线终端、无人驾驶(self driving)中的无线终端、远程医疗(remote medical)中的无线终端、智能电网(smart grid)中的无线终端、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端、智慧家庭(smart home)中的无线终端或者5G之后演进的通信网络中的终端设备等，本申请不作限制。
本申请公开的实施例将围绕包括多个设备、组件、模块等的系统来呈现本申请的各个方面、实施例或特征。应当理解和明白的是，各个系统可以包括另外的设备、组件、模块等，并且/或者可以并不包括结合附图讨论的所有设备、组件、模块等。此外，还可以使用这些方案的组合。
为了更好的理解本申请实施例公开的通信方法，对本申请实施例涉及的相关概念进行简单的介绍。
1.全连接神经网络：
全连接神经网络又可称为多层感知机(multilayer perceptron，MLP)。一个MLP包含一个输入层、一个输出层以及多个隐藏层，且每层包含数个节点，该节点可称为神经元。其中，相邻两层的神经元间两两相连。例如，图3a所示的MLP包括一个输入层、一个输出层以及两个隐藏层。其中，输入层包括4个神经元，每个隐藏层包括8个神经元，输出层包括6个神经元。
对于相邻两层中的神经元而言，下一层中神经元的输出h是，采用激活函数对所有与该神经元相连的上一层神经元x的加权和进行处理得到的。用矩阵可以表示为：
h＝f(wx+b)    (1)
其中，w为权重矩阵，b为偏置向量，f()为激活函数。
那么，神经网络的输出可以递归表达为：
y＝fn(wnfn-1(…)+bn)   (2)
可见，神经网络可以表示输入数据集合到输出数据集合的映射关系。神经网络的初始化通常是随机的，需要对其进行训练后投入使用。神经网络的训练是指利用已有数据从随机的w和b中确定这个映射关系的过程。
结合图3b，训练神经网络的具体方式包括：采用损失函数(loss function)对神经网络的输出结果进行评价，并将误差反向传播，通过梯度下降的方法来迭代优化w和b直到损失函数达到最小值。其中，梯度下降的过程可以表示为：
其中，θ为待优化参数(如w和b)，L为损失函数；η为学习效率，用于控制梯度下降的步长。
另外，反向传播的过程中利用了求偏导的链式法则，即前一层参数的梯度可以由后一层参数的梯度递推计算得到。例如，图3c中神经元j与神经元i之间权重wij的梯度可表示为：
其中，si是神经元i上的输入加权和。另外，公式(4)中，可称为中间层梯度，即将si看作中间层。
2.大模型(mega model)：
大模型是指通过大量数据预训练的一个参数量巨大的神经网络模型，如参数量巨大的全连接神经网络模型。大模型拥有极强的信息提取和表达能力，可以用于完成多种任务。例如，自然语言处理领域的大模型：生成型预训练变换器3(Generative Pre-trained Transformer 3，GPT-3)，包含1750亿的参数量，使用GPT-3可以完成翻译、文章撰写、搜索等各种自然语言处理相关的任务。
本申请实施例提供了一种通信方法，该通信方法可从一个或多个预训练神经网络模型中确定第一神经网络模型，并根据第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息调整第一神经网络模型，该调整后的第一神经网络模型可用于第一设备和第二设备之间进行通信。该方法能够以低复杂度和低信令开销，确定用于第一设备和第二设备之间通信的神经网络模型。
下面结合附图对本申请实施例提供的通信方法进行阐述。
请参阅图4，图4是本申请实施例提供的一种通信方法的交互示意图，该通信方法从第 三设备和第二设备之间交互的角度进行阐述。该通信方法包括以下步骤：
S101、第三设备从一个或多个预训练神经网络模型中确定第一神经网络模型。其中，一个或多个预训练神经网络模型可以是已训练好的一个或多个大模型。
在一种可选的实施方式中，第一神经网络模型是根据第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定的；其中，第一设备的通信系统参数包括第一设备支持的系统带宽和帧结构，第二设备的通信系统参数包括第二设备支持的系统带宽和帧结构。
可选的，通信系统参数还可包括载波频率、系统参数(numerology)、天线配置、参考信号位置等。其中，系统参数可包括子载波间隔、每个时隙(slot)的长度、每个时隙包括的符号(symbol)个数、每个符号中循环前缀(cyclic prefix，CP)的长度等。天线配置可包括天线端口数，参考信号位置可包括参考信号所在的时域资源位置、频域资源位置、空域资源位置等。也就是说，第一设备的通信系统参数表征了第一设备可工作的资源，第一设备的通信系统参数是第一设备可工作的资源的相关参数。第二设备的通信系统参数表征了第二设备可工作的资源，第二设备的通信系统参数是第二设备可工作的资源的相关参数。
另外，步骤S101中确定的第一神经网络模型用于确定第一设备和第二设备之间进行通信所采用的神经网络模型。其中，第一设备可以是第三设备，还可以是不同于第二设备和第三设备的其他设备。第一设备是不同于第二设备和第三设备的其他设备的情况下，第一设备可以是其他的网络设备或终端设备。
可选的，第一设备的通信系统参数和/或第二设备的通信系统参数可以是第三设备从预定义的多组通信系统参数中确定的通信系统参数，针对第一设备确定的一组通信系统参数适用于第一设备，针对第二设备确定的一组通信系统参数适用于第二设备。或者，第二设备的通信系统参数可以是第二设备发送给第三设备的。若第一设备是不同于第二设备和第三设备的其他设备，则第一设备的通信系统参数可以是第一设备发送给第三设备的。
可选的，第一神经网络模型是基于从一个或多个第二神经网络模型中选择的第三神经网络模型确定的；该一个或多个第二神经网络模型是从一个或多个预训练神经网络模型中确定的。其中，一个或多个第二神经网络模型中每个第二神经网络模型的输入维度与其对应的第一输入维度相同，和/或，每个第二神经网络模型的输出维度与其对应的第一输出维度相同。每个第二神经网络模型对应的第一输入维度和/或第一输出维度，是根据该第二神经网络模型适用的资源片(resource patch，RP)大小，以及第一设备的通信系统参数和/或第二设备的通信系统参数确定的。该实施方式可使得确定的第一神经网络模型能够适配于第一设备和第二设备的通信系统，有利于提高调整后的第一神经网络模型应用于第一设备和第二设备之间通信时的性能，进而可提高第一设备和第二设备进行通信的质量。另外，本申请实施例中，RP还可以称为资源包。
可选的，若第一设备和第二设备之间通信系统参数的值不同，则一个或多个预训练神经网络模型中每个预训练神经网络模型对应的第一输入维度和/或第一输出维度，可以是根据该预训练神经网络模型适用的RP大小，以及第一设备的通信系统参数和第二设备的通信系统参数中值最小的通信系统参数确定的。示例性地，通信系统参数包括系统带宽、帧结构和天线端口数，其中，第一设备和第二设备各自支持的帧结构对应的帧长度的值相同，系统带宽的值不同，天线端口数的值不同，那么，可根据值相同的帧长度、值最小的系统带宽、值最小的天线端口数以及预训练神经网络模型适用的RP大小，确定该预训练神经网络模型对应的第一输入维度和/或第一输出维度。
例如，第一设备为设备1，第二设备为设备2，其中，设备1支持的帧结构与设备2支持的帧结构对应的帧长度相同，均为帧长度1；设备1支持的系统带宽1小于设备2支持的系统带宽2，且设备1支持的天线端口数1大于设备2支持的天线端口数2，那么，可根据预训练神经网络模型适用的RP大小、帧长度1、系统带宽1和天线端口数2，确定该预训练神经网络模型对应的第一输入维度和/或第一输出维度。
可选的，每个预训练神经网络模型适用的RP大小可以是根据该预训练神经网络模型适用的RP的时域长度、RP的频域宽度确定的；或者，还可以是根据该预训练神经网络模型适用的RP的时域长度、RP的频域宽度、RP的空域宽度等确定的。另外，不同预训练神经网络模型适用的RP大小可能不同。其中，RP的时域长度的单位例如为毫秒(millisecond，ms)，RP的频域宽度的单位例如为千赫兹(kilohertz，kHz)，RP的空域宽度可以表示为RP占用的空间流数，其单位例如为流。
以通信系统参数包括系统带宽和帧结构为例，每个预训练神经网络模型对应的第一输入维度或若通信系统参数还包括天线端口数，每个预训练神经网络模型对应的第一输入维度或
其中，Bsys是第一设备支持的系统带宽和第二设备支持的系统带宽中值最小的系统带宽，Ssys是第一设备和第二设备分别支持的帧结构对应的帧长度中值最小的帧长度，n是第一设备支持的天线端口数和第二设备支持的天线端口数中值最小的天线端口数。特别地，若第一设备支持的系统带宽与第二设备支持的系统带宽相同，则Bsys是该相同的系统带宽的值；若第一设备支持的帧结构对应的帧长度与第二设备支持的帧结构对应的帧长度相同，则Ssys是该相同的帧长度的值；若第一设备支持的天线端口数与第二设备支持的天线端口数相同，则n是该相同的天线端口数的值。另外，Brp是该预训练神经网络模型适用的RP的频域宽度，Srp是该预训练神经网络模型适用的RP的时域长度，ceil()为向上取整函数。
例如，结合图5，图5中展示的资源网格为一个天线端口上的资源网络。如图5所示，若天线端口数n为1，说明这一个天线端口上可供第一设备和第二设备工作的资源网格可划分为2个RP，该RP为某预训练神经网络模型适用的RP，那么，该预训练神经网络模型对应的第一输入维度或第一输出维度可确定为2；若天线端口数n为2，说明这两个天线端口上可供第一设备和第二设备工作的资源网格一共可划分为4个RP，那么，该预训练神经网络模型对应的第一输入维度或第一输出维度可确定为4。
特别地，一般来说，网络设备的通信系统参数的值大于终端设备的通信系统参数的值，例如，网络设备支持的系统带宽大于终端设备支持的系统带宽，网络设备支持的帧结构对应的帧长度大于终端设备支持的帧结构对应的帧长度。那么，若第一设备和第二设备中的一个设备是网络设备，另一个设备是终端设备，则第一神经网络模型是根据第一设备和第二设备中终端设备的通信系统参数，从一个或多个预训练神经网络模型中确定的。这一情况下，上述公式中的Bsys是第一设备和第二设备中终端设备支持的系统带宽，Ssys是第一设备和第二设备中终端设备支持的帧结构对应的帧长度，n是第一设备和第二设备中终端设备支持的天线端口数。
可选的，第二神经网络模型支持第一业务类型；其中，第一业务类型是第一设备和第二设备需求的业务类型。该方式有利于使得确定的第一神经网络模型支持第一设备和第二设备需求的业务类型。可选的，第一业务类型可以是根据第一设备的业务信息和第二设备的业务 信息确定的。其中，第一设备和第二设备中任一设备的业务信息可以包括该设备需求的eMBB、URLLC和mMTC等业务类型中的一个；或者，任一设备的业务信息可以包括该设备对神经网络模型的通信时延、吞吐量等性能指标的需求，体现了该设备的性能需求。其中，神经网络模型的通信时延包括发送端基于神经网络模型处理待发送的信号的时间、传输处理后的信号的时间以及接收端基于神经网络模型处理接收的信号的时间，其单位例如为ms；神经网络模型的吞吐量是指在一定时间内基于神经网络模型发送和/或接收的数据量，其单位例如为兆比特每秒(million bits per second，Mbps)。
其中，第一设备是第三设备时，第一业务类型可以是第三设备根据自身的业务信息和第二设备发送的业务信息确定的。第一设备是不同于第二设备和第三设备的其他设备时，第一业务类型可以是第三设备根据第一设备发送给第三设备的业务信息和第二设备发送给第三设备的业务信息确定的。
关于每个预训练神经网络模型支持的业务类型和第一业务类型的表现形式包括实施方式1.1和实施方式1.2所述。
实施方式1.1，每个预训练神经网络模型支持的业务类型为eMBB、URLLC和mMTC等业务类型中的一个，第一业务类型为eMBB、URLLC和mMTC等业务类型中的一个。预训练神经网络模型支持的业务类型与第一业务类型相同时，说明该预训练神经网络模型支持第一业务类型，即支持第一设备和第二设备需求的业务类型。
另外，在第一设备和第二设备中每个设备的业务信息均包括eMBB、URLLC和mMTC等业务类型中的一个的情况下，第一设备的业务信息包括的业务类型和第二设备的业务信息包括的业务类型相同，该相同的业务类型即为第一业务类型。可选的，第一设备的业务信息和第二设备的业务信息包括的相同的业务类型可以是第一设备和第二设备之间预先约定的。
