Quantization parameter determination method and related device

Abstract

The invention discloses a quantization parameter determination method and a related device, and the method comprises the steps: receiving first data and second data transmitted by second equipment; the first data is obtained by quantizing the third data by the first quantizer through the quantization parameter configured in the mth round; the second data comprises at least one of channel state information and computing power information; determining fourth data based on the first data and the second data; the fourth data is the gradient of the input layer of the first neural network model; sending fourth data to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in the (m + 1) th round. According to the embodiment of the invention, the quantization parameter of the quantizer can be dynamically adjusted, and the communication overhead can be reduced under the condition that the processing precision of the AI task is ensured; or on the premise that the given communication overhead is met, the processing precision of the AI task is maximized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication and artificial intelligence, and particularly relates to a quantization parameter determination method and related apparatus. BACKGROUND

[0002] The Telecommunication Standardization Organization proposes a new vision for the sixth generation mobile communication technology (6th Generation Mobile Communication Technology, 6G for short), and points out that the 6G network needs to have the basic ability to support artificial intelligence (Artificial Intelligence, AI) tasks, such as distributed AI computing and reasoning, and end-to-end AI application. With the improvement of the prediction accuracy of AI models, the network is also deepening, and the consumption of model running on device memory will bring great challenges to device performance. In order to reduce the consumption of model inference on device memory and improve the inference speed, the prior art proposes a method of quantizing the model and input data. Common quantization methods include binary quantization, linear quantization, and quantization. A quantization scheme needs to determine several key parameters. The current quantization parameters are usually statically configured, which means that the existing quantization scheme is not suitable for the changing communication situation at any time, which is not conducive to reducing communication overhead. SUMMARY

[0003] Embodiments of the present application provide a quantization parameter determination method and related apparatus, which can dynamically adjust the quantization parameters of the quantizer, which is conducive to reducing the communication overhead while ensuring the processing accuracy of the AI task; or on the premise of meeting the given communication overhead, maximizing the processing accuracy of the AI task.

[0004] In a first aspect, an embodiment of the present application provides a quantization parameter determination method applied to a first device or a chip in the first device, and the method comprises the following steps:

[0005] receiving first data and second data sent by a second device; the first data is obtained by quantizing third data by a first quantizer through an mth round of configured quantization parameters; and the second data comprises at least one of channel state information and computing power information;

[0006] determining fourth data based on the first data and the second data; the fourth data is a gradient of an input layer of a first neural network model;

[0007] sending the fourth data to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in a (m+1)th round.

[0008] It can be seen that in the embodiment of the application, the first device can receive the first data quantized by the first quantizer and at least one of the channel state information and the computing power information corresponding to the second device sent by the second device. Based on the first data and at least one of the channel state information and the computing power information sent by the second device, the first device can determine the gradient of the input layer of the neural network model deployed by the first device and send the gradient to the second device. Then, the second device can calculate the gradient of the first quantizer based on the gradient, so that the second device can dynamically update at least one quantization parameter of the first quantizer, so that the quantization parameter of the first quantizer is adapted to the dynamically changing computing power state and channel state, thereby reducing the overhead of channel transmission as much as possible while ensuring the accuracy of end-edge collaborative AI tasks.

[0009] In a possible implementation, the third data is obtained by performing feature extraction on the fifth data through the second neural network model.

[0010] Based on the first data and the second data, the fourth data is determined, including:

[0011] The AI task processing result of the first data and the fifth data is obtained based on the first data.

[0012] The loss value of the loss function is determined by using the AI task processing result of the first data, the second data and the fifth data.

[0013] The fourth data is determined based on the loss value of the loss function.

[0014] In this implementation, based on at least one of the channel state information and the computing power information sent by the second device, in combination with the first data and the AI task processing result obtained according to the first data, the first device can calculate the overall loss of AI task processing and communication transmission data volume, so that the gradient of the input layer of the neural network model on the first device side can be obtained based on the loss, and then the gradient can be sent to the second device for gradient calculation and parameter updating of the local model.

[0015] In a possible implementation, the at least one first quantization parameter includes part or all of the number of breakpoints of the first quantizer, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factor.

[0016] Wherein, the breakpoint refers to a point for dividing the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to represent the approximation degree of the continuous function and the sign function.

[0017] In the implementation, the first quantizer supports dynamically updated quantization parameters, which can be part or all of the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factors, and the update configuration is relatively flexible.

[0018] In a possible implementation, when the at least one first quantization parameter is part of the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factors, before determining the fourth data based on the first data and the second data, the method further includes:

[0019] sending, to the second device, at least one second quantization parameter of the first quantizer, the at least one second quantization parameter being another part of the quantization parameters other than the at least one first quantization parameter from the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factors, and the at least one second quantization parameter being preconfigured by the first device.

[0020] In the implementation, the at least one second quantization parameter of the first quantizer can be preconfigured by the first device, and the first device can send the preconfigured at least one second quantization parameter to the second device, so that the second device does not need to update the at least one second quantization parameter in the training process.

[0021] In a possible implementation, when the at least one first quantization parameter is part of the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factors, before determining the fourth data based on the first data and the second data, the method further includes:

[0022] receiving, from the second device, at least one second quantization parameter of the first quantizer, the at least one second quantization parameter being another part of the quantization parameters other than the at least one first quantization parameter from the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factors, and the at least one second quantization parameter being preconfigured by the second device.

[0023] In the implementation, the at least one second quantization parameter of the first quantizer can be preconfigured by the second device, and the second device can send the preconfigured at least one second quantization parameter to the first device, to synchronize the quantization parameters.

[0024] In a possible implementation, the second device is one or more, and each second device corresponds to a group of first data, a group of fifth data, and a group of second data; and the AI task processing result of the fifth data is a probability that the first data corresponding to the one or more second devices is processed into a label of the fifth data.

[0025] The AI task processing result of the first data, the second data, and the fifth data is used to determine a loss value of the loss function, including:

[0026] Obtaining a KL divergence between a probability distribution of the first data corresponding to each second device and a reference probability distribution of the first data corresponding to each second device; wherein the dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device;

[0027] A loss value of the loss function is determined based on a probability that the first data corresponding to the one or more second devices is processed as a label of the fifth data and a KL divergence corresponding to each second device.

[0028] In this implementation, the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data can represent the difference in data size between the original input data and the quantized input data of each second device. The smaller the KL divergence, the closer the quantized data is to the reference probability distribution. By minimizing this KL divergence, the overhead of transmitting the quantized data on the kth link can be reduced. Including the probability that one or more first data corresponding to the second device will be processed as the label of the fifth data in the loss function can ensure the accuracy of the overall AI task. In other words, the design of the loss function is conducive to minimizing channel transmission overhead while ensuring the accuracy of the overall AI task. In addition, the dimensionality of the quantized first data is constrained by at least one of channel state information and computing power information. The loss function, which takes at least one of the channel state information and computing power information as input, can perceive the dimensionality of the first data under dynamically changing computing power and / or channel conditions. Based on the loss value of the loss function, the quantization parameter of the quantizer can be adjusted to adapt the quantization parameter to the dynamically changing computing power and channel conditions.

[0029] In one possible implementation, there are multiple second devices, and the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are different or the types of at least one second quantization parameter pre-configured by the first quantizers on the multiple second device sides are the same.

[0030] In this implementation, the type of at least one second quantization parameter pre-configured on the multiple second device sides may be the same or different, and the type of the pre-configured quantization parameter is relatively flexible.

[0031] In a second aspect, an embodiment of the present application provides a method for determining a quantization parameter, which is applied to a second device or a chip in the second device. The method includes:

[0032] Sending first data and second data to a first device; the first data is obtained by quantizing third data by a first quantizer using a quantization parameter configured in the mth round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine fourth data; the fourth data is a gradient of an input layer of the first neural network model;

[0033] receiving fourth data sent by the first device;

[0034] determining, based on the fourth data, at least one first quantization parameter of the first quantizer in the (m+1)th round.

[0035] It can be seen that in the embodiments of the present application, the second device can send the first data processed by the first quantizer and at least one of the channel state information and the computing power information corresponding to the second device to the first device. Based on the first data and at least one of the channel state information and the computing power information sent by the second device, the first device can determine the gradient of the input layer of the neural network model deployed by the first device and send the gradient to the second device. Then, the second device can calculate the gradient of the first quantizer based on the gradient, so that the second device can dynamically update at least one quantization parameter of the first quantizer to adapt the quantization parameter of the first quantizer to the dynamically changing computing power state and channel state, thereby reducing the overhead of channel transmission as much as possible while ensuring the accuracy of the end-edge collaborative AI task.

[0036] In a possible implementation, the at least one first quantization parameter includes part or all of the number of breakpoints of the first quantizer, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factor;

[0037] In this implementation, the breakpoint refers to a point for dividing the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to represent the approximation degree of the continuous function and the sign function.

[0038] In this implementation, the first quantizer supporting the dynamically updated quantization parameter can be part or all of the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factor, and the update configuration is relatively flexible.

[0039] In a possible implementation, when the at least one first quantization parameter is part of the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factor, before receiving the fourth data sent by the first device, the method further includes:

[0040] receiving at least one second quantization parameter of the first quantizer sent by the first device; the at least one second quantization parameter is another part of the quantization parameters except the at least one first quantization parameter among the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factor; and the at least one second quantization parameter is preconfigured by the first device.

