Communication method and apparatus

Abstract

The present application relates to the technical field of communications. Provided are a communication method and apparatus, which can improve the accuracy of an inference result, i.e., can obtain a better inference result. The method comprises: a receiving-side device receiving first information, and on the basis of the first information and a first inference model, obtaining an inference result, wherein the first information is compressed information, the first inference model is obtained by means of performing model training on the basis of a preconfigured model and a first module, and the first module is used for compressing and / or decompressing data.

Description

Communication methods and devices

[0001] This application claims priority to Chinese Patent Application No. 202411064900.2, filed with the State Intellectual Property Office of China on August 2, 2024, entitled "Communication Method and Apparatus", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of communication technology, and in particular to a communication method and apparatus. Background Technology

[0003] Model training typically employs two methods: one-sided training and two-sided training. In one-sided training, a single device (such as user equipment (UE) or base station (BS)) trains the model using training data. In two-sided training, two devices (such as UE and BS) train the model using the same training data. During inference, the data to be inferred can then be input into the trained model to obtain the inference result.

[0004] However, with the increasing maturity of technologies such as artificial intelligence (AI), autonomous driving, and scene reconstruction, models obtained through single-sided and double-sided training are currently unable to produce good inference results. Summary of the Invention

[0005] This application provides a communication method and apparatus that can improve the accuracy of reasoning results, i.e., obtain better reasoning results.

[0006] In a first aspect, embodiments of this application provide a communication method, which can be executed by a receiving device. Unless otherwise specified, "receiving device" in this application can refer to the receiving device itself, a component within the receiving device (e.g., a communication module, processor, circuit, chip, or chip system), or a logic module or software capable of implementing all or part of the functions of the receiving device. The method includes: receiving first information, wherein the first information is compressed information; and obtaining a reasoning result based on the first information and a first reasoning model, wherein the first reasoning model is obtained through model training based on a pre-configured model and a first module, and the first module is used to compress and / or decompress data.

[0007] Based on this scheme, after receiving compressed information (i.e., the first information) from the sending device, the receiving device can infer the first information using a first inference model trained on compressed and / or decompressed data (i.e., the first inference model is trained on a pre-configured model and a first module, where the first module is used to compress and / or decompress the data), and obtain the inference result. Since the first information is compressed data, inferring the data based on the first inference model trained on compressed and / or decompressed data yields better inference results and improves the accuracy of the inference results compared to inferring the data based on an inference model trained on uncompressed and / or undecompressed data.

[0008] In one possible design, the first information is obtained by compressing the intermediate features output by the second inference model; the first inference model is obtained by training the model based on the pre-configured model and the first module, including: the first inference model and the second inference model are obtained by training the first sub-model, the second sub-model and the first module in the pre-configured model, and the sum of the model layers of the first sub-model and the model layers of the second sub-model is equal to the model layers of the pre-configured model.

[0009] Based on this possible design, the models used for inference can be deployed on different devices. For example, the first inference model can be deployed on the receiving device, and the second inference model can be deployed on the sending device. Thus, the sending device can perform part of the inference task (i.e., obtain intermediate features), and the receiving device can complete the remaining inference (i.e., obtain the inference result). This reduces the computing power and resource consumption of the receiving device, allowing devices with lower computing power to participate in inference, thus broadening the applicability.

[0010] In one possible design, receiving the first information includes: receiving the first information from the transmitting device; before receiving the first information, the communication method further includes: receiving the first capability information, the first capability information indicating the inference capability of the transmitting device; determining the first inference model and / or the second inference model based on the first capability information; and sending the first indication information, the first indication information indicating the first inference model and / or the second inference model.

[0011] In one possible design, the method is executed by a receiving device. Before receiving the first information, the communication method further includes: acquiring second capability information, which indicates the inference capability of the receiving device; determining a first inference model and / or a second inference model based on the second capability information; and sending second indication information, which indicates the first inference model and / or the second inference model.

[0012] Based on the two possible designs mentioned above, the receiving device can determine the first inference model and / or the second inference model according to the inference capabilities of the sending device or the receiving device, and inform the sending device; this avoids the situation where the configured inference model exceeds the device's inference capabilities (e.g., the configured first inference model exceeds the receiving device's inference capabilities, or the configured second inference model exceeds the sending device's inference capabilities), which could lead to reduced accuracy of the inference results or even inference failure.

[0013] In one possible design, the reasoning result is obtained based on the first information and the first reasoning model, including: reasoning about the first information based on the first reasoning model to obtain the reasoning result.

[0014] Based on this possible design, after receiving the first information, the receiving device does not need to decompress the first information, but directly inputs it into the first inference model for inference to obtain the inference result; compared with the scheme of inferring the decompressed data, it can reduce the computing power of the first inference model and improve the inference efficiency.

[0015] In one possible design, the reasoning result is obtained based on the first information and the first reasoning model, including: decompressing the first information to obtain the data to be reasoned; and reasoning about the data to be reasoned based on the first reasoning model to obtain the reasoning result.

[0016] Based on this possible design, after receiving the first information, the receiving device decompresses the first information and inputs the decompressed data to be inferred into the first inference model for inference to obtain the inference result; compared with the scheme of inferring from undecompressed data, the accuracy of the inference result can be improved.

[0017] In one possible design, decompressing the first information to obtain the data to be inferred includes: using a first decompression module to decompress the first information to obtain the data to be inferred; the first module is related to the first decompression module.

[0018] In one possible design, the communication method further includes: obtaining configuration information, the configuration information indicating a first decompression method and / or a first decompression parameter, the first decompression method being the decompression method adopted by the first decompression module, and the first decompression parameter being the decompression parameter corresponding to the first decompression method.

[0019] Based on this possible design, during the decompression of the first information, the receiving device can perform decompression according to the decompression method and / or decompression parameters indicated by the configuration information; therefore, the sending end of the configuration information can flexibly set the first decompression method and / or the first decompression parameters, so that the receiving device can use different first decompression methods and / or first decompression parameters to achieve decompression of the first information.

[0020] In one possible design, obtaining configuration information includes: receiving configuration information.

[0021] In one possible design, obtaining configuration information includes: determining configuration information; the communication method also includes: sending configuration information.

[0022] In one possible design, the first module and the first decompression module use the same decompression method, and / or the type of the decompression parameter corresponding to the decompression method used by the first module is the same as the type of the first decompression parameter.

[0023] Based on the two possible designs mentioned above, the first module is related to the first decompression module. Therefore, the receiving device can also obtain the relevant parameters of the first decompression module (such as the first decompression method and / or the first decompression parameters) according to the relevant parameters of the first module, thereby saving the resource consumption used to carry configuration information.

[0024] Secondly, embodiments of this application provide a communication method, which can be executed by a transmitting device. Unless otherwise specified, "transmitting device" in this application can refer to the receiving device itself, a component within the transmitting device (e.g., a communication module, processor, circuit, chip, or chip system), or a logic module or software capable of implementing all or part of the transmitting device's functions. The method includes: inferring intermediate features from data to be transmitted based on a second inference model, wherein the second inference model is obtained through model training based on a first sub-model, a second sub-model, and a first module in a pre-configured model, the sum of the model layers of the first sub-model and the second sub-model equals the model layers of the pre-configured model, and the first module is used to compress and / or decompress data; compressing the intermediate data to obtain first information; and transmitting the first information.

[0025] Based on this scheme, the sending device can perform part of the inference task (i.e., obtain intermediate features), so that the receiving device can complete the remaining inference (i.e. obtain the inference result), thereby sharing the computing power and resource consumption of the receiving device, enabling devices with lower computing power to participate in inference, and making it more widely applicable.

[0026] In one possible design, before reasoning about the data to be transmitted according to the second inference model to obtain intermediate features, the communication method further includes: acquiring first capability information, which indicates the inference capability of the transmitting device; determining a first inference model and / or a second inference model based on the first capability information; and sending third indication information, which indicates the first inference model and / or the second inference model.

[0027] In one possible design, sending the first information includes: sending the first information to the receiving device; before reasoning about the data to be sent according to the second inference model to obtain intermediate features, the communication method further includes: receiving second capability information, the second capability information indicating the inference capability of the receiving device; determining the first inference model and / or the second inference model according to the second capability information; and sending fourth indication information, the fourth indication information indicating the first inference model and / or the second inference model.

[0028] In one possible design, compressing intermediate data to obtain first information includes: compressing intermediate data using a first compression module to obtain first information, wherein the first module is related to the first compression module.

[0029] In one possible design, the communication method further includes: obtaining configuration information, wherein the configuration information indicates a first compression method and / or a first compression parameter, wherein the first compression method is the compression method adopted by the first compression module, and the first compression parameter is the compression parameter corresponding to the first compression method.

[0030] In one possible design, obtaining configuration information includes: receiving configuration information.

[0031] In one possible design, obtaining configuration information includes: determining configuration information; the communication method also includes: sending configuration information.

[0032] In one possible design, the first module and the first compression module use the same compression method, and / or the compression parameter type corresponding to the compression method used by the first module is the same as the type of the first compression parameter.

[0033] The technical effects of any design in the second aspect can be referenced from the technical effects of the corresponding design in the first aspect, and will not be elaborated here.

[0034] In conjunction with the first or second aspect, in one possible design, the configuration information indicates a first decompression method and / or a first decompression parameter, including: the configuration information indicates a first compression method and / or a first compression parameter, wherein the first compression method corresponds to the first decompression method, and the first compression parameter corresponds to the first decompression parameter.

[0035] Based on this possible design, the configuration information can indicate the compression method that corresponds to the decompression method, thereby indicating the first decompression method; and / or, indicate the compression parameters corresponding to the compression method, thereby indicating the first decompression parameters, providing another possible implementation for the configuration information to indicate the first decompression method and / or the first decompression parameters.

[0036] In conjunction with the first or second aspect, in one possible design, when the configuration information indicates a first compression method, the configuration information includes a first index, the first compression method is the compression method corresponding to the first index in the first correspondence, and the first correspondence includes a correspondence between multiple compression methods and multiple indices.

[0037] In conjunction with the first or second aspect, in one possible design, when the configuration information indicates the first compression method, the configuration information includes a second index, the first compression method is the compression method corresponding to the second correspondence second index and the first data type, the first data type is the data type of the data to be inferred, and the second correspondence includes the correspondence between multiple data types, multiple compression methods and multiple indexes.

[0038] In conjunction with the first or second aspect, in one possible design, the configuration information also includes a first data type.

[0039] In conjunction with the first or second aspect, in one possible design, multiple compression methods include one or more of fixed quantization compression, 2D projection compression, dictionary compression, or sampling compression.

[0040] In conjunction with the first or second aspect, in one possible design, when the compression method is fixed quantization compression or 2D projection compression, the type of the compression parameter is the number of quantization bits; when the compression method is dictionary compression, the type of the compression parameter includes the weighting parameter corresponding to dictionary compression and / or the dimension of the dictionary matrix; when the compression method is sampling compression, the type of the compression parameter is the sampling rate and / or the sampling position.

[0041] Thirdly, a communication device is provided for implementing various methods. This communication device can be a receiving device in the first aspect, a transmitting device in the second aspect, or a device included in the receiving or transmitting device, such as a chip or chip system. The communication device includes modules, units, or means corresponding to the implementation of the methods, which can be implemented in hardware, software, or by hardware executing corresponding software. The hardware or software includes one or more modules or units corresponding to the functions.

[0042] In some possible designs, the communication device may include a processing module and a transceiver module. The processing module can be used to implement the processing functions in any of the above aspects and any possible implementations thereof. The transceiver module may include a receiving module and a transmitting module, respectively used to implement the receiving function and the transmitting function in any of the above aspects and any possible implementations thereof.

[0043] In some possible designs, the transceiver module can consist of transceiver circuits, transceivers, transceivers, or communication interfaces.

[0044] Fourthly, a communication device is provided, comprising: a processor and a memory; the memory is used to store computer instructions, which, when executed by the processor, cause the communication device to perform the method described in any aspect. The communication device may be a receiving device in the first aspect, a transmitting device in the second aspect, or a device included in the receiving or transmitting device, such as a chip or chip system. The communication device includes modules, units, or means corresponding to the implementation of the method, which may be implemented in hardware, software, or by hardware executing corresponding software. The hardware or software includes one or more modules or units corresponding to the function.

[0045] Fifthly, a communication device is provided, comprising: a processor and a communication interface; the communication interface being used to communicate with a module outside the communication device; the processor being used to execute computer programs or instructions to cause the communication device to perform the method described in any of the aspects. The communication device may be a receiving device in the first aspect, or a transmitting device in the second aspect, or a device included in the receiving or transmitting device, such as a chip or chip system. The communication device includes modules, units, or means corresponding to the implementation of the method, which may be implemented in hardware, software, or by hardware executing corresponding software. The hardware or software includes one or more modules or units corresponding to the functions.

[0046] A sixth aspect provides a communication device, comprising: at least one processor; the processor being configured to execute a computer program or instructions to cause the communication device to perform the method described in any aspect. The communication device may be a receiving device in the first aspect, or a transmitting device in the second aspect, or a device included in the receiving or transmitting device, such as a chip or chip system. The communication device includes modules, units, or means corresponding to the implementation of the method, which may be implemented in hardware, software, or by hardware executing corresponding software. The hardware or software includes one or more modules or units corresponding to the functions.

[0047] In some possible designs, the communication device includes a memory for storing necessary program instructions and data. This memory may be coupled to the processor, or it may be independent of the processor.

[0048] In some possible designs, when the device is a chip system, it can be composed of chips or contain chips and other discrete components.

[0049] It is understandable that when the communication device provided in any of the fifth to sixth aspects is a chip, the sending action / function of the communication device can be understood as outputting information, and the receiving action / function of the communication device can be understood as inputting information.

[0050] In a seventh aspect, a computer-readable storage medium is provided that stores a computer program or instructions that, when executed on a communication device, enable the communication device to perform the method described in any aspect.

[0051] In an eighth aspect, a computer program product containing instructions is provided, which, when run on a communication device, enables the communication device to perform the method described in either aspect.

[0052] A ninth aspect provides a communication system comprising a receiving device (or a means included in the receiving device, such as a chip or chip system) as described in the first aspect and a transmitting device (or a means included in the transmitting device, such as a chip or chip system) as described in the second aspect.

[0053] The technical effects of any of the design methods in aspects three through nine can be found in the technical effects of different design methods in aspects one or two above, and will not be repeated here. Attached Figure Description

[0054] Figure 1 is a schematic diagram of a neural network architecture provided in an embodiment of this application;

[0055] Figure 2 is a schematic diagram of a model training process provided in an embodiment of this application;

[0056] Figure 3 is a schematic diagram of a reasoning process provided in an embodiment of this application;

[0057] Figure 4 is a schematic diagram of the architecture of a communication system provided in an embodiment of this application;

[0058] Figure 5 is a schematic diagram of the architecture of another communication system provided in an embodiment of this application;

[0059] Figure 6 is a schematic diagram of the architecture of a communication device provided in an embodiment of this application;

[0060] Figure 7 is a flowchart illustrating a communication method provided in an embodiment of this application;

[0061] Figure 8 is a schematic diagram of another reasoning process provided in an embodiment of this application;

[0062] Figure 9 is a flowchart illustrating another communication method provided in an embodiment of this application;

[0063] Figure 10 is a flowchart illustrating another communication method provided in an embodiment of this application;

[0064] Figure 11 is a flowchart illustrating another communication method provided in an embodiment of this application;

[0065] Figure 12 is a schematic diagram illustrating the working principle of a compression method provided in an embodiment of this application;

[0066] Figure 13 is a schematic diagram illustrating the working principle of another compression method provided in an embodiment of this application;

[0067] Figure 14 is a schematic diagram illustrating the working principle of another compression method provided in an embodiment of this application.

[0068] Figure 15 is a schematic diagram of a model training process provided in an embodiment of this application;

[0069] Figure 16 is a schematic diagram of another model training process provided in an embodiment of this application;

[0070] Figure 17 is a schematic diagram of another model training process provided in an embodiment of this application;

[0071] Figure 18 is a schematic diagram of another model training process provided in an embodiment of this application;

[0072] Figure 19 is a schematic diagram of another model training process provided in an embodiment of this application;

[0073] Figure 20 is a schematic diagram of another model training process provided in an embodiment of this application;

[0074] Figure 21 is a schematic diagram of another model training process provided in an embodiment of this application;

[0075] Figure 22 is a schematic diagram of another model training process provided in an embodiment of this application;

[0076] Figure 23 is a schematic diagram of the architecture of another communication device provided in an embodiment of this application;

[0077] Figure 24 is a schematic diagram of the architecture of another communication device provided in an embodiment of this application. Detailed Implementation

[0078] In the description of this application, unless otherwise stated, " / " indicates that the objects before and after are in an "or" relationship. For example, A / B can mean A or B. "And / or" in this application is merely a description of the relationship between the related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. A and B can be singular or plural.