在第一设备和第二设备中每个设备的业务信息均包括该设备对神经网络模型的通信时延、吞吐量等性能指标的需求的情况下，第三设备可根据第一设备的业务信息和第二设备的业务信息，从eMBB、URLLC和mMTC等业务类型中确定第一业务类型，且确定的第一业务类型能够满足第一设备和第二设备的性能需求。
在第一设备的业务信息包括eMBB、URLLC和mMTC等业务类型中的一个，第二设备的业务信息包括第二设备对神经网络模型的通信时延、吞吐量等性能指标的需求的情况下，第三设备可判断第一设备的业务信息中包括的业务类型是否能够满足第二设备的性能需求；若满足，则将第一设备的业务信息中包括的业务类型作为第一业务类型。另外，第一设备的业务信息包括第一设备对神经网络模型的通信时延、吞吐量等性能指标的需求，第二设备的业务信息包括eMBB、URLLC和mMTC等业务类型中的一个的情况与之类似，不再赘述。
实施方式1.2，每个预训练神经网络模型支持的业务类型以一个或多个性能指标的形式表示，第一业务类型以一个或多个性能指标中各性能指标的取值范围来表示。预训练神经网络模型支持的业务类型中各性能指标的值在第一业务类型中该性能指标的取值范围内时，说明该预训练神经网络模型支持第一业务类型，即支持第一设备和第二设备需求的业务类型。
另外，在第一设备和第二设备中每个设备的业务信息均包括该设备对神经网络模型的通信时延、吞吐量等性能指标的需求的情况下，第一业务类型表示的各性能指标的取值范围包含于第一设备的性能需求和第二设备的性能需求的交集内。例如，第一设备的业务信息包括：神经网络模型的通信时延小于值#1，吞吐量大于值#2；第二设备的业务信息包括：神经网络模型的通信时延小于值#3，吞吐量大于值#4。那么，第一业务类型可表示为：神经网络模型的通信时延小于值#5，吞吐量大于值#6，其中，值#5小于或等于值#1和值#3中最小的一个， 值#6大于或等于值#2和值#4中最大的一个。
在第一设备和第二设备中每个设备的业务信息均包括eMBB、URLLC和mMTC等业务类型中的一个的情况下，第一设备的业务信息中包括的业务类型与第二设备的业务信息中包括的业务类型相同。第一业务类型表示的各性能指标的取值范围包含于，该相同的业务类型所支持的各性能指标的取值范围内。例如，第一设备的业务信息和第二设备的业务信息均包括业务类型1，业务类型1支持神经网络模型的通信时延小于值#1；那么，第一业务类型可表示为：神经网络模型的通信时延小于值#3，其中，值#3小于或等于值#1。
在第一设备的业务信息包括eMBB、URLLC和mMTC等业务类型中的一个，第二设备的业务信息包括第二设备对神经网络模型的通信时延、吞吐量等性能指标的需求的情况下，第一业务类型表示的各性能指标的取值范围包含于，第一设备的业务信息包括的业务类型支持的各性能指标的取值范围与第二设备的性能需求的交集内。例如，第一设备的业务信息包括业务类型1，业务类型1支持神经网络模型的通信时延小于值#1。第二设备的业务信息包括：神经网络模型的通信时延小于值#2。那么，第一业务类型可表示为：神经网络模型的通信时延小于值#3，其中，值#3小于或等于值#1和值#2中最小的一个。另外，第一设备的业务信息包括第一设备对神经网络模型的通信时延、吞吐量等性能指标的需求，第二设备的业务信息包括eMBB、URLLC和mMTC等业务类型中的一个的情况与之类似，不再赘述。
可选的，基于实施方式1.1和实施方式1.2，该通信方法还可包括：第二设备向第三设备发送业务信息和通信系统参数；第三设备根据第一设备的业务信息和第二设备的业务信息确定第一业务类型，再根据第一业务类型，以及第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定一个或多个第二神经网络模型，从一个或多个第二神经网络模型中确定第三神经网络模型。关于第二神经网络模型、第一设备的业务信息和通信系统参数、第二设备的业务信息和通信系统参数的阐述可参见前述的相关阐述，此处不再赘述。
其中，第三设备根据第一业务类型，以及第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定一个或多个第二神经网络模型，可包括：第三设备从一个或多个预训练神经网络模型中确定支持第一业务类型的预训练神经网络模型，再根据第一设备的通信系统参数和/或第二设备的通信系统参数，从支持第一业务类型的预训练神经网络模型中确定输入维度等于其对应的第一输入维度和/或输出维度等于其对应的第一输出维度的一个或多个第二神经网络模型。
或者，可包括：第三设备根据第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定输入维度等于其对应的第一输入维度和/或输出维度等于其对应的第一输出维度的预训练神经网络模型，再从确定的预训练神经网络模型中确定支持第一业务类型的一个或多个第二神经网络模型。
又或者，可包括：第三设备从一个或多个预训练神经网络模型中确定支持第一业务类型的预训练神经网络模型；同时，第三设备根据第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定输入维度等于其对应的第一输入维度和/或输出维度等于其对应的第一输出维度的预训练神经网络模型；接着，第三设备再将上述两个操作中得到的相同的预训练神经网络模型作为第二神经网络模型。例如，第三设备确定了支持第一业务类型的预训练神经网络模型1和预训练神经网络模型2；同时，第三设备确定了输入维度等于其对应的第一输入维度和/或输出维度等于其对应的第一输出维度的预训练神经网络模型2和预训练神经网络模型3；那么，将预训练神经网络模型2作为第二神经网 络模型。
在一种可选的实施方式中，如果第二神经网络模型为一个，第三神经网络模型即为该第二神经网络模型。如果第二神经网络模型为多个，第三神经网络模型是多个第二神经网络模型中参数量最大的第二神经网络模型；该方式可使得确定的第三神经网络模型是多个第二神经网络模型中性能最好的第二神经网络模型。
或者，第三神经网络模型是多个第二神经网络模型中运算量与第一运算量之间的绝对差值最小的第二神经网络模型，其中，第一运算量可以是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；该方式可使得确定的第三神经网络模型是：多个第二神经网络模型中运算量最接近于第一设备和第二设备的运算能力所支持的最大范围的第二神经网络模型。
又或者，第三神经网络模型是多个第二神经网络模型中参数量与第一参数量之间的绝对差值最小的第二神经网络模型，其中，第一参数量可以是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的；该方式可使得确定的第三神经网络模型是：多个第二神经网络模型中参数量最接近于第一设备和第二设备的存储能力所支持的最大范围的第二神经网络模型。
其中，每个第二神经网络模型的运算量与第一运算量之间的绝对差值是指该第二神经网络模型的运算量与第一运算量之间差值的绝对值；每个第二神经网络模型的参数量与第一参数量之间的绝对差值是指该第二神经网络模型的参数量与第一参数量之间差值的绝对值。
本申请实施例中，第一设备和第二设备中任一设备的算力可以表示为该设备每秒能够运算浮点的次数，其单位例如为每秒浮点运算次数(floating point operation per second，flop/s)。第一设备和第二设备中任一设备的时延要求是指该设备对神经网络模型的计算时延的要求，神经网络模型的计算时延是指神经网络模型计算一定浮点数所需的时间，其单位例如为秒(second，s)。第一设备和第二设备中任一设备的存储空间是指该设备基于神经网络模型通信时可使用的存储空间，其单位可以是字节(byte)。第一运算量的单位例如为flop，第一参数量的单位例如为byte。可选的，第二设备的算力、时延要求和存储空间可以是第二设备发送给第三设备的。若第一设备是不同于第二设备和第三设备的其他设备，第一设备的算力、时延要求和存储空间可以是第一设备发送给第三设备的。上述对第一运算量和第一参数量的相关阐述可适用于本申请实施例中任意提及第一运算量和第一参数量的位置处，下文中不再赘述。
可见，在只存在一个预训练神经网络模型的情况下，若该预训练神经网络模型的输入维度与其对应的第一输入维度相同，和/或该预训练神经网络模型的输出维度与其对应的第一输出维度相同，且该预训练神经网络模型支持第一业务类型，则该预训练神经网络模型即为第二神经网络模型，也为第三神经网络模型。在存在多个预训练神经网络模型的情况下，第三设备可从该多个预训练神经网络模型中确定一个或多个第二神经网络模型，再从该一个或多个第二神经网络模型中确定第三神经网络模型。
关于第三设备基于从一个或多个第二神经网络模型中选择的第三神经网络模型确定第一神经网络模型的相关阐述，可参见实施方式2.1至实施方式2.3所述。
实施方式2.1，第一神经网络模型是对第三神经网络模型进行蒸馏(distillation)得到的；第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量。对第三神经网络模型进行蒸馏可在尽量保证第三神经网络模型的性能不受影响的前提下，降低第三神经网络模型运算量和/或参数量，从而可使得得到的第一神经网络模型的运算 量在第一设备和第二设备的运算能力所支持的范围内，参数量在第一设备和第二设备的存储能力所支持的范围内。关于第一运算量和第一参数量的具体阐述可参见前述的相关阐述，此处不再赘述。
可选的，对第三神经网络模型进行蒸馏得到第一神经网络模型的过程中还可结合参考数据进行蒸馏，其中，参考数据反映了第一设备和第二设备的环境特征和数据特征。该方式可使得第一神经网络模型与第一设备和第二设备的通信场景更加适配，进而有利于提高调整后的第一神经网络模型应用于第一设备和第二设备之间通信时的性能，从而可提高第一设备和第二设备之间的通信质量。可选的，参考数据可以是根据第三设备统计的第一设备的历史数据和第二设备的历史数据确定的。例如，基于第一设备和第二设备之间信道的历史信道估计值可获取与该信道相关的参考数据；基于第一设备和第二设备之间的历史业务流，可获取与第一设备和第二设备之间的数据业务相关的参考数据。
实施方式2.2，第一神经网络模型为第三神经网络模型；第三神经网络模型的运算量小于或等于第一运算量，和/或，第三神经网络模型的参数量小于或等于第一参数量。可见，第三神经网络模型的运算量和参数量已经分别在第一设备和第二设备的运算能力和存储能力所支持的范围内了，那么，第一设备可直接将第三神经网络模型作为第一神经网络模型。关于第一运算量和第一参数量的具体阐述可参见前述的相关阐述，此处不再赘述。
另外，第三神经网络模型的运算量等于第一运算量，和/或，第三神经网络模型的参数量等于第一参数量的情况下，还可以不将第三神经网络模型作为第一神经网络模型，而是对第三神经网络模型进行蒸馏后得到第一神经网络模型。
可选的，基于实施方式2.1和实施方式2.2，第三设备存储了一个或多个预训练神经网络模型时，该通信方法还可包括：第三设备在第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量时，对第三神经网络模型进行蒸馏，得到第一神经网络模型；在第三神经网络模型的运算量小于或等于第一运算量，和/或，第三神经网络模型的参数量小于或等于第一参数量时，将第三神经网络模型作为第一神经网络模型。
实施方式2.3，第一神经网络模型是模型服务器根据接收的模型请求信息确定的；该模型请求信息包括第三神经网络模型的标识，以及第一运算量和/或第一参数量。其中，模型请求信息是第三设备发送给模型服务器的，第一神经网络模型是第三设备接收的来自模型服务器的。可见，该实施方式可应用于由模型服务器存储一个或多个预训练神经网络模型的情况下，第三设备可从模型服务器处获取第一神经网络模型。关于第一运算量和第一参数量的具体阐述可参见前述的相关阐述，此处不再赘述。
可选的，第三设备未存储一个或多个预训练神经网络模型，且模型服务器存储了一个或多个预训练神经网络模型时，该通信方法还包括：第三设备向模型服务器发送模型请求信息，模型请求信息包括第三神经网络模型的标识，以及第一运算量和/或第一参数量。模型服务器可根据模型请求信息中的第三神经网络模型的标识，从一个或多个预训练神经网络模型中确定第三神经网络模型，再根据第三神经网络模型、模型请求信息中的第一运算量和/或第一参数量确定第一神经网络模型之后，将第一神经网络模型发送给第三设备。相应的，第三设备接收来自模型服务器的第一神经网络模型。
具体地，模型服务器根据第三神经网络模型、模型请求信息中的第一运算量和第一参数量确定第一神经网络模型，可包括：模型服务器在第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量时，对第三神经网络模型进行蒸馏，得到第一神经网络模型；在第三神经网络模型的运算量小于或等于第一运算量，和/或，第三 神经网络模型的参数量小于或等于第一参数量时，将第三神经网络模型作为第一神经网络模型。