[0041] In the implementation, the at least one second quantization parameter of the first quantizer can be preconfigured by the first device, and the first device can send the preconfigured at least one second quantization parameter to the second device, so that the second device does not need to update the at least one second quantization parameter in the training process.

[0042] In a possible implementation, when the at least one first quantization parameter is part of the number of breakpoints, the breakpoint values, the quantization values at the breakpoints, and the approximation factors, before receiving the fourth data sent by the first device, the method further includes:

[0043] sending, to the first device, at least one second quantization parameter of the first quantizer, the at least one second quantization parameter being another part of the quantization parameters except the at least one first quantization parameter from the number of breakpoints, the breakpoint values, the quantization values at the breakpoints, and the approximation factors, and the at least one second quantization parameter being preconfigured by the second device.

[0044] In the implementation, the at least one second quantization parameter of the first quantizer can be preconfigured by the second device, and the second device can send the preconfigured at least one second quantization parameter to the first device to synchronize the quantization parameters.

[0045] In a possible implementation, the third data is obtained by performing feature extraction on the fifth data by using the second neural network model; the fourth data is determined based on a loss value of a loss function, and the loss value of the loss function is determined by using AI task processing results of the first data, the second data, and the fifth data, and the AI task processing result of the fifth data is obtained based on the first data.

[0046] In the implementation, the first device can calculate the overall loss of the AI task processing and the communication transmission data volume based on at least one of the channel state information and the computing power information sent by the second device, in combination with the first data and the AI task processing result obtained based on the first data, so as to obtain the gradient of the input layer of the neural network model on the first device side, and then the gradient can be sent to the second device to facilitate the second device to perform gradient calculation and parameter updating on the local model.

[0047] In a possible implementation, the second device is multiple, and the types of the at least one second quantization parameter preconfigured by the first quantizer of the multiple second devices are different or the types of the at least one second quantization parameter preconfigured by the first quantizer of the multiple second devices are the same.

[0048] In the implementation, the types of the at least one second quantization parameter preconfigured by the multiple second devices can be the same or different, and the types of the preconfigured quantization parameters are relatively flexible.

[0049] In a third aspect, an embodiment of the present application provides a quantization parameter determination apparatus applied to a first device or a chip in the first device, the apparatus comprising a first transceiver and a first processing unit;

[0050] The first transceiver is configured to receive first data and second data transmitted by a second device, wherein the first data is obtained by quantizing third data by a first quantizer using a mth quantization parameter, and the second data comprises at least one of channel state information and computing power information.

[0051] The first processing unit is configured to determine fourth data based on the first data and the second data, wherein the fourth data is a gradient of an input layer of a first neural network model.

[0052] The first transceiver is further configured to transmit the fourth data to the second device, wherein the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in a (m+1)th round.

[0053] In a possible implementation, the third data is obtained by performing feature extraction on fifth data by a second neural network model, and in the step of determining the fourth data based on the first data and the second data, the first processing unit is specifically configured to:

[0054] obtain an artificial intelligence (AI) task processing result of the fifth data based on the first data;

[0055] determine a loss value of a loss function using the AI task processing result of the first data, the second data and the fifth data; and

[0056] determine the fourth data based on the loss value of the loss function.

[0057] In a possible implementation, the at least one first quantization parameter comprises part or all of a number of breakpoints, a value of the breakpoints, a quantized value at the breakpoints and an approximation factor of the first quantizer.

[0058] The breakpoints are points for dividing the first quantizer into multiple value intervals, the first quantizer is obtained by relaxing a sign function in a second quantizer into a continuous function, and the approximation factor is used to represent an approximation degree of the continuous function and the sign function.

[0059] In a possible implementation, when the at least one first quantization parameter is part of the number of the breakpoints, the value of the breakpoints, the quantized value at the breakpoints and the approximation factor, the first transceiver is further configured to:

[0060] The first transceiving unit is further configured to send, to the second device, at least one second quantization parameter of the first quantizer, the at least one second quantization parameter being another part of the quantization parameters except the at least one first quantization parameter from among the number of breakpoints, the breakpoint values, the quantization values at the breakpoints, and the approximation factors, and the at least one second quantization parameter being preconfigured for the first device.

[0061] In a possible implementation, when the at least one first quantization parameter is part of the number of breakpoints, the breakpoint values, the quantization values at the breakpoints, and the approximation factors, the first transceiving unit is further configured to:

[0062] The first transceiving unit is further configured to receive, from the second device, at least one second quantization parameter of the first quantizer, the at least one second quantization parameter being another part of the quantization parameters except the at least one first quantization parameter from among the number of breakpoints, the breakpoint values, the quantization values at the breakpoints, and the approximation factors, and the at least one second quantization parameter being preconfigured for the second device.

[0063] In a possible implementation, the second device is one or more, and each second device corresponds to a group of first data, a group of fifth data, and a group of second data; the AI task processing result of the fifth data is a probability that the first data corresponding to the one or more second devices is processed into a label of the fifth data; and in determining a loss value of a loss function by using the AI task processing result of the first data, the second data, and the fifth data, the first processing unit is specifically configured to:

[0064] obtain a KL divergence between a probability distribution of the first data corresponding to each second device and a reference probability distribution of the first data corresponding to each second device, wherein a dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device;

[0065] determine the loss value of the loss function based on the probability that the first data corresponding to the one or more second devices is processed into the label of the fifth data and the KL divergence corresponding to each second device.

[0066] In a possible implementation, the second device is a plurality of, and the at least one second quantization parameter preconfigured for the first quantizer on the side of the plurality of second devices is of different types or the at least one second quantization parameter preconfigured for the first quantizer on the side of the plurality of second devices is of the same type.

[0067] It should be understood that, since the method embodiment and the device embodiment are different presentation forms of the same technical concept, the content of the first aspect of the embodiment of the present application should be synchronously adapted to the third aspect of the embodiment of the present application, and the same or similar beneficial effects can be achieved, which will not be described herein again.

[0068] In a fourth aspect, an embodiment of the present application provides a quantization parameter determination apparatus applied to a second device or a chip in the second device, the apparatus comprising a second transceiver and a second processor;

[0069] The second transceiver is configured to send first data and second data to the first device, the first data being obtained by quantizing third data by a first quantizer using a mth quantization parameter, and the second data comprising at least one of channel state information and computing power information, the first data and the second data being used by the first device to determine fourth data, the fourth data being a gradient of an input layer of a first neural network model, and the second transceiver is further configured to receive the fourth data sent by the first device;

[0070] The second processor is configured to determine at least one first quantization parameter of the first quantizer in a (m+1)th round based on the fourth data.

[0071] In a possible implementation, the at least one first quantization parameter comprises part or all of a number of breakpoints, a value of the breakpoints, a quantization value at the breakpoints, and an approximation factor of the first quantizer;

[0072] The breakpoints refer to points for dividing the first quantizer into multiple value intervals, the first quantizer is obtained by relaxing a sign function in a second quantizer into a continuous function, and the approximation factor is used to represent an approximation degree of the continuous function and the sign function.

[0073] In a possible implementation, when the at least one first quantization parameter is part of the number of the breakpoints, the value of the breakpoints, the quantization value at the breakpoints, and the approximation factor, the second transceiver is further configured to:

[0074] receive at least one second quantization parameter of the first quantizer sent by the first device, the at least one second quantization parameter being another part of the quantization parameter except the at least one first quantization parameter from the number of the breakpoints, the value of the breakpoints, the quantization value at the breakpoints, and the approximation factor, and the at least one second quantization parameter being preconfigured by the first device.

[0075] In a possible implementation, when the at least one first quantization parameter is part of the number of the breakpoints, the value of the breakpoints, the quantization value at the breakpoints, and the approximation factor, the second transceiver is further configured to:

[0076] send at least one second quantization parameter of the first quantizer to the first device, the at least one second quantization parameter being another part of the quantization parameter except the at least one first quantization parameter from the number of the breakpoints, the value of the breakpoints, the quantization value at the breakpoints, and the approximation factor, and the at least one second quantization parameter being preconfigured by the second device.

[0077] In a possible implementation, the third data is obtained by performing feature extraction on the fifth data by using the second neural network model; and the fourth data is determined based on a loss value of the loss function, and the loss value of the loss function is determined by using AI task processing results of the first data, the second data, and the fifth data, and the AI task processing result of the fifth data is obtained based on the first data.

[0078] In a possible implementation, the second device is multiple, and the types of the at least one second quantization parameter preconfigured for the first quantizer on the side of the multiple second devices are different or the types of the at least one second quantization parameter preconfigured for the first quantizer on the side of the multiple second devices are the same.

[0079] It should be understood that, since the method embodiments and the device embodiments are different present forms of the same technical concept, the content of the second aspect of the embodiments of the present application should be synchronously adapted to the fourth aspect of the embodiments of the present application, and the same or similar beneficial effects can be achieved, which will not be described here.

[0080] In a fifth aspect, the embodiments of the present application provide a communication device, including a processor, a memory, a communication interface, and one or more programs, the one or more programs are stored in the memory and configured to be executed by the processor to cooperate with the communication interface to implement the method in any one of the embodiments of the first aspect or the second aspect.