[0079] In the description of this application, unless otherwise stated, "multiple" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0080] Furthermore, to facilitate a clear description of the technical solutions in the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0081] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0082] It is understood that the term "embodiment" used throughout the specification means that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this application. Therefore, various embodiments throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It is understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0083] It is understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0084] It is understood that in this application, "...when" and "if" both refer to the corresponding processing that will be carried out under certain objective circumstances, and are not limited to a specific time, nor do they require a judgment action to be performed during implementation, nor do they imply any other limitations.

[0085] It is understood that some optional features in the embodiments of this application can be implemented independently in certain scenarios without relying on other features, such as the current solution on which they are based, to solve the corresponding technical problems and achieve the corresponding effects. Alternatively, they can be combined with other features as needed in certain scenarios. Correspondingly, the apparatus given in the embodiments of this application can also implement these features or functions, which will not be elaborated here.

[0086] It is understood that in this application, "instruction" can include direct and indirect instructions, as well as explicit and implicit instructions. When describing "a certain instruction information instructs A" or "instruction information of A," it can include whether the instruction information directly or indirectly instructs A, but does not necessarily mean that the instruction information carries A. The information indicated by a certain piece of information is called the information to be instructed. In the specific implementation process, there are many ways to instruct the information to be instructed, such as, but not limited to, directly instructing the information to be instructed, such as the information to be instructed itself or its index. It can also indirectly instruct the information to be instructed by instructing other information, where there is a relationship between the other information and the information to be instructed. It can also instruct only a part of the information to be instructed, while the other parts are known or pre-agreed upon. For example, the instruction of specific information can be achieved by using a pre-agreed (e.g., protocol-defined) arrangement of various information, thereby reducing instruction overhead to some extent. At the same time, the common parts of various information can be identified and uniformly indicated to reduce the instruction overhead caused by individually indicating the same information. Furthermore, the specific instruction method can also be any existing instruction method, such as, but not limited to, the above-mentioned instruction methods and their various combinations. As described above, for example, when multiple pieces of information of the same type need to be indicated, the indication methods for different pieces of information may differ. In specific implementation, the required indication method can be selected according to specific needs. This application embodiment does not limit the selected indication method; therefore, the indication methods involved in this application embodiment should be understood to cover various methods that enable the party to be indicated to obtain the information to be indicated. The information to be indicated can be sent as a whole or divided into multiple sub-information pieces and sent separately. Furthermore, the sending period or timing of these sub-information pieces can be the same or different. This application does not limit the specific sending method. The sending period or timing of these sub-information pieces can be predefined, for example, predefined according to a protocol, or configured by the transmitting device by sending configuration information to the receiving device.

[0087] In this application, "send" and "receive" indicate the direction of signal transmission. For example, "send information to XX" can be understood as the destination of the information being XX, which can include direct transmission via the air interface or indirect transmission via the air interface from other units or modules. "Receive information from YY" can be understood as the source of the information being YY, which can include direct reception from YY via the air interface or indirect reception from YY via the air interface from other units or modules. "Send" can also be understood as the "output" of a chip interface, and "receive" can also be understood as the "input" of a chip interface. In other words, sending and receiving can occur between devices, such as between network devices and terminal devices, or within a device, such as between components, modules, chips, software modules, or hardware modules within the device via buses, traces, or interfaces.

[0088] In this application, unless otherwise specified, the same or similar parts between the various embodiments can be referred to each other. Unless otherwise specified or logically conflicting, the terminology and / or descriptions between different embodiments are consistent and can be mutually referenced. Different embodiments can be combined to form new embodiments based on their inherent logical relationships. The following descriptions of the embodiments of this application do not constitute a limitation on the scope of protection of this application.

[0089] To facilitate understanding of the technical solutions in the embodiments of this application, a brief introduction to the relevant technologies of this application is given below:

[0090] Artificial intelligence (AI) and AI models:

[0091] Artificial intelligence (AI) has been widely applied in various fields. AI models are implemented using various algorithms, such as neural network models, decision trees, support vector machines, and Bayesian classifiers. This application uses a neural network model as an example for illustration, but it does not limit the specific type of AI model.

[0092] A neural network model is a mathematical or computational model that mimics the structure and function of biological neural networks (i.e., the central nervous system of animals, especially the brain). Figure 1 shows a schematic diagram of a neural network architecture. As shown in Figure 1, the architecture of a neural network mainly consists of neurons and the connections between them. A neuron is the most basic unit of a neural network; each circle in Figure 1 represents a neuron. Each neuron can connect with one or more other neurons to form a network. In other words, a neural network is a complex network formed by a large number of simple neurons extensively interconnected.

[0093] A neural network can include an input layer, an output layer, and hidden layers. Each neuron in the input layer receives various feature information from the sample data. That is, the input layer only receives information from the external environment; each neuron in the input layer acts as an independent variable, without performing any calculations, only transmitting information to the next layer. There can be at least one hidden layer between the input and output layers; that is, a neural network can have multiple hidden layers. Hidden layers are used to analyze information. The function used by each neuron in a hidden layer during computation connects the variables of the previous and next layers, making them more suitable for the data. Finally, the output layer generates the final result. For example, in a classification neural network, each neuron in the output layer corresponds to a specific classification.

[0094] Currently, common hidden layer operations include convolution, pooling, and fully connected layers. Hidden layers used to implement convolution operations are called convolutional layers, which are primarily used for feature extraction. Common convolution operations include 3x3 convolution and 5x5 convolution. Hidden layers used to implement pooling operations are called pooling layers, which are mainly used to compress features and simplify the computational complexity of neural networks. Hidden layers used to implement fully connected operations are called fully connected layers, which are primarily used to connect all features.

[0095] Model training is typically implemented using two methods: one-sided training and two-sided training. In one-sided training, a single device trains an initial AI model using training data to obtain a trained AI model. Specifically, the training data can be sent from another device to this device, or it can be generated by this device itself. Taking the example of training data being sent from another device to this device, as shown in Figure 2(a), device #1 can send training data to device #2, allowing device #2 to train the AI ​​model using the training data to obtain a trained AI model.

[0096] In the two-sided training process, two devices train the initial AI model using training data to obtain a trained AI model. Specifically, as shown in Figure 2(b), device #1 trains the first X layers (the first X layers refer to the first X hidden layers, and the initial model includes Y hidden layers) of the initial AI model using training data to obtain intermediate features. These intermediate features are then sent to device #2. Device #2 trains the last YX layers (the last YX layers refer to the last YX hidden layers) of the initial AI model using these intermediate features, thus obtaining the trained AI model. In other words, in the two-sided training process, the initial AI model can be divided into two parts: the first X layers of the initial AI model form one part, and the last YX layers form the other part. One of these two parts is deployed in device #1, and the other part is deployed in device #2.

[0097] In the model training process, after inputting training data into the initial AI model, the AI ​​model can output intermediate data. By comparing the training data with this intermediate data, the loss during training is determined. The AI ​​model can then be updated in reverse based on the loss (e.g., adjusting the AI ​​model based on the loss, and then training the adjusted AI model again using the training data). After multiple training iterations, a well-trained AI model is obtained. Then, during inference, the data to be inferred can be input into this well-trained AI model, which will then output the inference result.

[0098] However, with the increasing maturity of technologies such as AI, autonomous driving, and scene reconstruction, during the inference process, as shown in Figure 3, the sending device typically compresses the data to be inferred before sending it to the receiving device. The receiving device then decompresses the data and inputs the decompressed data into the trained AI model. In other words, the data input to the AI ​​model is compressed and decompressed. However, the training data does not undergo compression and / or decompression during model training. Consequently, the AI ​​model trained based on this process cannot produce satisfactory inference results, leading to reduced accuracy.

[0099] In view of this, embodiments of this application provide a communication method and apparatus. After receiving compressed information (i.e., first information) from a sending device, the receiving device can perform inference on the data to be inferred based on a first inference model trained using compressed and / or decompressed data (i.e., the first inference model is trained based on a pre-configured model and a first module, the first module being used to compress and / or decompress the data), and obtain an inference result.

[0100] Since the first information is data obtained after compression and decompression, the first inference model trained on the data to be inferred can produce better inference results and improve the accuracy of the inference results compared to the inference model trained on the data that has not undergone compression and / or decompression.

[0101] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0102] The communication method provided in this application embodiment can be used in any communication system, such as a third-generation partnership project (3GPP) communication system, for example, a long-term evolution (LTE) system; or a fifth-generation (5G) mobile communication system; a hybrid LTE and 5G network system; a new radio (NR) system; a vehicle-to-everything (V2X) system; a device-to-device (D2D) communication system; a machine-to-machine (M2M) communication system; an Internet of Things (IoT) system; a narrowband Internet of Things (NB-IoT) system; a global system for mobile communications (GSM); an enhanced data rate for GSM evolution (EDGE) system; a wideband code division multiple access (WCDMA) system; and a code division multiple access 2000 system. Access, CDMA2000, Time Division-Synchronization Code Division Multiple Access (TD-SCDMA), Enhanced Mobile Broadband (eMBB), Ultra-Reliable and Low-Latency Communication (URLLC), Enhanced Machine-Type Communication (eMTC), and various types of future communication systems, or non-terrestrial network (NTN) systems (such as satellite communication systems), non-3GPP communication systems, etc., are not restricted.

[0103] The communication method provided in this application can be applied to various communication scenarios, such as one or more of the following communication scenarios: radio frequency map (RF map), sensing, etc., without limitation.

[0104] The communication system provided in the embodiments of this application will be described below using Figure 4 as an example.

[0105] Figure 4 is a schematic diagram of a communication system provided in an embodiment of this application. As shown in Figure 4, the communication system may include at least one transmitting device and at least one receiving device.

[0106] The sending device can perform inference on the data to be inferred based on the trained model and send the inference results to the receiving device; correspondingly, the receiving device can process the inference results after receiving them.

[0107] For example, the sending device can be a terminal device, and the receiving device can be a network device. Alternatively, the sending device can be a terminal device, and the receiving device can be a terminal device. Alternatively, the sending device can be a network device, and the receiving device can be a network device. Alternatively, the sending device can be a network device, and the receiving device can be a terminal device.

[0108] Optionally, the network device in this application is a device that connects a terminal device to a wireless network. The network device can be a node in a wireless access network, also known as a base station, or a radio access network (RAN) node (or device).

[0109] For example, network equipment may include evolved base stations (NodeBs, eNBs, or e-NodeBs) in LTE systems or evolved LTE-A systems, such as traditional macro base stations (eNBs) and micro base stations (eNBs) in heterogeneous network scenarios. Alternatively, it may include transmission reception points (TRPs), home base stations (e.g., home evolved NodeBs, or home Node Bs, HNBs), base band units (BBUs), base band pools, or wireless fidelity (WiFi) access points (APs). Alternatively, it may include base stations in non-terrestrial networks (NTNs), i.e., those deployed on high-altitude platforms or satellites. In NTNs, network equipment may function as Layer 1 (L1) relays, base stations, distributed units (DUs), or integrated access and backhaul (IAB) nodes. Alternatively, it can be a gateway station or a ground station. Alternatively, the network device can be a device that implements base station functions in IoT, such as a device that implements base station functions in V2X, D2D, or machine-to-machine (M2M) communication. Alternatively, it can include in-vehicle devices or wearable devices. Alternatively, it can include network devices in 5G networks or public land mobile networks (PLMNs) that evolve from 5G. The embodiments of this application are not limited.

[0110] In some implementations, a network device can be understood as the network device itself, or a component in the network device (e.g., a communication device, communication module, processor, circuit, chip, or chip system), or it can be a logic module or software that can implement all or part of the functions of the network device.

[0111] In some embodiments, the network device may also include a communication module, circuit, or chip that performs the corresponding communication function. The network device may also be configured with program instructions for performing the corresponding communication function and corresponding program instructions. The network device in this application may also be a logical node, logical module, or software capable of implementing all or part of the functions of a network device.

[0112] In some possible scenarios, the network device in this application embodiment can also be a module or unit capable of implementing some functions of a base station. For example, the network device may include a centralized unit (CU) and a distributed unit (DU). This includes RAN equipment for CU and DU nodes that separates the protocol layer of the base station (gNB) in the NR system. Some protocol layer functions are centrally controlled by the CU, while the remaining partial or complete protocol layer functions are distributed in the DU, which is centrally controlled by the CU. Furthermore, the CU can be divided into a control plane (CU-CP) and a user plane (CU-UP). The CU-CP is responsible for control plane functions, mainly including radio resource control (RRC) and the corresponding packet data convergence protocol (PDCP) (PDCP-C). PDCP-C is mainly responsible for encryption / decryption, integrity protection, and data transmission of control plane data. CU-UP is responsible for user plane functions, mainly including the Service Data Adaptation Protocol (SDAP) and the corresponding Packet Data Convergence Protocol User (PDCP-U). SDAP is primarily responsible for processing core network data and mapping flows to bearers. PDCP-U is mainly responsible for data plane encryption / decryption, integrity protection, header compression, sequence number maintenance, and data transmission. CU-CP and CU-UP are connected via the E1 interface. CU-CP represents the gNB connecting to the core network via the NG interface and to the DU via the F1 interface control plane (F1-C). CU-UP connects to the DU via the F1 interface user plane (F1-U). Alternatively, PDCP-C may also be included in CU-UP.

[0113] It is understood that CU (including CU-CP or CU-UP) or DU may have different names in different systems, but those skilled in the art will understand their meaning. For example, in an open radio access network (O-RAN) system, CU can also be called an open centralized unit (O-CU), DU can also be called an open distributed unit (O-DU), CU-CP can also be called an open centralized unit-control plane (O-CU-CP), and CU-UP can also be called an open centralized unit user plane (O-CU-UP). For ease of description, this application uses CU, CU-CP, CU-UP, and DU as examples. Network devices may also include active antenna units (AAU). CU implements some of the functions of gNB, and DU implements some of the functions of gNB. For example, CU is responsible for handling non-real-time protocols and services, implementing the functions of the RRC layer. The DU (User Unit) is responsible for handling physical layer protocols and real-time services, implementing the functions of the radio link control (RLC), media access control (MAC), and physical (PHY) layers. In some deployments, the CU can also be divided into a centralized unit control plane (CU-CP) node and a centralized unit user plane (CU-UP) node. The CU-CP handles control plane functions, while the CU-UP handles user plane functions.

[0114] Optionally, the base station in this application embodiment may include various forms of base stations, such as: macro base station, micro base station (also known as small station), relay station, access point, home base station, TRP, transmission point (TP), mobile switching center, etc. This application embodiment does not specifically limit these.

[0115] Optionally, the terminal device in this application embodiment can be a user-side device used to implement wireless communication functions, such as a terminal or a chip that can be used in the terminal. The terminal can be a user equipment (UE), access terminal, satellite terminal, terminal unit, terminal station, mobile station, mobile station, remote station, remote terminal, mobile device, wireless communication device, terminal agent, or terminal apparatus in a 5G network or a PLMN evolved from 5G.Terminals can be cellular phones, cordless phones, session initiation protocol (SIP) phones, wireless local loop (WLL) stations, smartphones (such as mobile phones), personal digital assistants (PDAs), handheld devices with wireless communication capabilities, computing devices or other processing devices connected to a wireless modem, in-vehicle devices (such as cars, bicycles, electric vehicles, airplanes, ships, trains, high-speed trains, etc.) or wearable devices (such as smartwatches, smart bracelets, pedometers, smart glasses, etc.), intelligent robots, robotic arms, workshop equipment, virtual reality (VR) terminal devices, augmented reality (AR) terminal devices, wireless terminals in industrial control, wireless terminals in self-driving, wireless terminals in telemedicine or telehealth services, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, and wireless terminals in smart homes. Wireless terminals in the home (e.g., refrigerators, televisions, air conditioners, electricity meters, etc.), wireless data cards, tablet computers, laptops, handheld computers, mobile internet devices (MID), wireless modems, handsets, laptop computers, machine type communication (MTC) terminals, point of sale (POS) machines, customer-premises equipment (CPE), light user equipment (light UE), reduced capability user equipment (REDCAP UE), and flying equipment (e.g., intelligent robots, hot air balloons, drones, airplanes), etc.Alternatively, a terminal can be a communication-enabled terminal (or a device that functions as a terminal) in the Internet of Things (IoT), such as a terminal in a vehicle-to-everything (V2X) system (i.e., a vehicle device, such as a complete vehicle unit, onboard module, onboard chip, onboard unit (OBU), or telematics box (T-BOX, etc.), a terminal in a device-to-device (D2D) system, or a terminal in a machine-to-machine (M2M) communication system. Terminals can be mobile or fixed.