可选的，预训练神经网络模型的信息包括以下一个或多个：标识(也可以称为身份标识符(identifier，ID))、业务类型、RP大小、输入维度、输出维度、参数量和运算量。其中，不同的预训练神经网络模型具有不同的标识。可选的，预训练神经网络模型的信息是预定义的。例如，第三设备可以从标准文本、产品说明中获取定义的每个预训练神经网络模型的信息。或者，预训练神经网络模型的信息是从模型服务器获取的。例如，模型服务器主动将预训练神经网络模型的信息发送给第三设备；或由第三设备向模型服务器请求获取预训练神经网络模型的信息，再由模型服务器发送给第三设备。可见，第三神经网络模型的标识可以是第三设备基于一个或多个预训练神经网络模型中每个预训练神经网络模型的信息确定的。
本申请实施例中，预训练神经网络模型可包括子模型，具体地，预训练神经网络模型可包括发射机子模型和接收机子模型。其中，发射机子模型用于处理待发送的信号以及发送处理后的信号，接收机子模型用于接收信号以及对接收的信号进行处理。此处采用“发射机子模型”和“接收机子模型”来命名这两种子模型的方式仅是一种示例性的命名方式，还可以采用“第一子模型”和“第二子模型”的方式命名，此处不作限定。后文以“发射机子模型”和“接收机子模型”的命名方式为例进行阐述。
预训练神经网络模型的输入维度是指预训练神经网络模型中发射机子模型的输入维度，输出维度是指预训练神经网络模型中接收机子模型的输出维度，参数量是指发射机子模型的参数量和接收机子模型的参数量之和，运算量是指发射机子模型的运算量和接收机子模型的运算量之和。另外，本申请实施例中，对某预训练神经网络模型的蒸馏、调整等处理包括：对该预训练神经网络模型中的发射机子模块和接收机子模型均进行处理。例如，对第三神经网络模型进行蒸馏包括对第三神经网络模型中的发射机子模型和接收机子模型进行蒸馏。可见，第一神经网络模型也包括发射机子模型和接收机子模型。其中，若第一神经网络模型是对第三神经网络模型进行蒸馏得到的，第一神经网络模型中的发射机子模型是第三神经网络模型中的发射机子模型进行蒸馏得到的，第一神经网络模型中的接收机子模型是第三神经网络模型中的接收机子模型进行蒸馏得到的。
本申请实施例还提供了表1，示例性地展示了一个或多个预训练神经网络模型中每个预训练神经网络模型的信息。
表1
另外，除了表1所示的将多个预训练神经网络模型中每个预训练神经网络模型的信息在同一个表中体现的方式以外，还可以根据预训练神经网络模型的任意类别的信息对多个预训练神经网络模型进行分类，并将同一类别的预训练神经网络模型的信息置于同一表中体现，不同类别的预训练神经网络模型的信息置于不同表中体现。例如，根据预训练神经网络模型支持的业务类型将多个预训练神经网络模型的信息分置于多个表中体现，同一表中包括的预 训练神经网络模型支持的业务类型相同，不同表中包括的预训练神经网络模型支持的业务类型不同，如表2和表3所示。表2中的预训练神经网络模型均支持eMBB，表3中的预训练神经网络模型均支持mMTC。可选的，表2和表3的表项中可能不包含业务类型，也可能包含业务类型，此处不做限定。又例如，根据预训练神经网络模型适用的RP的频域宽度将多个预训练神经网络模型的信息分置于多个表中体现，等等。关于一个或多个预训练神经网络模型中每个预训练神经网络模型的信息的体现方式，此处不作限定。
表2
表3
S102、第三设备根据第一信息调整第一神经网络模型，第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息。
其中，第三设备跟据第一信息调整第一神经网络模型包括：第三设备对第一神经网络模型中的发射机子模型和接收机子模型进行调整。那么，调整后的第一神经网络模型中的发射机子模型是第一神经网络模型中的发射机子模型进行调整得到的，调整后的第一神经网络模型中的接收机子模型是第一神经网络模型中的接收机子模型进行调整得到的。
在一种可选的实施方式中，调整后的第一神经网络模型是根据第一信息中的信道状态信息，对第四神经网络模型进行训练得到的；第四神经网络模型是根据第一神经网络模型适用的RP大小，以及第一信息中的通信资源信息，对第一神经网络模型的输入维度和/或输出维度进行调整得到的。可选的，第一信息中的通信资源信息包括用于第一设备和第二设备之间进行通信的资源，该资源可包括时域资源、频域资源、空域资源等资源。与前文中通信系统参数所表征的设备可工作的资源可能不同，用于第一设备和第二设备之间进行通信的资源是第一设备和第二设备可工作的资源中第一设备和第二设备可共同工作的部分或全部资源，也是第一设备和第二设备在基于神经网络模型进行的通信中实际可使用的资源。可选的，用于第一设备和第二设备之间进行通信的资源可以是第三设备为第一设备和第二设备分配的，还可以是不同于第三设备的其他设备为第一设备和第二设备分配的，此处不作限定。另外，关于第一神经网络模型适用的RP大小的阐述可参见前述的相关阐述，此处不再赘述。
可选的，若第一设备和第二设备采用神经网络模型进行的通信是单向通信，即第一设备和第二设备中仅一个设备在通信中作为发送方，另一设备在通信中作为接收方，第一信息中的信道状态信息可以是第一设备和第二设备中作为接收方的设备进行信道估计得到的。若第一设备和第二设备采用神经网络模型进行的通信是双向通信，即第一设备在通信中既作为发送方又作为接收方，第二设备在通信中也是既作为发送方又作为接收方，第一信息中的信道状态信息包括第一设备和第二设备分别进行信道估计得到的信息。其中，第二设备进行信道估计得到的信息可以是第二设备发送给第三设备的。若第一设备是不同于第二设备和第三设备的其他设备，则第一设备进行信道估计得到的信息可以是第一设备发送给第三设备的。
可选的，该通信方法还可包括：若根据第一神经网络模型适用的RP大小以及第一信息中的通信资源信息确定的第二输入维度小于第一神经网络模型的输入维度，则第三设备将第一神经网络模型的输入维度调整为第二输入维度，即第四神经网络模型的输入维度为第二输入维度；否则，第四神经网络模型的输入维度等于第一神经网络模型的输入维度，即无需对第一神经网络模型的输入维度做调整。
可选的，该通信方法还可包括：若根据第一神经网络模型适用的RP大小以及第一信息中的通信资源信息确定的第二输出维度小于第一神经网络模型的输出维度，则第三设备将第一神经网络模型的输出维度调整为第二输出维度，即第四神经网络模型的输出维度为第二输出维度；否则，第四神经网络模型的输出维度等于第一神经网络模型的输出维度，即无需对第一神经网络模型的输出维度做调整。
以对第一神经网络模型的输入维度进行调整为例，结合图6a，第一神经网络模型的输入维度为4，若根据第一神经网络模型适用的RP大小，以及第一信息中的通信资源信息确定的第二输入维度为2，则将第一神经网络模型的输入维度调整为2，即第四神经网络模型的输入维度为2。
具体地，若第一神经网络模型适用的RP大小是根据其适用的RP的时域长度S1和RP的频域宽度B1确定的，通信资源信息包括用于第一设备和第二设备之间进行通信的时域资源和频域资源，那么，第二输入维度或其中，S2是为第一设备和第二设备分配的时域资源的长度，B2是为第一设备和第二设备分配的频域资源的宽度，ceil()为向上取整函数。
在另一种可选的实施方式中，调整后的第一神经网络模型是根据第一信息中的信道状态信息，对第四神经网络模型进行训练得到的；第四神经网络模型是根据第一神经网络模型适用的RP大小，以及第一信息中的通信资源信息，对第一神经网络模型的输入值和/或输出值进行调整得到的。其中，关于第一神经网络模型适用的RP大小、第一信息中的通信资源信息和信道状态信息的阐述可参见前述的相关阐述，此处不再赘述。
可选的，该通信方法还可包括：若根据第一神经网络模型适用的RP大小以及第一信息中的通信资源信息确定的第二输入维度小于第一神经网络模型的输入维度，则第三设备将第一神经网络模型的M1个输入值置零，即第四神经网络模型存在M1个输入值为0。其中，M1等于第一经网络模型的输入维度减第二输入维度得到的值。可选的，该通信方法还可包括：若根据第一神经网络模型适用的RP大小以及第一信息中的通信资源信息确定的第二输出维度小于第一神经网络模型的输出维度，则第三设备将第一神经网络模型的M2个输出值置零，即第四神经网络模型存在M2个输出值为0。其中，M2等于第一经网络模型的输出维度减第二输出维度得到的值。关于第二输入维度和第二输出维度的具体阐述可参见前述的相关阐述，此处不再赘述。
以对第一神经网络模型的输入维度进行调整为例，结合图6b，第一神经网络模型的输入维度为4，若根据第一神经网络模型适用的RP大小，以及第一信息中的通信资源信息确定的第二输入维度为2，说明第二输入维度小于第一神经网络模型的输入维度，则将第一神经网络模型的2个输入值置零，即第四神经网络模型的两个输入值为0。
可选的，该通信方法还可包括：第三设备确定第一神经网络模型的每个输入和/或每个输出在第一资源中对应的RP，不同输入和/或输出对应的RP不同。其中，第一资源是第一设备的通信系统参数所表征的资源和第二设备的通信系统参数所表征的资源中第一设备和第二设备可共同工作的资源。若第二输入维度小于第一神经网络模型的输入维度，则将第一资源中 未被分配给第一设备和第二设备的资源所对应的输入的输入值置零，和/或，若第二输出维度小于第一神经网络模型的输出维度，则将第一资源中未被分配给第一设备和第二设备的资源所对应的输出的输出值置零。例如，结合图6c所示，第一神经网络模型的输入维度为4，且在第一设备和第二设备可共同工作的资源(即第一资源)中，第一神经网络模型的输入#1对应RP#1、输入#2对应RP#2、输入#3对应RP#3、输入#4对应RP#4，用于第一设备和第二设备之间进行通信的资源包括RP#1和RP#2，不包括RP#3和RP#4，那么，将RP#3对应的第一神经网络模型的输入#3的输入值置零，以及将RP#4对应的第一神经网络模型的输入#4的输入值置零。
可选的，第一信息还可包括第一设备的设备状态信息和/或第二设备的设备状态信息。在上述两种实施方式中，在对第四神经网络模型进行训练得到调整后的第一神经网络模型的过程中，除了根据第一信息中的信道状态信息对第四神经网络模型进行训练，还可结合第一设备的设备状态信息和/或第二设备的设备状态信息对第四神经网络模型进行训练。其中，第一设备和第二设备中任一设备的设备状态信息可包括该设备的负载(例如中央处理器(central processing unit，CPU)负载、内存等)和电量等。可选的，在对第四神经网络模型进行训练得到调整后的第一神经网络模型的过程中，还可对第四神经网络模型进行蒸馏，以进一步减少第四神经网络模型的参数量和运算量。
S103、第三设备发送调整后的第一神经网络模型；相应的，第二设备接收该调整后的第一神经网络模型。或者，第三设备发送调整后的第一神经网络模型中的子模型；相应的，第二设备接收该调整后的第一神经网络模型中的子模型。其中，调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
S104、第二设备基于调整后的第一神经网络模型进行通信；或者，第二设备基于调整后的第一神经网络模型中的子模型进行通信。
下面对第三设备发送的是调整后的第一神经网络模型，或是调整后的第一神经网络模型中的子模型，这两种实施方式分别进行阐述。
实施方式3.1，第三设备发送的是调整后的第一神经网络模型。
其中，若第一设备是第三设备，则第三设备向第二设备发送调整后的第一神经网络模型。若第一设备是不同于第二设备和第三设备的其他设备，则第三设备向第一设备和第二设备分别发送调整后的第一神经网络模型。
若在第一设备与第二设备的通信中，第一设备作为发送方，第二设备作为接收方，那么，第一设备可从调整后的第一神经网络模型中确定发射机子模型，并基于该发射机子模型处理待发送的信号以及向第二设备发送信号。第二设备可从调整后的第一神经网络模型中确定接收机子模型，并基于该接收机子模型接收来自第一设备的信号以及对接收的信号进行处理。第二设备在通信中作为发送方，第一设备在通信中作为接收方的情况与之类似，不再赘述。
实施方式3.2，第三设备发送的是调整后的第一神经网络模型中的子模型。
其中，在第一设备是第三设备的情况下，若第二设备在通信中作为发送方，第三设备向第二设备发送的调整后的第一神经网络模型中的子模型是：调整后的第一神经网络模型中的发射机子模型；若第二设备在通信中作为接收方，第三设备向第二设备发送的调整后的第一神经网络模型中的子模型是：调整后的第一神经网络模型中的接收机子模型。