[0081] In a sixth aspect, the embodiments of the present application provide a chip, including a processor configured to call and run a computer program from a memory, so that a device installed with the chip executes the method in any one of the embodiments of the first aspect or the second aspect.

[0082] In a seventh aspect, the embodiments of the present application provide a computer readable storage medium, which stores a computer program for execution by a device, and the computer program is executed to implement the method in any one of the embodiments of the first aspect or the second aspect.

[0083] In an eighth aspect, the embodiments of the present application provide a computer program product, when the computer program product is run by a communication device, so that the communication device executes the method in any one of the embodiments of the first aspect or the second aspect. BRIEF DESCRIPTION OF DRAWINGS

[0084] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the background art, the drawings needed to be used in the embodiments of the present application or the background art will be described below.

[0085] Figure 1 A schematic diagram of end-edge collaborative deployment of a neural network model;

[0086] Figure 2A schematic diagram of a system architecture provided for an embodiment of the present application;

[0087] Figure 3 A flowchart of a quantization parameter determination method provided for an embodiment of the present application;

[0088] Figure 3A A schematic diagram of a multi-device same quantization configuration provided for an embodiment of the present application;

[0089] Figure 3B A schematic diagram of a multi-device different quantization configuration provided for an embodiment of the present application;

[0090] Figure 4 A flowchart of another quantization parameter determination method provided for an embodiment of the present application;

[0091] Figure 5 A schematic diagram of a base station pre-configuring part of quantization parameters provided for an embodiment of the present application;

[0092] Figure 6 A flowchart of another quantization parameter determination method provided for an embodiment of the present application;

[0093] Figure 7 A schematic diagram of a terminal pre-configuring part of quantization parameters provided for an embodiment of the present application;

[0094] Figure 8A A schematic diagram of a relationship between an inference error rate of an edge-aided AI model and a wireless channel data transmission delay provided for an embodiment of the present application;

[0095] Figure 8B A schematic diagram of a relationship between an inference error rate of an edge-aided AI model and a number of edge devices participating in the collaboration provided for an embodiment of the present application;

[0096] Figure 8C A schematic diagram of a relationship between an inference error rate of an edge-aided AI model and a peak signal-to-noise ratio provided for an embodiment of the present application;

[0097] Figure 9A A schematic diagram of a relationship between an inference error rate of another edge-aided AI model and a wireless channel data transmission delay provided for an embodiment of the present application;

[0098] Figure 9B A schematic diagram of a relationship between an inference error rate of another edge-aided AI model and a number of edge devices participating in the collaboration provided for an embodiment of the present application;

[0099] Figure 9C A schematic diagram of a relationship between an inference error rate of another edge-aided AI model and a peak signal-to-noise ratio provided for an embodiment of the present application;

[0100] Figure 10 A structure diagram of a quantization parameter determination apparatus provided by an embodiment of the present application is shown in FIG. 2.

[0101] Figure 11 A structure diagram of another quantization parameter determination apparatus provided by an embodiment of the present application is shown in FIG. 3.

[0102] Figure 12 A structure diagram of a communication device provided by an embodiment of the present application is shown in FIG. 4. DETAILED DESCRIPTION

[0103] The terms "first", "second", "third", and "fourth" and the like in the description and in the claims of the present application are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the terms so used are interchangeable under appropriate circumstances such that the descriptive terms used herein are to be interpreted in the context as exercised by those of ordinary skill in the art. Furthermore, the term "including" and "comprising" as well as their conjugates are intended to cover non-exclusive inclusion. For example, processes, methods, articles, or apparatuses that comprise a list of steps or elements are not necessarily limited to those steps or elements but can include other not expressly listed steps or elements, or can also include additional or alternative steps or elements inherent in such processes, methods, articles, or apparatuses. The term "exemplary" is used in the sense of serving as an example, instance, or illustration.

[0104] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily all referring to a common or identical embodiment. It is appreciated that embodiments described herein can be combined with other embodiments.

[0105] As used in this description, the terms "component," "module," "system", and the like are intended to refer to a computer-related entity, either hardware, firmware, a combination of hardware and software, software, or software in execution. For example, a component can be, but is not limited to being, a process running on a processor, a processor, an object, an executable, a thread of execution, a program, and / or a computer. By way of illustration, both an application running on a computing device and the computing device can be a component. One or more components can reside within a process and / or thread of execution and a component can be localized, partially localized, and / or distributed across two or more computers. Also, these components can execute from various computer readable media having various data structures stored thereon. The components can communicate by way of local and / or remote processes such as in accordance with a signal having one or more data packets (e.g., data from one component interacting with another component in a local system, distributed system, and / or across a network such as the Internet with other systems via the signal).

[0106] Firstly, the related terms and related technical backgrounds in the present application are briefly introduced, so as to facilitate the understanding of those skilled in the art.

[0107] (1) Channel State Information-Reference Signal, CSI-RS;

[0108] (2) Distributed Information Bottleneck, DIB;

[0109] (3) Floating Point, FP;

[0110] (4) Information Bottleneck, IB;

[0111] (5) Integer, INT;

[0112] (6) Long Term Evolution, LTE;

[0113] (7) Medium Access Control, MAC;

[0114] (8) Machine Learning, ML;

[0115] (9) Non-access stratum, NAS;

[0116] (10) Packet Data Convergence Control, PDCP;

[0117] (11) Physical Layer, PHY;

[0118] (12) Peak signal to noise ratio, PSNR;

[0119] (13) Radio Access Network, RAN;

[0120] (14) Radio Link Control, RLC;

[0121] (15) Radio Resource Control, RRC;

[0122] (16) Reference Signal Received Power, RSRP;

[0123] (17) Reference Signal Received Quality, RSRQ;

[0124] (18) Service Data Adaptation Protocol, SDAP;

[0125] (19) Signal to Interference & Noise Ratio, SINR;

[0126] (20) Signal to Noise Ratio, SNR;

[0127] (21) Sounding Reference Signal, SRS;

[0128] (22) Transmission Reception Point, TRP;

[0129] (23) User Equipment, UE;

[0130] (24) New Radio, NG;

[0131] (25) The Fifth Generation Mobile Communications, 5G;

[0132] (26) Model Quantization: Generally refers to the conversion of floating-point operations of neural networks to fixed-point operations. Model quantization usually involves low-precision models, regular-precision models, and mixed-precision models. The weight numerical format of low-precision models is FP16 (half-precision floating-point type) or INT8 (8-bit fixed-point integer type), but low-precision usually refers to INT8. The weight numerical format of regular-precision models is FP32 (32-bit floating-point type or single-precision floating-point type). The mixed-precision model uses both FP32 and FP16 weight numerical formats. FP16 is used to save memory size, but some parameters or operators must use FP32 format to maintain accuracy.

[0133] (27) Radio Access Network Protocol Layers in NR:

[0134] L1 layer (Layer 1): PHY, mainly including modulation scheme, waveform, channel code, multi-antenna technology, frame structure, duplex scheme, control and reference signals.

[0135] L2 layer (Layer 2): Radio network layer. NR's L2 includes the MAC layer, RLC layer, PDCP layer, and SDAP layer.

[0136] L3 layer (Layer 3): includes the NAS layer and the RRC layer. The L3 layer belongs to the control plane protocol.

[0137] Regarding AI models in future 6G networks, relevant technical reports indicate that they can be divided into multiple parts based on specific tasks and environments. The purpose is to offload the computationally intensive and energy-intensive parts to the network side, while leaving the data privacy-sensitive and latency-sensitive parts on the device side. The device side executes specific neural network layers and transmits intermediate data to the network side. The network side executes the remaining layers and feeds back the inference results to the device side, such as Figure 1 As shown in the figure. However, on the device side, the number of network layers in the deployed neural network models continues to increase as the accuracy requirements become increasingly higher. Each model run consumes a significant amount of device memory, battery, bandwidth, and other factors. To address this issue, there are currently two main solutions: one is to design a more efficient network architecture to achieve acceptable accuracy with a relatively small model size; the other is to reduce the network size through compression and encoding. The most widely used compression method is model quantization. Model quantization can reduce the model's storage space, memory usage, and device power consumption. It can also accelerate inference and improve hardware compatibility.

[0138] Among existing schemes, binary quantization methods can use simple bitwise operations to simultaneously calculate large numbers, with theoretical performance improvements of 5 to 128 times for 1-bit calculations. This only introduces one additional quantization operation, which can benefit from the acceleration of Single Instruction Multiple Data (SIMD). Linear quantization can be divided into symmetric and asymmetric quantization, both of which support 8-bit calculations and offer efficiency improvements ranging from 1 to 16 times. The additional quantization and dequantization operations introduced by linear quantization are standard vector operations and can therefore also be accelerated using SIMD, resulting in minimal additional computational time. Logarithmic quantization is a relatively specialized quantization method that multiplies two exponentials with the same base, equivalent to adding their exponentials, reducing computational effort and converting additions into index calculations. Currently, a quantization scheme in this field is typically determined by three key parameters: the number of breakpoints, the amplitude, and the breakpoint value. A breakpoint is a point that divides the quantizer or quantization function into multiple value intervals, while the amplitude refers to the quantization value at the breakpoint (i.e., the value being quantized or mapped). The relevant quantization schemes basically adopt static configuration for the above three quantization parameters, while the computing power status and channel status corresponding to the terminal device are dynamically changing. The statically configured quantization parameters are not compatible with the dynamically changing computing power and channels. Future networks need to reasonably quantize wireless data to flexibly adjust the number of data transmission bits and reduce communication overhead, computing memory and power consumption. How to design a quantization scheme so that it can adapt to the real-time changing computing power status and wireless channel status, and the introduction of this quantization will not cause a significant performance degradation in the accuracy of AI tasks, is the problem to be solved by this application.