[0116] In some embodiments, the terminal device may also be a device or module that is connected to the communication system shown above and has corresponding communication functions. The terminal device typically includes a communication module that performs the corresponding communication functions, or a chip responsible for communication functions within the terminal device, such as a modem chip (also known as a baseband chip), or a system-on-chip (SoC) chip or system-in-a-package (SIP) chip containing a modem module. The terminal device also contains program instructions for performing the corresponding communication functions.

[0117] In some embodiments, the terminal device may also be a device or module that is connected to the communication system shown above and has corresponding communication functions. The terminal device typically contains a communication module, circuit, or chip that performs the corresponding communication functions, and the terminal device is also configured with program instructions for performing the corresponding communication functions.

[0118] Optionally, the roles of network devices and terminal devices can be relative. For example, in Figure 5, terminal devices #9 and #10, since terminal device #10 needs to access network device #1 through terminal device #9, terminal device #9 can be configured as a network device relative to terminal device #10; while relative to network device #1, terminal device #9 is a terminal device. That is, network device #1 and terminal device #9 communicate through a wireless air interface protocol. Optionally, network device #1 and terminal device #9 can also communicate through a network device-to-network device interface protocol. In this case, terminal device #9 also acts as a network device relative to network device #1.

[0119] Optionally, communication between network devices and terminal devices, between network devices, or between terminal devices can be conducted using licensed spectrum, unlicensed spectrum, or both. Alternatively, communication between network devices and terminal devices, between network devices, or between terminal devices can be conducted using spectrum below 6 GHz, spectrum above 6 GHz, or both. The embodiments of this application do not limit the spectrum resources used for wireless communication.

[0120] In the embodiments of this application, the functions of the network device can also be executed by modules (such as chips) within the network device, or by a control subsystem that includes network device functions. The control subsystem that includes network device functions can be a control center in the aforementioned application scenarios such as smart grids, industrial control, intelligent transportation, and smart cities. Similarly, the functions of the terminal device can be executed by modules (such as chips, modems, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or software (such as program code in memory), or by a device that includes terminal device functions; there are no limitations on this.

[0121] In specific implementation, as shown in Figure 4, each communication device (such as a transmitting device and a receiving device) can adopt the composition structure shown in Figure 6, or include the components shown in Figure 6. Figure 6 is a schematic diagram of the composition of a communication device 600 provided in an embodiment of this application. The communication device 600 can be a transmitting device or a chip or system-on-a-chip in the transmitting device; it can also be a receiving device or a chip or system-on-a-chip in the receiving device. As shown in Figure 6, the communication device 600 includes a processor 601, a communication interface 602, and a communication line 603.

[0122] Furthermore, the communication device 600 may also include a memory 604. The processor 601, memory 604, and communication interface 602 can be connected via a communication line 603.

[0123] The processor 601 can be a central processing unit (CPU), a network processor (NP), a digital signal processor (DSP), a microprocessor, a microcontroller, a programmable logic device (PLD), or any combination thereof. The processor 601 can also be other devices with processing capabilities, such as circuits, devices, or software modules, without limitation.

[0124] Communication interface 602 is used to communicate with other devices or other communication networks. These other communication networks can be Ethernet, radio access network (RAN), wireless local area network (WLAN), etc. Communication interface 602 can be a module, circuit, transceiver, or any device capable of enabling communication.

[0125] Communication line 603 is used to connect different components in communication device 600, enabling communication between them. Communication line 603 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used in Figure 6, but this does not indicate that there is only one bus or one type of bus.

[0126] The memory 604 may be a device with storage function for storing instructions and / or data. The instructions may be computer programs.

[0127] For example, memory 604 may be read-only memory (ROM) or other types of static storage devices that can store static information and / or instructions; it may also be random access memory (RAM) or other types of dynamic storage devices that can store information and / or instructions; it may also be electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, etc., without limitation.

[0128] It should be noted that the memory 604 can exist independently of the processor 601, or it can be integrated with the processor 601. The memory 604 can be used to store instructions, program code, or some data, etc. The memory 604 can be located inside or outside the communication device 600, without limitation. The processor 601 is used to execute the instructions stored in the memory 604 to implement the communication method provided in the following embodiments of this application.

[0129] In one example, processor 601 may include one or more CPUs, such as CPU0 and CPU1 in Figure 6.

[0130] As an optional implementation, the communication device 600 may include multiple processors, for example, in addition to the processor 601 in FIG. 6, it may also include a processor 607.

[0131] As an optional implementation, the communication device 600 also includes an output device 605 and an input device 606. Exemplarily, the input device 606 is a liquid crystal display (LCD), a light-emitting diode (LED) display device, a cathode ray tube (CRT) display device, or a projector, etc. For example, the input device 606 can be a keyboard, mouse, microphone, joystick, touchscreen device, or sensing device, etc. The output device 605 is a display screen, a speaker, etc.

[0132] It should be noted that the communication device 600 can be a desktop computer, a portable computer, a web server, a mobile phone, a tablet computer, a wireless terminal, an embedded device, a chip system, or a device with a similar structure to that shown in Figure 6. Furthermore, the composition shown in Figure 6 does not constitute a limitation on the communication device. In addition to the components shown in Figure 6, the communication device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0133] In this embodiment of the application, the chip system may be composed of chips or may include chips and other discrete devices.

[0134] Furthermore, the actions, terms, etc., involved in the various embodiments of this application can be referenced interchangeably without limitation. The message names or parameter names in the messages exchanged between the various devices in the embodiments of this application are merely examples, and other names may be used in specific implementations without limitation.

[0135] The communication method provided in the embodiments of this application is described below with reference to the accompanying drawings. It is understood that in the embodiments of this application, the sending end device or the receiving end device can perform some or all of the steps in the embodiments of this application. These steps or operations are merely examples, and the embodiments of this application can also perform other operations or variations of various operations. Furthermore, the steps can be performed in different orders as presented in the embodiments of this application, and it is not necessary to perform all the operations in the embodiments of this application.

[0136] Referring to Figure 7, a flowchart of a communication method provided in this application is shown. The communication method includes the following steps S701 to S702:

[0137] S701, the transmitting device sends first information to the receiving device; correspondingly, the receiving device receives the first information from the transmitting device. The first information is compressed information.

[0138] For example, the sending device can compress the data to be sent to the receiving device before sending the data to be sent to the receiving device to obtain first information, and then send the first information to the receiving device; thereby reducing the resource consumption of the air interface and improving the efficiency of air interface transmission.

[0139] S702, the receiving device obtains the inference result based on the first information and the first inference model. The first inference model is obtained through model training based on a pre-configured model and a first module, and the first module is used to compress and / or decompress data.

[0140] As an example, the receiving device obtains a reasoning result based on the first information and the first reasoning model, including: the receiving device reasoning about the first information based on the first reasoning model to obtain the reasoning result.

[0141] In other words, after receiving the first information, the receiving device does not need to decompress the first information, but directly inputs it into the first inference model for inference to obtain the inference result.

[0142] Based on this example, after receiving the first information, the receiving device does not need to decompress the first information, but directly inputs it into the first inference model for inference to obtain the inference result. Compared with the scheme of inferring from the decompressed data, this can reduce the computing power of the first inference model and improve the inference efficiency.

[0143] As another example, the receiving device obtains a reasoning result based on the first information and the first reasoning model, including: the receiving device decompressing the first information to obtain the data to be reasoned; and the receiving device reasoning about the data to be reasoned based on the first reasoning model to obtain the reasoning result.

[0144] For example, since the first information is compressed, the receiving device can recover it by decompressing it after receiving the first information, thereby obtaining the data to be inferred.

[0145] Optionally, the module used to decompress the first information can be called the first decompression module; therefore, the receiving device decompresses the first information to obtain the data to be inferred, including: the receiving device uses the first decompression module to decompress the first information to obtain the data to be inferred.

[0146] Based on this example, after receiving the first information, the receiving device decompresses the first information and inputs the decompressed data to be inferred into the first inference model for inference to obtain the inference result; compared with the scheme of inferring from undecompressed data, the accuracy of the inference result can be improved.

[0147] Combining the two examples above, the first inference model is exemplarily obtained through model training based on the pre-configured model and the first module. This can be understood as: adding the first module during model training allows it to be trained together with the pre-configured model, resulting in the trained model, i.e., the first inference model. Furthermore, the first module is used to compress and / or decompress data. This can be understood as: the first module is used to compress data during model training, or the first module is used to decompress data during model training, or the first module is used to both compress and decompress data during model training.

[0148] This application provides a communication method in which, after receiving compressed information (i.e., first information) from a sending device, the receiving device can perform inference on the data to be inferred based on a first inference model trained using compressed and / or decompressed data (i.e., the first inference model is trained based on a pre-configured model and a first module, where the first module is used to compress and / or decompress the data). Since the data to be inferred is data obtained after compression and decompression, inferring based on the first inference model trained on compressed and / or decompressed data yields better inference results and improves the accuracy of the inference results compared to inferring based on an inference model trained on uncompressed and / or undecompressed data.

[0149] The above is a general description of the communication method provided in this application. The following sections describe steps S701 to S702 respectively:

[0150] (I) Regarding step S701:

[0151] As an example, the first information can be data that has not undergone reasoning. That is, after the sending device obtains the data to be sent, it can compress the data to be sent to obtain the first information, and then send the first information to the receiving device.

[0152] For example, the data to be sent can be raw data, or any data that has not undergone reasoning; this application does not limit this.

[0153] For example, in this example, the inference process of the data to be sent can be as shown in Figure 8(a), that is, the sending device performs compression, and the receiving device performs decompression and inference.

[0154] Optionally, the sending device compresses the data to be sent to obtain first information, including: the sending device compresses the data to be sent using a first compression module to obtain first information.

[0155] For example, the compression method used by the first compression module is the first compression method, and the compression parameters corresponding to the first compression method are the first compression parameters. Specifically, the implementation of the first compression method and the first compression parameters can be found in the relevant descriptions of the following embodiments, and will not be repeated here.

[0156] Based on this example, inference is performed by a single device (i.e., the inference process is performed by the receiving device). Compared to the process where multiple devices perform inference, this reduces the interaction between multiple devices involved in inference, saves resources used by multiple devices for interaction, and improves the efficiency of inference.

[0157] As another example, the first piece of information could be data that has undergone reasoning, meaning that the sending device has also reasoned about the data to be sent.

[0158] For example, in this case, inference is performed by two devices (i.e., both the sending device and the receiving device perform inference). The sending device uses a second inference model, while the receiving device uses a first inference model.

[0159] At this point, both the first and second inference models are trained based on the first sub-model, the second sub-model, and the first module. The pre-configured model includes the first and second sub-models, and the sum of the model layers of the first and second sub-models equals the model layers of the pre-configured model. In other words, the first inference model is trained based on the pre-configured model and the first module, and includes: the first and second inference models are trained based on the first and second sub-models in the pre-configured model, as well as the first module.

[0160] For example, the number of model layers in this application refers to the number of hidden layers. That is, the sum of the number of hidden layers in the first sub-model and the number of hidden layers in the second sub-model equals the number of hidden layers in the pre-configured model. In this case, the pre-configured model can also be considered as being split into a first sub-model and a second sub-model. For example, if the pre-configured model has Y layers, the second sub-model can include the first X layers of the pre-configured model, and the first sub-model can include the last YX layers of the pre-configured model. Here, Y and X are both positive integers greater than 0.

[0161] Optionally, since the first information is compressed, and given the characteristics of air interface transmission between communication devices, compression can be performed before data transmission. Therefore, the inference process for the data to be transmitted can be as shown in Figure 8(b), whereby the sending device, after outputting intermediate features based on the second inference model, can compress these intermediate features to obtain the first information, and then send the first information to the receiving device. In other words, the first information is obtained by compressing the intermediate features output by the second inference model.

[0162] For example, when the reasoning process of the data to be sent is as shown in Figure 8(b), before step S701, as shown in Figure 9, the communication method may also include steps S703 to S704:

[0163] S703. The transmitting device infers the data to be transmitted based on the second inference model to obtain intermediate features.

[0164] S704. The transmitting device compresses the intermediate features to obtain the first information.

[0165] Optionally, the sending device compresses the intermediate features to obtain the first information, including: the sending device compresses the intermediate features using a first compression module to obtain the first information.

[0166] For example, the compression method used by the first compression module is the first compression method, and the compression parameters corresponding to the first compression method are the first compression parameters. Specifically, the implementation of the first compression method and the first compression parameters can be found in the relevant descriptions of the following embodiments, and will not be repeated here.

[0167] Based on this example, inference is performed by multiple devices (i.e., the inference process is performed by the sending and receiving devices). Compared to the process where inference is performed by a single device, the computing power required for each device to perform inference is reduced, meaning that devices with lower computing power can also participate in the inference process of multiple devices, making it more widely applicable.

[0168] Combining the two examples above, optionally, the data to be sent can be generated by the sending device itself; or, it can be obtained by the sending device from another device (i.e., the other device sends the data to be sent to the sending device); or, it can be obtained by the sending device in any possible way, which is not limited in this application.

[0169] For example, the data to be sent can be compressed data or uncompressed data. Specifically, when the data to be sent is compressed data, one possible implementation is that the data to be sent received by the sending device is the compressed data; another implementation is that the sending device compresses the uncompressed data to generate the data to be sent.

[0170] Optionally, the sum of the number of model layers in the first inference model and the number of model layers in the second inference model is equal to the number of model layers in the pre-configured model.

[0171] Optionally, the first inference model and / or the second inference model may be determined and communicated to the other party by the transmitting device, or may be determined and communicated to the other party by the receiving device, or may be indicated by other devices besides the transmitting device or the receiving device that have the function of determining the thrust model. This application does not limit this.

[0172] As one possible implementation, when the first inference model and / or the second inference model are determined by the receiving device, the receiving device can determine the number of model layers it can support based on the inference capabilities of the sending device or the receiving device, and then determine and inform the sending device of the first inference model and / or the second inference model.

[0173] For example, the inference capability of a transmitting device may include its currently idle transmission resources and / or current computing power. Similarly, the inference capability of a receiving device may also include its currently idle transmission resources and / or current computing power.

[0174] As an example, when the receiving device determines the number of model layers it can support based on its own inference capabilities, and then determines and informs the sending device of the first inference model and / or the second inference model, before step S703, as shown in Figure 10(a), the communication method may include the following steps S705 to S707:

[0175] S705. The transmitting device sends first capability information to the receiving device, and correspondingly, the receiving device receives the first capability information from the transmitting device. The first capability information indicates the inference capability of the transmitting device.

[0176] Optionally, the inference capability of the sending device can refer to the resources and / or computing power that the sending device can use for inference, or in other words, the ability of the sending device to support inference.

[0177] For example, the first capability information may indicate the currently idle transmission resources and / or current computing power of the transmitting device, or the first capability information may indicate the number of model layers that the transmitting device can currently support. The number of model layers that the transmitting device can currently support is determined based on the currently idle transmission resources and / or current computing power of the transmitting device. Specifically, the currently idle transmission resources and / or current computing power of the transmitting device are positively correlated with the number of model layers that the transmitting device can currently support.

[0178] S706. The receiving device determines a first inference model and / or a second inference model based on the first capability information.

[0179] For example, the receiving device can determine the number of model layers that the sending device can currently support based on the first capability information, thereby determining the number of model layers in the second inference model. The number of model layers in the second inference model is less than or equal to the number of model layers that the sending device can currently support. Furthermore, since the sum of the number of model layers in the first inference model and the second inference model equals the number of model layers in the pre-configured model, the number of model layers in the first inference model can be further determined based on the number of model layers in the second inference model and the number of model layers in the pre-configured model.

[0180] S707, the receiving device sends first indication information to the sending device; correspondingly, the sending device receives the first indication information from the receiving device. The first indication information indicates a first inference model and / or a second inference model.

[0181] For example, the first indication information may include the structure and / or parameters of the first inference model, and / or the structure and / or parameters of the second inference model. For instance, the parameters of the first inference model and / or the second inference model may include the number of model layers of the first inference model and / or the model parameters used to constitute the first inference model and / or the second inference model.

[0182] Alternatively, the first indication information may also indicate the reference model, and the model layer number of the first inference model and / or the second inference model; thus, the first inference model is layer X of the reference model, where X is the model layer number of the first inference model, indicating which layer X of the reference model the first inference model belongs to. Similarly, the second inference model is layer Y of the reference model, where Y is the model layer number of the second inference model, indicating which layer Y of the reference model the second inference model belongs to. X and Y are both positive integers.