在第一设备是不同于第二设备和第三设备的其他设备的情况下，若第一设备在通信中作为发送方且第二设备在通信中作为接收方，第三设备向第一设备发送调整后的第一神经网络模型中发射机子模型，向第二设备发送调整后的第一神经网络模型中接收机子模型。若第二 设备在通信中作为发送方且第一设备在通信中作为接收方，第三设备向第二设备发送调整后的第一神经网络模型中发射机子模型，向第一设备发送调整后的第一神经网络模型中接收机子模型。
另外，在第一设备和第二设备之间的通信为双向通信，即第一设备既作为发送方又作为接收方，第二设备也是既作为发送方又作为接收方的情况下，若双向通信包括的反向的两个单向通信中，第一设备和第二设备支持的系统带宽的最小值相同、支持的帧结构对应的帧长度的最小值相同、用于第一设备和第二设备之间进行通信的资源的大小相同，那么，第一设备和第二设备可基于确定的调整后的第一神经网络模型进行双向通信。其中，第一设备作为发送方时采用调整后的第一神经网络模型中发射机子模型进行通信，作为接收方时采用调整后的第一神经网络模型中接收机子模型进行通信。第二设备分别作为发送方和接收方的情况与之类似，不再赘述。否则，第三设备需针对反向的两个单向通信分别执行本申请实施例提供的通信方法，以分别得到每个单向通信下用于第一设备和第二设备之间通信的神经网络模型。
在一种可选的实施方式中，第三设备发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型之后，该通信方法还可包括：第三设备获取第一设备和第二设备基于调整后的第一神经网络模型进行通信时的性能指标，该性能指标包括以下一个或多个：吞吐量、误码率、丢包率、通信时延等。若性能指标不满足预设条件，第三设备获取第五神经网络模型；该第五神经网络模型是第一设备和第二设备对调整后的第一神经网络模型进行联合训练得到的。可见，性能指标不满足预设条件可说明基于调整后的第一神经网络模型进行的通信质量低，即调整后的第一神经网络模型应用于第一设备和第二设备之间通信时的性能低。那么，由第一设备和第二设备对调整后的第一神经网络模型进行联合训练，可提高调整后的第一神经网络模型的性能。
其中，若性能指标是吞吐量，预设条件为吞吐量大于第一值；若性能指标是误码率，预设条件为误码率小于第二值；若性能指标是丢包率，预设条件为丢包率小于第三值；若性能指标是通信时延，预设条件为通信时延小于第四值；关于其他性能指标对应的预设条件与前述类似，此处不再过多阐述。可选的，第一值可以是预设的，还可以是协议中定义的。其他性能指标对应的预设条件中阈值(如上述的第二值、第三值、第四值等)的确定方式与第一值类似，不再赘述。
其中，第一设备和第二设备对调整后的第一神经网络模型进行联合训练，可包括：第一设备和第二设备中作为接收方的设备采用损失函数计算调整后的第一神经网络模型中接收机子模型的参数的梯度值，并基于计算得到的梯度值更新该接收机子模型的参数；作为接收方的设备还将该接收机子模型的参数的梯度值发送给第一设备和第二设备中作为发送方的设备。那么，作为发送方的设备可基于接收的梯度值计算调整后的第一神经网络模型中发射机子模型的参数的梯度值，并基于计算得到的梯度值更新该发射机子模型的参数。也就是说，第五神经网络模型中发射机子模型是调整后的第一神经网络模型中发射机子模型的参数进行更新得到的，第五神经网络模型中接收机子模型是调整后的第一神经网络模型中接收机子模型的参数进行更新得到的。
可选的，基于调整后的第一神经网络模型进行通信时的性能指标可以是第三设备不定期获取的或是周期性获取的。那么，第三设备可得到第一设备和第二设备基于调整后的第一神经网络模型进行通信时的性能指标的变化趋势，从而可得到调整后的第一神经网络模型应用于第一设备和第二设备之间通信时的性能的变化趋势。该方式有利于在确定调整后的第一神 经网络模型的性能逐渐降低时，第一设备和第二设备能够及时对调整后的第一神经网络模型进行联合训练，以减少调整后的第一神经网络模型的性能降低对第一设备和第二设备之间的通信质量所造成的影响。
可选的，若第三设备存储了一个或多个预训练神经网络模型，该通信方法还包括：第三设备根据第五神经网络模型，对一个或多个预训练神经网络模型中的第三神经网络模型进行更新。若第三设备未存储一个或多个预训练神经网络模型，且模型服务器存储了一个或多个预训练神经网络模型，该通信方法还可包括：第三设备向模型服务器发送第五神经网络模型；模型服务器根据第五神经网络模型，对一个或多个预训练神经网络模型中的第三神经网络模型进行更新。可选的，第三设备或模型服务器对一个或多个预训练神经网络模型中的第三神经网络模型进行更新的操作可以是采用迁移学习的方式实现的。
另外，第三设备和第二设备之间可预先建立普通空口。第一设备是不同于第二设备和第三设备的其他设备的情况下，第三设备还与第一设备之间预先建立普通空口，第一设备与第二设备之间可预先建立普通空口。若通信系统中还存在模型服务器，第三设备与模型服务器之间也可预先建立普通空口。普通空口可以是基于信道编码、调制、资源映射、预编码等模块中的一个或多个建立的，还可以是基于AI技术中的神经网络建立的，此处不作限定。
该通信方法中，第一设备和第二设备之间基于调整后的第一神经网络模型进行通信之前，不同设备之间交互的信息可通过普通空口传输。其中，不同设备之间交互的信息可包括以下的部分或全部：第二设备向第三设备发送的通信系统参数、业务信息、算力、时延要求、存储空间、信道状态信息、设备状态信息等、模型服务器向第三设备发送的每个预训练神经网络模型的信息或第一神经网络模型等、第三设备向第二设备发送的调整后的第一神经网络模型或调整后的第一神经网络模型中的子模型等。第一设备是不同于第二设备和第三设备的其他设备的情况下，不同设备之间交互的信息还可包括：第一设备向第三设备发送的通信系统参数、业务信息、算力、时延要求、存储空间、信道状态信息、设备状态信息等、第三设备向第一设备发送的调整后的第一神经网络模型或调整后的第一神经网络模型中的子模型等。
第一设备和第二设备在获取了调整后的第一神经网络模型之后，可基于该调整后的第一神经网络模型建立AI空口，该AI空口与上述的普通空口不同，该AI空口可用于传输第一设备和第二设备基于调整后的第一神经网络模型进行通信的信号(如数据、信息、控制信令等)。或者，第一设备和第二设备基于调整后的第一神经网络模型进行通信的信号还可直接复用普通空口进行传输。此处不作限定。
综上所述，该通信方法中，第三设备从一个或多个预训练神经网络模型中确定第一神经网络模型；根据两个设备之间进行通信的通信资源信息和/或信道状态信息，调整第一神经网络模型，并发送调整后的第一神经网络模型，该调整后的第一神经网络模型用于两个设备之间进行通信。那么，第一设备和第二设备之间可基于调整后的第一神经网络模型进行通信。可见，第一设备和第二设备之间进行通信所采用的神经网络模型，是根据已训练好的预训练神经网络模型进行调整得到的，且确定的神经网络模型适配于第一设备和第二设备的通信场景。与第一设备和第二设备重新训练神经网络模型以用于通信的方式相比，可降低神经网络模型确定的复杂度，还减少了设备间为确定神经网络模型而交互的信令开销。另外，与第一设备和第二设备之间采用多个模块(如包括编码、调制、多接入、波形、射频等处理模块，各模块还可能包含多个子模块)处理和传输信号的方式相比，避免了单独设计和优化各模块而对整体性能所造成的影响。
以第一设备是第三设备为例，本申请实施例还提供了另一种通信方法，如图7所示。图7所示的通信方法为图4所示的通信方法中的一种具体实施方法。该通信方法包括以下步骤：
S201、第二设备向第三设备发送第二设备的通信系统参数和业务信息；相应的，第三设备接收第二设备的通信系统参数和业务信息。其中，第二设备的业务信息可包括第二设备需求的eMBB、URLLC和mMTC等业务类型中的一个；或者，第二设备的业务信息可包括第二设备对神经网络模型的通信时延、吞吐量等性能指标的需求。
S202、第三设备根据自身的业务信息和第二设备的业务信息，确定第一业务类型。其中，第三设备的业务信息可包括第三设备需求的eMBB、URLLC和mMTC等业务类型中的一个；或者，第三设备的业务信息可包括第三设备对神经网络模型的通信时延、吞吐量等性能指标的需求。
S203、第三设备根据第一业务类型、自身的通信系统参数以及第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定一个或多个第二神经网络模型。
其中，每个第二神经网络模型均支持第一业务类型；每个第二神经网络模型的输入维度与其对应的第一输入维度相同，和/或，每个第二神经网络模型的输出维度与其对应的第一输出维度相同。其中，每个第二神经网络模型对应的第一输入维度和/或第一输出维度，是根据该第二神经网络模型适用的RP大小，以及第三设备的通信系统参数和/或第二设备的通信系统参数确定的。
S204、第三设备从一个或多个第二神经网络模型中确定第三神经网络模型。
其中，第二神经网络模型为一个时，将该第二神经网络模型作为第三神经网络模型。第二神经网络模型为多个时，将多个第二神经网络模型中参数量最大的第二神经网络模型作为第三神经网络模型；或者，将多个第二神经网络模型中运算量与第一运算量之间的绝对差值最小的第二神经网络模型作为第三神经网络模型；或者，将多个第二神经网络模型中参数量与第一参数量之间的绝对差值最小的第二神经网络模型作为第三神经网络模型。
S205、第三设备向模型服务器发送模型请求信息，模型请求信息包括第三神经网络模型的标识、第一运算量和第一参数量。其中，第一运算量可以是根据第三设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量可以是根据第三设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
S206、模型服务器根据第三神经网络模型的标识，从存储的一个或多个预训练神经网络模型中确定第三神经网络模型。
S207a、模型服务器在第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量时，对第三神经网络模型进行蒸馏，得到第一神经网络模型。
S207b、模型服务器在第三神经网络模型的运算量小于或等于第一运算量，和/或，第三神经网络模型的参数量小于或等于第一参数量时，将第三神经网络模型作为第一神经网络模型。
其中，S207a、S207b是并列的步骤，具体执行哪一个步骤可以根据所判断的条件的结果来确定。
S208、模型服务器向第三设备发送第一神经网络模型；相应的，第三设备接收来自模型服务器的第一神经网络模型。
S209、第三设备根据第一信息调整第一神经网络模型；该第一信息包括第二设备和第三设备之间进行通信的通信资源信息和/或信道状态信息。
S210、第三设备向第二设备发送调整后的第一神经网络模型。相应的，第二设备接收调 整后的第一神经网络模型。
S211、第三设备和第二设备之间基于调整后的第一神经网络模型进行通信。
S212、第三设备监测基于调整后的第一神经网络模型进行通信时的性能指标。
S213、在性能指标不满足预设条件时，第三设备与第二设备联合训练调整后的第一神经网络模型，得到第五神经网络模型。
S214、第三设备向模型服务器发送第五神经网络模型；相应的，模型服务器接收该第五神经网络模型。
S215、模型服务器根据第五神经网络模型对存储的第三神经网络模型进行更新。
关于上述各步骤的具体阐述可参见图4所示的通信方法中的相关阐述，此处不再赘述。
第一设备和第二设备是网络设备和终端设备时，本申请实施例提供的通信方法可确定用于网络设备和终端设备之间进行上行通信的神经网络模型，也可确定用于网络设备和终端设备之间进行下行通信的神经网络模型。第一设备和第二设备是两个终端设备时，该通信方法可确定用于两个终端设备之间进行设备到设备通信(device-to-device，D2D)的神经网络模型。
另外，在存在多组设备均采用神经网络模型进行通信的情况下，本申请实施例提供的通信方法还可确定每组设备所采用的神经网络模型，其中，多组设备中每组设备包括通信的两个设备，且不同组设备包括的两个设备中存在至少一个设备不同。例如，组1包括第一设备1和第二设备1，组2包括第一设备1和第二设备2，组3包括第一设备2和第二设备3。
具体地，在存在N1组设备且每组设备包括通信的两个设备的情况下，第三设备可针对N1组设备中每组设备分别从一个或多个预训练神经网络模型中确定该组设备对应的第三神经网络模型。若存在N2组设备对应的第三神经网络模型相同，且由模型服务器存储一个或多个预训练神经网络模型，则模型服务器可执行一次根据第三神经网络的标识获取第三神经网络模型的操作，而不必针对N2组设备分别执行获取第三神经网络模型的操作。第三设备针对N1组设备中每组设备分别确定该组设备对应的第一神经网络模型，以及执行图4中的步骤S102、S103和S104。其中，N1、N2均为大于1的整数，且N2小于或等于N1。
