A communication method and system

Abstract

A communication method and system are disclosed. The method may include: receiving first data and inputting the first data into a first model for processing to obtain the output of the first model; obtaining a first intermediate gradient and normalized coefficients based on the output of the first model; and sending first information and second information, wherein the first information indicates the first intermediate gradient and the second information indicates the normalized coefficients. In this embodiment, the received first data is processed to calculate the first intermediate gradient and normalized coefficients, and then the first intermediate gradient and normalized coefficients are sent. This method can help obtain the original intermediate gradient based on the first intermediate gradient and normalized coefficients, thereby improving the efficiency of model training.

Description

Communication method and system Technical Field

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

[0002] Artificial intelligence (AI) technology has been very successfully applied in image processing and natural language processing. Its increasing maturity will significantly drive the evolution of future mobile communication network technologies. Currently, academic research is extensively focused on applying AI technology to both the network layer (e.g., network optimization, mobility management, resource allocation, etc.) and the physical layer (e.g., channel coding, channel prediction, receivers, etc.). Furthermore, with the advent of the era of large models, deep learning models such as BERT and GPT are capable of completing increasingly complex tasks while achieving superior performance. However, large models with massive parameters require significant computing resources for both training and inference, making them infeasible for typical terminals. Therefore, for next-generation wireless communications, networks will expand traditional access services to include computing power, data, and algorithms, providing diverse AI support capabilities. This will make AI training and inference more efficient and real-time, while effectively protecting user data privacy.

[0003] When AI technology is applied to a communication system consisting of a first communication device and a second communication device, the first and second communication devices are primarily composed of neural networks. Before use, the communication system needs to train the parameters of the neural networks within the first and / or second communication devices. The training process includes forward reasoning of training data, calculation of loss functions, and reverse transmission of gradients. Forward reasoning of training data refers to the process of the second communication device sending training data to the first communication device, and reverse transmission of gradients refers to the first communication device transmitting the gradients of the first communication device back to the second communication device.

[0004] However, when the aforementioned AI technology is combined with a wireless communication system, how the second communication device obtains the gradient transmitted back by the first communication device is an urgent problem to be solved.

[0005] Summary of the Invention

[0006] The embodiments of the present application provide a communication method and system that can improve the efficiency of model training.

[0007] In a first aspect, an embodiment of the present application provides a communication method, which is implemented in a terminal or a chip or circuit for a terminal, or a network device or a chip or circuit for a network device, including:

[0008] receiving first data, and inputting the first data into a first model for processing to obtain an output of the first model;

[0009] Obtaining a first intermediate gradient and a normalization coefficient based on an output of the first model;

[0010] First information and second information are sent respectively, where the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient.

[0011] In an embodiment of the present application, processing is performed based on the received first data to calculate a first intermediate gradient and a normalization coefficient. The first intermediate gradient and normalization coefficient are then transmitted. This approach facilitates obtaining the original intermediate gradient based on the first intermediate gradient and the normalization coefficient, thereby improving the efficiency of model training.

[0012] In a possible implementation, the normalization coefficient includes a first coefficient and a second coefficient, the first coefficient is indicated using a first period, and the second coefficient is indicated using a second period, wherein the first period is greater than the second period.

[0013] The hierarchical indication of the normalization coefficient can reduce the transmission overhead.

[0014] In a possible implementation, the first period and / or the second period is configured by at least one of a radio resource control layer RRC, a MAC control element MAC CE, downlink control information DCI, uplink control information UCI, and sidelink control information SCI.

[0015] In one possible implementation, the first cycle is related to hyperparameters of model training.

[0016] In a possible implementation manner, the granularity of the normalization coefficient is configured by at least one of RRC, MAC CE, and DCI.

[0017] By setting the granularity of the normalization coefficient, the accuracy of the normalization coefficient indication can be improved.

[0018] In one possible implementation, the granularity of the normalization coefficient is determined based on at least one of the following:

[0019] One or more subbands, one or more resource blocks RB, one or more resource block group bundles RBG, one or more control channel elements CCE, one or more aggregation levels, and processing granularity in a neural network structure.

[0020] In a possible implementation, the normalization coefficient is one or more scalars, and each of the one or more scalars corresponds to one or more granularities.

[0021] In another possible implementation manner, the normalization coefficient is a codeword in a preset codebook.

[0022] In another possible implementation, the normalization coefficient is a eigenvalue after compression of the neural network.

[0023] In one possible implementation, the normalization coefficient is related to at least one of the following parameters:

[0024] The training time of the first model, the number of epochs or iterations of the training process of the first model, the loss function of the first model, and the learning rate of the first model.

[0025] By using the adaptive adjustment method of the normalization coefficient, the accuracy of the normalization coefficient indication can be improved.

[0026] Optionally, the normalization coefficients decrease with increasing training time and epoch number, and decrease with decreasing loss function and learning rate. Of course, some fluctuation is also allowed.

[0027] In a possible implementation, the method further includes:

[0028] Send third information and fourth information, where the third information includes a first bit set, the first bit set indicating the quantization normalization coefficient of the first model during the i-th training, and the fourth information includes a second bit set, the second bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, where i and k are both positive integers.

[0029] It can be understood that the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, that is, the quantization bit width of the normalization coefficient during the i-th training is greater than the quantization bit width of the normalization coefficient during the i+k-th training.

[0030] The setting of the bit number change based on the normalization coefficient can reduce signaling overhead and more effectively utilize control physical resources.

[0031] In a possible implementation, the change in the number of bits is implemented based on different formats of the scheduled UCI.

[0032] In another possible implementation, the change in the number of bits is based on RRC reconfiguration.

[0033] In a possible implementation, the method further includes:

[0034] Send fifth and sixth information, where the fifth information includes a third bit set, and the third bit set indicates the quantization normalization coefficient of the first model during the i-th training; the sixth information includes a fourth bit set and a fifth bit set, where the fourth bit set indicates the preset content, and the fifth bit set indicates the quantization normalization coefficient of the first model during the i+k-th training, where i and k are both positive integers.

[0035] The setting of partial bit multiplexing based on the normalization coefficient can reduce signaling overhead and more effectively utilize control physical resources.

[0036] In a possible implementation, the method further includes: training the first model to obtain a second model.

[0037] In a second aspect, an embodiment of the present application provides a communication method, which is applied to a network device or a chip or circuit implemented for a network device, or a terminal or a chip or circuit implemented for a terminal, including:

[0038] sending first data, where the first data is obtained by inputting the second data into a third model for processing;

[0039] receiving first information and second information, wherein the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient;

[0040] A fourth model is obtained based on the first intermediate gradient, the normalization coefficient, and the third model.

[0041] In the embodiment of the present application, model training is performed based on the received first intermediate gradient and normalization coefficient. By adopting this method, the original intermediate gradient can be obtained based on the first intermediate gradient and the normalization coefficient, thereby improving the efficiency of model training.

[0042] In a possible implementation, obtaining the fourth model based on the first intermediate gradient, the normalization coefficient, and the third model includes:

[0043] obtaining a second intermediate gradient based on the first intermediate gradient and the normalization coefficient;

[0044] The third model is processed based on the second intermediate gradient to obtain a fourth model.