[0139] See Figure 2 , Figure 2 A schematic diagram of a system architecture provided in an embodiment of the present application is shown as follows: Figure 2 As shown in FIG, K terminal devices and network devices are included. The K terminal devices and network devices form a multi-link end-edge collaborative task processing architecture, where K is greater than or equal to 1 (when K is equal to 1, the multi-link end-edge collaborative task processing architecture degenerates into a single-link end-edge collaborative task processing architecture). The K terminal devices are deployed with an end-side neural network model, and the network devices are deployed with an edge-side neural network model. The original data X1,…,X2 with the target variable Y (e.g., label) K After feature extraction by K end-side neural network models, intermediate features Z1,…,Z are output. K , for example, its form can be tokens or embeddings. Data Z1,…,Z K After quantization by the quantizer, the data is obtained The data is transmitted to the network device through the wireless channel. Due to the existence of noise or interference, the data received by the network device becomes data And vector With vector Z k The network device aggregates the K groups of data and performs the remaining AI tasks (e.g., classification) through the side neural network model to obtain the final task result. Data Stream It can be viewed as a Markov chain. The end-side neural network model and quantizer respectively extract features and compress the raw data. The quantized data is sent to the wireless channel. The edge-side neural network model performs AI tasks, such as inference, based on the data received from the wireless channel to obtain the final task results.

[0140] In the study of multi-link end-edge collaborative task processing without introducing a quantization mechanism, the researchers of this application used DIB theory for mathematical modeling and designed a mechanism for generating, updating, and deploying a multi-link end-edge collaborative model. The theoretical formula can be expressed as:

[0141]

[0142] in, Indicates the data received by the network device side A collection of Indicates that the base station side has received data The mutual information between the actual task result Y and the actual task result Y, the larger the mutual information value, the better the data transmitted through the wireless channel. The more information that can be included in the summary to obtain the correct AI task result Y, the higher the accuracy of the AI ​​task result can be understood. Represents the original input data X of the kth device k Data from the kth wireless link The smaller the mutual information, the less data is transmitted in the k-th wireless link, that is, the smaller the wireless communication overhead of the k-th link. represents the sum of data transmitted by K wireless links, The smaller it is, the smaller the total wireless communication overhead of K links is. The configurable parameter β can control the ratio between the two items, and balance the accuracy of AI task results and the total wireless communication overhead. Therefore, minimizing the above formula This means: while ensuring that the wireless communication overhead of the K links is as small as possible, ensure that the accuracy of the AI ​​task results is maximized (or while ensuring the accuracy of the AI ​​task results, minimize the wireless communication overhead of the K links).

[0143] The above modeling can serve as the algorithmic theoretical basis for generating a multi-link end-edge collaborative model. The input of the algorithm is: M pairs of data and labels where M is the size of the batch data. The dimension of the data tokens or embedding transmitted in the wireless link is denoted as N i i.e., the vector or Z i (i = 1, 2, …, K). The output of the algorithm is: K sets of end-side neural network models (such as: φ1, φ2, …, φ K ) and an edge-side neural network model (such as: ψ), i.e., a total of K + 1 AI models are output. Therefore, the loss function of the algorithm can be represented as:

[0144]

[0145] In the present disclosure, the terminal device includes, but is not limited to, a UE, a subscriber unit, a subscriber station, a mobile station, a mobile station, a remote station, a remote terminal device, a mobile terminal device, a user terminal device, a wireless electronic device, a user agent, a user equipment, a cellular phone, a cordless phone, a session initiation protocol (SIP) phone, a wireless local loop (WLL) station, a personal digital assistant (PDA), a handheld device with wireless communication function, a computing device, a processing device connected to a wireless modem, a vehicle-mounted communication module or other embedded communication module, a wireless terminal in industrial control, a wireless terminal in unmanned, a wireless terminal in remote medical, a wireless terminal in smart grid, a wearable device, a terminal device in Internet of Things, a household appliance, a virtual reality device, a terminal device in future 5G network, or a terminal device in future evolved public land mobile network (PLMN), and the like.

[0146] The network device mentioned in the present application includes various forms of base stations, such as macro base stations, micro base stations (also known as small stations), relay stations, access points, and the like. In systems using different wireless access technologies, the names of devices with base station functions may be different, such as gNB in 5GS, eNB or eNodeB in LTE systems, Node B in 3rd generation (3G) systems, and the like. In some deployments, the network device can be a central unit (CU) and a distributed unit (DU), and the like. In other deployments, the CU can also be divided into a CU-control plane (CP) and a CU-user plane (UP), and the like. In still other deployments, the network device can also be a radio unit (RU), an open radio access network (ORAN) architecture, and the like. The embodiments of the present application do not limit the deployment mode of the network device. For example, when the network device is an ORAN architecture, the network device shown in the embodiments of the present application can be an access network device in the ORAN, or a module in the access network device, and the like. In the ORAN architecture, the CU can also be referred to as an open (O)-CU, the DU can also be referred to as an O-DU, the CU-CP can also be referred to as an O-CU-CP, the CU-UP can also be referred to as an O-CU-UP, and the RU can also be referred to as an O-RU. In still other deployments, the network device can also be a core network device, and the like.

[0147] On the basis of modeling using the DIB theory, the technical solutions provided by the present application will be described in detail in combination with specific embodiments.

[0148] Please refer to Figure 3 , Figure 3 The flowchart of a quantization parameter determination method provided by the embodiments of the present application is shown in Figure 3 , which includes steps 301-304:

[0149] 301: The second device sends the first data and the second data to the first device.

[0150] Correspondingly, the first device receives the first data and the second data sent by the second device.

[0151] In an embodiment of the present application, the second device may be a terminal device or a chip in a terminal device; the first device may be a terminal device or a network device. For example, in a conventional communication scenario, the first device may be a base station, and in a sidelink communication scenario, the first device may also be a terminal device.

[0152] The first data is obtained by quantizing the third data by the first quantizer using the quantization parameters configured in the mth round. The third data is obtained by extracting features from the fifth data using the second neural network model (for example, a device-side neural network model). The second neural network model and the first quantizer are deployed on the second device. The fifth data is the input data of the second neural network model, such as X. K .like Figure 2 In the architecture shown, the neural network model in the Kth terminal device is K Perform feature extraction to obtain the third data Z K , Z K The first data is quantized by the quantizer on the Kth terminal device side It should be understood that due to the presence of channel noise, the first data received by the first device side is represented as

[0153] Exemplarily, the first quantizer can be expressed as:

[0154]

[0155] Among them, let us record an ordinary vector z k,i ,i=1,…,d represents vector z k Each dimension of z k,i The third data Z inputted by the first quantizer on the K-th terminal device side is represented by K The value in the i-th dimension, is a set of real numbers. Where T represents the number of quantization parameter breakpoints; a k,t Indicates the amplitude of the t-th breakpoint, that is, the quantized value of the t-th breakpoint; b k,t The first quantizer is obtained by relaxing the sign function sgn in the second quantizer into a continuous function arctan, and γ is the approximation factor, which represents the degree of approximation between the continuous function arctan and the sign function sgn.

[0156] Exemplarily, the second quantizer can be expressed as:

[0157]

[0158] The second quantizer and the first quantizer can both convert z kEach dimension of the first data is quantized into one of T+1 quantization values. To enable the quantization parameter of the quantizer to be dynamically adjusted according to the edge computing state and the channel state, and without affecting the accuracy of the AI task result, it is necessary to first ensure that the loss function after adding quantization can calculate the gradient, that is, does not affect the back propagation of the gradient during model training. However, the second quantizer is constructed based on the sign function sgn, and the discontinuity of the sign function sgn will cause the gradient calculation to be unable to be performed. Therefore, it is necessary to perform relaxation transformation on the sign function sgn based on mathematical theory, so that the sign function sgn is relaxed into a continuous function.

[0159] The second data includes at least one of channel state information and computing power information, and the computing power information of the second device can be the number of floating point operations per second, and the channel state information can be a downlink channel measurement result or an uplink reference signal.

[0160] 302: The first device determines fourth data based on the first data and the second data.

[0161] In the embodiments of the present application, the fourth data is the gradient of the input layer of the first neural network model (which can be an edge-side neural network model), and the first neural network model is deployed on the first device.

[0162] For example, the first device inputs the first data into the first neural network model to perform the remaining AI task (such as reasoning or prediction, etc.), and obtains the AI task processing result (such as reasoning result or prediction value, etc.) of the original input data (i.e., the fifth data). The first device determines the loss value of the loss function by using the first data, the second data, and the AI task processing result of the fifth data, and performs gradient calculation of each layer of the first neural network model based on the loss value of the loss function, thereby obtaining the fourth data.