结合图8a，以第一设备是第三设备，第三设备是网络设备，第二设备是终端设备为例。存在n组设备(n为大于1的整数)，该n组设备中每组设备均包括第三设备和第二设备且不同组设备包括的第二设备不同，并且，n组设备中每组设备包括的第三设备和第二设备之间进行蜂窝网上行通信。
第三设备针对该n组设备确定了相同的第三神经网络模型(即模型A)，那么，针对n组设备中的每组设备，模型服务器可根据模型A，以及该组设备对应的第一运算量和/或第一参数量确定第一神经网络模型，并将确定的第一神经网络模型发送给第三设备。第三设备可得到n组设备中每组设备对应的第一神经网络模型，针对每组设备分别执行：根据该组设备包括的两个设备之间进行通信的通信资源信息和/或信道状态信息，调整第一神经网络模型，得到调整后的第一神经网络模型；即得到对应第三设备和第二设备1确定的模型B1、对应第三设备和第二设备2确定的模型B2、…、对应第三设备和第二设备n确定的模型Bn。接着，第三设备可向第二设备1发送模型B1中的发射机子模型，向第二设备2发送模型B2中的发射机子模型，…，向第二设备n发送模型Bn中的发射机子模型。那么，在第三设备与第二设备1进行上行通信时，第二设备1基于模型B1中发射机子模型处理待发送的信号以及发送信号，第三设备基于模型B1中接收子模型接收信号以及对接收的信号进行处理。第三设备与第二设备2、…、第二设备n分别进行上行通信的情况与前述的第三设备和第二设备1 之间进行上行通信的情况类似，此处不再过多阐述。
同理地，前述的n组设备中每组设备包括的第三设备和第二设备之间进行蜂窝网下行通信的情况下，第三设备可针对n组设备分别确定调整后的第一神经网络模型，并向每组设备中的第二设备发送调整后的第一神经网络模型中的接收机子模型。在第三设备与每个第二设备进行下行通信时，第三设备基于调整后的第一神经网络模型中发射机子模型处理待发送的信号以及发送信号，第二设备基于调整后的第一神经网络模型中接收机子模型接收信号以及对接收的信号进行处理。
结合图8b，以第一设备是不同于第二设备和第三设备的其他设备，第三设备是网络设备，第一设备和第二设备均是终端设备为例。存在m组设备(m为大于1的整数)，该m组设备中每组设备包括第一设备和第二设备，且不同组设备之间存在至少一个设备不同。该m组设备中每组设备包括的第一设备和第二设备之间进行的通信为D2D。
第三设备针对该m组设备确定了相同的第三神经网络模型(即模型C)，那么，针对m组设备中的每组设备，模型服务器可根据模型C，以及该组设备对应的第一运算量和/或第一参数量确定第一神经网络模型，并将确定的第一神经网络模型发送给第三设备。第三设备可得到m组设备中每组设备对应的第一神经网络模型，针对每组设备分别执行：根据该组设备包括的第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息，调整第一神经网络模型，得到调整后的第一神经网络模型；即得到对应第一设备1和第二设备1确定的模型D1、对应第一设备2和第二设备2确定的模型D2、…、对应第一设备m和第二设备m确定的模型Dm。接着，第三设备可向作为发送方的第一设备1发送模型D1中的发射机子模型，向作为接收方的第二设备1发送模型D2中的发射机子模型；…；向作为发送方的第一设备m发送模型Dm中的发射机子模型，向作为接收方的第二设备m发送模型Dm中的接收机子模型。那么，第一设备1可基于模型D1中发射机子模型处理待发送的信号以及发送信号，第二设备1基于模型D1中接收子模型接收信号以及对接收的信号进行处理。m组设备中其他组设备包括的第一设备和第二设备之间的D2D与前述的第一设备1和第二设备1之间进行D2D的情况类似，此处不再过多阐述。
本申请实施例提供的通信方法中第三设备的相关操作还可通过设计一种协议栈的方式来实现。一种实施方式中，可在协议栈的媒体接入控制(medio access control，MAC)层中增加“模型获取与分发”功能模块，如图9所示。该功能模块具有上述方法实施例中第三设备的功能，即该功能模块可具有模型选择、模型蒸馏、模型调整、模型下发以及模型更新等功能。具体地，在MAC实体的控制器中增加“模型获取与分发”功能模块。另一实施方式中，可在协议栈中新增一个协议层，该协议层中包括“模型获取与分发”功能模块。在协议栈中，该新增的协议层可以与MAC层位于同一层级。
其中，“模型获取与分发”功能模块的模型选择功能可包括：从一个或多个预训练神经网络模型中确定第三神经网络模型。
模型蒸馏功能可包括：在第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量时，对第三神经网络模型进行蒸馏，得到第一神经网络模型；否者，将第三神经网络模型作为第一神经网络模型。
模型调整功能可包括：根据第一信息调整第一神经网络模型；该第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息。
模型下发功能可包括：发送调整后的第一神经网络模型，或者，发送调整后的第一神经 网络模型中的子模型；该调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
模型更新功能包括：获取第一设备和第二设备基于调整后的第一神经网络模型进行通信时的性能指标；在性能指标不满足预设条件时，获取第一设备和第二设备对调整后的第一神经网络模型进行联合训练得到的第五神经网络模型，并根据该第五神经网络模型对一个或多个预训练神经网络模型中的第三神经网络模型进行更新。
关于上述模型选择、模型蒸馏、模型调整、模型下发以及模型更新等功能的具体阐述可参见上述方法实施例中的相关阐述，此处不再赘述。
为了实现上述本申请实施例提供的方法中的各功能，第三设备或第二设备可以包括硬件结构和/或软件模块，以硬件结构、软件模块、或硬件结构加软件模块的形式来实现上述各功能。上述各功能中的某个功能以硬件结构、软件模块、还是硬件结构加软件模块的方式来执行，取决于技术方案的特定应用和设计约束条件。
如图10所示，本申请实施例提供了一种通信装置1000。该通信装置1000可以是第三设备的部件(例如，集成电路，芯片等等)，也可以是第二设备的部件(例如，集成电路，芯片等等)。该通信装置1000也可以是其他通信单元，用于实现本申请方法实施例中的方法。该通信装置1000可以包括：通信单元1001和处理单元1002。其中，处理单元1002用于控制通信单元1001进行数据/信令收发。可选的，通信装置1000还可以包括存储单元1003。
在一种可能的设计中，处理单元1002，用于从一个或多个预训练神经网络模型中确定第一神经网络模型。
处理单元1002，还用于根据第一信息调整第一神经网络模型；第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息。
通信单元1001，用于发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型；调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
一种可选的实施方式中，第一神经网络模型是根据第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定的。
第一设备的通信系统参数包括第一设备支持的系统带宽和帧结构，第二设备的通信系统参数包括第二设备支持的系统带宽和帧结构。
一种可选的实施方式中，第一神经网络模型是基于从一个或多个第二神经网络模型中选择的第三神经网络模型确定的；一个或多个第二神经网络模型是从一个或多个预训练神经网络模型中确定的。
一个或多个第二神经网络模型中每个第二神经网络模型的输入维度与其对应的第一输入维度相同，和/或，每个第二神经网络模型的输出维度与其对应的第一输出维度相同；每个第二神经网络模型对应的第一输入维度和/或第一输出维度，是根据该第二神经网络模型适用的RP大小，以及第一设备的通信系统参数和/或第二设备的通信系统参数确定的。
可选的，RP大小是根据RP的时域长度、RP的频域宽度确定的。
一种可选的实施方式中，第二神经网络模型支持第一业务类型；第一业务类型是第一设备和第二设备需求的业务类型。
一种可选的实施方式中，如果第二神经网络模型为多个，第三神经网络模型是多个第二神经网络模型中参数量最大的第二神经网络模型。
或者，第三神经网络模型是多个第二神经网络模型中运算量与第一运算量之间的绝对差值最小的第二神经网络模型，第一运算量是根据第一设备的算力和时延要求，和/或，第二设 备的算力和时延要求确定的。
或者，第三神经网络模型是多个第二神经网络模型中参数量与第一参数量之间的绝对差值最小的第二神经网络模型，第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
一种可选的实施方式中，第一神经网络模型是对第三神经网络模型进行蒸馏得到的；第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量。
第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
一种可选的实施方式中，第一神经网络模型为第三神经网络模型；第三神经网络模型的运算量小于或等于第一运算量，和/或，第三神经网络模型的参数量小于或等于第一参数量。
第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
一种可选的实施方式中，第一神经网络模型是模型服务器根据接收的模型请求信息确定的；模型请求信息包括第三神经网络模型的标识，以及第一运算量和/或第一参数量。
第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的，第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
一种可选的实施方式中，一个或多个预训练神经网络模型中每个预训练神经网络模型的信息是预定义的，或是从模型服务器获取的。
每个预训练神经网络模型的信息包括以下一个或多个：标识、业务类型、RP大小、输入维度、输出维度、参数量和运算量。
一种可选的实施方式中，调整后的第一神经网络模型是根据第一信息中的信道状态信息，对第四神经网络模型进行训练得到的。
第四神经网络模型是根据第一神经网络模型适用的RP大小，以及第一信息中的通信资源信息，对第一神经网络模型的输入维度和/或输出维度进行调整得到的。
本申请实施例和上述所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述所示实施例的描述，不再赘述。
在另一种可能的设计中，通信单元1001，用于接收调整后的第一神经网络模型，或者，接收调整后的第一神经网络模型中的子模型；调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
调整后的第一神经网络模型是基于第一信息对第一神经网络模型调整得到的，第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；第一神经网络设备是从一个或多个预训练神经网络模型中确定的；
通信单元1001，还用于基于调整后的第一神经网络模型进行通信，或者，基于调整后的第一神经网络模型中的子模型进行通信。
一种可选的实施方式中，第一神经网络模型是根据第一设备的通信系统参数和/或第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定的。
第一设备的通信系统参数包括第二设备支持的系统带宽和帧结构，第二设备的通信系统参数包括第二设备支持的系统带宽和帧结构。
一种可选的实施方式中，通信单元1001，还用于发送通信系统参数。
一种可选的实施方式中，第一神经网络模型是基于从一个或多个第二神经网络模型中选择的第三神经网络模型确定的；一个或多个第二神经网络模型是从一个或多个预训练神经网络模型中确定的。
一个或多个第二神经网络模型中每个第二神经网络模型的输入维度与其对应的第一输入维度相同，和/或，每个第二神经网络模型的输出维度与其对应的第一输出维度相同。每个第二神经网络模型对应的第一输入维度和/或第一输出维度，是根据该第二神经网络模型适用的RP大小，以及第一设备的通信系统参数和/或第二设备的通信系统参数确定的。
可选的，RP大小是根据RP的时域长度、RP的频域宽度确定的。