[0045] In a possible implementation, the normalization coefficient includes a first coefficient and a second coefficient, the first coefficient is indicated using a first period, and the second coefficient is indicated using a second period, wherein the first period is greater than the second period.

[0046] The hierarchical indication of the normalization coefficient can reduce the transmission overhead.

[0047] In a possible implementation manner, the first period and / or the second period is configured by at least one of RRC, MAC CE, DCI, UCI, and SCI.

[0048] In one possible implementation, the first cycle is related to hyperparameters of model training.

[0049] In a possible implementation manner, the granularity of the normalization coefficient is configured by at least one of RRC, MAC CE, and DCI.

[0050] By setting the granularity of the normalization coefficient, the accuracy of the normalization coefficient indication can be improved.

[0051] In one possible implementation, the granularity of the normalization coefficient is determined based on at least one of the following:

[0052] One or more subbands, one or more resource blocks RB, one or more resource block group bundles RBG, one or more control channel elements CCE, one or more aggregation levels, and processing granularity in a neural network structure.

[0053] In a possible implementation, the normalization coefficient is one or more scalars, and each of the one or more scalars corresponds to one or more granularities.

[0054] In another possible implementation manner, the normalization coefficient is a codeword in a preset codebook.

[0055] In another possible implementation, the normalization coefficient is a eigenvalue after compression of the neural network.

[0056] In one possible implementation, the normalization coefficient is related to at least one of the following parameters:

[0057] The training time of the first model, the number of epochs or iterations of the training process of the first model, the loss function of the first model, and the learning rate of the first model.

[0058] By using the adaptive adjustment method of the normalization coefficient, the accuracy of the normalization coefficient indication can be improved.

[0059] Optionally, the normalization coefficients decrease with increasing training time and epoch number, and decrease with decreasing loss function and learning rate. Of course, some fluctuation is also allowed.

[0060] In a possible implementation, the method further includes:

[0061] Receive third information and fourth information, where the third information includes a first bit set, the first bit set indicating the quantization normalization coefficient of the first model during the i-th training, and the fourth information includes a second bit set, the second bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, where i and k are both positive integers.

[0062] It can be understood that the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, that is, the quantization bit width of the normalization coefficient during the i-th training is greater than the quantization bit width of the normalization coefficient during the i+k-th training.

[0063] The setting of the bit number change based on the normalization coefficient can reduce signaling overhead and more effectively utilize control physical resources.

[0064] In a possible implementation, the change in the number of bits is implemented based on different formats of the scheduled UCI.

[0065] In another possible implementation, the change in the number of bits is based on RRC reconfiguration.

[0066] In a possible implementation, the method further includes:

[0067] Receive fifth and sixth information, the fifth information including a third bit set, the third bit set indicating the quantization normalization coefficient of the first model during the i-th training, the sixth information including a fourth bit set and a fifth bit set, the fourth bit set indicating preset content, the fifth bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, wherein i and k are both positive integers.

[0068] The setting of partial bit multiplexing based on the normalization coefficient can reduce signaling overhead and more effectively utilize control physical resources.

[0069] In a possible implementation, the method further includes:

[0070] Decoding the second information to obtain a decoding result;

[0071] If the decoded result represents a decoding error, the normalization coefficient is recorded as a first value or a second value, the first value is used to indicate that the normalization coefficient is not considered, and the second value is used to indicate that the current model training is terminated.

[0072] Based on the above normalized indicator coefficient error tolerance setting, the robustness of training can be improved.

[0073] In another possible implementation, the method further includes:

[0074] Decoding the second information to obtain a decoding result;

[0075] If the decoded result indicates a decoding error, the normalization coefficient is recorded as a third value, where the third value is a filtered value of at least two historical normalization coefficients.

[0076] Based on the above normalized indicator coefficient error tolerance setting, the robustness of training can be improved.

[0077] In a third aspect, an embodiment of the present application provides a communication device, which may include a transceiver module and a processing module, wherein:

[0078] a transceiver module, configured to receive first data, and input the first data into a first model for processing to obtain an output of the first model;

[0079] a processing module, configured to obtain a first intermediate gradient and a normalization coefficient based on an output of the first model;

[0080] The transceiver module is further configured to send first information and second information respectively, where the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient.

[0081] In a possible implementation, the normalization coefficient includes a first coefficient and a second coefficient, the first coefficient is indicated using a first period, and the second coefficient is indicated using a second period, wherein the first period is greater than the second period.

[0082] In a possible implementation manner, the first period and / or the second period is configured by at least one of RRC, MAC CE, DCI, UCI, and SCI.

[0083] In one possible implementation, the first cycle is related to hyperparameters of model training.

[0084] In a possible implementation manner, the granularity of the normalization coefficient is configured by at least one of RRC, MAC CE, and DCI.

[0085] In one possible implementation, the granularity of the normalization coefficient is determined based on at least one of the following:

[0086] One or more subbands, one or more resource blocks RB, one or more resource block group bundles RBG, one or more control channel elements CCE, one or more aggregation levels, and processing granularity in a neural network structure.

[0087] In a possible implementation, the normalization coefficient is one or more scalars, and each of the one or more scalars corresponds to one or more granularities.

[0088] In another possible implementation manner, the normalization coefficient is a codeword in a preset codebook.

[0089] In another possible implementation, the normalization coefficient is a eigenvalue after compression of the neural network.

[0090] In one possible implementation, the normalization coefficient is related to at least one of the following parameters:

[0091] The training time of the first model, the number of epochs or iterations of the training process of the first model, the loss function of the first model, and the learning rate of the first model.

[0092] In a possible implementation, the transceiver module is further configured to:

[0093] Send third information and fourth information, where the third information includes a first bit set, the first bit set indicating the quantization normalization coefficient of the first model during the i-th training, and the fourth information includes a second bit set, the second bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, where i and k are both positive integers.

[0094] In a possible implementation, the change in the number of bits is implemented based on different formats of the scheduled UCI.

[0095] In another possible implementation, the change in the number of bits is based on RRC reconfiguration.

[0096] In a possible implementation, the transceiver module is further configured to:

[0097] Send fifth and sixth information, where the fifth information includes a third bit set, and the third bit set indicates the quantization normalization coefficient of the first model during the i-th training; the sixth information includes a fourth bit set and a fifth bit set, where the fourth bit set indicates the preset content, and the fifth bit set indicates the quantization normalization coefficient of the first model during the i+k-th training, where i and k are both positive integers.

[0098] In a possible implementation, the processing module is further used to train the first model to obtain a second model.

[0099] In a fourth aspect, an embodiment of the present application provides a communication device, which may include a transceiver module and a processing module, wherein:

[0100] a transceiver module, configured to send first data, where the first data is obtained by inputting the second data into a third model for processing;

[0101] The transceiver module is further configured to receive first information and second information, wherein the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient;

[0102] A processing module is configured to obtain a fourth model based on the first intermediate gradient, the normalization coefficient, and the third model.

[0103] In a possible implementation, the processing module is configured to:

[0104] obtaining a second intermediate gradient based on the first intermediate gradient and the normalization coefficient;

[0105] The third model is processed based on the second intermediate gradient to obtain a fourth model.

[0106] In a possible implementation, the normalization coefficient includes a first coefficient and a second coefficient, the first coefficient is indicated using a first period, and the second coefficient is indicated using a second period, wherein the first period is greater than the second period.