[0163] In this implementation, the first device can calculate the overall loss of AI task processing and communication transmission data volume based on at least one of the channel state information and the computing power information sent by the second device, in combination with the first data and the AI task processing result obtained according to the first data, so as to obtain the gradient of the input layer of the neural network model on the first device side, and then send it to the second device, so as to facilitate the second device to perform gradient calculation and parameter update on the local model.

[0164] The loss function can be referred to as a first loss function, and the first loss function is constructed based on the first quantizer and a second loss function, and the second loss function is constructed based on the DIB theory, that is, the second loss function is The loss function. Specifically, the expression of the first quantizer is substituted into The loss function is derived based on mathematical theories such as variational method and Log-sum inequality, and a loss function after introducing quantization (i.e., a first loss function) is obtained. The first loss function can be expressed as:

[0165]

[0166] wherein C QML IB represents the first loss function, x 1:K represents a set of original input data of the K second devices in each dimension, y represents a true value (such as a label) corresponding to the original input data in each dimension, represents a probability that the original input data in each dimension is processed (such as predicted) to be the true value y, represents a probability that each dimension of the original data is quantized to , represents a probability that the first neural network model ψ processes (such as predicts) the to be the true value y, is a processing loss of the first neural network model ψ, represents a KL divergence between a probability distribution of the quantized first data after adding quantization and a reference probability distribution of the first data, represents a reference probability distribution of the first data corresponding to the kth second device, β k represents a trade-off coefficient of the accuracy of the AI task of the kth link and the reduction of transmission overhead.

[0167] For example, the second device is one or more (such as Figure 2 ​corresponding to the first device, and a set of fifth data corresponding to the one or more second devices. The first loss function is based on an expression that can be seen as follows: the AI task processing result of the fifth data is the probability that the first data corresponding to the one or more second devices is processed into the label (or called the true value) of the fifth data (or the probability that the AI task processing result of the fifth data is correct). The first device can obtain the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data, based on the probability that the first data corresponding to the one or more second devices is processed into the label of the fifth data, and the KL divergence corresponding to each second device, to determine the loss value of the loss function. Specifically, the loss value of the first loss function is obtained by summing the KL divergence corresponding to each second device and adding the probability that the first data corresponding to the one or more second devices is processed into the label of the fifth data. The dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device, i.e., at least one of the channel state information and the computing power information sent by the second device determines the dimension of the first data sent by each second device, that is, the quantizer learns what kind of channel state and / or computing power state the first data of what dimension is output, i.e., how much the intermediate feature is quantized.

[0168] In this implementation, the KL divergence between the probability distribution of the first data corresponding to each second device and the reference probability distribution of the first data can represent the difference in data quantity between the original input data of each second device and the quantized input data. The smaller the KL divergence, the closer the quantized data is to the reference probability distribution. By minimizing the KL divergence, the overhead of transmitting the quantized data over the kth link can be reduced. Adding the probability that the first data corresponding to the one or more second devices is processed into the label of the fifth data in the loss function can ensure the accuracy of the overall AI task, i.e., the design of the loss function is beneficial to reducing the transmission overhead of the channel while ensuring the accuracy of the overall AI task. In addition, the dimension of the quantized first data is constrained by at least one of the channel state information and the computing power information, and the loss function takes at least one of the channel state information and the computing power information as input, which can perceive the dimension of the first data under the dynamically changing computing power state and / or channel state, so that the loss value of the loss function can adjust the quantization parameter of the quantizer to adapt the quantization parameter of the quantizer to the dynamically changing computing power and channel.

[0169] 303: The first device sends the fourth data to the second device.

[0170] Correspondingly, the second device receives the fourth data.

[0171] 304: The second device determines at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0172] In the embodiments of the present application, the second device performs gradient calculation of the local model (including the first quantizer and the second neural network model) based on the received fourth data, updates at least one first quantization parameter of the first quantizer in the (m+1)th round based on the calculated gradient, and updates the parameters of each layer of the second neural network model in the (m+1)th round.

[0173] The at least one first quantization parameter includes part or all of the number of breakpoints of the first quantizer The value of the breakpoint The quantization value at the breakpoint And the approximation factor Part or all of the above, j = 1, …, K, that is, for the first quantizer, the dynamically updated quantization parameters supported can be Part or all of the above, if the dynamically updated quantization parameters supported are part, another part of the quantization parameters can be pre-configured.

[0174] In this implementation, the dynamically updated quantization parameters supported by the first quantizer can be part or all of the number of breakpoints, the value of the breakpoint, the quantization value at the breakpoint, and the approximation factor, and the update configuration is relatively flexible.

[0175] It should be noted that steps 301-304 describe the update mechanism of the quantization parameters of the first quantizer in a single round of training. After multiple rounds of data interaction between the first device and the second device, when the first device determines that the loss value of the first loss function converges, the first device can indicate the optimal quantization parameters to the second device. The final output of the embodiments of the present application is K second neural network models, K first quantizers, and a first neural network model. The loss value of the first loss function converges can be that the number of training rounds reaches a preset number of rounds or the loss value is less than or equal to a preset threshold.

[0176] For example, when the loss value of the first loss function converges, the first device can indicate a set of quantization parameters to multiple second devices. The values of the set of quantization parameters configured by the multiple second devices can be the same (i.e., the same quantization configuration is used for K links), or the values of the set of quantization parameters configured by the multiple second devices can be different (i.e., different quantization configurations are used for K links). For example, the first device can indicate the final quantization parameters of each second device to the corresponding second device through first signaling. The first signaling can be L1 layer signaling (such as downlink control information signaling) or L2 layer signaling (such as MAC layer control unit signaling).

[0177] To better understand the scheme of K links using the same quantization configuration, the following takes the end-edge collaborative interaction of K UEs and a base station as an example to briefly describe the parameter updating mechanism of the quantizer. As shown in Figure 3A , the following steps can be included:

[0178] 1: The UE reports quantized data, channel state information, and computing power information;

[0179] 2: The base station calculates the loss value of the loss function based on the quantized data, channel state information, and computing power information, and determines the gradient of the input layer of the first neural network model based on the loss value;

[0180] 3: The base station sends the gradient of the input layer to the UE;

[0181] 4: The UE updates the parameters of the second neural network model and the

[0182] 5: The UE and the base station repeat the updating process of steps 1-4 for multiple rounds until the value of the loss function is minimized;

[0183] 6: The base station issues the optimal quantization scheme {a * ,b * ,T * ,γ *} to the UE;

[0184] 7: The UE completes the configuration of the local first quantizer.

[0185] To better understand the scheme of K links using different quantization configurations, the following takes the end-edge collaborative interaction of K UEs and a base station as an example to briefly describe the parameter updating mechanism of the quantizer. As shown in Figure 3B , the following steps can be included:

[0186] 1: The UE reports quantized data, channel state information, and computing power information;

[0187] 2: The base station calculates the loss value of the loss function based on the quantized data, channel state information, and computing power information, and determines the gradient of the input layer of the first neural network model based on the loss value;

[0188] 3: The base station sends the gradient of the input layer to the UE;

[0189] 4: The UE updates the parameters of the second neural network model and the

[0190] 5: The UE and the base station repeat the updating process of steps 1-4 for multiple rounds until the value of the loss function is minimized;

[0191] 6-1: The base station issues the optimal quantization scheme

[0192] …

[0193] 6-K: The base station issues the optimal quantization scheme to the UE K

[0194] 7: The UE completes the configuration of the local first quantizer.

[0195] It can be seen that in the embodiments of the present application, the second device can send the first data processed by the first quantizer to the first device, and at least one of the channel state information and the computing power information corresponding to the second device. Based on the first data, and at least one of the channel state information and the computing power information sent by the second device, the first device can determine the gradient of the input layer of the neural network model deployed by the first device, and send the gradient to the second device, and then the second device can calculate the gradient of the first quantizer based on the gradient, so that the second device can dynamically update at least one quantization parameter of the first quantizer, so that the quantization parameter of the first quantizer adapts to the dynamically changing computing power state and channel state, and thus the overhead of channel transmission is reduced as much as possible while ensuring the accuracy of end-edge collaborative AI tasks.

[0196] Please refer to Figure 4 , Figure 4 Another flowchart of a quantization parameter determination method provided by the embodiments of the present application is shown in Figure 4 , which includes steps 401-405:

[0197] 401: The second device sends the first data and the second data to the first device.

[0198] 402: The first device sends at least one second quantization parameter of the first quantizer to the second device.

[0199] In the embodiments of the present application, the at least one second quantization parameter is another part of the quantization parameters in the except the at least one first quantization parameter, and the at least one second quantization parameter is pre-configured by the first device. For example: the base station pre-configures and for the terminal device, and the base station issues and to the terminal device, and the terminal device configures and for the local first quantizer.

[0200] It should be noted that steps 401 and 402 do not have a time limit, and the first device can send at least one second quantization parameter to the second device before step 401.

[0201] Exemplarily, the first device may send at least one second quantization parameter of the first quantizer to the second device via second signaling, where the second signaling may be L3 layer signaling, such as RRC signaling.

[0202] 403: The first device determines fourth data based on the first data and the second data.

[0203] 404: The first device sends fourth data to the second device.

[0204] 405: The second device determines at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0205] For example, in the case where the first device has preconfigured at least one second quantization parameter for the second device, the second device only needs to update at least one parameter other than the second quantization parameter in each round of training, for example: and In this case, the second device only needs to update That's it.