[0183] For example, the sending device and the receiving device can pre-agree on multiple reference models, so that one of the models can be indicated by the first indication information to determine the first inference model and / or the second inference model. Specifically, the multiple reference models can be determined by the sending device and informed to the receiving device, or they can be determined by the receiving device and informed to the sending device, or they can be predefined by the protocol.

[0184] Specifically, the receiving device can inform the sending device of its determined first inference model and / or second inference model, ensuring that both ends are aware of the first and / or second inference models. Specifically, when the first indication information indicates the second inference model, the sending device can determine the second inference model based on the first indication information; when the first indication information indicates the first inference model, after receiving the first indication information, the sending device can determine the second inference model based on the first inference model and a pre-configured model. For example, the number of model layers in the second inference model can be determined based on the number of model layers in the first inference model and the number of model layers in the pre-configured model.

[0185] As another example, when the receiving device determines the number of model layers it can support based on its own inference capabilities, and then determines and informs the sending device of the first inference model and / or the second inference model, before step S703, as shown in Figure 10(b), the communication method may include the following steps S708 to S710:

[0186] S708. The receiving device acquires the second capability information. The second capability information indicates the inference capability of the receiving device.

[0187] Optionally, the inference capability of the receiving device can refer to the resources and / or computing power that the receiving device can use for inference, or in other words, the ability of the receiving device to support inference.

[0188] For example, the second capability information may indicate the currently available transmission resources and / or current computing power of the receiving device, or it may indicate the number of model layers that the receiving device can currently support. The number of model layers that the receiving device can currently support is determined based on its currently available transmission resources and / or current computing power. Specifically, the currently available transmission resources and / or current computing power of the receiving device are positively correlated with the number of model layers that the receiving device can currently support.

[0189] S709. The receiving device determines the first inference model and / or the second inference model based on the second capability information.

[0190] For example, the receiving device can determine the number of model layers it can currently support based on the second capability information, thereby determining the number of model layers in the first inference model. The number of model layers in the first inference model is less than or equal to the number of model layers the receiving device can currently support. Furthermore, since the sum of the number of model layers in the first inference model and the second inference model equals the number of model layers in the pre-configured model, the number of model layers in the second inference model can be further determined based on the number of model layers in the first inference model and the number of model layers in the pre-configured model.

[0191] S710, the receiving device sends second indication information to the sending device, and correspondingly, the sending device receives the second indication information from the receiving device. The second indication information indicates the first inference model and / or the second inference model.

[0192] Specifically, the second indication information may include the structure and / or parameters of the first inference model, and / or the structure and / or parameters of the second inference model. For example, the parameters of the first inference model and / or the second inference model may include the number of model layers of the first inference model and / or the model parameters used to constitute the first inference model and / or the second inference model.

[0193] Alternatively, the second indication information may also indicate the reference model and the model layer number of the first inference model and / or the second inference model; thus, the first inference model is layer X of the reference model, where X is the model layer number of the first inference model, indicating which layer X of the reference model the first inference model belongs to. Similarly, the second inference model is layer Y of the reference model, where Y is the model layer number of the second inference model, indicating which layer Y of the reference model the second inference model belongs to. X and Y are both positive integers. Specifically, the implementation of the reference model can be found in the relevant description of the above embodiments, and will not be repeated here.

[0194] For example, the receiving device can inform the sending device of its determined first inference model and / or second inference model, ensuring that both ends are aware of the first and / or second inference models. Specifically, when the second indication information indicates the second inference model, the receiving device can determine the second inference model based on the second indication information; when the second indication information indicates the first inference model, the receiving device, after receiving the second indication information, can determine the second inference model based on the first inference model and the pre-configured model. For example, the number of model layers in the second inference model can be determined based on the number of model layers in the first inference model and the number of model layers in the pre-configured model.

[0195] As a first possible implementation, when the first inference model and / or the second inference model are determined by the transmitting device, the transmitting device can determine the number of model layers it can support based on the inference capabilities of the transmitting device or the receiving device, and then determine and inform the receiving device of the first inference model and / or the second inference model.

[0196] For example, the inference capability of a transmitting device may include its currently idle transmission resources and / or current computing power. Similarly, the inference capability of a receiving device may also include its currently idle transmission resources and / or current computing power.

[0197] As an example, when the sending device determines the number of model layers it can support based on its own inference capabilities, and then determines and informs the receiving device of the first inference model and / or the second inference model, before step S703, as shown in Figure 10(c), the communication method may include the following steps S711 to S713:

[0198] S711. The transmitting device acquires first capability information. The first capability information indicates the inference capability of the transmitting device.

[0199] For example, the implementation of the first capability information can be found in the relevant description in step S705 above, and will not be repeated here.

[0200] S712. The transmitting device determines a first inference model and / or a second inference model based on the first capability information.

[0201] For example, the sending device determines the first inference model and / or the second inference model in a manner similar to the receiving device determining the first inference model and / or the second inference model in step S706 above. For details, please refer to the relevant description of step S706 above, which will not be repeated here.

[0202] S713. The transmitting device sends third indication information to the receiving device; correspondingly, the receiving device receives the third indication information from the transmitting device. The third indication information indicates the first inference model and / or the second inference model.

[0203] For example, the implementation of the third instruction information is similar to the implementation of the first instruction information in step S708 above. For details, please refer to the relevant description of the first instruction information above, which will not be repeated here.

[0204] For example, the sending device can inform the receiving device of its determined first inference model and / or second inference model, ensuring that both ends are aware of the first and / or second inference models. Specifically, when the third indication information indicates the first inference model, the receiving device can obtain the first inference model based on the third indication information; when the third indication information indicates the second inference model, after receiving the third indication information, the receiving device can determine the first inference model based on the second inference model and the pre-configured model. For example, the number of model layers in the first inference model can be determined based on the number of model layers in the second inference model and the number of model layers in the pre-configured model.

[0205] As another example, when the sending device determines the number of model layers it can support based on the inference capability of the receiving device, and then determines and informs the receiving device of the first inference model and / or the second inference model, before step S703, as shown in (d) of FIG10, the communication method may include the following steps S714 to S716:

[0206] S714. The receiving device sends second capability information to the sending device; correspondingly, the sending device receives the second capability information from the receiving device. The second capability information indicates the inference capability of the receiving device.

[0207] For example, the implementation of the second capability information can be found in the relevant description in step S708 above, and will not be repeated here.

[0208] S715. The transmitting device determines the first inference model and / or the second inference model based on the second capability information.

[0209] For example, the sending device determines the first inference model and / or the second inference model in a manner similar to the receiving device determining the first inference model and / or the second inference model in step S709 above. For details, please refer to the relevant description in step S708 above, which will not be repeated here.

[0210] S716. The transmitting device sends a fourth indication information to the receiving device; correspondingly, the receiving device receives the fourth indication information from the transmitting device. The fourth indication information indicates the first inference model and / or the second inference model.

[0211] For example, the implementation of the fourth instruction information is similar to the implementation of the second instruction information in step S710 above. For details, please refer to the relevant description of the second instruction information above, which will not be repeated here.

[0212] For example, the sending device can inform the receiving device of its determined first inference model and / or second inference model, ensuring that both ends are aware of the first and / or second inference models. Specifically, when the fourth indication information indicates the first inference model, the receiving device can obtain the first inference model based on the fourth indication information; when the fourth indication information indicates the second inference model, after receiving the fourth indication information, the receiving device can determine the first inference model based on the second inference model and the pre-configured model. For example, the number of model layers in the first inference model can be determined based on the number of model layers in the second inference model and the number of model layers in the pre-configured model.

[0213] (II) Regarding step S702:

[0214] Optionally, when the receiving device decompresses the first information and then inputs it into the first inference model, the receiving device decompresses the first information to obtain the data to be inferred, including: the receiving device decompresses the first information using the first compression module to obtain the data to be inferred.

[0215] For example, prior to step S702, as shown in FIG11, the communication method may further include step S717:

[0216] S717, The receiving device obtains configuration information. This configuration information indicates the relevant parameters of the first decompression module.

[0217] Optionally, the configuration information can be generated autonomously by the receiving device. In this case, step S717 can be replaced by: the receiving device determining the configuration information. Alternatively, the sending device can determine and inform the receiving device. In this case, step S717 can be replaced by: the sending device sending the configuration information to the receiving device; correspondingly, the receiving device receiving the configuration information from the sending device. Alternatively, other devices besides the sending and receiving devices (such as the device training the first inference model) can inform the receiving device. In this case, step S717 can be replaced by: the other device sending the configuration information to the receiving device; correspondingly, the receiving device receiving the configuration information from the other device. Alternatively, the receiving device can obtain the configuration information in any possible way, and this application does not impose any restrictions.

[0218] For example, the relevant parameters of the first decompression module include, but are not limited to: the first decompression method and the first decompression parameters; wherein, the first decompression method is the decompression method adopted by the first decompression module, and the first decompression parameters are the decompression parameters corresponding to the first decompression method.

[0219] Optionally, the first module is related to the first decompression module.

[0220] Specifically, when the first module is used to decompress data (e.g., the first module is used to decompress data, or the first module is used to compress and decompress data), the first module and the first decompression module use the same decompression method; and / or, the type of the decompression parameter corresponding to the decompression method used by the first module is the same as the type of the first decompression parameter. That is, the decompression method used by the first module is the first decompression method, and / or, the decompression parameter corresponding to the decompression method of the first module is the first decompression parameter. For example, the implementation of the first decompression method and the first decompression parameter can be found in the relevant descriptions of the following embodiments, and will not be repeated here.

[0221] Alternatively, when the first module is used to compress data, the decompression method corresponding to the compression method used by the first module is the same as the decompression method used by the first decompression module; and / or, the type of the decompression parameter corresponding to the decompression method used by the first module is the same as the type of the first decompression parameter.

[0222] For example, when the first module can be used to decompress data, when the receiving device knows the relevant parameters of the first module (such as the decompression method used by the first module, and / or the type of decompression parameters corresponding to the decompression method), the configuration information can also indicate the first decompression method and / or the first decompression parameters by indicating the relationship between the first module and the first decompression module.

[0223] Specifically, the configuration information can be represented by 2 bits. When the 2 bits are 00, it indicates that the first module and the first decompression module use the same decompression method, and the type of the decompression parameter corresponding to the decompression method used by the first module is the same as the type of the first decompression parameter; when the 2 bits are 01, it indicates that the first module and the first decompression module use the same decompression method; when the 2 bits are 10, it indicates that the type of the decompression parameter corresponding to the decompression method used by the first module is the same as the type of the first decompression parameter. It should be understood that the above examples exemplify the relationship between the first module and the first decompression module represented by different values ​​of the 2 bits. In fact, the different values ​​of the 2 bits and the relationship between the first module and the first decompression module can also include other correspondences besides the above examples, which are not limited in this application.

[0224] Based on this exemplary scheme, when the receiving device knows the relevant parameters of the first module, the configuration information may include indication information indicating the relationship between the first module and the first decompression module; compared with the scheme where the configuration information directly indicates the first decompression module and / or the first decompression parameters, the indication information occupies fewer resources; therefore, when the receiving end receives configuration information from other devices, it can reduce the resource consumption when the two devices interact.

[0225] Optionally, when the first module can be used to decompress data, the difference between the decompression accuracy of the first module and the decompression accuracy of the first decompression module is a first value.

[0226] For example, the decompression accuracy of the first module can be higher than that of the first decompression module (i.e., the first value is greater than 0), or the decompression accuracy of the first module can be lower than that of the first decompression module (i.e., the first value is less than 0), or the decompression accuracy of the first module can be equal to that of the first decompression module (i.e., the first value is equal to 0).

[0227] For example, the first value may be determined autonomously by the receiving device, or it may be determined and informed to the receiving device by the sending device, or it may be determined and informed to the receiving device by a device other than the sending device and the receiving device.

[0228] Specifically, the first value can be indicated by configuration information, meaning the configuration information also indicates the first value. Alternatively, the first value can be indicated by other information besides configuration information. For example, it can be indicated by second indication information. Thus, when the first value is determined by the transmitting device or a device other than the transmitting and receiving devices, before step S702, the communication device may further include step S718: the transmitting device or the other device sends second indication information to the receiving device, and correspondingly, the receiving device receives the second indication information from the transmitting device or the other device. The second indication information indicates the first value. Alternatively, the first value can be predefined by a protocol. Alternatively, the receiving device can determine the first value based on any possible method, which is not limited in this application.

[0229] It should be understood that, typically, compression methods and decompression methods exist in pairs (or, there is a correspondence between compression and decompression methods; that is, the decompression method corresponding to a compression method can be determined based on the compression method, and vice versa). Similarly, the compression parameters corresponding to a compression method and the decompression parameters corresponding to the decompression method (i.e., the decompression parameters corresponding to that compression method) also exist in pairs (that is, the compression parameters corresponding to a compression method correspond to the decompression parameters corresponding to that compression method). For example, a first compression method corresponds to a first decompression method, and a first compression parameter corresponds to a first decompression parameter; that is, a first decompression module corresponds to a first compression module. Therefore, a first compression method can represent a first decompression method, and a first compression parameter can represent a first decompression parameter; thus, configuration information can indicate the first decompression method by indicating the first compression method; and / or, configuration information can indicate the first decompression parameter by indicating the first compression parameter. In other words, configuration information indicating the first decompression method and / or the first decompression parameter includes: configuration information indicating the first compression method and / or the first compression parameter.

[0230] Optionally, the first compression method can be any one of fixed quantization compression, 2D projection compression, dictionary compression, or sampling compression.

[0231] For example, when the first compression method is fixed quantization compression, the first decompression method is the decompression paired with fixed quantization compression; when the first compression method is 2D projection compression, the first decompression method is the decompression configured with 2D projection compression; when the first compression method is dictionary compression, the first decompression method is the decompression paired with dictionary compression; when the first compression method is sampling compression, the first decompression method is the decompression configured with sampling compression.

[0232] For example, fixed quantization compression refers to quantizing multidimensional data along each dimension to obtain quantized data for each dimension, thereby compressing the multidimensional data. The compression parameter for each dimension is of type quantization bit count; that is, quantization is performed according to the quantization bit count corresponding to that dimension, thus achieving data compression in that dimension. For instance, taking three-dimensional data of x, y, and z as an example, quantization can be performed separately along the x, y, and z axes to obtain compressed data (i.e., the quantization bits corresponding to the x, y, and z axes respectively).

[0233] For example, 2D projection compression refers to the transformation of N-dimensional data, where N is a positive integer greater than 2. Specifically, N-dimensional data can include spatial location dimensions, as well as parameter dimensions such as intensity, color, and delay. For instance, common radar data includes point XYZ position, scanning vertical / horizontal angle, intensity, delay, and related attributes such as color, which can constitute a multi-dimensional point cloud. Since radar collects data with uniformly distributed beams, its positional information can also be represented by the beam's vertical and horizontal indexes.

[0234] Compression processing includes: using 2D data from N-dimensional data as positional information and the remaining N-2 dimensions as numerical information, resulting in compressed data based on the positional and numerical information. N-dimensional data is three-dimensional data, consisting of three spatial positional dimensions: x, y, and z. x and y are used as positional information, and z as numerical information. Alternatively, the quantized values ​​of x and y can be used as positional information, and the quantized value of z as numerical information.

[0235] Compression processing can also include projection. The receiving device obtains the projection mode configuration and performs projection according to the configuration. The projection mode configuration can be pre-configured by the receiving device, or configured and communicated to the receiving device by another device (such as the sending device), or it can be defined by the protocol. Projection, also known as mapping, converts data into a different coordinate system representation. Specifically, a cylindrical mapping can be used, where one dimension remains unchanged, and the other two dimensions are converted to polar coordinates. For example, mapping the x, y, and z coordinates to... R, z coordinates, where Alternatively, a spherical mapping can be used to map the x, y, and z coordinates to... R coordinates, where Alternatively, a planar mapping can be used, projecting the xyz coordinate system onto the XOY / XOZ / YOZ plane. Optionally, the xyz coordinates can be rotated and / or translated before projection.

[0236] As shown in Figure 12(a), x, y, z can be mapped to... R coordinates. And for The data on the R-axis is quantized. Data on coordinates (or, The quantized data on the coordinate system can form a 2D plane as shown in Figure 12(b). The quantized data on the R coordinate system can be used as the corresponding fill value on the 2D plane to obtain a compressed 2D plane as shown in Figure 12(c).

[0237] For example, dictionary compression refers to weighted encoding of data based on a dictionary matrix, so that the data can be represented by k basis vectors in the dictionary matrix, where k is the weighting coefficient. The dictionary matrix is ​​a D*L two-dimensional dictionary, where D represents the dimension of the basis vectors and L represents the number of basis vectors. Specifically, the value of L can be pre-defined (e.g., predefined by a protocol, or pre-agreed upon between the devices involved in inference (e.g., the sending and receiving devices)). Values ​​of L include, but are not limited to, 8, 9, 16, and 32. As shown in Figure 13, taking L=9, k=3, and the k basis vectors d2, d5, and d8 as an example, sample Y can be represented by X.