[0107] In a possible implementation manner, the first period and / or the second period is configured by at least one of RRC, MAC CE, DCI, UCI, and SCI.

[0108] In one possible implementation, the first cycle is related to hyperparameters of model training.

[0109] In a possible implementation manner, the granularity of the normalization coefficient is configured by at least one of RRC, MAC CE, and DCI.

[0110] In one possible implementation, the granularity of the normalization coefficient is determined based on at least one of the following:

[0111] One or more subbands, one or more resource blocks RB, one or more resource block group bundles RBG, one or more control channel elements CCE, one or more aggregation levels, and processing granularity in a neural network structure.

[0112] In a possible implementation, the normalization coefficient is one or more scalars, and each of the one or more scalars corresponds to one or more granularities.

[0113] In another possible implementation manner, the normalization coefficient is a codeword in a preset codebook.

[0114] In another possible implementation, the normalization coefficient is a eigenvalue after compression of the neural network.

[0115] In one possible implementation, the normalization coefficient is related to at least one of the following parameters:

[0116] The training time of the first model, the number of epochs or iterations of the training process of the first model, the loss function of the first model, and the learning rate of the first model.

[0117] In a possible implementation, the transceiver module is further configured to:

[0118] Receive third information and fourth information, where the third information includes a first bit set, the first bit set indicating the quantization normalization coefficient of the first model during the i-th training, and the fourth information includes a second bit set, the second bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, where i and k are both positive integers.

[0119] In a possible implementation, the change in the number of bits is implemented based on different formats of the scheduled UCI.

[0120] In another possible implementation, the change in the number of bits is based on RRC reconfiguration.

[0121] In a possible implementation, the transceiver module is further configured to:

[0122] Receive fifth and sixth information, the fifth information including a third bit set, the third bit set indicating the quantization normalization coefficient of the first model during the i-th training, the sixth information including a fourth bit set and a fifth bit set, the fourth bit set indicating preset content, the fifth bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, wherein i and k are both positive integers.

[0123] In a possible implementation, the processing module is further configured to:

[0124] Decoding the second information to obtain a decoding result;

[0125] If the decoded result represents a decoding error, the normalization coefficient is recorded as a first value or a second value, the first value is used to indicate that the normalization coefficient is not considered, and the second value is used to indicate that the current model training is terminated.

[0126] In another possible implementation, the processing module is further configured to:

[0127] Decoding the second information to obtain a decoding result;

[0128] If the decoded result indicates a decoding error, the normalization coefficient is recorded as a third value, where the third value is a filtered value of at least two historical normalization coefficients.

[0129] With respect to the third or fourth aspect above, in one possible implementation, the processing module may be a processor, and the transceiver module may be a communication module, a transceiver, or a communication interface. It is understood that the transceiver module may be a transceiver in the device, for example, implemented by an antenna, a feeder, and a codec in the device. Alternatively, if the communication device is a chip provided in the device, the transceiver module may be an input / output interface of the chip, such as an input / output circuit, a pin, etc.

[0130] In a fifth aspect, an embodiment of the present application provides a communication device, which includes one or more processors; wherein the one or more processors are used to execute computer programs stored in one or more memories, so that the communication device implements the method as described in any one of the first aspects, or implements the method as described in any one of the second aspects.

[0131] In a possible implementation manner, the communication device further includes the one or more memories.

[0132] In a possible implementation, the communication device is a chip or a chip system.

[0133] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores instructions. When the instructions are executed by a processor, the method described in any one of the first aspects is implemented, or the method described in any one of the second aspects is implemented.

[0134] In a seventh aspect, an embodiment of the present application provides a computer program product, comprising a computer program, which, when executed, implements the method as described in any one of the first aspects, or implements the method as described in any one of the second aspects.

[0135] In an eighth aspect, an embodiment of the present application provides a communication system, comprising an apparatus as described in any one of the third aspects and an apparatus as described in any one of the fourth aspects.

[0136] In a ninth aspect, the present application provides a computer program product. When the computer program product is run on a computer, the computer is used to implement the method described in any one of the first aspects and / or the method described in any one of the second aspects.

[0137] The beneficial effects of some implementation methods of the technical solutions provided in the second to ninth aspects of this application can refer to the beneficial effects of the technical solution of the first aspect, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0138] The following is a brief introduction to the drawings used in describing the embodiments.

[0139] FIG1 is a schematic diagram of the architecture of a communication system provided in an embodiment of the present application;

[0140] FIG2 is a flow chart of a communication method provided in an embodiment of the present application;

[0141] FIG3 is a communication diagram provided in an embodiment of the present application;

[0142] FIG4 is a schematic diagram of a hierarchical indication provided in an embodiment of the present application;

[0143] FIG5 is a schematic diagram of a multiplexing normalization coefficient indication provided by an embodiment of the present application;

[0144] FIG6 is a schematic diagram of a normalization coefficient fault tolerance method provided in an embodiment of the present application;

[0145] FIG7 is a schematic structural diagram of a communication device provided in an embodiment of the present application;

[0146] FIG8 is a schematic structural diagram of another communication device provided in an embodiment of the present application;

[0147] FIG9 is a schematic structural diagram of another communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0148] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0149] It should be noted that in this application, "send" can be understood as "output" and "receive" can be understood as "input". "Send information to A", where "to A" only indicates the direction of information transmission, A is the destination, and does not limit "sending information to A" to direct transmission on the air interface. "Sending information to A" includes sending information directly to A, and also includes sending information indirectly to A through a transmitter, so "sending information to A" can also be understood as "outputting information to A". Similarly, "receiving information from A" indicates that the source of the information is A, including receiving information directly from A, and also including receiving information indirectly from A through a receiver, so "receiving information from A" can also be understood as "inputting information from A".

[0150] It is understood that, in this application, "indication" can include direct indication, indirect indication, explicit indication, and implicit indication. When describing a certain indication information as indicating A, it can be understood that the indication information carries A, directly indicates A, or indirectly indicates A. In this application, the information indicated by the indication information is referred to as the information to be indicated. In the specific implementation process, there are many ways to indicate the information to be indicated. For example, but not limited to, the information to be indicated can be directly indicated, such as the information to be indicated itself or an index of the information to be indicated, or it can be indirectly indicated by indicating other information, where there is an association between the other information and the information to be indicated. It is also possible to indicate only a part of the information to be indicated, while the rest of the information to be indicated is known or agreed in advance. For example, it is also possible to indicate specific information by using a pre-agreed (e.g., protocol-specified) order of arrangement of various information, thereby reducing the indication overhead to a certain extent. The information to be indicated can be sent as a whole or divided into multiple sub-information and sent separately, and the sending period and / or sending time of these sub-information can be the same or different. The specific sending method is not limited in this application. The sending period and / or sending timing of these sub-information may be predefined, for example, predefined according to a protocol, or may be configured by the transmitting end device by sending configuration information to the receiving end device.

[0151] The system architecture of the embodiment of the present application will be described in detail below with reference to the accompanying drawings.