[0206] Among them, the specific implementation of step 401 and steps 403-405 is as follows: Figure 3 The embodiments shown have been described in detail and can achieve the same or similar beneficial effects.

[0207] It can be seen that at least one second quantization parameter of the first quantizer can be pre-configured by the first device. The first device can send the configured at least one second quantization parameter to the second device, so the second device does not need to update the at least one second quantization parameter during the training process.

[0208] For better understanding Figure 4 In the embodiment shown in FIG. 1 , the parameter update mechanism of the quantizer is briefly described below by taking the end-edge cooperative interaction of K UEs and a base station as an example. Figure 5 As shown, the following steps may be included:

[0209] 1: UE reports quantized data, channel state information, and computing power information;

[0210] 2: Base station sends and

[0211] 3: The base station calculates the loss value of the loss function based on the quantized data, channel state information, and computing power information, and determines the gradient of the input layer of the first neural network model based on the loss value;

[0212] 4: The base station sends the gradient of the input layer to the UE;

[0213] 5: UE updates the first quantizer based on the gradient

[0214] 6: UE repeats the update procedure with the base station of steps 1-5 until the value of the loss function is minimized.

[0215] 7: The base station issues the optimal quantization scheme to the UE

[0216] 8: The UE completes the configuration of the local first quantizer.

[0217] It should be noted that the types of the at least one second quantization parameter preconfigured by the first quantizer of the plurality of second devices are different or the types of the at least one second quantization parameter preconfigured by the first quantizer of the plurality of second devices are the same. For example, the base station can preconfigure and for the K UEs. and for the UE1. and for the UE2.

[0218] In this implementation, the types of the at least one second quantization parameter preconfigured by the plurality of second devices can be the same or different, and the types of the quantization preconfigured parameters are relatively flexible.

[0219] Please refer to Figure 6 , Figure 6 Another flowchart of a quantization parameter determination method provided by an embodiment of the present application is shown in FIG. 6, and the method includes steps 601-605. Figure 6

[0220] 601: The second device sends first data and second data to the first device.

[0221] 602: The second device sends at least one second quantization parameter of the first quantizer to the first device.

[0222] In the embodiment of the present application, the at least one second quantization parameter is another part of the quantization parameters in except the at least one first quantization parameter, and the at least one second quantization parameter is preconfigured by the second device. For example, the terminal device preconfigures and to the base station for end-edge synchronization, and the quantization scheme finally configured by the base station for the terminal device can not include and and

[0223] ​​It should be noted that steps 601 and 602 are not limited in time sequence, and the second device can also send the at least one second quantization parameter to the first device before step 601.

[0224] 603: The first device determines fourth data based on the first data and the second data.

[0225] 604: The first device sends the fourth data to the second device.

[0226] 605: The second device determines at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0227] The specific implementation of steps 601, 603-605 is described in the embodiments shown in Figure 3 , and the same or similar beneficial effects can be achieved.

[0228] As can be seen, the at least one second quantization parameter of the first quantizer can be preconfigured by the second device, and the second device can send the preconfigured at least one second quantization parameter to the first device for synchronization of the quantization parameter.

[0229] To better understand the scheme in the embodiments shown in Figure 6 , the parameter updating mechanism of the quantizer is briefly described below with an example of end-edge collaborative interaction of K UEs and a base station. As shown in Figure 7 , it can include the following steps:

[0230] 1: The UE reports quantized data, channel state information and computing power information;

[0231] 2: The UE reports the preconfigured and

[0232] 3: The base station calculates the loss value of the loss function based on the quantized data, channel state information and computing power information, and determines the gradient of the input layer of the first neural network model based on the loss value;

[0233] 4: The base station sends the gradient of the input layer to the UE;

[0234] 5: The UE updates the

[0235] 6: The UE and the base station repeat the updating process of steps 1-5 for multiple rounds until the value of the loss function is minimized;

[0236] 7: The base station issues the optimal quantization scheme to the UE

[0237] 8: The UE completes the configuration of the local first quantizer.

[0238] It should be noted that the types of the at least one second quantization parameter pre-configured by the first quantizer on the side of the plurality of second devices are different or the types of the at least one second quantization parameter pre-configured by the first quantizer on the side of the plurality of second devices are the same.

[0239] In order to measure the influence of the quantization mechanism proposed in the embodiments of the present application on the performance of the AI model, the present application analyzes the loss function before and after introducing quantization. The approximate KL divergence after introducing quantization is The difference between the actual KL divergence D KL The theoretical upper bound of the difference is derived as follows:

[0240]

[0241] Where C and δ are constants related to the wireless channel, and d is the dimension of the intermediate quantity Based on the derivation, it can be concluded that the difference between the loss function before and after adding quantization is bounded, that is, the influence of the design of the quantization scheme on the performance of the end-edge AI model is bounded and controllable.

[0242] For the quantization mechanism provided in the embodiments of the present application, relevant personnel have tested it on the MNIST (Mixed National Institute of Standards and Technology database, a computer vision dataset) dataset and the CIFAR (Canadian Institute for Advanced Research, an image dataset widely used in the field of computer vision). Figure 8A - Figure 8C The test results on the MNIST dataset are shown in the following figure, where the dotted line 1 is the test result of the end-edge collaborative model based on the information bottleneck theory + dynamic quantization mechanism designed by the present scheme, the dotted line 2 is the test result of the end-edge collaborative model based on the information bottleneck theory (without adding the dynamic quantization mechanism), the dotted line 3 is the test result of the end-edge collaborative model based on the information bottleneck theory in the prior art + linear quantization scheme, and the dotted line 4 is the test result of the end-edge collaborative model based on the information bottleneck theory in the prior art (without adding the linear quantization scheme). Among them, Figure 8A represents the relationship between the inference error rate of the end-edge collaborative AI model and the data transmission delay of the wireless channel, Figure 8B represents the relationship between the inference error rate of the end-edge collaborative AI model and the number of end-side devices participating in collaboration, Figure 8C represents the relationship between the inference error rate of the end-edge collaborative AI model and the peak signal-to-noise ratio (PSNR). Figure 9A - Figure 9CFig. 1 is a test result on a CIFAR dataset, wherein the line 1 is a test result of an end-edge collaborative model based on the information bottleneck theory and a dynamic quantization mechanism, the line 2 is a test result of an end-edge collaborative model based on the information bottleneck theory (without adding a dynamic quantization mechanism), the line 3 is a test result of an end-edge collaborative model based on the information bottleneck theory and a linear quantization scheme in the prior art, and the line 4 is a test result of an end-edge collaborative model based on the information bottleneck theory (without adding a linear quantization scheme) in the prior art. Wherein, Figure 9A Fig. 2 shows a relationship between an inference error rate of an end-edge collaborative AI model and a wireless channel data transmission delay, Figure 9B Fig. 3 shows a relationship between an inference error rate of an end-edge collaborative AI model and a number of end-side devices participating in collaboration, Figure 9C Fig. 4 shows a relationship between an inference error rate of an end-edge collaborative AI model and a PSNR. It can be seen from the simulation result that the addition of quantization will cause a performance loss of the AI model, but the performance of the end-edge collaborative model generation scheme based on the DIB theory and the dynamic quantization mechanism is better than that of the best scheme without adding quantization in the prior art, which indicates that the design of the DIB+dynamic quantizer can still ensure the high-quality completion of the AI task, and the design of the dynamic updating quantizer does not affect the AI task performance of the original end-edge model designed based on the DIB theory. At the same time, the dynamic quantization mechanism is suitable for the future wireless network by adapting to the dynamically changing end-edge computing power state and wireless channel state.

[0243] The above describes the method of the embodiments of the present application, and the device of the embodiments of the present application is provided below.

[0244] Please refer to Figure 10 , Figure 10 Fig. 5 is a structural schematic diagram of a quantization parameter determination device provided by the embodiments of the present application. As shown in Figure 10 , the device includes a first transceiver unit 1001 and a first processing unit 1002. Wherein:

[0245] The first transceiver unit 1001 is configured to receive first data and second data sent by a second device; the first data is obtained by quantizing third data by a first quantizer through an mth round of configured quantization parameters; and the second data includes at least one of channel state information and computing power information.

[0246] The first processing unit 1002 is configured to determine fourth data based on the first data and the second data; and the fourth data is a gradient of an input layer of a first neural network model.

[0247] The first transceiver unit 1001 is further configured to send the fourth data to the second device; and the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in an (m+1)th round.

[0248] It can be seen that in Figure 10 In the device shown, it is possible to receive first data sent by a second device after quantization processing by a first quantizer, and at least one of the channel state information and computing power information corresponding to the second device. Based on the first data and at least one of the channel state information and computing power information sent by the second device, the device can determine the gradient of the input layer of the neural network model it deploys and send the gradient to the second device. The second device can then calculate the gradient of the first quantizer based on the gradient, so that the second device can dynamically update at least one quantization parameter of the first quantizer to adapt the quantization parameter of the first quantizer to the dynamically changing computing power state and channel state, thereby minimizing the channel transmission overhead while ensuring the accuracy of the end-edge collaborative AI task.