一种可选的实施方式中，第二神经网络模型支持第一业务类型；第一业务类型是第一设备和第二设备需求的业务类型。
一种可选的实施方式中，通信单元1001，还用于发送需求的业务类型。
一种可选的实施方式中，如果第二神经网络模型为多个，第三神经网络模型是多个第二神经网络模型中参数量最大的第二神经网络模型。
或者，第三神经网络模型是多个第二神经网络模型中运算量与第一运算量之间的绝对差值最小的第二神经网络模型，第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的。
或者，第三神经网络模型是多个第二神经网络模型中参数量与第一参数量之间的绝对差值最小的第二神经网络模型，第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
一种可选的实施方式中，通信单元1001，还用于发送算力和时延要求，或者发送算力和存储空间，或者发送算力、时延要求和存储空间。
一种可选的实施方式中，第一神经网络模型是对第三神经网络模型进行蒸馏得到的；第三神经网络模型的运算量大于第一运算量，和/或，第三神经网络模型的参数量大于第一参数量。
第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
一种可选的实施方式中，第一神经网络模型为第三神经网络模型；第三神经网络模型的运算量小于或等于第一运算量，和/或，第三神经网络模型的参数量小于或等于第一参数量。
第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的；第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
一种可选的实施方式中，第一神经网络模型是模型服务器根据接收的模型请求信息确定的；模型请求信息包括第三神经网络模型的标识，以及第一运算量和/或第一参数量。
第一运算量是根据第一设备的算力和时延要求，和/或，第二设备的算力和时延要求确定的，第一参数量是根据第一设备的算力和存储空间，和/或，第二设备的算力和存储空间确定的。
可选的，一个或多个预训练神经网络模型中每个神经网络模型的信息是预定义的，或是 从模型服务器获取的；每个预训练神经网络模型的信息包括以下一个或多个：标识、业务类型、RP大小、输入维度、输出维度、参数量和运算量。
一种可选的实施方式中，调整后的第一神经网络模型是根据第一信息中的信道状态信息，对第四神经网络模型进行训练得到的。
第四神经网络模型是根据第一神经网络模型适用的RP大小，以及第一信息中的通信资源信息，对第一神经网络模型的输入维度和/或输出维度进行调整得到的。
本申请实施例和上述所示方法实施例基于同一构思，其带来的技术效果也相同，具体原理请参照上述所示实施例的描述，不再赘述。
本申请实施例还提供一种通信装置1100，如图11所示。通信装置1100可以是第三设备或第二设备，也可以是支持第三设备实现上述方法的芯片、芯片系统、或处理器等，还可以是支持第二设备实现上述方法的芯片、芯片系统、或处理器等。该装置可用于实现上述方法实施例中描述的方法，具体可以参见上述方法实施例中的说明。
所述通信装置1100可以包括一个或多个处理器1101。所述处理器1101可以是通用处理器或者专用处理器等。例如可以是基带处理器、数字信号处理器、专用集成电路、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件或中央处理器(Central Processing Unit，CPU)。基带处理器可以用于对通信协议以及通信数据进行处理，中央处理器可以用于对通信装置(如，基站、基带芯片，终端、终端芯片，分布单元(distributed unit，DU)或集中单元(centralized unit，CU)等)进行控制，执行软件程序，处理软件程序的数据。
可选的，通信装置1100中可以包括一个或多个存储器1102，其上可以存有指令1104，所述指令可在处理器1101上被运行，使得通信装置1100执行上述方法实施例中描述的方法。可选的，存储器1102中还可以存储有数据。处理器1101和存储器1102可以单独设置，也可以集成在一起。
存储器1102可包括但不限于硬盘(hard disk drive，HDD)或固态硬盘(solid-state drive，SSD)等非易失性存储器，随机存储记忆体(random access memory，RAM)、可擦除可编程只读存储器(erasable programmable ROM，EPROM)、ROM或便携式只读存储器(compact disc read-only memory，CD-ROM)等等。
可选的，所述通信装置1100还可以包括收发器1105、天线1106。所述收发器1105可以称为收发单元、收发机、或收发电路等，用于实现收发功能。收发器1105可以包括接收器和发送器，接收器可以称为接收机或接收电路等，用于实现接收功能；发送器可以称为发送机或发送电路等，用于实现发送功能。
通信装置1100为第三设备：收发器1105用于执行上述图4所示通信方法中的S103，以及用于执行图9所示通信方法中的S201、S205、S208、S210、S211、S213和S214。处理器1101用于执行上述图4所示通信方法中的S101、S102，以及用于执行图9所示通信方法中的S202至S204、S209、S212、S213。
通信装置1100为第二设备：收发器1105用于执行上述图4所示通信方法中的S103，以及用于执行图9所示通信方法中的S201、S210、S211、S213。处理器1101用于执行上述图4所示通信方法中的S104，以及用于执行图9所示通信方法中的S213。
另一种可能的设计中，处理器1101中可以包括用于实现接收和发送功能的收发器。例如该收发器可以是收发电路，或者是接口，或者是接口电路。用于实现接收和发送功能的收发电路、接口或接口电路可以是分开的，也可以集成在一起。上述收发电路、接口或接口电路 可以用于代码/数据的读写，或者，上述收发电路、接口或接口电路可以用于信号的传输或传递。
又一种可能的设计中，可选的，处理器1101可以存有指令1103，指令1103在处理器1101上运行，可使得所述通信装置1100执行上述方法实施例中描述的方法。指令1103可能固化在处理器1101中，该种情况下，处理器1101可能由硬件实现。
又一种可能的设计中，通信装置1100可以包括电路，所述电路可以实现前述方法实施例中发送或接收或者通信的功能。本申请实施例中描述的处理器和收发器可实现在集成电路(integrated circuit，IC)、模拟IC、射频集成电路(radio frequency integrated circuit，RFIC)、混合信号IC、专用集成电路(application specific integrated circuit，ASIC)、印刷电路板(printed circuit board，PCB)、电子设备等上。该处理器和收发器也可以用各种IC工艺技术来制造，例如互补金属氧化物半导体(complementary metal oxide semiconductor，CMOS)、N型金属氧化物半导体(nMetal-oxide-semiconductor，NMOS)、P型金属氧化物半导体(positive channel metal oxide semiconductor，PMOS)、双极结型晶体管(bipolar junction transistor，BJT)、双极CMOS(BiCMOS)、硅锗(SiGe)、砷化镓(GaAs)等。
以上实施例描述中的通信装置可以是第一设备或第二设备，但本申请实施例中描述的通信装置的范围并不限于此，而且通信装置的结构可以不受图11的限制。通信装置可以是独立的设备或者可以是较大设备的一部分。例如所述通信装置可以是：
(1)独立的集成电路IC，或芯片，或，芯片系统或子系统；
(2)具有一个或多个IC的集合，可选的，该IC集合也可以包括用于存储数据，指令的存储部件；
(3)ASIC，例如调制解调器(modulator)；
(4)可嵌入在其他设备内的模块；
(5)接收机、终端、智能终端、蜂窝电话、无线设备、手持机、移动单元、车载设备、网络设备、云设备、人工智能设备等等；
(6)其他等等。
对于通信装置可以是芯片或芯片系统的情况，可参见图12所示的芯片的结构示意图。图12所示的芯片1200包括处理器1201和接口1202。其中，处理器1201的数量可以是一个或多个，接口1202的数量可以是多个。该处理器1201可以是逻辑电路，该接口1202可以是输入输出接口、输入接口或输出接口。所述芯片1200还可包括存储器1203。
一种设计中，对于芯片用于实现本申请实施例中第一设备的功能的情况：
处理器1201，用于从一个或多个预训练神经网络模型中确定第一神经网络模型。
处理器1201，还用于根据第一信息调整第一神经网络模型；第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息。
接口1202，用于发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型；调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
另一种设计中，对于芯片用于实现本申请实施例中第二设备的功能的情况：
接口1202，用于接收调整后的第一神经网络模型，或者，接收调整后的第一神经网络模型中的子模型；调整后的第一神经网络模型用于第一设备和第二设备之间进行通信。
调整后的第一神经网络模型是基于第一信息对第一神经网络模型调整得到的，第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；第一神经网络设备是从一个或多个预训练神经网络模型中确定的。
接口1202，还用于基于调整后的第一神经网络模型进行通信，或者，基于调整后的第一神经网络模型中的子模型进行通信。
本申请实施例中通信装置1100、芯片1200还可执行上述通信装置1000所述的实现方式。本领域技术人员还可以了解到本申请实施例列出的各种说明性逻辑块(illustrative logical block)和步骤(step)可以通过电子硬件、电脑软件，或两者的结合进行实现。这样的功能是通过硬件还是软件来实现取决于特定的应用和整个系统的设计要求。本领域技术人员可以对于每种特定的应用，可以使用各种方法实现所述的功能，但这种实现不应被理解为超出本申请实施例保护的范围。
本申请实施例和上述图4所示的通信方法和图9所示的通信方法基于同一构思，其带来的技术效果也相同，具体原理请参照上述图4所示的通信方法和图9所示的通信方法中的描述，不再赘述。
本领域技术人员还可以了解到本申请实施例列出的各种说明性逻辑块(illustrative logical block)和步骤(step)可以通过电子硬件、电脑软件，或两者的结合进行实现。这样的功能是通过硬件还是软件来实现取决于特定的应用和整个系统的设计要求。本领域技术人员可以对于每种特定的应用，可以使用各种方法实现所述的功能，但这种实现不应被理解为超出本申请实施例保护的范围。
本申请还提供了一种计算机可读存储介质，用于储存计算机软件指令，当所述指令被通信装置执行时，实现上述任一方法实施例的功能。
本申请还提供了一种计算机程序产品，用于储存计算机软件指令，当所述指令被通信装置执行时，实现上述任一方法实施例的功能。
本申请还提供了一种计算机程序，当其在计算机上运行时，实现上述任一方法实施例的功能。
本申请还提供了一种通信系统，该系统包括上述方面的至少一个第一设备、至少一个第二设备。在另一种可能的设计中，该系统还包括上述方面的至少一个模型服务器。又一种可能的设计中，该系统还可以包括本申请提供的方案中与第一设备、第二设备进行交互的其他设备。
上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机指令时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(digital subscriber line，DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质(例如，软盘、硬盘、磁带)、光介质(例如，高密度数字视频光盘(digital video disc，DVD))、或者半导体介质(例如，SSD)等。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。