[0152] Among them, the terminal 101 can also be called a terminal device, user equipment (UE), mobile station, mobile terminal, etc. The terminal can be widely used in various scenarios, such as device-to-device (D2D), vehicle-to-everything (V2X) communication, machine-type communication (MTC), Internet of Things (IOT), virtual reality, augmented reality, industrial control, autonomous driving, telemedicine, smart grid, smart furniture, smart office, smart wearable, smart transportation, smart city, etc. The terminal can be a mobile phone, tablet computer, computer with wireless transceiver function, wearable device, vehicle, drone, helicopter, airplane, ship, robot, robotic arm, smart home device, etc. The embodiments of this application do not limit the specific technology and specific device form adopted by the terminal.

[0153] The network device 102 can be a base station (BS), an evolved NodeB (eNodeB), a transmission reception point (TRP), a next-generation NodeB (gNB) in a fifth-generation (5G) mobile communication system, a next-generation base station in a sixth-generation (6G) mobile communication system, a base station in a future mobile communication system, or an access node in a WiFi system, etc.; it can also be a module or unit that performs some of the functions of a base station, for example, a centralized unit (CU) or a distributed unit (DU). The network device 102 can be a macro base station, a micro base station or an indoor station, a relay node or a donor node, etc. The embodiments of the present application do not limit the specific technology and specific device form adopted by the network device 102. For ease of description, the following description uses a base station as an example of a network device.

[0154] Base stations and terminals can be fixed or mobile. They can be deployed on land, indoors or outdoors, handheld or vehicle-mounted; on water; or in the air on aircraft, balloons, and satellites. The embodiments of this application do not limit the application scenarios of base stations and terminals.

[0155] It should be noted that the embodiments of the present application do not limit the number of the above-mentioned network devices and the number of terminals. Among them, base stations and terminals, base stations and base stations, and terminals and terminals can communicate through authorized spectrum, unlicensed spectrum, and authorized spectrum and unlicensed spectrum at the same time; they can communicate through spectrum below 6 gigahertz (GHz), spectrum above 6 GHz, and spectrum below 6 GHz and spectrum above 6 GHz at the same time. The embodiments of the present application do not limit the spectrum resources used for wireless communications.

[0156] In the embodiments of the present application, the functions of the base station may also be performed by a module (such as a chip) in the base station, or by a control subsystem that includes the base station functions. The control subsystem that includes the base station functions here may be a control center in the above-mentioned application scenarios such as smart grid, industrial control, smart transportation, and smart city. The functions of the terminal may also be performed by a module (such as a chip or modem) in the terminal, or by a device that includes the terminal functions.

[0157] The above describes the architecture of the embodiment of the present application. The following describes the method of the embodiment of the present application in detail.

[0158] Referring to Figure 2, it is a flow chart of a communication method provided by an embodiment of the present application. Optionally, the method can be applied to the aforementioned communication system, such as the communication system shown in Figure 1. The communication method shown in Figure 2 may include steps 201-206. It should be understood that this application is described in the order of 201-206 for the convenience of description, and is not intended to limit execution to the above order. The embodiment of the present application does not limit the order of execution, execution time, number of executions, etc. of the above one or more steps. The following takes the execution subject of steps 201, 205 and 206 of the communication method as an example, and the execution subject of steps 202-204 as UE for description. This application is also applicable to other execution subjects. Steps 201-206 are as follows:

[0159] 201. A base station sends first data.

[0160] The first data is obtained by the base station inputting the second data into the third model for processing.

[0161] For example, a third model (Tx-side sub-model 1 as shown in FIG3 ) is deployed on the base station side. The second data is input into the third model for processing to obtain the first data. The base station then maps the first data to a downlink (DL) data channel and transmits it. Of course, other transmission methods may also be used, and this solution does not limit this.

[0162] 202. The UE receives first data and inputs the first data into a first model for processing to obtain an output of the first model.

[0163] The UE receives the first data. A first model (such as the Rx-side sub-model 2 shown in FIG3 ) is deployed on the UE side. The first data is input into the first model for processing to obtain an output of the first model. Exemplarily, the output of the first model may be third data, etc.

[0164] 203. The UE obtains a first intermediate gradient and a normalization coefficient based on the output of the first model.

[0165] The specific method of neural network training is to evaluate the output of the neural network using a loss function, propagate the error backward, and iteratively optimize the model parameters through gradient descent until the loss function reaches a minimum. The backward propagation process can be used to calculate the gradient of the parameters of the previous layer by the chain rule of partial derivatives, that is, the gradient of the parameters of the next layer can be recursively calculated.

[0166] The intermediate gradient is one or more terms in the gradient expression of the neural network parameters, or the product of multiple terms. When the intermediate gradient is multiple terms in the gradient expression of the neural network parameters, the multiple terms need to be separately transmitted back to, for example, the base station.

[0167] For example, the gradient of a model is related to the loss function, the model parameters, the model input, and the model output. Based on the output of the first model, combined with its input and its loss function, an intermediate gradient can be obtained.

[0168] The normalization coefficient can be understood as a coefficient for performing power normalization processing on the signal.

[0169] It is understandable that in the embodiment of the present application, the first intermediate gradient is a power-normalized first intermediate gradient. That is, the UE performs power normalization on the first intermediate gradient and then maps the power-normalized first intermediate gradient to air interface resources and sends it to the base station.

[0170] In an optional implementation manner, the UE performing power normalization on the first intermediate gradient includes: performing normalization processing on average powers of all gradient symbols included in the first intermediate gradient.

[0171] In another optional implementation manner, the UE performing power normalization on the first intermediate gradient includes: performing normalization processing on the maximum power of all gradient symbols included in the first intermediate gradient.

[0172] The above implementations are merely examples and are not limited in this solution.

[0173] 204. The UE sends first information and second information respectively, where the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient.

[0174] Exemplarily, the UE sends the first intermediate gradient via an uplink (UL) data channel, and sends the normalization coefficient via an uplink UL control channel. Of course, other transmission methods may also be used, and this solution does not limit the transmission method.

[0175] The UE sends the first intermediate gradient and the normalization coefficient to the base station so that the base station can restore the intermediate gradient and then perform model training.

[0176] In a possible implementation, the UE trains the first model and obtains a trained second model through multiple iterative trainings.

[0177] To reduce transmission overhead and improve the accuracy of feedback coefficients, the normalized coefficients can be indicated in a hierarchical manner. For example, the hierarchical indication may employ at least two levels of indication. The specific indication signaling may indicate the quantized value of the coefficient or an index, such as by using a table lookup to obtain the specific quantized value.

[0178] In a possible implementation, the normalization coefficient includes a first coefficient and a second coefficient, wherein the first coefficient is indicated using a first period, and the second coefficient is indicated using a second period, wherein the first period is greater than the second period.

[0179] That is to say, the normalization coefficient in this example is composed of two parts. As shown in Figure 4, the normalization coefficient includes normalization coefficient 1 and normalization coefficient 2. Among them, normalization coefficient 1 is a long-cycle indication. It can be indicated after the preset number of trainings are completed (or before the start, etc.). Its indication period can be pre-configured by at least one of the radio resource control layer (Radio Resource Control, RRC), MAC control element (MAC Control Element, MAC CE), downlink control information (Downlink Control Information, DCI), uplink control information (Uplink Control Information, UCI), and sidelink control information (Sidelink Control Information, SCI), and is related to training hyperparameters such as batch size batchsize and the number of Epochs in the training process. Of course, it can also be updated by reporting on the UE side.