[0238] For example, sampling compression refers to sampling data according to a sampling rate and / or sampling location to obtain multiple sample data, thereby compressing the data. Examples include sampling the location of training data or sampling the importance of feature data. Specifically, as shown in Figure 14, sampling the data yields seven sample data, which are then compressed into these seven sample data. Furthermore, a subset of these seven sample data can be selected, meaning the radio frequency map can be compressed into this subset of sample data. For instance, sample data matching a certain feature can be set to 1, indicating that the sample data is retained; sample data not matching the feature can be set to 0, indicating that the sample data is discarded.

[0239] Based on the above, when the first compression method is fixed quantization compression or 2D projection compression, the type of the first compression parameter is the number of quantization bits; when the first compression method is dictionary compression, the type of the first compression parameter includes the weighting parameter k corresponding to dictionary compression and / or the dimension (D, L) of the dictionary matrix; when the first compression method is sampling compression, the type of the first compression parameter is the sampling rate and / or the sampling position.

[0240] In some implementations, the configuration information may indicate the first compression method based on the following two methods:

[0241] As an example, the configuration information may include a first index, and the first compression method is the compression method corresponding to the first index in the first correspondence. The first correspondence includes the correspondence between multiple compression methods and multiple indexes.

[0242] For example, the compression methods include, but are not limited to, one or more of the following: fixed quantization compression, 2D projection compression, dictionary compression, or sampling compression. In this case, the first correspondence includes the contents shown in Table 1 below:

[0243] Table 1

[0244] As shown in Table 1 above, when the first index is 1, it indicates that the first compression method is fixed quantization compression; when the first index is 2, it indicates that the first compression method is 2D projection compression; when the first index is 3, it indicates that the first compression method is dictionary compression; and when the first index is 4, it indicates that the first compression method is sampling compression.

[0245] It should be understood that Table 1 above exemplarily lists the correspondence between indexes and compression methods. In reality, the correspondence between indexes and compression methods can also be other than those in Table 1. For example, index 1 corresponds to sampling compression, index 2 corresponds to field compression, etc. Furthermore, Table 1 exemplarily lists four compression methods. In reality, the first correspondence may also include fewer than four compression methods or more than four compression methods, and this application does not impose any restrictions.

[0246] As another example, the configuration information may include a second index and a first data type, and the first compression method is the compression method corresponding to the second index and the first data type in the second correspondence. Here, the first data type is the data type of the data to be inferred, and the second correspondence is a correspondence between multiple data types, multiple compression methods, and multiple indexes.

[0247] Specifically, the data type of the data to be inferred (i.e., the first data type) can also be a default or preset type, and the configuration information can also indicate the second index; further, the first compression method is determined based on the second index and the first data type.

[0248] For example, multiple data types include, but are not limited to: perceptual data, RF map data, etc.; multiple compression methods include, but are not limited to: fixed quantization compression, 2D projection compression, dictionary compression, or one or more of sampling compression. In this case, the second correspondence includes the contents shown in Table 2 below:

[0249] Table 2

[0250] As shown in Table 2 above, when the second index is 0 and the first data type is perceptual data, it indicates that the first compression method is fixed quantization compression; when the second index is 1 and the first data type is perceptual data, it indicates that the first compression method is 2D projection compression; when the second index is 0 and the first data type is RF map data, it indicates that the first compression method is dictionary compression; when the second index is 1 and the first data type is perceptual data, it indicates that the first compression method is sampling compression.

[0251] It should be understood that Table 2 above exemplarily lists some implementations of data types and the correspondence between indexes, data types, and compression methods. In reality, data types may also include other implementations besides those in Table 2, and / or the correspondence between indexes, data types, and compression methods may also be other correspondences besides those in Table 2. Furthermore, Table 2 exemplarily lists four compression methods; in reality, the second correspondence may also include fewer than four compression methods or more than four compression methods. Similarly, Table 2 lists two quantity types; in reality, the second correspondence may also include more than two data types, and this application does not impose any restrictions.

[0252] In some implementations, the receiving device can determine the type of the first compression parameter (i.e., the type of compression parameter corresponding to the first compression method), and further determine the value of the first compression parameter.

[0253] Combining the two examples above, the first compression method can also be pre-configured or default. In this case, the receiving device does not need to obtain configuration information, that is, step S717 is not executed.

[0254] (1) Regarding the type of the first compression parameter:

[0255] Optionally, the receiving device may determine the type of the first compression parameter based on the first compression method.

[0256] For example, as can be seen from the foregoing, different compression methods correspond to different types of compression parameters; therefore, after the receiving device determines the first compression method based on the first index or the second index included in the configuration information, it can determine the type of compression parameters corresponding to that compression method.

[0257] For example, the correspondence between different compression methods and their corresponding compression parameter types is pre-configured or defaulted, that is, the compression parameter types corresponding to different compression methods are pre-configured or defaulted. Thus, after the receiving device determines the first compression method through the indication of configuration information (such as the first index, or the second index and the first data type), it also knows the type of the first compression parameter.

[0258] Specifically, taking the pre-configured correspondence between different compression methods and their corresponding compression parameter types as an example, the correspondence between different compression methods and their corresponding compression parameter types can be merged with the first correspondence (i.e., the correspondence between different indices and different compression methods). In this case, the merged correspondence is the correspondence between multiple indices, multiple compression methods, and multiple compression parameter types. This merged correspondence can include the content shown in Table 3 below, or in other words, the first correspondence shown in Table 1 can be replaced with the content shown in Table 3 below:

[0259] Table 3

[0260] For example, the configuration information can indicate the first compression method and the type of the first compression parameter through an index in Indicator Table 3. For instance, when the configuration information includes index 1, it indicates that the first compression method is fixed quantization compression and the type of the first compression parameter is quantization bit count; when the configuration information includes index 2, it indicates that the first compression method is 2D projection compression and the type of the first compression parameter is quantization bit count; when the configuration information includes index 3, it indicates that the first compression method is dictionary compression and the type of the first compression parameter is dictionary dimension (D*L) and / or sparsity; when the configuration information includes index 4, it indicates that the first compression method is sampling compression and the type of the first compression parameter is sampling rate and / or sampling position.

[0261] Alternatively, the correspondence between different compression methods and their corresponding compression parameter types can be combined with the second correspondence (i.e., the correspondence between different indices, different data types, and different compression methods). In this case, the combined correspondence is the correspondence between multiple indices, multiple compression methods, multiple compression parameter types, and multiple data types. This combined correspondence can include the content shown in Table 4 below, or in other words, the second correspondence shown in Table 2 can be replaced with the content shown in Table 4 below:

[0262] Table 4

[0263] For example, the configuration information can indicate the compression method and / or the type of the first compression parameter of the first module by indicating an index and / or a data type in Table 4. When the data type is default or pre-configured, the configuration information can indicate an index, that is, the compression method and the type of the first compression parameter of the first module. When the compression method of the first module is default or pre-configured, the configuration information can indicate a data type, that is, the compression method and the type of the first compression parameter of the first module. For example, taking the configuration information indicating an index and a data type as an example, when the configuration information indicates index 0 and sensed data, it means that the first compression method is fixed quantization compression, and the type of the first compression parameter is the number of quantization bits; when the configuration information indicates index 1 and sensed data, it means that the first compression method is 2D projection compression, and the type of the first compression parameter is the number of quantization bits; when the configuration information indicates index 0 and RF map data, it means that the first compression method is dictionary compression, and the type of the first compression parameter is dictionary dimension (D*L) and / or sparsity; when the configuration information indicates index 1 and RF map data, it means that the first compression method is sampling compression, and the type of the first compression parameter is sampling rate and / or sampling position.

[0264] It should be understood that Tables 3 and 4 above exemplarily list the correspondence between indexes, compression methods, compression parameter types (and data types); in fact, the correspondence between the above parameters can also be implemented in other forms besides Tables 3 or 4 above. This application does not limit this, as long as an index can correspond to a unique compression method and a unique compression parameter type, or an index and a data type can correspond to a unique compression method and a unique compression parameter type.

[0265] (2) Regarding the value of the first compression parameter (i.e., the first compression parameter):

[0266] Optionally, the values ​​of corresponding compression parameters can be pre-configured for different compression methods, so that after the receiving device determines the first compression method through the configuration information, it also knows the value of the first compression parameter.

[0267] Alternatively, the configuration information can also indicate index A, where the value of the first compression parameter is the value corresponding to index A in the corresponding relationship A. Here, the corresponding relationship A is the correspondence between multiple values ​​of the first compression parameter and multiple indices.

[0268] For example, for each type of compression parameter, a pre-configured correspondence between multiple values ​​of the compression parameter and multiple indices can be established. Thus, the receiving device can determine the value of the first compression parameter based on the indication in the configuration information (i.e., index A).

[0269] Specifically, the correspondence A can include the following two possible implementations:

[0270] In one possible implementation, each index in the correspondence A corresponds to a value of the compression parameter.

[0271] For example, when the first compression method is fixed quantization compression or 2D projection compression, the type of the first compression parameter is the number of quantization bits; in this case, the correspondence A can be a correspondence between multiple indices and multiple values ​​of the number of quantization bits. Specifically, the correspondence A can include the contents shown in Table 5A below:

[0272] Table 5A

[0273] For example, based on the foregoing, when using fixed quantization compression for multidimensional data, each dimension of the multidimensional data needs to be quantized and compressed separately. Thus, index A can be an index value from Table 5A above, and the number of quantization bits corresponding to index A (i.e., the value of the compression parameter) is the same as the number of quantization bits corresponding to each dimension. In other words, the number of quantization bits corresponding to each dimension is the same. Alternatively, the configuration information can also include indexes B, C, ..., such that indexes A, B, C, ... each correspond to an index value from Table 5A above, and each index corresponds to a dimension of the multidimensional data, thereby determining the number of quantization bits corresponding to each dimension.

[0274] Alternatively, each index in correspondence A corresponds to a value of the compression parameter, including: each index in correspondence A corresponds to a value of the first compression parameter in each dimension. Therefore, when the first compression method is fixed quantization compression or 2D projection compression, i.e., the type of the first compression parameter is quantization bit count, taking three-dimensional data as an example, correspondence A can include the contents shown in Table 5B below:

[0275] Table 5B

[0276] As shown in Table 5B, when index A is any of the index values ​​in Table 5B, its corresponding number of quantization bits includes three values, which are the number of quantization bits corresponding to each dimension of the multidimensional data. Specifically, the correspondence between the three values ​​and the three dimensions can be pre-defined, or it can be a default value; this application does not impose any restrictions.

[0277] For example, when the first compression method is dictionary compression, the type of the first compression parameter is dictionary dimension and / or sparsity; wherein, the dictionary dimension can use a default value or a fixed value, and the index A is used to indicate a value of sparsity; that is, the correspondence A is the correspondence between multiple values ​​of sparsity and multiple indices; specifically, the correspondence A may include the contents shown in Table 6A below:

[0278] Table 6A

[0279] As shown in Table 6A, when index A is any of the index values ​​in Table 6A above, the corresponding sparsity value is the sparsity value in the first compression parameter (or, in other words, the sparsity of the first compression module).

[0280] Alternatively, sparsity can use a default value or a fixed value, in which case index A is used to indicate a value of the dictionary dimension; that is, the correspondence A is the correspondence between multiple values ​​of the dictionary dimension and multiple indices; specifically, the correspondence A can include the contents shown in Table 6B:

[0281] Table 6B

[0282] As shown in Table 6B, when index A is any of the index values ​​in Table 6B above, the corresponding dictionary dimension value is the dictionary dimension value in the first compression parameter (or, the dictionary dimension of the first compression module).

[0283] Alternatively, index A can be used to indicate a value for the dictionary dimension and a value for the sparsity; that is, the correspondence A is a correspondence between multiple values ​​for the dictionary dimension, multiple values ​​for the sparsity, and multiple indexes; specifically, the correspondence A can include the content shown in Table 6C:

[0284] Table 6C

[0285] As shown in Table 6C, when index A is any of the index values ​​in Table 6C above, the corresponding dictionary dimension value and sparsity value are the values ​​of the first compression parameter (or, the compression parameter of the first compression module).

[0286] For example, when the compression method of the first module is sample compression, the type of the first compression parameter is sampling rate and / or sampling position; taking the type of the first compression parameter as sampling rate as an example, the correspondence A can be a correspondence between multiple indices and multiple values ​​of the sampling rate. Specifically, the correspondence A can include the contents shown in Table 7 below:

[0287] Table 7

[0288] As shown in Table 7, when index A is any of the index values ​​in Table 7 above, its corresponding sampling rate is the first compression parameter.

[0289] In another possible implementation, each index in the correspondence A corresponds to multiple values, and the value of the first compression parameter is one of these multiple values.

[0290] For example, when the first compression method is fixed quantization compression or 2D projection compression, the type of the first compression parameter is the number of quantization bits; in this case, the correspondence A can be a correspondence between multiple indices and multiple values ​​of the number of quantization bits. Specifically, the correspondence A can include the contents shown in Table 8 below:

[0291] Table 8

[0292] As shown in Table 8, when index A is any of the index values ​​in Table 8, the corresponding number of quantization bits includes multiple values ​​(e.g., indices 0 to 2 correspond to 4 values ​​respectively, index 3 corresponds to 3 values, and index 4 corresponds to 2 values). In this case, the receiving device can select one value from these multiple values ​​as the value of the first compression parameter. For example, the receiving device can randomly select one value from these multiple values ​​as the value of the first compression parameter (e.g., in each iteration of training, one value is selected from these multiple values ​​as the value of the first compression parameter), or the receiving device can determine a value as the value of the first compression parameter based on its own capabilities, or the receiving device can determine the value of the first compression parameter in any possible way, which is not limited in this application.

[0293] For example, when the first compression method is dictionary compression, the type of the first compression parameter is dictionary dimension and / or sparsity; wherein, the dictionary dimension can use a default value or a fixed value, and the index A is used to indicate multiple values ​​of sparsity; that is, the correspondence A is the correspondence between multiple values ​​of sparsity and multiple indices; specifically, the correspondence A can include the contents shown in Table 9A below:

[0294] Table 9A

[0295] As shown in Table 9A, when index A is any of the index values ​​in Table 9A, its corresponding sparsity includes multiple values ​​(e.g., indices 0 to 2 correspond to 3 values ​​each, and index 3 corresponds to 2 values). In this case, the receiving device can select one value from these multiple values ​​as the sparsity value in the first compression parameter. For example, the receiving device can randomly select one value from these multiple values ​​as the sparsity value in the first compression parameter, or the receiving device can determine a value based on its own capabilities as the sparsity value in the first compression parameter, or the receiving device can determine the sparsity value in the first compression parameter in any possible way; this application does not impose any restrictions.

[0296] Alternatively, sparsity can use a default value or a fixed value, in which case index A is used to indicate multiple values ​​of the dictionary dimension; that is, the correspondence A is the correspondence between multiple values ​​of the dictionary dimension and multiple indices; specifically, the correspondence A can include the content shown in Table 9B:

[0297] Table 9B

[0298] As shown in Table 9B, when index A is any of the index values ​​in Table 9B, its corresponding dictionary dimension includes multiple values. In this case, the receiving device can select one value from these multiple values ​​as the value of the dictionary dimension in the first compression parameter. For example, the receiving device can randomly select one value from these multiple values ​​as the value of the dictionary dimension in the first compression parameter, or the receiving device can determine a value as the value of the dictionary dimension in the first compression parameter based on its own capabilities, or the receiving device can determine the value of the dictionary dimension in the first compression parameter in any possible way; this application does not impose any restrictions.

[0299] Alternatively, index A can be used to indicate one value of the dictionary dimension and multiple values ​​of sparsity; that is, the correspondence A is the correspondence between multiple values ​​of the dictionary dimension, multiple values ​​of sparsity, and multiple indexes; specifically, the correspondence A can include the contents shown in Table 9C:

[0300] Table 9C

[0301] As shown in Table 6C, when index A is any of the index values ​​in Table 6C, the corresponding dictionary dimension value is the same as the dictionary dimension value in the first compression parameter. Furthermore, this index value also corresponds to multiple sparsity values, allowing the receiving device to determine one value from these multiple values ​​as the sparsity value in the first compression parameter. Specifically, the implementation of the receiving device selecting one value from multiple values ​​can be found in the relevant description of the above embodiments, and will not be repeated here.