Claims (28)

  1. 一种通信方法，其特征在于，所述方法包括：
    从一个或多个预训练神经网络模型中确定第一神经网络模型；
    根据第一信息调整所述第一神经网络模型；所述第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；
    发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型；
    所述调整后的第一神经网络模型用于所述第一设备和所述第二设备之间进行通信。
  2. 一种通信方法，其特征在于，所述方法包括：
    接收调整后的第一神经网络模型，或者，接收调整后的第一神经网络模型中的子模型；所述调整后的第一神经网络模型用于第一设备和第二设备之间进行通信；
    所述调整后的第一神经网络模型是基于第一信息对第一神经网络模型调整得到的，所述第一信息包括所述第一设备和所述第二设备之间进行通信的通信资源信息和/或信道状态信息；所述第一神经网络设备是从一个或多个预训练神经网络模型中确定的；
    基于所述调整后的第一神经网络模型进行通信，或者，基于所述调整后的第一神经网络模型中的子模型进行通信。
  3. 根据权利要求1或2所述的方法，其特征在于，
    所述第一神经网络模型是根据所述第一设备的通信系统参数和/或所述第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定的；所述第一设备的通信系统参数包括所述第一设备支持的系统带宽和帧结构，所述第二设备的通信系统参数包括所述第二设备支持的系统带宽和帧结构。
  4. 根据权利要求3所述的方法，其特征在于，
    所述第一神经网络模型是基于从一个或多个第二神经网络模型中选择的第三神经网络模型确定的；
    所述一个或多个第二神经网络模型是从所述一个或多个预训练神经网络模型中确定的；
    所述一个或多个第二神经网络模型中每个第二神经网络模型的输入维度与其对应的第一输入维度相同，和/或，所述每个第二神经网络模型的输出维度与其对应的第一输出维度相同；
    所述每个第二神经网络模型对应的第一输入维度和/或第一输出维度，是根据所述第二神经网络模型适用的资源片RP大小，以及所述第一设备的通信系统参数和/或所述第二设备的通信系统参数确定的。
  5. 根据权利要求4所述的方法，其特征在于，所述RP大小是根据所述RP的时域长度和所述RP的频域宽度确定的。
  6. 根据权利要求4或5所述的方法，其特征在于，所述第二神经网络模型支持第一业务类型；所述第一业务类型是所述第一设备和所述第二设备需求的业务类型。
  7. 根据权利要求4至6任一项所述的方法，其特征在于，
    所述第一神经网络模型是对所述第三神经网络模型进行蒸馏得到的；
    所述第三神经网络模型的运算量大于第一运算量，和/或，所述第三神经网络模型的参数量大于第一参数量；
    所述第一运算量是根据所述第一设备的算力和时延要求，和/或，所述第二设备的算力和实时延要求确定的；所述第一参数量是根据所述第一设备的算力和存储空间，和/或，所述第二设备的算力和存储空间确定的。
  8. 根据权利要求4至6任一项所述的方法，其特征在于，
    所述第一神经网络模型为所述第三神经网络模型；
    所述第三神经网络模型的运算量小于或等于第一运算量，和/或，所述第三神经网络模型的参数量小于或等于第一参数量；
    所述第一运算量是根据所述第一设备的算力和时延要求，和/或，所述第二设备的算力和实时延要求确定的；所述第一参数量是根据所述第一设备的算力和存储空间，和/或，所述第二设备的算力和存储空间确定的。
  9. 根据权利要求4至6任一项所述的方法，其特征在于，
    所述第一神经网络模型是模型服务器根据接收的模型请求信息确定的；
    所述模型请求信息包括所述第三神经网络模型的标识，以及第一运算量和/或第一参数量；
    所述第一运算量是根据所述第一设备的算力和时延要求，和/或，所述第二设备的算力和实时延要求确定的，所述第一参数量是根据所述两个设备的算力和存储空间，和/或，所述第二设备的算力和存储空间确定的。
  10. 根据权利要求9所述的方法，其特征在于，
    所述一个或多个预训练神经网络模型中每个预训练神经网络模型的信息是预定义的，或是从所述模型服务器获取的；
    每个预训练神经网络模型的信息包括以下一个或多个：标识、业务类型、RP大小、输入维度、输出维度、参数量和运算量。
  11. 根据权利要求1至10任一项所述的方法，其特征在于，
    所述调整后的第一神经网络模型是根据所述第一信息中的所述信道状态信息，对第四神经网络模型进行训练得到的；
    所述第四神经网络模型是根据所述第一神经网络模型适用的RP大小，以及所述第一信息中的所述通信资源信息，对所述第一神经网络模型的输入维度和/或输出维度进行调整得到的。
  12. 一种通信装置，其特征在于，所述装置包括：
    处理单元，用于从一个或多个预训练神经网络模型中确定第一神经网络模型；
    所述处理单元，还用于根据第一信息调整所述第一神经网络模型；所述第一信息包括第一设备和第二设备之间进行通信的通信资源信息和/或信道状态信息；
    通信单元，用于发送调整后的第一神经网络模型，或者，发送调整后的第一神经网络模型中的子模型；所述调整后的第一神经网络模型用于所述第一设备和所述第二设备之间进行通信。
  13. 一种通信装置，其特征在于，所述装置包括：
    通信单元，用于接收调整后的第一神经网络模型，或者，接收调整后的第一神经网络模型中的子模型；所述调整后的第一神经网络模型用于第一设备和第二设备之间进行通信；
    所述调整后的第一神经网络模型是基于第一信息对第一神经网络模型调整得到的，所述第一信息包括所述第一设备和所述第二设备之间进行通信的通信资源信息和/或信道状态信息；所述第一神经网络设备是从一个或多个预训练神经网络模型中确定的；
    所述通信单元，还用于基于所述调整后的第一神经网络模型进行通信，或者，基于所述调整后的第一神经网络模型中的子模型进行通信。
  14. 根据权利要求12或13所述的装置，其特征在于，
    所述第一神经网络模型是根据所述第一设备的通信系统参数和/或所述第二设备的通信系统参数，从一个或多个预训练神经网络模型中确定的；所述第一设备的通信系统参数包括所述第一设备支持的系统带宽和帧结构，所述第二设备的通信系统参数包括所述第二设备支持的系统带宽和帧结构。
  15. 根据权利要求14所述的装置，其特征在于，
    所述第一神经网络模型是基于从一个或多个第二神经网络模型中选择的第三神经网络模型确定的；
    所述一个或多个第二神经网络模型是从所述一个或多个预训练神经网络模型中确定的；
    所述一个或多个第二神经网络模型中每个第二神经网络模型的输入维度与其对应的第一输入维度相同，和/或，所述每个第二神经网络模型的输出维度与其对应的第一输出维度相同；
    所述每个第二神经网络模型对应的第一输入维度和/或第一输出维度，是根据所述第二神经网络模型适用的资源片RP大小，以及所述第一设备的通信系统参数和/或所述第二设备的通信系统参数确定的。
  16. 根据权利要求15所述的方法，其特征在于，所述RP大小是根据所述RP的时域长度和所述RP的频域宽度确定的。
  17. 根据权利要求15或16所述的装置，其特征在于，所述第二神经网络模型支持第一业务类型；所述第一业务类型是所述第一设备和所述第二设备需求的业务类型。
  18. 根据权利要求15至17任一项所述的装置，其特征在于，
    所述第一神经网络模型是对所述第三神经网络模型进行蒸馏得到的；
    所述第三神经网络模型的运算量大于第一运算量，和/或，所述第三神经网络模型的参数量大于第一参数量；
    所述第一运算量是根据所述第一设备的算力和时延要求，和/或，所述第二设备的算力和时延要求确定的；所述第一参数量是根据所述第一设备的算力和存储空间，和/或，所述第二设备的算力和存储空间确定的。
  19. 根据权利要求15至17任一项所述的装置，其特征在于，
    所述第一神经网络模型为所述第三神经网络模型；
    所述第三神经网络模型的运算量小于或等于第一运算量，和/或，所述第三神经网络模型的参数量小于或等于第一参数量；
    所述第一运算量是根据所述第一设备的算力和时延要求，和/或，所述第二设备的算力和时延要求确定的；所述第一参数量是根据所述第一设备的算力和存储空间确定的，和/或，所述第二设备的算力和存储空间。
  20. 根据权利要求15至17任一项所述的装置，其特征在于，
    所述第一神经网络模型是模型服务器根据接收的模型请求信息确定的；
    所述模型请求信息包括所述第三神经网络模型的标识，以及第一运算量和/或第一参数量；
    所述第一运算量是根据所述第一设备的算力和时延要求，和/或，所述第二设备的算力和时延要求确定的，所述第一参数量是根据所述第一设备的算力和存储空间确定的，和/或，所述第二设备的算力和存储空间。
  21. 根据权利要求20所述的装置，其特征在于，
    所述一个或多个预训练神经网络模型中每个预训练神经网络模型的信息是预定义的，或是从所述模型服务器获取的；
    每个预训练神经网络模型的信息包括以下一个或多个：标识、业务类型、RP大小、输入维度、输出维度、参数量和运算量。
  22. 根据权利要求13至21任一项所述的装置，其特征在于，
    所述调整后的第一神经网络模型是根据所述第一信息中的所述信道状态信息，对第四神经网络模型进行训练得到的；
    所述第四神经网络模型是根据所述第一神经网络模型适用的RP大小，以及所述第一信息中的所述通信资源信息，对所述第一神经网络模型的输入维度和/或输出维度进行调整得到的。
  23. 一种通信装置，其特征在于，包括处理器和收发器，所述收发器用于与其它通信装置进行通信；所述处理器用于运行程序，以使得所述通信装置实现权利要求1、3至11任一项所述的方法，或者，以使得所述通信装置实现权利要求2至11任一项所述的方法。
  24. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质用于存储有指令，当其在计算机上运行时，使得权利要求1、3至11任一项所述的方法被执行；或者权利要求2至11任一项所述的方法被执行。
  25. 一种包含指令的计算机程序产品，其特征在于，当其在计算机上运行时，使得权利要求1、3至11任一项所述的方法被执行；或者权利要求2至11任一项所述的方法被执行。
  26. 一种通信系统，其特征在于，包括用于实现权利要求1、3至11任一项所述的方法的装置，以及包括用于实现权利要求2至11任一项所述的方法的装置。
  27. 一种处理器，其特征在于，用于实现权利要求1、3至11任一项所述的方法，或者，用于实现权利要求2至11任一项所述的方法的装置。
  28. 一种芯片系统，该芯片系统包括处理器和接口，所述接口用于获取程序或指令，所述处理器用于调用所述程序或指令以实现权利要求1、3至11任一项所述的方法，或者，实现权利要求2至11任一项所述的方法。
PCT/CN2023/086876 2022-04-11 2023-04-07 一种通信方法及相关装置 WO2023197950A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210375312.5 2022-04-11
CN202210375312.5A CN116963187A (zh) 2022-04-11 2022-04-11 一种通信方法及相关装置