[0180] The normalization factor 2 indicates a short period. For example, as shown in Figure 3, each training session is indicated separately. The indication period can be pre-configured by at least one of the following methods: the radio resource control layer (RRC), the MAC control element (MAC CE), or the downlink control information (DCI), and this solution does not impose any restrictions on this.

[0181] The current normalization coefficient value can be calculated by combining normalization coefficient 1 and normalization coefficient 2. The calculation method can be a simple mathematical calculation or other preset algorithms, and this solution does not impose any restrictions on this.

[0182] In one possible implementation, normalization coefficient 1 may be the high-order bit of the quantized normalization coefficient, and normalization coefficient 2 may be the low-order bit. Alternatively, normalization coefficient 1 may be the filtered mean value of the normalization coefficient, and normalization coefficient 2 may be the difference value.

[0183] The hierarchical indication of the normalization coefficient can reduce the transmission overhead.

[0184] In a possible implementation manner, the granularity of the normalization coefficient is configured by at least one of RRC, MAC CE, and DCI.

[0185] The granularity of the normalization coefficient may be pre-configured by, for example, RRC, MAC CE, DCI, etc.

[0186] Exemplarily, all sent gradients obtain the same normalized value, that is, the normalization coefficient is a scalar. Alternatively, the normalization coefficient can also be multiple scalars. Each of the one or more scalars can correspond to one or more granularities.

[0187] Of course, the normalization coefficient can also be fed back at a finer granularity, such as by feeding back the normalization coefficient based on one or more subbands, one or more resource blocks (RBs), one or more resource block groups (RBGs), one or more control channel elements (CCEs), one or more aggregation levels, the processing granularity in the neural network structure, and other information. That is, the normalization coefficient is a vector. This vector can directly correspond to the minimum granularity, or it can be a codeword in a preset codebook, or it can be a eigenvalue after neural network compression. This solution does not impose any restrictions on this.

[0188] By setting the granularity of the normalization coefficient, the accuracy of the normalization coefficient indication can be improved.

[0189] In one possible implementation, the normalization coefficient is related to at least one of the following parameters:

[0190] The training time of the first model, the number of epochs or iterations of the training process of the first model, the loss function of the first model, and the learning rate of the first model.

[0191] As the training progresses, the gradient of the model update gradually becomes smaller. In order to reduce the instruction overhead, the normalization coefficient can be adaptively adjusted according to the training status.

[0192] For example, the normalization coefficient is adaptively adjusted according to the training time. t It can be adjusted as follows: scale t =f(t)·scale0;

[0193] Where t is the model training time, f(t) is the function of the training time t, and scale0 is a preset value or a periodically reported normalization coefficient. Optionally, in this example, the base station side and the UE side perform training synchronously.

[0194] For example, the normalization coefficient is adaptively adjusted according to the number of training epochs or iterations. t It can be adjusted as follows: scale t =f(n)·scale0;

[0195] Wherein, n is the number of training epochs or iterations, and f(n) is a function of the number of training epochs or iterations n. Optionally, in this example, the base station side and the UE side perform training synchronously.

[0196] For example, the normalization coefficient is adaptively adjusted according to the training loss function. t It can be adjusted as follows: scale t =f(l)·scale0;

[0197] Where l is the loss value, and f(l) is the function of the loss value l. In this example, the normalization coefficient scale can be obtained based on the loss value reported by the UE side for this training. t .

[0198] For example, the normalization coefficient is adaptively adjusted according to the learning rate. t It can be adjusted as follows: scale t =f(lr)·scale0;

[0199] Where lr is the learning rate, and f(lr) is the function of the learning rate lr. In this example, the learning rate can be changed in a pre-configured manner or dynamically indicated by the base station.

[0200] In a possible implementation, the normalization coefficient may be adaptively adjusted according to the at least two variables mentioned above, but this solution does not impose any limitation on this.

[0201] Optionally, functions f(t) and f(n) are decreasing functions, while functions f(l) and f(lr) are increasing functions. That is, the normalization coefficients decrease with increasing training time and epoch number, and decrease with decreasing loss function and learning rate. Of course, some fluctuation is also permitted.

[0202] This example only uses the above parameters as an example. It can also be adjusted according to other parameters, and this solution does not limit this.

[0203] By using the adaptive adjustment method of the normalization coefficient, the accuracy of the normalization coefficient indication can be improved.

[0204] As the training progresses, the gradient of the model update gradually becomes smaller. In order to more effectively utilize the control physical resources, this solution also provides a method for expressing the normalization coefficient.

[0205] In a possible implementation, the method further includes:

[0206] The UE sends third information and fourth information, where the third information includes a first bit set, where the first bit set indicates the quantization normalization coefficient of the first model during the i-th training, and the fourth information includes a second bit set, where the second bit set indicates the quantization normalization coefficient of the first model during the i+k-th training. The number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, where i and k are both positive integers.

[0207] Correspondingly, the base station receives the third information and the fourth information.

[0208] It can be understood that the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, that is, the quantization bit width of the normalization coefficient during the i-th training is greater than the quantization bit width of the normalization coefficient during the i+k-th training.

[0209] That is, by specifying several preset indication lengths, for example, the normalization coefficient can be represented by several different lengths of bits. For example, the normalization coefficient is initially represented by 8 bits, and then later by 4 bits as training progresses.

[0210] The change in the number of bits may be achieved by directly scheduling different formats of uplink control information (UCI), or by using RRC reconfiguration, etc. This solution does not impose any restrictions on this.

[0211] The setting of the bit number change based on the normalization coefficient can reduce signaling overhead and more effectively utilize control physical resources.

[0212] In another possible implementation, the method further includes:

[0213] Send fifth and sixth information, where the fifth information includes a third bit set, and the third bit set indicates the quantization normalization coefficient of the first model during the i-th training; the sixth information includes a fourth bit set and a fifth bit set, where the fourth bit set indicates the preset content, and the fifth bit set indicates the quantization normalization coefficient of the first model during the i+k-th training, where i and k are both positive integers.

[0214] Correspondingly, the base station receives the fifth information and the sixth information.

[0215] For example, during early training, the quantization bit width of the normalization coefficient during the i-th training may be 16 bits, indicating the normalization coefficient. During late training, the quantization bit width corresponding to the fourth bit set is 6 bits, which can indicate a model test instruction, a model save instruction, etc. The quantization bit width of the normalization coefficient during the i+k-th training is also 6 bits, indicating the normalization coefficient.

[0216] In other words, some bits in the normalization coefficient are reused for representation. When model training converges, there is a need to test and save the model. This can keep the UCI length unchanged but reinterpret the received fields.

[0217] As shown in Figure 5, the normalization coefficient is represented by 6 bits: "1," "1," "0," "1," "0," and "1." When training converges, the base station reinterprets the received 6 bits. The first bit, "1," indicates a model test, the second bit indicates a model save, and the last four bits (shown in the small dashed box in Figure 5) still represent the normalization coefficient.

[0218] The setting of partial bit multiplexing based on the normalization coefficient can reduce signaling overhead and more effectively utilize control physical resources.

[0219] 205. The base station receives the first information and the second information.

[0220] 206. The base station obtains a fourth model based on the first intermediate gradient, the normalization coefficient, and the third model.