[0302] It should be noted that Table 9C above uses an example of an index value corresponding to one value of the dictionary dimension and multiple values ​​of sparsity to illustrate the implementation of the correspondence relationship A. In fact, the correspondence relationship A can also be an index value corresponding to multiple values ​​of the dictionary dimension and one value of sparsity, or an index value corresponding to multiple values ​​of the dictionary dimension and multiple values ​​of sparsity. In this case, the implementation of the first compression parameter can be found in the relevant description of the above embodiment, and will not be repeated here.

[0303] For example, when the compression method of the first module is sample compression, the type of the first compression parameter is sampling rate and / or sampling position; taking the type of the first compression parameter as sampling rate as an example, the correspondence A can be a correspondence between multiple indices and multiple values ​​of the sampling rate. Specifically, the correspondence A can include the contents shown in Table 10 below:

[0304] Table 10

[0305] As shown in Table 10, when index A is any of the index values ​​in Table 10, it corresponds to multiple values ​​of the sampling rate. In this case, the receiving device can select one of these multiple values ​​as the value of the first compression parameter. For example, the receiving device can randomly select one of these multiple values ​​as the value of the first compression parameter, or the receiving device can determine a value as the value of the first compression parameter based on its own capabilities, or the receiving device can determine the value of the first compression parameter in any possible way; this application does not impose any restrictions.

[0306] It should be understood that Tables 1 to 10 above exemplarily list some values ​​of the first compression parameter in this application, and do not mean that the first compression parameter only includes the values ​​shown in any of the tables 1 to 10 above. In fact, the first compression parameter may have more or fewer values, or any other possible values ​​besides the examples above, and this application does not limit them.

[0307] Combining the two possible implementation methods described above, optionally, after determining the first compression method, the receiving device can determine the decompression method corresponding to the first compression method as the first decompression method. And / or, after determining the first compression parameters, the receiving device can determine the decompression parameters corresponding to the first compression parameters as the first decompression parameters. Further, after determining the first decompression method and / or the first decompression parameters, the receiving device can perform decompression according to the first decompression method and / or the first decompression parameters to obtain the data to be inferred.

[0308] In some implementations, as can be seen from the above embodiments, the first information is the information output by the first compression module; that is, the sending device also needs to know the relevant parameters of the first compression module before performing compression. For example, the relevant parameters of the first compression module include, but are not limited to, the first decompression method and the first decompression parameters.

[0309] For example, before the sending device performs compression, it can obtain relevant parameters of the first compression module. Specifically, the relevant parameters of the first compression module can be generated autonomously by the sending device. Alternatively, the receiving device can determine and inform the sending device, in which case the receiving device sends the relevant parameters of the first compression module to the sending device; correspondingly, the sending device receives the relevant parameters of the first compression module from the receiving device. Alternatively, other devices besides the sending and receiving devices (such as the device training the first inference model) can inform the sending device, in which case the other device sends the relevant parameters of the first compression module to the sending device; correspondingly, the sending device receives the relevant parameters of the first compression module from the other device. Alternatively, the sending device can obtain the relevant parameters of the first compression module in any possible way, which is not limited in this application.

[0310] Optionally, when the first module is used to compress data, or to compress and decompress data, the first module is related to the first compression module. Specifically, the first module and the first compression module use the same compression method; and / or, the type of decompression parameters corresponding to the compression method used by the first module is the same as the type of the first compression parameters. That is to say, the compression method used by the first module is the first compression method, and / or, the compression parameters corresponding to the compression method of the first module are the first compression parameters.

[0311] For example, when the sending device knows the compression method used by the first module and / or the type of compression parameters corresponding to the compression method, it can also indicate the first compression method and / or the first compression parameters by indicating the relationship between the first module and the first compression module.

[0312] Optionally, the difference between the compression accuracy of the first module and the compression accuracy of the first compression module is a second value.

[0313] For example, the compression precision of the first module can be higher than that of the first compression module (i.e., the second value is greater than 0), or the compression precision of the first module can be lower than that of the first compression module (i.e., the second value is less than 0), or the compression precision of the first module can be equal to that of the first compression module (i.e., the second value is equal to 0). Specifically, the first value and the second value can be the same or different.

[0314] For example, the second value may be determined autonomously by the sending device, or it may be determined and informed to the sending device by the receiving device, or it may be determined and informed to the sending device by a device other than the sending and receiving devices. Alternatively, the second value may be predefined by a protocol. Or, the sending device may determine the second value in any possible way, which is not limited in this application.

[0315] Specifically, when the second value is communicated to the sending device by the receiving end or other device, the second value and the relevant parameters of the first compression module can be carried in the same signaling. Alternatively, the second value and the relevant parameters of the first compression module can be carried in different signaling. The implementation of the second value is similar to that of the first value described above; please refer to the relevant description of the first value for details, which will not be repeated here.

[0316] Specifically, the implementation of the sending device determining the first compression method and / or the first compression parameters is similar to the implementation of the receiving device obtaining the first compression method and / or the first compression parameters in the above embodiments. For details, please refer to the relevant descriptions of the first compression method and the first compression parameters above, which will not be repeated here.

[0317] Optionally, when the first module is used to compress or decompress data, the first inference model is obtained by training a model based on a pre-configured model, the first module, and the second module. Specifically, when the first module is used to compress data, the second module is used to decompress data; and when the first module is used to decompress data, the second module is used to compress data.

[0318] Typically, the model training process is similar to the inference process. When a single model (i.e., a pre-configured model) is used for training during the model training phase, the same trained single model (i.e., the first inference model) is also used for inference during the inference phase. Similarly, when multiple models (i.e., the first sub-model and the second sub-model) are used for training during the model training phase, the same trained multiple models (i.e., the first inference model and the second inference model) are used for inference during the inference phase. In other words, the deployment of the inference model (such as the first inference model, or the first and second inference models) follows the same model training process.

[0319] As an example, when training is performed using a single model (i.e., a pre-configured model) during the model training phase, inference is performed by a first inference model during the inference phase. That is, the inference process at this time includes the steps shown in Figure 8(a), where the number of model layers in the first inference model is equal to the number of model layers in the pre-configured model.

[0320] Optionally, in this example, model training is performed by a single device; that is, a one-sided training method is used to train the pre-configured model. This results in a trained model, i.e., the first inference model.

[0321] For example, the single device can be a sending device, or a receiving device, or any other device with model training capabilities (such as a server), excluding sending and receiving devices; this application does not impose any restrictions. Furthermore, when the single device is a sending device or any other device with model training capabilities besides sending and receiving devices, after obtaining the first inference model, the single device can inform the receiving device of the first inference model (such as the relevant configuration of the first inference model), enabling the receiving device to perform inference on the data to be inferred based on the first inference model and obtain the inference result; thus enabling the receiving device to execute step S703.

[0322] Specifically, taking the original data in model training as training data and the device that performs model training (i.e., the single device) as the first device, the first device can train the pre-configured model based on the training data and the first module to obtain the first inference model.

[0323] Optionally, the configuration of the first module can be determined autonomously by the first device, or it can be determined and communicated to the first device by the second device. Specifically, the second device can be any device other than the first device, including the sending device, the receiving device, or any other device with model training capabilities (such as a server). Alternatively, the configuration of the first module can be predetermined by the protocol.

[0324] For example, when the configuration of the first module is determined by the second device and communicated to the first device, the model training process may include steps S1501 to S1502 as shown in Figure 15:

[0325] S1501, the second device sends configuration information #1 to the first device; correspondingly, the first device receives configuration information #1 from the second device. Configuration information #1 indicates the configuration of the first module.

[0326] Optionally, configuration information #1 can indicate the function of the first module, for example, the first module is used to compress data, or the first module can be used to decompress data, or the first module is used to compress and decompress data.

[0327] For example, when the first module is used to compress or decompress data, the model training may also include a second module; wherein, when the first module is used to compress data, the second module is used to decompress data; and when the first module is used to decompress data, the second module is used to compress data. In this case, the first inference model is obtained through model training based on the first module and the pre-configured model, including: the first inference model is obtained through model training based on the first module, the second module, and the pre-configured model.

[0328] Optionally, configuration information #1 can also indicate relevant parameters of the first module.

[0329] For example, when the first module is used to compress data, the relevant parameters of the first module may include the compression method of the first module and / or the compression parameters corresponding to the compression method; based on the foregoing, when the first module is used to compress data, the first module is related to the first compression module; at this time, it can also be considered that the configuration information #1 indicates the first compression method and / or the first compression parameters.

[0330] When the first module is used to decompress data, the relevant parameters of the first module may include the decompression method of the first module, and / or the decompression parameters corresponding to the decompression method. Based on the foregoing, when the first module is used to decompress data, the first module is related to the first decompression module. In this case, configuration information #1 can also be considered as indicating the first decompression method and / or the first decompression parameters. Furthermore, based on the foregoing, the first decompression method can be represented by a first compression method, and the first decompression parameters can be represented by first decompression parameters. Therefore, configuration information #1 can indicate the first decompression method by indicating the first compression method; and / or, configuration information #1 can indicate the first decompression parameters by indicating the first compression parameters.

[0331] When the first module is used to compress and decompress data, configuration information #1 can indicate a first compression method and / or a first compression parameter. Thus, the first device can determine a first decompression method based on the first compression method and determine the first decompression parameter based on the first compression parameter. Specifically, the implementation of the first device determining the first decompression method based on the first compression method and determining the first decompression parameter based on the first compression parameter can be described in the relevant descriptions of the above embodiments, and will not be repeated here.

[0332] S1502. The first device trains the pre-configured model based on configuration information #1 using the first module and training data to obtain the first inference model.

[0333] For example, the first device trains the pre-configured model based on configuration information #1 using the first module and training data to obtain the first inference model. This can be understood as: the first module is added during the training of the pre-configured model using training data, or the first module also participates in the training process based on the training data and the pre-configured model.

[0334] Optionally, the training data can be generated autonomously by the first device, or it can be provided to the first device by another device besides the first device. For example, when the training data is provided to the first device by the second device, the model training process may further include step S1503 as shown in Figure 15 before step S1502:

[0335] S1503, the second device sends training data to the first device, and correspondingly, the first device receives training data from the second device.

[0336] For example, the embodiments of this application do not limit the order between steps S1503 and S1501; for example, step S1503 may be executed before step S1501, or step S1503 may be executed after step S1501, or step S1503 may be executed simultaneously with step S1501.

[0337] For example, based on different implementations of the first module, the first device obtains the first inference model in the following three scenarios:

[0338] Scenario 1: The first module is used to compress and decompress data.

[0339] Optionally, during model training, at least one first module may be added; wherein each of the at least one first module may be deployed at the deployment location of the first module shown in Figure 16(a). For example, it may be deployed before the pre-configured model, deployed after the pre-configured model, etc. For instance, since model training is performed by the first device, at least one first module may be deployed in the first device. Furthermore, when training data is sent to the first device from the second device, and a first module of the at least one first module is deployed before the pre-configured model, that first module may be deployed in either the first device or the second device.

[0340] Specifically, taking the addition of a first module during model training as an example, when the first module is deployed before the pre-configured model, the model training process can be as shown in Figure 16(b). When the first module is deployed in the first device, the first device can input the training data into the first module, causing the training data to undergo compression and decompression to obtain the data to be trained. This data is then input into the pre-configured model, outputting training result #1. Further, the training label corresponding to the current task can be compared with the training result #1, and the loss of training result #1 relative to the training label can be calculated. The pre-configured model is then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained. Specifically, different tasks can include AI reconstruction, detection, segmentation, etc. For example, the training label corresponding to AI reconstruction is the training data.

[0341] Alternatively, when training data is sent from the second device to the first device, and the first module is deployed in the second device, the second device can input the training data into the first module, causing the training data to undergo compression and decompression to obtain the data to be trained. This data to be trained is then sent to the first device, which can then input this data to the pre-configured model and output the first training result. Further, the training label can be compared with the training result #1 to calculate the loss of training result #1 relative to the training data. The pre-configured model is then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained.

[0342] When the first module is deployed before the pre-configured model, the model training process can be as shown in Figure 16(c). The first device can input training data into the pre-configured model, output training result #2, and then input training result #2 into the first module, so that training result #2 undergoes compression and decompression to obtain training result #3. Then, the training result #3 is compared with the training label corresponding to the current task, and the loss of training result #3 relative to the training label is calculated. The pre-configured model is then adjusted according to the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained.

[0343] It should be understood that the above examples illustrate the implementation of the first inference model by assuming that the deployment position of the first module remains unchanged during multiple iterations of model training (as shown in Figure 16(b), the first module is deployed before the pre-configured model, or as shown in Figure 16(c), the first module is deployed after the pre-configured model). In reality, the deployment position of the first module can change during multiple iterations. For example, in the Nth iteration (or after the Nth reverse update), the first module is deployed before the pre-configured model, and in the N+1th iteration (or after the N+1th reverse update), the first module is deployed after the pre-configured model. Here, N is a positive integer greater than or equal to 1.

[0344] Optionally, the deployment location of at least one first module can be determined autonomously by the first device, or determined and informed by the second device, or predefined by a protocol. Alternatively, the first device can obtain the deployment location of the first module in any possible way, and this application does not impose any restrictions.

[0345] For example, taking the addition of a first module during model training as an example, when the deployment location of the first module is determined by the second device and communicated to the first device, the model training process may further include step S1504 as shown in Figure 17 before step S1502:

[0346] S1504, the second device sends configuration information #2 to the first device, and correspondingly, the first device receives configuration information #2 from the second device. Configuration information #2 indicates the deployment location of the first module.

[0347] For example, taking the deployment location shown in Figure 16(a) as an example, configuration information #2 can be indicated by 1 bit; when the 1 bit is 1, it indicates that the first module is deployed before the pre-configured model, and when the 1 bit is 0, it indicates that the first module is deployed after the pre-configured model. Alternatively, the 1 bit is 0, indicating that the first module is deployed before the pre-configured model, and the 1 bit is 1, indicating that the first module is deployed after the pre-configured model.

[0348] Furthermore, whether the first module is deployed on the first device or the second device can be determined autonomously by the first device, or it can be indicated by configuration information #2. For example, when the first device autonomously determines whether to deploy the first module on the first device or the second device, the first device can make this determination based on its own capabilities and those of the second device (such as remaining resources, computing power, etc.). For instance, it might choose to deploy the first module on the device with stronger capabilities, or, if its own capabilities are sufficient to execute the first module, it might choose to deploy the first module on itself.

[0349] When configuration information #2 also indicates whether the first module is deployed in the first device or the second device, the configuration information can be represented by 2 bits; when the 2 bits are 11, it means that the first module is deployed before the pre-configuration model and the first module is deployed in the first device; when the 2 bits are 10, it means that the first module is deployed before the pre-configuration model and the first module is deployed in the second device; when the 2 bits are 01, it means that the first module is deployed after the pre-configuration model and the first module is deployed in the first device.

[0350] Optionally, the first module can be deployed before the pre-configured model, or it can be replaced by deploying the first module after the training data.

[0351] For example, the above examples illustrate different implementations of configuration information #2 indicating the deployment location of the first module. In fact, configuration information #2 can also indicate the deployment location of the first module in any possible form other than those described above, and this application does not limit it.

[0352] For example, configuration information #1 and configuration information #2 can be located in the same signaling (in this case, step S1501 and step S1504 are the same step), or configuration information #1 and configuration information #2 can be located in different signaling.

[0353] For example, the embodiments of this application do not limit the order between steps S1504 and S1503; for example, step S1504 may be executed before step S1503, or step S1504 may be executed after step S1503, or step S1504 may be executed simultaneously with step S1503.

[0354] It should be noted that, when the pre-configured model is determined based on scenario one, during the inference process, the receiving device needs to decompress the first information before inputting it into the first inference model.

[0355] Scenario 2: The first module is used to compress data.

[0356] As an example, a second module is also needed during model training to decompress the data.

[0357] For example, the first module and the second module are corresponding, that is, the compression method used by the first module corresponds to the decompression method used by the second module, and the compression parameters used by the first module correspond to the decompression parameters used by the second module; therefore, it can also be considered that the first module and the second module exist in pairs.

[0358] Optionally, during model training, at least one pair of first modules and second modules can be added; wherein, the first module and second module of the at least one pair of first modules and second modules can both be deployed before the pre-configured model; or, the first module and the second module can both be deployed after the pre-configured model.