[0249] In one possible implementation, the third data is obtained by performing feature extraction on the fifth data using the second neural network model; in determining the fourth data based on the first data and the second data, the first processing unit 1002 is specifically configured to:

[0250] Obtaining an artificial intelligence (AI) task processing result of fifth data based on the first data;

[0251] Determine a loss value of a loss function using AI task processing results of the first data, the second data, and the fifth data;

[0252] The fourth data is determined based on the loss value of the loss function.

[0253] In a possible implementation, the at least one first quantization parameter includes part or all of the number of breakpoints, values ​​of the breakpoints, quantization values ​​at the breakpoints, and approximation factors of the first quantizer;

[0254] The breakpoint refers to the point where the first quantizer is divided into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

[0255] In a possible implementation, when the at least one first quantization parameter is the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and a portion of the approximation factor, the first transceiver unit 1001 is further configured to:

[0256] At least one second quantization parameter of the first quantizer is sent to the second device; the at least one second quantization parameter is the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and another part of the quantization parameters in the approximation factor except the at least one first quantization parameter; the at least one second quantization parameter is preconfigured by the first device.

[0257] In a possible implementation, when the at least one first quantization parameter is part of the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factors, the first transceiver 1001 is further configured to:

[0258] receive at least one second quantization parameter of the first quantizer sent by the second device; the at least one second quantization parameter is another part of the quantization parameters except the at least one first quantization parameter from the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factors; and the at least one second quantization parameter is preconfigured for the second device.

[0259] In a possible implementation, the second device is one or more, and each second device corresponds to a group of first data, a group of fifth data, and a group of second data; the AI task processing result of the fifth data is a probability that the first data corresponding to the one or more second devices is processed into a label of the fifth data; and in determining a loss value of a loss function by using the AI task processing result of the first data, the second data, and the fifth data, the first processing unit 1002 is specifically configured to:

[0260] obtain a KL divergence between a probability distribution of the first data corresponding to each second device and a reference probability distribution of the first data corresponding to each second device; wherein a dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device;

[0261] determine the loss value of the loss function based on the probability that the first data corresponding to the one or more second devices is processed into the label of the fifth data and the KL divergence corresponding to each second device.

[0262] In a possible implementation, the second device is multiple, and the types of the at least one second quantization parameter preconfigured for the first quantizer on the side of the multiple second devices are different or the types of the at least one second quantization parameter preconfigured for the first quantizer on the side of the multiple second devices are the same.

[0263] It should be noted that, Figure 10 the implementation of each unit described can also correspond to the respective description of the embodiments shown in Figures 3 to 7 , and, Figure 10 the beneficial effects brought by the described quantization parameter determination apparatus can refer to the respective description of the embodiments shown in Figures 3 to 7 , which will not be described here.

[0264] Please refer to Figure 11 , Figure 11 for another structure diagram of a quantization parameter determination apparatus provided by the embodiments of the present application. As Figure 11 shown, the apparatus includes a second transceiver 1101 and a second processing unit 1102. Wherein:

[0265] The second transceiver 1101 is configured to send first data and second data to the first device; the first data is obtained by quantizing the third data by the first quantizer through the mth round of configured quantization parameters; the second data includes at least one of channel state information and computing power information; the first data and the second data are used for the first device to determine fourth data; the fourth data is the gradient of the input layer of the first neural network model; and the fourth data sent by the first device is received.

[0266] The second processing unit 1102 is configured to determine at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0267] It can be seen that, in the device shown in Figure 11 In the device shown, the first data processed by the first quantizer can be sent to the first device, and at least one of the channel state information and the computing power information corresponding to the second device. Based on the first data and at least one of the channel state information and the computing power information sent by the second device, the first device can determine the gradient of the input layer of the neural network model deployed by the first device, and send the gradient to the second device. Then, the device can calculate the gradient of the first quantizer based on the gradient, so that the device can dynamically update at least one quantization parameter of the first quantizer, so that the quantization parameter of the first quantizer is adapted to the dynamically changing computing power state and channel state, thereby reducing the overhead of channel transmission as much as possible while ensuring the accuracy of end-edge collaborative AI tasks.

[0268] In a possible implementation, the at least one first quantization parameter includes part or all of the number of breakpoints of the first quantizer, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factor;

[0269] Wherein, the breakpoint refers to a point for dividing the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to represent the approximation degree of the continuous function and the sign function.

[0270] In a possible implementation, when the at least one first quantization parameter is part of the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factor, the second transceiver 1101 is further configured to:

[0271] Receive at least one second quantization parameter of the first quantizer sent by the first device; the at least one second quantization parameter is another part of the quantization parameter except the at least one first quantization parameter in the number of breakpoints, the values of the breakpoints, the quantization values at the breakpoints, and the approximation factor; and the at least one second quantization parameter is pre-configured by the first device.

[0272] In a possible implementation, when the at least one first quantization parameter is part of the number of breakpoints, the breakpoint values, the quantization values at the breakpoints, and the approximation factors, the second transceiver 1101 is further configured to:

[0273] send, to the first device, at least one second quantization parameter of the first quantizer; the at least one second quantization parameter is another part of the quantization parameters except the at least one first quantization parameter from the number of breakpoints, the breakpoint values, the quantization values at the breakpoints, and the approximation factors; and the at least one second quantization parameter is preconfigured for the second device.

[0274] In a possible implementation, the third data is obtained by performing feature extraction on the fifth data by using the second neural network model; and the fourth data is determined based on a loss value of a loss function, the loss value of the loss function is determined by using AI task processing results of the first data, the second data, and the fifth data, and the AI task processing result of the fifth data is obtained based on the first data.

[0275] In a possible implementation, the second device is multiple, and the types of the at least one second quantization parameter preconfigured for the first quantizer on the side of the multiple second devices are different or the types of the at least one second quantization parameter preconfigured for the first quantizer on the side of the multiple second devices are the same.

[0276] It should be noted that, Figure 11 the implementation of each unit described can also correspond to the Figures 3 to 7 respective description of the embodiments shown. Moreover, Figure 11 the beneficial effects brought by the described quantization parameter determination apparatus can refer to the Figures 3 to 7 respective description of the embodiments shown, which will not be described here.

[0277] Based on the description of the above method embodiments and apparatus embodiments, the embodiments of the present application further provide a communication device. Please refer to Figure 12 , Figure 12 for a structural schematic diagram of a communication device provided by the embodiments of the present application. The communication device at least includes a processor 1201, a memory 1202, and a communication interface 1203, and the processor 1201, the memory 1202, and the communication interface 1203 are connected to each other through a bus 1204. The communication device can be used to execute the related steps of the transmission block segmentation method. The communication device can be a sending end device or a chip in the sending end device in a wireless communication system, or can be a receiving end device or a chip in the receiving end device, such as a terminal device or a network device. The processor 1201 in the communication device is configured to read the computer program code stored in the above memory 1202, and execute the method of any one of the embodiments shown in Figures 3 to 7 .

[0278] The memory 1202 includes, but is not limited to, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read only memory (EPROM), or a compact disc read-only memory (CD-ROM), which is used to store relevant computer programs and data.

[0279] The processor 1201 can be one or more central processing units (CPUs). When the processor 1201 is a CPU, the CPU can be a single-core CPU or a multi-core CPU.

[0280] For example, when the communication device is the first device, the processor 1201 in the communication device can be configured to read one or more programs stored in the memory 1202, and perform the following operations:

[0281] receive first data and second data sent by the second device; the first data is obtained by quantizing the third data by using a quantization parameter configured in the mth round of the first quantizer; and the second data includes at least one of channel state information and computing power information;

[0282] determine fourth data based on the first data and the second data; the fourth data is a gradient of an input layer of the first neural network model;

[0283] send the fourth data to the second device; and the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in the (m+1)th round.

[0284] For example, when the communication device is the second device, the processor 1201 in the communication device can be configured to read one or more programs stored in the memory 1202, and perform the following operations:

[0285] send first data and second data to the first device; the first data is obtained by quantizing the third data by using a quantization parameter configured in the mth round of the first quantizer; and the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine fourth data; and the fourth data is a gradient of an input layer of the first neural network model;

[0286] receive fourth data sent by the first device;

[0287] determine at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

[0288] It should be noted that the implementation of each operation can also correspond to the description of the method of any one of the embodiments shown in Figures 3 to 7 .

[0289] It should be noted that although the communication device shown in Figure 12 only shows the processor 1201, the memory 1202, the communication interface 1203 and the bus 1204, in the specific implementation process, those skilled in the art should understand that the communication device also includes other devices necessary for normal operation. At the same time, according to the specific needs, those skilled in the art should understand that the communication device can also include hardware devices that realize other additional functions. In addition, those skilled in the art should understand that the communication device can also only include devices necessary for the implementation of the embodiments of the present application, and does not necessarily include all the devices shown in Figure 12 .

[0290] The embodiments of the present application also provide a chip, comprising: a processor, configured to call and run a computer program from a memory, so that a device installed with the chip executes the method described in any one of the above Figures 3 to 7 embodiments. The chip can be a chip in a communication device.