[0221] Exemplarily, the base station can recover a second intermediate gradient based on the first intermediate gradient and the normalization coefficient. This second intermediate gradient can be understood as the original intermediate gradient. That is, the gradient obtained by recovering the received intermediate gradient based on the normalization coefficient during uplink reception. The base station then processes the third model based on the second intermediate gradient to obtain a fourth model.

[0222] Specifically, the model gradient is updated based on the second intermediate gradient, and then the model is updated, and a trained fourth model can be obtained through iterative training.

[0223] Alternatively, in a possible implementation manner, the base station decodes the second information to obtain a decoded result.

[0224] If the decoded result represents a decoding error, the base station records the normalization coefficient as a first value or a second value, where the first value is used to indicate that the normalization coefficient is not considered, and the second value is used to indicate that the current model training is terminated.

[0225] For example, after receiving the normalization coefficient, if the normalization coefficient is decoded incorrectly, the base station (BS) may set the normalization coefficient to 1, ignoring the impact of the normalization coefficient and updating the transmitter and receiver normally. Alternatively, the base station (BS) may set the normalization coefficient to 0, as shown in Figure 6 , causing the base station to abandon transmitter training and allowing the receiver to update normally. Of course, other values ​​can also be set, and this solution does not limit this.

[0226] In another possible implementation, if the decoded result indicates a decoding error, the base station records the normalization coefficient as a third value, where the third value is a filtered value of at least two historical normalization coefficients.

[0227] Exemplarily, the third value is a filtered mean value of previous trainings, and the base station uses the filtered mean value to train the transmitter, while the receiver side is updated normally.

[0228] Based on the above normalized indicator coefficient error tolerance setting, the robustness of training can be improved.

[0229] In an embodiment of the present application, a UE processes first data from a base station to calculate a first intermediate gradient and a normalization coefficient. The UE sends the first intermediate gradient and normalization coefficient to the base station. The base station performs model training based on the received first intermediate gradient and normalization coefficient. Using this approach, the base station can obtain the original intermediate gradient based on the first intermediate gradient and normalization coefficient, thereby improving the efficiency of model training.

[0230] It should be noted that this example introduces the interaction between the base station and the UE as an example. It can also be the above-mentioned interaction between network devices and terminals such as between base stations, between UEs, or between UEs and base stations. This solution does not limit this.

[0231] It should be noted that in the various embodiments of the present application, unless otherwise specified or there is a logical conflict, the terms and / or descriptions between the various embodiments are consistent and can be referenced by each other. The technical features in different embodiments can be combined to form new embodiments according to their internal logical relationships.

[0232] The above describes in detail the method of the embodiment of the present application, and the following provides the device of the embodiment of the present application. It will be understood that in the various device embodiments of the present application, the division of multiple units or modules is only a logical division based on function, and is not intended to limit the specific structure of the device. In a specific implementation, some functional modules may be subdivided into more small functional modules, and some functional modules may be combined into one functional module, but no matter whether these functional modules are subdivided or combined, the general process performed by the device is the same. For example, some devices include a receiving unit and a sending unit. In some designs, the sending unit and the receiving unit can also be integrated into a communication unit, which can implement the functions implemented by the receiving unit and the sending unit. Typically, each unit corresponds to its own program code (or program instructions), and when the program code corresponding to each of these units runs on the processor, the unit is controlled by the processing unit to execute the corresponding process to implement the corresponding function.

[0233] The embodiments of the present application also provide an apparatus for implementing any of the above methods. For example, a communication apparatus is provided that includes a unit (or means) for implementing each step performed by the UE in any of the above methods. For another example, another communication apparatus is provided that includes a unit (or means) for implementing each step performed by the base station in any of the above methods.

[0234] For example, referring to FIG7 , which is a schematic diagram of the structure of a communication device provided by an embodiment of the present application, the communication recommendation device is used to implement the aforementioned communication method, such as the communication method shown in FIG2 .

[0235] As shown in FIG7 , the apparatus may include a transceiver module 701 and a processing module 702 , wherein:

[0236] The transceiver module 701 is configured to receive first data and input the first data into a first model for processing to obtain an output of the first model;

[0237] A processing module 702 is configured to obtain a first intermediate gradient and a normalization coefficient based on an output of the first model;

[0238] The transceiver module 701 is further configured to send first information and second information respectively, where the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient.

[0239] In a possible implementation, the normalization coefficient includes a first coefficient and a second coefficient, the first coefficient is indicated using a first period, and the second coefficient is indicated using a second period, wherein the first period is greater than the second period.

[0240] In a possible implementation manner, the first period and / or the second period is configured by at least one of RRC, MAC CE, DCI, UCI, and SCI.

[0241] In one possible implementation, the first cycle is related to hyperparameters of model training.

[0242] In a possible implementation manner, the granularity of the normalization coefficient is configured by at least one of RRC, MAC CE, and DCI.

[0243] In one possible implementation, the granularity of the normalization coefficient is determined based on at least one of the following:

[0244] One or more subbands, one or more resource blocks RB, one or more resource block group bundles RBG, one or more control channel elements CCE, one or more aggregation levels, and processing granularity in a neural network structure.

[0245] In a possible implementation, the normalization coefficient is one or more scalars, and each of the one or more scalars corresponds to one or more granularities.

[0246] In another possible implementation manner, the normalization coefficient is a codeword in a preset codebook.

[0247] In another possible implementation, the normalization coefficient is a eigenvalue after compression of the neural network.

[0248] In one possible implementation, the normalization coefficient is related to at least one of the following parameters:

[0249] The training time of the first model, the number of epochs or iterations of the training process of the first model, the loss function of the first model, and the learning rate of the first model.

[0250] In a possible implementation, the transceiver module 701 is further configured to:

[0251] Send third information and fourth information, where the third information includes a first bit set, the first bit set indicating the quantization normalization coefficient of the first model during the i-th training, and the fourth information includes a second bit set, the second bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, where i and k are both positive integers.

[0252] In a possible implementation, the change in the number of bits is implemented based on different formats of the scheduled UCI.

[0253] In another possible implementation, the change in the number of bits is based on RRC reconfiguration.

[0254] In a possible implementation, the transceiver module 701 is further configured to:

[0255] Send fifth and sixth information, where the fifth information includes a third bit set, and the third bit set indicates the quantization normalization coefficient of the first model during the i-th training; the sixth information includes a fourth bit set and a fifth bit set, where the fourth bit set indicates the preset content, and the fifth bit set indicates the quantization normalization coefficient of the first model during the i+k-th training, where i and k are both positive integers.

[0256] In a possible implementation, the processing module 702 is further configured to train the first model to obtain a second model.

[0257] For a detailed introduction to each of the above modules, please refer to the description of the aforementioned embodiments, which will not be repeated here.

[0258] 8 , which is a schematic diagram of the structure of another communication device provided by an embodiment of the present application. The communication recommendation device is used to implement the aforementioned communication method, such as the communication method shown in FIG8 .

[0259] As shown in FIG8 , the apparatus may include a transceiver module 801 and a processing module 802 , wherein:

[0260] The transceiver module 801 is configured to send first data, where the first data is obtained by inputting the second data into the third model for processing;

[0261] The transceiver module 801 is further configured to receive first information and second information, wherein the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient;

[0262] The processing module 802 is configured to obtain a fourth model based on the first intermediate gradient, the normalization coefficient, and the third model.