[0359] Specifically, taking the addition of a first module and a second module during model training as an example, when both the first and second modules are deployed before the pre-configured model, the model training process can be as shown in Figure 18(a). Compression is performed by the first module, and decompression is performed by the second module. That is, the training data is input into the first module, causing it to output compressed data #1. This compressed data #1 is then input into the second module, causing it to output the data to be trained. This data is then input into the pre-configured model, outputting the training result #1. Further, the training label corresponding to the current task can be compared with the training result #1, and the loss of the training result #1 relative to the training label can be calculated. The pre-configured model is then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained. Specifically, when both the first module and the second module are deployed in the first device, the process shown in Figure 18(a) is executed by the first device; when both the first module and the second module are deployed in the second device, the compression and decompression process shown in Figure 18(a) is executed by the second device, so that the second device sends the decompressed training data to the first device, and the first device can execute subsequent processes; when the first module is deployed in the second device and the second module is deployed in the first device, the compression process shown in Figure 18(a) is executed by the second device, that is, the second device sends the compressed data #1 to the first device, so that the first device can execute subsequent processes.

[0360] When both the first and second modules are deployed after the pre-configured model, the model training process can be as shown in Figure 18(b). Compression is performed by the first module, and decompression is performed by the second module. If both modules are deployed in the first device, the first device can input training data into the pre-configured model, output training result #2, and then input training result #2 into the first module, causing it to undergo compression to obtain compressed data #2. This compressed data #2 is then input into the second module, decompressed to obtain training result #3, and then compared with the training label corresponding to the current task to calculate the loss of training result #3 relative to the training label. The pre-configured model is then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained.

[0361] If the first module is located in the first device and the second module is located in the second device, the first device can input training data into the pre-configured model, output training result #2, and then input training result #2 into the first module, where it undergoes compression to obtain compressed data #2. This compressed data #2 is then sent to the second device, where it can input compressed data #3 into the second module. After decompression, training result #3 is obtained, and this result is compared with the training label corresponding to the current task to calculate the loss of training result #3 relative to the training label. The pre-configured model is then adjusted based on the loss, and a reverse update is performed. After multiple iterations, a trained pre-configured model, i.e., the first inference model, is obtained.

[0362] Optionally, the deployment location of the first module and / or the deployment location of the second module may be determined autonomously by the first device, or determined by the second device and communicated to the first device, or predefined by a protocol. Alternatively, the first device may obtain the deployment location of the first module and / or the deployment location of the second module in any possible way, and this application is not limited thereto.

[0363] For example, when the deployment location of the first module and / or the deployment location of the second module are determined by the second device and communicated to the first device, the above configuration information #2 can indicate the deployment location of the first module and / or the deployment location of the second module.

[0364] Specifically, if configuration information #2 only indicates the deployment position of the first module and / or the second module during model training, then configuration information #2 can be represented by 2 bits; when the 2 bits are 00, it means that both the first and second modules are deployed before the pre-configured model; when the 2 bits are 01, it means that both the first and second modules are deployed after the pre-configured model; and when the 2 bits are 10, it means that the first module is deployed before the pre-configured model and the second module is deployed after the pre-configured model.

[0365] For example, the above examples illustrate one possible correspondence between the value of the 2-bit and the deployment positions of the first module and / or the second module during model training. Other possible correspondences besides the above examples may also exist between the value of the 2-bit and the deployment positions of the first module and / or the second module during model training. Furthermore, while the above examples illustrate different implementations of configuration information #2 indicating the deployment positions of the first module and / or the second module, in reality, configuration information #2 can also indicate the deployment positions of the first module and / or the second module in any possible form other than those described above, and this application does not impose any limitations.

[0366] It should be noted that, when determining the pre-configured model based on this example, during the inference process, the receiving device needs to decompress the first information before inputting it into the first inference model.

[0367] As another example, no second module parameters are needed during model training, meaning the data during model training does not need to be decompressed.

[0368] For example, in this case, the first module can be deployed before the pre-configured model. Specifically, the model training process can be as shown in Figure 18(c): compression is performed by the first module, and decompression is performed by the second module. That is, the training data is input into the first module, compressed to obtain compressed data #3, and then the compressed data #3 is input into the pre-configured model to output the training result #4. Further, the training label corresponding to the current task can be compared with the training result #4 to calculate the loss of the training result #4 relative to the training label. Then, the pre-configured model is adjusted according to the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained. When the first module is deployed in the first device, the process shown in Figure 18(c) is performed by the first device; when the first module is deployed in the second device, the compression process shown in Figure 18(c) is performed by the second device, so that the second device sends the compressed data #3 to the first device, and the first device can execute the subsequent processes.

[0369] It should be noted that, based on this example, when the pre-configured model is determined, the receiving device does not need to decompress the first information during the inference process and can directly input the first information into the first inference model.

[0370] Scenario 3: The first module is used to decompress the data. In this case, the model training process also requires parameters from the second module, which is used to compress the data.

[0371] For example, the implementation of the first module is the same as the implementation of the second module in the above scenario two, and the implementation of the second module is the same as the implementation of the first module in the above scenario two. For details, please refer to the relevant description in the above scenario two, which will not be repeated here.

[0372] As another example, when multiple models (i.e., the first sub-model and the second sub-model) are used for training during the model training phase, the first inference model and the second inference model perform inference during the inference phase. That is, the inference process at this time includes the following as shown in Figure 8(b), where the sum of the number of model layers of the first inference model and the number of model layers of the second inference model is equal to the number of model layers of the configured model.

[0373] Optionally, in this example, taking the original data from model training as the training data, a first sub-model and a second sub-model can be trained based on the first module and the training data to obtain the trained first sub-model, i.e., the first inference model; and the trained second sub-model, i.e., the second inference model. Specifically, model training can be performed by a single device, that is, a one-sided training method is used to train the first and second sub-models; or, model training can be performed by multiple devices (i.e., two devices, such as the first device and the second device), that is, a two-sided training method is used to train the first and second sub-models.

[0374] Furthermore, after determining that the first inference model and the second inference model can be used, the receiving device is informed of the first inference model (such as the relevant configuration of the first inference model), enabling the receiving device to perform inference on the data to be inferred based on the first inference model and obtain the inference result; thus enabling the receiving device to execute step S703. The sending device is informed of the second inference model (such as the relevant configuration of the second inference model), enabling the sending device to perform inference based on the second inference model, thus enabling the sending device to execute step S705.

[0375] Optionally, the configuration of the first module can be determined autonomously by the first device, or it can be determined and communicated to the first device by the second device. Specifically, the second device can be any device other than the first device, including the sending device, the receiving device, or any other device with model training capabilities (such as a server). Alternatively, the configuration of the first module can be predetermined by the protocol.

[0376] For example, when the configuration of the first module is determined by the second device and communicated to the first device, the model training process can include the following two possible implementations based on different model training methods:

[0377] In one possible implementation, when training the first sub-model and the second sub-model using a one-sided training method (i.e., by the first device), the model training process may include steps S1901 to S1902 as shown in Figure 19(a):

[0378] S1901, the second device sends configuration information #1 to the first device; correspondingly, the first device receives configuration information #1 from the second device. Configuration information #1 indicates the configuration of the first module.

[0379] Optionally, configuration information #1 can indicate the function of the first module, for example, the first module is used to compress data, or the first module can be used to decompress data, or the first module is used to compress and decompress data.

[0380] For example, when the first module is used to compress or decompress data, the model training may also include a second module; the implementation of the first and second modules can be found in the relevant description in step S1701 above, and will not be repeated here.

[0381] Optionally, configuration information #1 can also indicate relevant parameters of the first module. For example, the implementation of configuration information #1 can be found in the relevant description of step S1701 above, and will not be repeated here.

[0382] S1902, the first device, based on configuration information #1, trains the first sub-model and the second sub-model using the first module and training data, thereby obtaining the first inference model and the second inference model.

[0383] For example, the first device trains the pre-configured model based on configuration information #1 using the first module and training data to obtain the first inference model. This can be understood as: the first module is added during the training of the pre-configured model using training data, or the first module also participates in the training process based on the training data and the pre-configured model.

[0384] In another possible implementation, when the first sub-model and the second sub-model are trained using a two-sided training method (i.e., by the first device and the second device), the model training process may further include S1903 to S1904 as shown in Figure 19(b):

[0385] S1903, First and Second Devices Obtain Configuration Information #1.

[0386] For example, the implementation of configuration information #1 can be found in the relevant description of step S1701 above, and will not be repeated here.

[0387] S1904. The first device and the second device train the first sub-model and the second sub-model based on configuration information #1, the first module, and training data. This yields the first inference model and the second inference model.

[0388] Combining the two possible implementation methods described above, optionally, the training data can be generated autonomously by the first device, or it can be provided to the first device by another device besides the first device. For example, when the training data is provided to the first device by the second device, before step S1902, the model training process may further include step S1905 as shown in Figure 19(a), or before step S1904, the model training process may further include step S1905 as shown in Figure 19(b).

[0389] S1905, The second device sends training data to the first device, and correspondingly, the first device receives training data from the second device.

[0390] For example, the embodiments of this application do not limit the order of steps S1905 and S1901 (or S1903); for example, step S1905 may be executed before step S1901 (or step S1903), or step S1905 may be executed after step S1901 (or step S1903), or step S1905 may be executed simultaneously with step S1901.

[0391] For example, combining the two possible implementations mentioned above, and based on different implementations of the first module, the first device obtains the first inference model and the second inference model based on the following three scenarios:

[0392] Scenario 1: The first module is used to compress and decompress data.

[0393] Optionally, during model training, at least one first module can be added; wherein each of the at least one first module can be deployed at the deployment location of the first module as shown in Figure 20(a). For example, it can be deployed before the second inference model; the first module can be deployed after the first inference model; the first module can be deployed after the second inference model and before the first inference model, etc. In addition, when training data is sent from the second device to the first device, and a first module of at least one first module is deployed before the pre-configured model, the first module can be deployed in either the first device or the second device.

[0394] For example, taking the addition of a first module during model training as an example, when the first module is deployed before the second inference model, the model training process can be as shown in Figure 20(b). Training data can be input into the first module, where it undergoes compression and decompression to obtain the data to be trained. This data is then input into the second sub-model, outputting intermediate feature #1. Intermediate feature #1 is then input into the first sub-model to output training result #5. Further, the training label corresponding to the current task can be compared with the training result #5, and the loss of training result #5 relative to the training label can be calculated. The first and second sub-models are then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained first sub-model (i.e., the first inference model) and the trained second sub-model (i.e., the second inference model) are obtained.

[0395] Specifically, when using a one-sided training method (i.e., model training is performed by a single device, as shown in Figure 19(a) of the model training process), both the first and second sub-models are deployed on the first device. In this case, if the first module is deployed on the first device, the model training process shown in Figure 20(b) above is executed by the first device. If the first module is deployed on the second device, in the model training process shown in Figure 20(b) above, the second device can input the training data into the first module, so that the training data undergoes compression and decompression to obtain the data to be trained, and sends the data to be trained to the first device, thereby enabling the first device to execute the remaining processes.

[0396] When a two-sided training method is adopted (i.e., model training is performed by multiple devices, as shown in Figure 19(b) of the model training process), the first sub-model is deployed in the first device and the second sub-model is deployed in the second device. At this time, the second device can input the training data into the first module, so that the training data undergoes compression and decompression to obtain the data to be trained, and then input the data to be trained into the second sub-model to output intermediate feature #1; and send the intermediate feature #1 to the first device so that the first device can execute the remaining steps.

[0397] For example, when the first module is deployed after the first inference model, the model training process can be as shown in Figure 20(c). Training data can be input into the second sub-model, outputting intermediate feature #2. Then, intermediate feature #2 is input into the first sub-model to output training result #6. Finally, training result #6 is input into the first module, where it undergoes compression and decompression to obtain training result #7. Further, the training label corresponding to the current task can be compared with training result #7, and the loss of training result #7 relative to the training label can be calculated. The first and second sub-models are then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained first sub-model (i.e., the first inference model) and the trained second sub-model (i.e., the second inference model) are obtained.

[0398] Specifically, when using a single-sided training method (i.e., model training is performed by a single device, as shown in Figure 19(a) of the model training process), both the first and second sub-models are deployed on the first device. The first module is also deployed on the first device; that is, the model training process shown in Figure 20(c) is executed entirely by the first device. When using a two-sided training method (i.e., model training is performed by multiple devices, as shown in Figure 19(b) of the model training process), the first sub-model is deployed on the first device, and the second sub-model is deployed on the second device. In this case, in the model training process shown in Figure 20(c), the second device can input training data into the first module, causing the training data to undergo compression and decompression to obtain the data to be trained. This data is then input into the second sub-model, outputting intermediate feature #1. The intermediate feature #1 is then sent to the first device, enabling the first device to execute the remaining steps.

[0399] For example, taking the addition of a first module during model training as an example, when the first module is deployed before the first inference model and after the second inference model, the model training process can be as shown in Figure 20(d). Training data can be input into the second sub-model, outputting intermediate feature #2. Then, intermediate feature #2 is input into the first module, undergoing compression and decompression to obtain intermediate feature #3. This intermediate feature #3 is then input into the first sub-module, outputting training result #8. Further, the training label corresponding to the current task can be compared with the training result #8, calculating the loss of training result #8 relative to the training label. The first and second sub-models are then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained first sub-model (i.e., the first inference model) and the trained second sub-model (i.e., the second inference model) are obtained.

[0400] Specifically, when using a single-sided training method (i.e., model training is performed by a single device, as shown in Figure 19(a) of the model training process), both the first and second sub-models are deployed in the first device; at this time, the first module is also deployed in the first device, and the model training process shown in Figure 20(d) above is performed by the first device. When using a two-sided training method (i.e., model training is performed by multiple devices, as shown in Figure 19(b) of the model training process), the first sub-model is deployed in the first device, and the second sub-model is deployed in the second device; at this time, if the first device is deployed in the first device, in the model training process shown in Figure 20(d) above, the second device can input the training data into the second sub-model, output intermediate feature #2, and then send the intermediate feature #2 to the first device, which will then perform the remaining steps. If the first device is deployed in the second device, then in the model training process shown in (d) of Figure 20 above, the second device can input the training data into the second sub-model, output intermediate feature #2, and then input the intermediate feature #2 into the first module, so that the intermediate feature #2 undergoes compression and decompression to obtain intermediate feature #3, and then sends the intermediate feature #3 to the first device, so that the first device can execute the remaining steps.

[0401] It should be understood that the above examples illustrate the implementation of the first inference model by showing that the first module is deployed in a fixed position during multiple iterations of model training (as shown in Figure 20(b), the first module is deployed before the second sub-model; or, as shown in Figure 20(c), the first module is deployed after the first sub-model; or, as shown in Figure 20(d), the first module is deployed after the second sub-model and before the first sub-model). In reality, the deployment position of the first module can change during multiple iterations. For example, in the Nth iteration (or after the Nth reverse update), the first module is deployed before the second sub-model; in the N+1th iteration (or after the N+1th reverse update), the first module is deployed after the first sub-model, and so on. Here, N is a positive integer greater than or equal to 1.

[0402] For example, based on the foregoing, the first sub-model and the second sub-model are split from the pre-configured model. Therefore, when the first module is deployed after the second sub-model and before the first sub-model, it can also be considered that the first module is deployed in the middle of the pre-configured model.

[0403] Optionally, the deployment location of at least one first module may be determined by the first device and communicated to the second device, or it may be determined by the second device and communicated to the first device, or it may be predefined by a protocol. Alternatively, the first device may obtain the deployment location of the first module in any possible way, and this application is not limited thereto.

[0404] For example, taking the addition of a first module during model training as an example, when the deployment location of the first module is determined by the second device and communicated to the first device, before step S1902, the model training process may also include step S1906 as shown in Figure 21(a) or Figure 21(b):

[0405] S1906, the second device sends configuration information #3 to the first device, and correspondingly, the first device receives configuration information #3 from the second device. Configuration information #3 indicates the deployment location of the first module.

[0406] For example, taking the deployment location shown in Figure 20(a) and the pre-configured model layer number as Y, different mappings between indexes and different deployment locations can be pre-configured. For instance, this mapping may include the content shown in Table 11 below:

[0407] Table 11

[0408] In this context, configuration information #3 indicates that for any index value in Table 11 above, the first module can be deployed after the position corresponding to that index value. For example, the first module can be deployed after the training data, meaning it is deployed before the second sub-model; another example is that the first module can be deployed after layer y, where y = 1, 2, ..., Y-1; meaning the first module is deployed after the second sub-model and before the first sub-model (i.e., the first module is deployed in the middle of the pre-configured model), where the second sub-model has y layers and the first sub-model has Yy layers. Yet another example is that the first module can be deployed after the output of the pre-configured model, meaning it is deployed after the first sub-model.

[0409] Furthermore, whether the first module is deployed on the first device or the second device can be mutually agreed upon by the first and second devices, such as as indicated by configuration information #3. Specifically, the first or second device can be determined based on the capabilities of the two devices (such as remaining resources, computing power, etc.). For example, the device with stronger capabilities can be selected to deploy the first module. In addition, configuration information #3 can indicate whether the first module is deployed on the first or second device through different bit values. The specific implementation is similar to the implementation of configuration information #1 indicating whether the first module is deployed on the first or second device, and will not be repeated here.