[0291] The embodiments of the present application also provide a computer readable storage medium (Memory), which stores a computer program, when the computer program is run, the method described in any one of the above Figures 3 to 7 embodiments is implemented. It can be understood that the computer readable storage medium here can include the built-in storage medium in the device, of course, it can also include the expansion storage medium supported by the device. The computer readable storage medium provides a storage space, which stores the operating system of the device. And in the storage space, one or more computer programs suitable for being loaded and executed by the processor of the device are also stored. It should be noted that the computer readable storage medium here can be a high-speed RAM, or a non-volatile memory, such as at least one disk memory; optionally, it can also be at least one computer readable storage medium located away from the aforementioned processor.

[0292] The embodiments of the present application also provide a computer program product, which comprises: computer program code, when the computer program code is run by a communication device, Figures 3 to 7 Figures 3 to 7 Figures 3 to 7 Figures 3 to 7 Figures 3 to 7 the method flow described in any one of the above embodiments is implemented.

[0293] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0294] It should be understood that the processor mentioned in the embodiments of the present application can be a CPU, and can also be other general-purpose processors, Digital Signal Processors (DSP), Application Specific Integrated Circuits (ASIC), Field Programmable Gate Arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0295] It should also be understood that the memory mentioned in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a ROM, a Programmable ROM (PROM), an EPROM, an Electrically EPROM (EEPROM) or a flash memory. The volatile memory can be a RAM used as an external cache. By way of example but not limitation, many forms of RAM can be used, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous Dynamic RAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM) and Direct Rambus RAM (DRRAM).

[0296] It should be noted that when the processor is a general-purpose processor, DSP, ASIC, FPGA or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, the memory (storage module) is integrated in the processor.

[0297] It should be noted that the memory described herein is intended to include, but not limited to, these and any other suitable types of memory.

[0298] It should be understood that in various embodiments of the present application, the size of the sequence number of each process described above does not mean the order of execution, the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0299] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the division of the above-described device embodiments is only illustrative, and for example, the division of the units can be changed, and some units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the display or discussion of the coupling or direct coupling or communication connection between the units can be through some interfaces, and can be indirect coupling or communication connection between the units, and can be electrical, mechanical or other forms.

[0300] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0301] In addition, each functional unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can be physically present separately, or two or more units can be integrated into one unit. The integrated unit, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium.

[0302] In the present application, "at least one" means one or more, and "multiple" means two or more. The "and / or" describes the association relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the following three cases: A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. In the textual description of the present application, the character " / ", generally indicates that the front and rear associated objects are in an "or" relationship.

[0303] The steps in the method embodiments of the present application can be adjusted, combined and deleted according to actual needs.

[0304] The modules in the device embodiments of the present application can be combined, divided and deleted according to actual needs.

[0305] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the foregoing embodiments of the present application have been described in detail, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for determining a quantization parameter, characterized in that: Applied to a first device; the method includes: receiving first data and second data sent by a second device; wherein the first data is obtained by quantizing the third data by the first quantizer using the quantization parameter configured in the mth round; and the second data includes at least one of channel state information and computing power information; Determine fourth data based on the first data and the second data; the fourth data is the gradient of the input layer of the first neural network model; The fourth data is sent to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in the (m+1)th round.

2. The method according to claim 1, characterized in that The third data is obtained by performing feature extraction on the fifth data using a second neural network model; The determining fourth data based on the first data and the second data includes: Obtaining an artificial intelligence (AI) task processing result of the fifth data based on the first data; Determine a loss value of a loss function using a result of AI task processing of the first data, the second data, and the fifth data; The fourth data is determined based on a loss value of the loss function.

3. The method according to claim 1 or 2, characterized in that The at least one first quantization parameter includes part or all of the number of breakpoints of the first quantizer, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and the approximation factor; Among them, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

4. The method according to claim 3, characterized in that When the at least one first quantization parameter is the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and a portion of the approximation factors, before determining the fourth data based on the first data and the second data, the method further includes: Send at least one second quantization parameter of the first quantizer to the second device; the at least one second quantization parameter is the number of the breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and another part of the quantization parameters in the approximation factor except the at least one first quantization parameter; the at least one second quantization parameter is preconfigured by the first device.

5. The method according to claim 3, characterized in that When the at least one first quantization parameter is the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and a portion of the approximation factors, before determining the fourth data based on the first data and the second data, the method further includes: Receive at least one second quantization parameter of the first quantizer sent by the second device; the at least one second quantization parameter is the number of the breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and another part of the quantization parameters in the approximation factor except the at least one first quantization parameter; the at least one second quantization parameter is pre-configured by the second device.

6. The method according to claim 2, characterized in that There are one or more second devices, each of which corresponds to a set of the first data, a set of the fifth data, and a set of the second data; the AI ​​task processing result of the fifth data is a probability that the first data corresponding to one or more second devices is processed as a label of the fifth data; The determining the loss value of the loss function by using the AI ​​task processing result of the first data, the second data, and the fifth data includes: Obtaining a KL divergence between a probability distribution of the first data corresponding to each second device and a reference probability distribution of the first data corresponding to each second device; wherein a dimension of the first data corresponding to each second device is determined by the second data corresponding to each second device; A loss value of the loss function is determined based on a probability that the first data corresponding to one or more second devices is processed as a label of the fifth data and the KL divergence corresponding to each second device.

7. The method according to claim 4 or 5, characterized in that There are multiple second devices, and the types of the at least one second quantization parameter pre-configured by the first quantizers on multiple second device sides are different or the types of the at least one second quantization parameter pre-configured by the first quantizers on multiple second device sides are the same.

8. A method for determining a quantization parameter, characterized in that: Applied to a second device; the method includes: Sending first data and second data to a first device; the first data is obtained by quantizing third data by a first quantizer using a quantization parameter configured in the mth round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine fourth data; the fourth data is a gradient of an input layer of a first neural network model; receiving the fourth data sent by the first device; At least one first quantization parameter of the first quantizer in the (m+1)th round is determined based on the fourth data.

9. The method according to claim 8, characterized in that The at least one first quantization parameter includes part or all of the number of breakpoints of the first quantizer, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and the approximation factor; Among them, the breakpoint refers to the point that divides the first quantizer into multiple value intervals; the first quantizer is obtained by relaxing the sign function in the second quantizer into a continuous function, and the approximation factor is used to characterize the degree of approximation between the continuous function and the sign function.

10. The method according to claim 9, characterized in that When the at least one first quantization parameter is the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and a portion of the approximation factors, before receiving the fourth data sent by the first device, the method further includes: Receive at least one second quantization parameter of the first quantizer sent by the first device; the at least one second quantization parameter is the number of the breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and another part of the quantization parameters in the approximation factor except the at least one first quantization parameter; the at least one second quantization parameter is preconfigured by the first device.

11. The method according to claim 9, characterized in that When the at least one first quantization parameter is the number of breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and a portion of the approximation factors, before receiving the fourth data sent by the first device, the method further includes: Send at least one second quantization parameter of the first quantizer to the first device; the at least one second quantization parameter is the number of the breakpoints, the values ​​of the breakpoints, the quantization values ​​at the breakpoints, and another part of the quantization parameters in the approximation factor except the at least one first quantization parameter; the at least one second quantization parameter is preconfigured by the second device.

12. The method according to any one of claims 8 to 11, characterized in that The third data is obtained by performing feature extraction on the fifth data through the second neural network model; the fourth data is determined based on the loss value of the loss function, and the loss value of the loss function is determined by the AI ​​task processing result of the first data, the second data and the fifth data, and the AI ​​task processing result of the fifth data is obtained based on the first data.

13. The method according to claim 10 or 11, characterized in that There are multiple second devices, and the types of the at least one second quantization parameter pre-configured by the first quantizers on multiple second device sides are different or the types of the at least one second quantization parameter pre-configured by the first quantizers on multiple second device sides are the same.

14. A device for determining a quantization parameter, characterized in that: Applied to a first device; the device comprises a first transceiver unit and a first processing unit; The first transceiver unit is configured to receive first data and second data sent by a second device; the first data is obtained by quantizing third data by the first quantizer using the quantization parameter configured in the mth round; and the second data includes at least one of channel state information and computing power information; The first processing unit is configured to determine fourth data based on the first data and the second data; the fourth data being a gradient of an input layer of the first neural network model; The first transceiver unit is further used to send the fourth data to the second device; the fourth data is used by the second device to determine at least one first quantization parameter of the first quantizer in the (m+1)th round.

15. A device for determining a quantization parameter, characterized in that: Applied to the second device, the apparatus comprises a second transceiver unit and a second processing unit The second transceiver unit is configured to send first data and second data to the first device; the first data is obtained by quantizing the third data by the first quantizer using the quantization parameter configured in the mth round; the second data includes at least one of channel state information and computing power information; the first data and the second data are used by the first device to determine fourth data; the fourth data is the gradient of the input layer of the first neural network model; and receive the fourth data sent by the first device; The second processing unit is configured to determine at least one first quantization parameter of the first quantizer in the (m+1)th round based on the fourth data.

16. A communication device, characterized in that: The method comprises a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to cooperate with the communication interface to implement the method as described in any one of claims 1-7 or 8-13 when executed by the processor.

17. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program for execution by a device, and when the computer program is executed, the method according to any one of claims 1 to 7 or 8 to 13 is implemented.

18. A computer program product, characterized in that When the computer program product is executed by a communication device, the communication device executes the method according to any one of claims 1 to 7 or 8 to 13.

Images