[0263] In a possible implementation, the processing module 802 is configured to:

[0264] obtaining a second intermediate gradient based on the first intermediate gradient and the normalization coefficient;

[0265] The third model is processed based on the second intermediate gradient to obtain a fourth model.

[0266] In a possible implementation, the normalization coefficient includes a first coefficient and a second coefficient, the first coefficient is indicated using a first period, and the second coefficient is indicated using a second period, wherein the first period is greater than the second period.

[0267] In a possible implementation manner, the first period and / or the second period is configured by at least one of RRC, MAC CE, and DCI.

[0268] In one possible implementation, the first cycle is related to hyperparameters of model training.

[0269] In a possible implementation manner, the granularity of the normalization coefficient is configured by at least one of RRC, MAC CE, and DCI.

[0270] In one possible implementation, the granularity of the normalization coefficient is determined based on at least one of the following:

[0271] One or more subbands, one or more resource blocks RB, one or more resource block group bundles RBG, one or more control channel elements CCE, one or more aggregation levels, and processing granularity in a neural network structure.

[0272] In a possible implementation, the normalization coefficient is one or more scalars, and each of the one or more scalars corresponds to one or more granularities.

[0273] In another possible implementation manner, the normalization coefficient is a codeword in a preset codebook.

[0274] In another possible implementation, the normalization coefficient is a eigenvalue after compression of the neural network.

[0275] In one possible implementation, the normalization coefficient is related to at least one of the following parameters:

[0276] The training time of the first model, the number of epochs or iterations of the training process of the first model, the loss function of the first model, and the learning rate of the first model.

[0277] In a possible implementation, the transceiver module 801 is further configured to:

[0278] Receive third information and fourth information, where the third information includes a first bit set, the first bit set indicating the quantization normalization coefficient of the first model during the i-th training, and the fourth information includes a second bit set, the second bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, where i and k are both positive integers.

[0279] In a possible implementation, the change in the number of bits is implemented based on different formats of the scheduled UCI.

[0280] In another possible implementation, the change in the number of bits is based on RRC reconfiguration.

[0281] In a possible implementation, the transceiver module 801 is further configured to:

[0282] Receive fifth and sixth information, the fifth information including a third bit set, the third bit set indicating the quantization normalization coefficient of the first model during the i-th training, the sixth information including a fourth bit set and a fifth bit set, the fourth bit set indicating preset content, the fifth bit set indicating the quantization normalization coefficient of the first model during the i+k-th training, wherein i and k are both positive integers.

[0283] In a possible implementation, the processing module 802 is further configured to:

[0284] Decoding the second information to obtain a decoding result;

[0285] If the decoded result represents a decoding error, the normalization coefficient is recorded as a first value or a second value, the first value is used to indicate that the normalization coefficient is not considered, and the second value is used to indicate that the current model training is terminated.

[0286] In another possible implementation, the processing module 802 is further configured to:

[0287] Decoding the second information to obtain a decoding result;

[0288] If the decoded result indicates a decoding error, the normalization coefficient is recorded as a third value, where the third value is a filtered value of at least two historical normalization coefficients.

[0289] For a detailed introduction to each of the above modules, please refer to the description of the aforementioned embodiments, which will not be repeated here.

[0290] It should be understood that the division of the various units in the above communication device is merely a division of logical functions. In actual implementation, they may be fully or partially integrated into a single physical entity, or they may be physically separated. In addition, the units in the communication device may be implemented in the form of a processor calling software; for example, the communication device includes a processor, the processor is connected to a memory, and the memory stores instructions. The processor calls the instructions stored in the memory to implement any of the above methods or to implement the functions of the various units of the device, where the processor is, for example, a general-purpose processor, such as a central processing unit (CPU) or a microprocessor, and the memory is a memory within the device or a memory outside the device. Alternatively, the units in the device can be implemented in the form of hardware circuits, and the functions of some or all of the units can be realized by designing the hardware circuits. The hardware circuit can be understood as one or more processors. For example, in one implementation, the hardware circuit is an application-specific integrated circuit (ASIC), which realizes the functions of some or all of the above units by designing the logical relationship of the components in the circuit. For another example, in another implementation, the hardware circuit can be implemented by a programmable logic device (PLD). Taking a field programmable gate array (FPGA) as an example, it can include a large number of logic gate circuits, and the connection relationship between the logic gate circuits is configured by configuring the configuration file, thereby realizing the functions of some or all of the above units. All units of the above devices can be implemented in the form of software called by the processor, or in the form of hardware circuits, or in part by software called by the processor, and the rest by hardware circuits.

[0291] 9 is a schematic diagram of the hardware structure of another communication device provided in an embodiment of the present application. The communication device 900 shown in FIG9 (the device 900 may be a computer device) includes a memory 901, a processor 902, a communication interface 903, and a bus 904. The memory 901, the processor 902, and the communication interface 903 are connected to each other via the bus 904.

[0292] The memory 901 may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM).

[0293] The memory 901 can store programs. When the program stored in the memory 901 is executed by the processor 902, the processor 902 and the communication interface 903 are used to perform the various steps of the communication method of the embodiment of the present application.

[0294] The processor 902 is a circuit with signal processing capabilities. In one implementation, the processor 902 can be a circuit with instruction reading and execution capabilities, such as a central processing unit (CPU), a microprocessor, a graphics processing unit (GPU) (which can be understood as a microprocessor), or a digital signal processor (DSP). In another implementation, the processor 902 can implement certain functions through the logical relationship of a hardware circuit. The logical relationship of the hardware circuit is fixed or reconfigurable. For example, the processor 902 is a hardware circuit implemented by an ASIC or a programmable logic device (PLD), such as an FPGA. In a reconfigurable hardware circuit, the process of the processor loading a configuration file and implementing the hardware circuit configuration can be understood as the process of the processor loading instructions to implement the functions of some or all of the above units. In addition, it can also be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as a neural network processing unit (NPU), a tensor processing unit (TPU), a deep learning processing unit (DPU), etc. The processor 902 is used to execute relevant programs to implement the functions required to be performed by the units in the communication device of the embodiment of the present application, or to execute the communication method of the method embodiment of the present application.

[0295] It can be seen that each unit in the above device can be one or more processors (or processing circuits) configured to implement the above method, such as: CPU, GPU, NPU, TPU, DPU, microprocessor, DSP, ASIC, FPGA, or a combination of at least two of these processor forms.

[0296] In addition, the various units in the above devices can be fully or partially integrated together, or can be implemented independently. In one implementation, these units are integrated together and implemented in the form of a system-on-a-chip (SOC). The SOC may include at least one processor for implementing any of the above methods or implementing the functions of the various units of the device. The type of the at least one processor can be different, for example, including a CPU and FPGA, a CPU and an artificial intelligence processor, a CPU and a GPU, etc.

[0297] The communication interface 903 uses a transceiver device such as, but not limited to, a transceiver to implement communication between the apparatus 900 and other devices or a communication network. For example, data can be obtained through the communication interface 903 .

[0298] The bus 904 may include a path for transmitting information between various components of the device 900 (eg, the memory 901 , the processor 902 , and the communication interface 903 ).