Publications (1)

Publication Number Publication Date
WO2023197950A1 true WO2023197950A1 (zh) 2023-10-19

Family

ID=88329065

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/086876 WO2023197950A1 (zh) 2022-04-11 2023-04-07 一种通信方法及相关装置

Country Status (2)

Country Link
CN (1) CN116963187A (zh)
WO (1) WO2023197950A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210049451A1 (en) * 2019-08-14 2021-02-18 Google Llc Communicating a Neural Network Formation Configuration
US20210194548A1 (en) * 2019-12-20 2021-06-24 Qualcomm Incorporated Neural network and antenna configuration indication
CN113225197A (zh) * 2020-01-21 2021-08-06 株式会社Ntt都科摩 基于神经网络模型的通信系统及其配置方法
CN113543186A (zh) * 2020-04-13 2021-10-22 中国移动通信有限公司研究院 传输方法及设备
CN115841135A (zh) * 2021-12-24 2023-03-24 中兴通讯股份有限公司 神经网络的生成、指示信息的发送方法，通信节点及介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210049451A1 (en) * 2019-08-14 2021-02-18 Google Llc Communicating a Neural Network Formation Configuration
US20210194548A1 (en) * 2019-12-20 2021-06-24 Qualcomm Incorporated Neural network and antenna configuration indication
CN113225197A (zh) * 2020-01-21 2021-08-06 株式会社Ntt都科摩 基于神经网络模型的通信系统及其配置方法
CN113543186A (zh) * 2020-04-13 2021-10-22 中国移动通信有限公司研究院 传输方法及设备
CN115841135A (zh) * 2021-12-24 2023-03-24 中兴通讯股份有限公司 神经网络的生成、指示信息的发送方法，通信节点及介质

Also Published As

Publication number Publication date
CN116963187A (zh) 2023-10-27

Similar Documents

Publication Publication Date Title
US20230232213A1 (en) Information transmission methods and apparatuses, and communication devices and storage medium
WO2023020502A1 (zh) 数据处理方法及装置
US20240179603A1 (en) Communication method and apparatus
CN117716678A (zh) 经由具有可调定时的前传接口发送数据
WO2024041117A1 (zh) 一种计算任务的分割方法及相关装置
WO2023197950A1 (zh) 一种通信方法及相关装置
WO2023088465A1 (zh) 一种模型训练方法及相关装置
US20220369332A1 (en) Scheduling method and apparatus
JP6945089B1 (ja) ネットワークスライス構成
CN116938658A (zh) 数据处理方法、装置及系统
CN116456359A (zh) 一种通信方法、装置及系统
US20210014760A1 (en) Device steering based on profile and historical behavior
WO2023185890A1 (zh) 一种数据处理方法及相关装置
WO2024026891A1 (zh) 基于人工智能ai的信道估计方法和装置
CN108260135B (zh) 一种系统信息传输方法、用户终端和传输节点
WO2019100808A1 (zh) 捆绑大小确定方法、用户终端和网络侧设备
WO2024060002A1 (zh) 通信方法以及相关装置
EP4322065A1 (en) Gradient transmission method and related apparatus
WO2024037455A1 (zh) 资源单位的确定方法及装置
WO2022100514A1 (zh) 决策方法和决策装置
US20230259742A1 (en) Communication method, apparatus, and system
WO2023179577A1 (zh) 一种通信方法及相关装置
WO2024131889A1 (zh) 一种通信方法、装置、芯片及模组设备
WO2023202629A1 (zh) 一种时隙格式指示方法及通信装置
CN117580183B (zh) 网络优先级的协商方法和相关设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23787594

Country of ref document: EP

Kind code of ref document: A1