[0299] It should be noted that although the device 900 shown in FIG9 only shows a memory, a processor, and a communication interface, during the specific implementation process, those skilled in the art will understand that the device 900 also includes other components necessary for normal operation. At the same time, according to specific needs, those skilled in the art will understand that the device 900 may also include hardware components that implement other additional functions. In addition, those skilled in the art will understand that the device 900 may also include only the components necessary to implement the embodiments of the present application, and does not necessarily include all the components shown in FIG9.

[0300] An embodiment of the present application further provides a communication system, which includes the communication device in the embodiment shown in FIG. 7 and the communication device in the embodiment shown in FIG. 8 .

[0301] An embodiment of the present application further provides a chip, which includes a processor and an interface, wherein the processor is configured to read instructions through the interface to execute one or more steps in the communication method in the aforementioned embodiment.

[0302] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. The computer program is used to implement one or more steps in the communication method in the aforementioned embodiment.

[0303] An embodiment of the present application further provides a computer program product. When the computer program product is run on a computer, the computer is caused to execute one or more steps of the communication method in the aforementioned embodiment.

[0304] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, devices and units can refer to the specific descriptions of the corresponding steps in the aforementioned method embodiments and will not be repeated here.

[0305] It should be understood that in the description of this application, unless otherwise specified, " / " indicates that the objects associated with each other are in an "or" relationship. For example, A / B can mean A or B; where A and B can be singular or plural. Also, in the description of this application, unless otherwise specified, "multiple" means two or more than two. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or plural. In addition, to facilitate the clear description of the technical solutions of the embodiments of this application, in the embodiments of this application, words such as "first" and "second" are used to distinguish between identical or similar items with substantially the same functions and effects. Those skilled in the art will understand that words such as "first" and "second" do not limit the quantity or execution order, and words such as "first" and "second" do not necessarily mean different. For example, the first software communication package and the second software communication package may be the same software communication package. At the same time, in the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a concrete manner for ease of understanding.

[0306] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The mutual coupling, direct coupling, or communication connection shown or discussed can be through some interface, indirect coupling or communication connection of devices or units, and can be electrical, mechanical or other forms.

[0307] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0308] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. 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 a computer-readable storage medium or transmitted via the computer-readable storage medium. The computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated therein. The available medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic medium such as a floppy disk, a hard disk, a tape, a magnetic disk, or an optical medium such as a digital versatile disc (DVD), or a semiconductor medium such as a solid state disk (SSD).

[0309] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A communication method, characterized in that: include: receiving first data, and inputting the first data into a first model for processing to obtain an output of the first model; Obtaining a first intermediate gradient and a normalization coefficient based on an output of the first model; First information and second information are sent respectively, wherein the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient.

2. The method according to claim 1, characterized in that The normalization coefficient includes a first coefficient and a second coefficient, the first coefficient is indicated by a first period, and the second coefficient is indicated by a second period, wherein the first period is greater than the second period.

3. The method according to claim 2, characterized in that The first period and / or the second period is configured by at least one of a radio resource control layer RRC, a MAC control element MAC CE, downlink control information DCI, uplink control information UCI, and sidelink control information SCI.

4. The method according to claim 2 or 3, characterized in that: The first cycle is related to the hyperparameters of model training.

5. The method according to any one of claims 1 to 4, characterized in that: The granularity of the normalization coefficient is configured by at least one of RRC, MAC CE, and DCI.

6. The method according to any one of claims 1 to 4, characterized in that: The granularity of the normalization coefficient is determined based on at least one of the following: One or more subbands, one or more resource blocks RB, one or more resource block group RBG bundling, one or more control channel elements CCE, one or more aggregation levels, processing granularity in the neural network structure.

7. The method according to claim 6, characterized in that The normalization coefficient is one or more scalars, and each of the one or more scalars corresponds to one or more granularities.

8. The method according to claim 6, characterized in that The normalization coefficient is a codeword in a preset codebook.

9. The method according to claim 6, characterized in that The normalization coefficient is the eigenvalue after the neural network is compressed.

10. The method according to any one of claims 1 to 9, characterized in that: The normalization coefficient is related to at least one of the following parameters: The training time of the first model, the number of epochs or iterations of the training process of the first model, the loss function of the first model, and the learning rate of the first model.

11. The method according to claim 10, characterized in that The method further comprises: Send third information and fourth information, wherein the third information includes a first bit set, the first bit set indicates the quantization normalization coefficient of the first model when performing the i-th training, and the fourth information includes a second bit set, the second bit set indicates the quantization normalization coefficient of the first model when performing the i+k-th training, the number of bits corresponding to the first bit set is greater than the number of bits corresponding to the second bit set, wherein i and k are both positive integers.

12. The method according to claim 11, characterized in that The change in the number of bits is achieved based on different formats of scheduling uplink control information UCI.

13. The method according to claim 11, characterized in that The change in the number of bits is based on RRC reconfiguration.

14. The method according to claim 10, characterized in that The method further comprises: Send fifth information and sixth information, the fifth information including a third bit set, the third bit set indicating that the first model is The quantization normalization coefficient when performing the i-th training, the sixth information includes a fourth bit set and a fifth bit set, the fourth bit set indicates a preset content, and the fifth bit set indicates the quantization normalization coefficient of the first model when performing the i+k-th training, wherein i and k are both positive integers.

15. The method according to any one of claims 1 to 14, characterized in that The method further comprises: The first model is trained to obtain a second model.

16. A communication method, characterized in that: include: Sending first data, where the first data is obtained by inputting the second data into a third model for processing; receiving first information and second information, wherein the first information is used to indicate the first intermediate gradient, and the second information is used to indicate the normalization coefficient; A fourth model is obtained based on the first intermediate gradient, the normalization coefficient and the third model.

17. The method according to claim 16, characterized in that The obtaining a fourth model based on the first intermediate gradient, the normalization coefficient and the third model comprises: Obtaining a second intermediate gradient based on the first intermediate gradient and the normalization coefficient; The third model is processed based on the second intermediate gradient to obtain a fourth model.

18. The method according to claim 16 or 17, characterized in that The normalization coefficient includes a first coefficient and a second coefficient, the first coefficient is indicated by a first period, and the second coefficient is indicated by a second period, wherein the first period is greater than the second period.

19. The method according to any one of claims 16 to 18, characterized in that The method further comprises: Decoding the second information to obtain a decoded result; If the decoded result represents a decoding error, the normalization coefficient is recorded as a first value or a second value, the first value is used to indicate that the normalization coefficient is not considered, and the second value is used to indicate that the current model training is terminated.

20. The method according to any one of claims 16 to 18, characterized in that The method further comprises: Decoding the second information to obtain a decoded result; If the decoded result indicates a decoding error, the normalization coefficient is recorded as a third value, and the third value is a filtered value of at least two historical normalization coefficients.

21. A communication device, characterized in that: The apparatus comprises means for performing the method of any one of claims 1-20.

22. A communication system, characterized in that: The method comprises a first communication device and a second communication device, wherein the first communication device is used to execute the method according to any one of claims 1 to 15, and the second communication device is used to execute the method according to any one of claims 16 to 20.

23. A communication device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 20 when executing the computer program.

24. A computer-readable storage medium, wherein a computer program or instruction is stored in the computer-readable storage medium, and when the computer program or instruction is executed, the method according to any one of claims 1 to 20 is executed.

Images