[0410] For example, the above examples illustrate different implementations of configuration information #3 indicating the deployment location of the first module. In fact, configuration information #3 can also indicate the deployment location of the first module in any possible form other than those described above, and this application does not limit it.

[0411] For example, configuration information #1 and configuration information #3 can be located in the same signaling, or configuration information #1 and configuration information #3 can be located in different signaling.

[0412] For example, the embodiments of this application do not limit the order of steps S1906 and S1901 (or S1903); for example, step S1906 may be executed before step S1901 (or step S1903), or step S1906 may be executed after step S1901 (or step S1903), or step S1906 may be executed simultaneously with step S1901 (or step S1903).

[0413] Furthermore, Table 11 above can also be applied to a scheme in which a single model (i.e., a pre-configured model) is used to perform training during the model training phase (as shown in Figure 16(a) above). In this case, configuration information #2 may include index 0 or index Y. When configuration information #2 includes index 0, it means that the first module is deployed after the training data, that is, the first module is deployed before the pre-configured model. When configuration information #2 includes index 1, it means that the first module is deployed after the pre-configured model.

[0414] It should be noted that, when the pre-configured model is determined based on scenario one, during the inference process, the receiving device needs to decompress the first information before inputting it into the first inference model.

[0415] Scenario 2: The first module is used to compress data. In this case, the model training process also requires parameters from the second module, which is used to decompress the data.

[0416] For example, the first module and the second module are corresponding, that is, the compression method used by the first module corresponds to the decompression method used by the second module, and the compression parameters used by the first module correspond to the decompression parameters used by the second module; therefore, it can also be considered that the first module and the second module exist in pairs.

[0417] Optionally, during model training, at least one pair of first modules and second modules can be added; wherein, the pair of first modules and second modules can both be deployed before the second sub-model; or, the pair of first modules and second modules can both be deployed after the first sub-model; or, the pair of first modules and second modules can both be deployed before the first sub-model and after the second sub-model.

[0418] For example, taking the addition of a first module and a second module during model training as an example, when both the first module and the second module are deployed before the second sub-model, the model training process can be as shown in Figure 22(a). Compression is performed by the first module, and decompression is performed by the second module. Specifically, training data is input into the first module, causing it to output compressed data #1. This compressed data #1 is then input into the second module, causing it to output data to be trained. This data is then input into the second sub-model, outputting intermediate feature #3. Intermediate feature #3 is then input into the first sub-model, outputting training result #9. Further, the training label corresponding to the current task can be compared with the training result #9, and the loss of training result #9 relative to the training label can be calculated. The pre-configured model is then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained.

[0419] Specifically, when using a single-sided training method (i.e., model training is performed by a single device, as shown in Figure 19(a) of the model training process), both the first sub-model and the second sub-model are deployed on the first device. In this case, if both the first module and the second module are deployed on the first device, the process shown in Figure 22(a) is executed by the first device. If both the first module and the second module are deployed on the second device, the compression and decompression shown in Figure 22(a) are performed by the second device, so that the second device sends the decompressed training data to the first device, and the first device can then execute the subsequent process. If the first module is deployed on the second device and the second module is deployed on the first device, the compression shown in Figure 22(a) is performed by the second device, that is, the second device sends the compressed data #1 to the first device, and the first device can then execute the subsequent process. When a two-sided training method is adopted (i.e., model training is performed by multiple devices, as shown in Figure 19(b) of the model training process), the first sub-model is deployed in the first device and the second sub-model is deployed in the second device. At this time, both the first module and the second module are deployed in the second device. Therefore, the second device can send the intermediate feature #3 output by the second sub-model to the first device, so that the first device can execute the subsequent process.

[0420] For example, taking the addition of a first module and a second module during model training as an example, when both the first module and the second module are deployed after the first sub-model, the model training process can be as shown in Figure 22(b). Compression is performed by the first module, and decompression is performed by the second module. At this time, the training data is input into the second sub-model, outputting intermediate feature #2. Then, intermediate feature #2 is input into the first sub-model, outputting training result #6. Training result #6 is then input into the first module, causing training result #7 to undergo compression to obtain training result #10. Training result #10 is then input into the second module, causing training result #10 to undergo decompression to obtain training result #11. This training result #11 is then compared with the training label corresponding to the current task, and the loss of training result #11 relative to the training label is calculated. The pre-configured model is then adjusted based on the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained.

[0421] Specifically, when using a one-sided training method (i.e., model training is performed by a single device, as shown in Figure 19(a) of the model training process), both the first and second sub-models are deployed on the first device. In this case, if the first module and the second module are also deployed on the first device, that is, the model training process shown in Figure 22(b) is executed by the first device. If the first module is deployed on the first device and the second module is deployed on the second device, the first device can send the training result #10 output by the first module to the second device, which will then execute the subsequent processes.

[0422] When a two-sided training method is adopted (i.e., model training is performed by multiple devices, as shown in Figure 19(b) of the model training process), the first sub-model is deployed in the first device and the second sub-model is deployed in the second device; the first module and the second module are also deployed in the first device. At this time, in the model training process shown in Figure 22(b) above, the second device can send the intermediate feature #2 output by the second sub-model to the first device, so that the first device can execute the subsequent process.

[0423] For example, taking the addition of a first module and a second module during model training as an example, when both the first and second modules are deployed before the first sub-model and after the second sub-model (i.e., both the first and second modules are deployed in the middle of the pre-configured model), the model training process can be as shown in Figure 22(c). Compression is performed by the first module, and decompression is performed by the second module. That is, the training data is input into the second sub-model, which outputs intermediate feature #2. Then, intermediate feature #2 is input into the first module, which outputs intermediate feature #4. Then, intermediate feature #4 is input into the second module, which outputs intermediate feature #5. Then, intermediate feature #5 is input into the first sub-model to obtain the training result #12. Further, the training label corresponding to the current task can be compared with the training result #12 to calculate the loss of the training result #12 relative to the training label. Then, the pre-configured model is adjusted according to the loss, and a reverse update is performed. After multiple iterations, the trained pre-configured model, i.e., the first inference model, is obtained.

[0424] Specifically, when using a single-sided training method (i.e., model training is performed by a single device, as shown in Figure 19(a) of the model training process), both the first and second sub-models are deployed on the first device. In this case, the first and second modules are also deployed on the first device; that is, the model training process shown in Figure 22(c) is executed entirely by the first device. When using a two-sided training method (i.e., model training is performed by multiple devices, as shown in Figure 19(b) of the model training process), the first sub-model is deployed on the first device, and the second sub-model is deployed on the second device. In this case, in the model training process shown in Figure 22(c), if both the first and second modules are deployed on the first device, the second device can send the intermediate feature #2 output by the second sub-model to the first device, allowing the first device to execute subsequent processes. If both the first and second modules are deployed on the second device, the second device can send the intermediate feature #5 output by the second module to the first device, allowing the first device to execute subsequent processes. If the first module is deployed in the second device, and the second module is deployed in the first device, then the second device can send the intermediate feature #4 output by the first module to the first device, so that the first device can execute the subsequent process.

[0425] Furthermore, the first module can be deployed before the second sub-model, and the second module can be deployed before the first sub-model and after the second sub-model; or, the first module can be deployed before the second sub-model, and the second module can be deployed after the first sub-model; or, the first module can also be deployed before the first sub-model and after the second sub-model, and the second module can be deployed after the first sub-model. In this case, the implementation of the model training process is similar to the model training process shown in Figure 22 (Figure 22(a) to (c)), and the details can be found in the relevant description in Figure 22, which will not be repeated here.

[0426] Optionally, the deployment locations of at least one pair of first and second modules can be determined by the first device and communicated to the second device, or determined by the second device and communicated to the first device, or predefined by a protocol. Alternatively, the first device can obtain the deployment location of the first module in any possible way, and this application does not impose any restrictions.

[0427] Specifically, the deployment locations of at least one pair of first and second modules are implemented in a manner similar to the deployment location of the first module in Scenario 1 above. For details, please refer to the relevant description in Scenario 1 above, which will not be repeated here.

[0428] It should be noted that, when the pre-configured model is determined based on scenario two, during the inference process, the receiving device needs to decompress the first information before inputting it into the first inference model.

[0429] Scenario 3: The first module is used to decompress the data. In this case, the model training process also requires parameters from the second module, which is used to compress the data.

[0430] For example, the implementation of the first module is the same as the implementation of the second module in the above scenario two, and the implementation of the second module is the same as the implementation of the first module in the above scenario two. For details, please refer to the relevant description in the above scenario two, which will not be repeated here.

[0431] Combining the three scenarios mentioned above, in this application, the pre-configured model (or the first sub-model and the second sub-model) can be an initial model; or it can be a model that has already been trained. In this case, training the pre-configured model is equivalent to fine-tuning the model.

[0432] Furthermore, adding compression and decompression models (such as the first module, or the first and second modules) during model training may lead to non-differentiable back-updates during training, causing model training failure. Therefore, consider configuring a back-update operation, such as a jump, or configuring an approximately differentiable function, or assuming that the differentiable function in the back-update phase is approximately differentiable, to avoid model training failure.

[0433] For example, this reverse operation can be indicated by configuration information (such as one of configuration information #1 to configuration information #3). This configuration information can indicate an index, and the update operation corresponding to that index is the reverse update operation. Specifically, the correspondence between different indices and different update operations can include the content shown in Table 12 below:

[0434] Table 12

[0435] It should be understood that Table 12 above exemplarily lists some possible implementations of the reverse update operation in this application, and does not mean that the reverse update operation only includes the implementations shown in Table 12 above. In fact, the implementation of the reverse update operation can also be other implementations besides those in Table 12 above. This application does not limit it, as long as the model training can be completed.

[0436] It should be noted that the correspondences described above are merely illustrative and do not represent that the relevant parameters only include the contents shown in the above correspondences; the relevant parameters may also include other relevant values ​​besides those in the above correspondences. Furthermore, while the above examples use tables to illustrate the correspondences, in reality, the correspondences can be represented in the form of lists, sets, etc., and this application does not impose any restrictions.

[0437] It should be noted that the various embodiments of this application can be implemented independently or in combination, without limitation. Unless otherwise specified or in conflict, the terminology and / or descriptions between the different embodiments provided in this application are consistent and can be referenced mutually. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0438] The foregoing primarily describes the solutions provided in this application from the perspective of device-to-device interaction. It is understood that each device, in order to achieve the aforementioned functions, includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, based on the algorithm steps of the examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0439] It is understood that, in order to achieve the aforementioned functions, the communication device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0440] This application embodiment can divide each device into functional modules according to the above method example. For example, each function can be divided into a separate functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0441] Figure 23 shows a schematic diagram of a communication device 2300. The communication device 2300 includes a processing module 2301 and a transceiver module 2302. This communication device can be used to implement the functions of the aforementioned transmitting or receiving devices.

[0442] In some embodiments, the communication device 2300 may further include a storage module (not shown in FIG23) for storing program instructions and data.

[0443] In some embodiments, the transceiver module 2302, also referred to as a transceiver unit, is used to implement sending and / or receiving functions. The transceiver module 2302 may consist of a transceiver circuit, a transceiver, a transceiver unit, or a communication interface.

[0444] In some embodiments, the transceiver module 2302 may include a receiving module and a sending module, respectively configured to perform the receiving or sending actions in the above method embodiments (i.e., the receiving and sending steps performed by the sending or receiving device), and / or other processes to support the technology described herein; the processing module 2301 may be configured to perform the processing actions in the above method embodiments (i.e., the processing steps (e.g., determining, acquiring, etc.) performed by the sending or receiving device), and / or other processes to support the technology described herein.

[0445] When the communication device 2300 is used to implement the functions of the aforementioned transmitting device:

[0446] In some embodiments, the transceiver module 2302 is configured to receive first information, which is compressed information. The processing module 2301 is configured to obtain a reasoning result based on the first information and a first reasoning model, wherein the first reasoning model is obtained by training a pre-configured model and a first module, and the first module is used to compress and / or decompress data.

[0447] Optionally, the transceiver module 2302 is further configured to receive first capability information, which indicates the inference capability of the transmitting device. The processing module 2301 is further configured to determine a first inference model and / or a second inference model based on the first capability information. The transceiver module 2302 is further configured to send first indication information, which indicates the first inference model and / or the second inference model.

[0448] Optionally, the processing module 2301 is further configured to acquire second capability information, which indicates the inference capability of the receiving device. The processing module 2301 is also configured to determine a first inference model and / or a second inference model based on the second capability information. The transceiver module 2302 is further configured to send second indication information, which indicates the first inference model and / or the second inference model.

[0449] Optionally, the processing module 2301 is also used to reason about the first information based on the first reasoning model to obtain the reasoning result.

[0450] Optionally, the processing module 2301 is also used to decompress the first information to obtain the data to be reasoned; and to reason about the data to be reasoned according to the first reasoning model to obtain the reasoning result.

[0451] Optionally, the processing module 2301 is further configured to obtain configuration information, the configuration information indicating a first decompression method and / or a first decompression parameter, the first decompression method being the decompression method adopted by the first decompression module, and the first decompression parameter being the decompression parameter corresponding to the first decompression method.

[0452] Optionally, the transceiver module 2302 is also used to receive configuration information.

[0453] Optionally, the processing module 2301 is also used to determine configuration information; the transceiver module 2302 is also used to send configuration information.

[0454] Optionally, the processing module 2301 is further configured to decompress the first information using the first decompression module to obtain the data to be inferred, wherein the first module is related to the first decompression module.

[0455] When the communication device 2300 is used to implement the functions of the aforementioned receiving device:

[0456] In some embodiments, the processing module 2301 is used to infer intermediate features from the data to be sent based on the second inference model; and to compress the intermediate data to obtain first information. The second inference model is obtained through model training based on the first sub-model, the second sub-model, and the first module in the pre-configured model. The sum of the model layers of the first sub-model and the second sub-model is equal to the model layers of the pre-configured model. The first module is used to compress and / or decompress the data. The transceiver module 2302 is used to send the first information.

[0457] Optionally, the processing module 2301 is further configured to acquire first capability information, which indicates the inference capability of the transmitting device; and determine a first inference model and / or a second inference model based on the first capability information; the transceiver module 2302 is further configured to send third indication information, which indicates the first inference model and / or the second inference model.

[0458] Optionally, the transceiver module 2302 is further configured to receive second capability information, which indicates the inference capability of the receiving device; the processing module 2301 is further configured to determine a first inference model and / or a second inference model based on the second capability information; and the transceiver module 2302 is further configured to send fourth indication information, which indicates the first inference model and / or the second inference model.

[0459] Optionally, the processing module 2301 is also used to compress the intermediate data using the first compression module to obtain the first information, wherein the first module is related to the first decompression module.

[0460] Optionally, the processing module 2301 is also used to obtain configuration information, the configuration information indicating a first compression method and / or a first compression parameter, the first compression method being the compression method adopted by the first compression module, and the first compression parameter being the compression parameter corresponding to the first compression method.

[0461] All relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0462] In this application, the communication device (such as a transmitting device or a receiving device) 2300 is presented in an integrated manner, divided into various functional modules. Here, "module" may refer to an application-specific integrated circuit (ASIC), a circuit, a processor and memory that executes one or more software or firmware programs, integrated logic circuits, and / or other devices that can provide the above functions.

[0463] In some embodiments, those skilled in the art will recognize that the communication device 2300 can take the form of the communication device 600 shown in FIG6 in terms of hardware implementation.

[0464] As an example, the function / implementation process of the processing module 2301 in Figure 23 can be implemented by the processor 601 in the communication device 600 shown in Figure 6 calling computer execution instructions stored in the memory 604. The function / implementation process of the transceiver module 2302 in Figure 23 can be implemented by the communication interface 602 in the communication device 600 shown in Figure 6.

[0465] In some embodiments, when the communication device 2300 in FIG23 is a chip or chip system, the function / implementation process of the transceiver module 2302 can be implemented through the input / output interface (or communication interface) of the chip or chip system, and the function / implementation process of the processing module 2301 can be implemented through the processor (or processing circuit) of the chip or chip system.

[0466] Since the communication device 2300 provided in this embodiment can execute the above method, the technical effects it can achieve can be referred to the above method embodiment, and will not be repeated here.

[0467] As a possible product form, the transmitting or receiving device described in the embodiments of this application can also be implemented using one or more FPGAs, programmable logic devices (PLDs), controllers, state machines, gate logic, discrete hardware components, any other suitable circuits, or any combination of circuits capable of performing the various functions described throughout this application.

[0468] As another possible product form, the transmitting or receiving device described in this application embodiment can be implemented using a general bus architecture. For ease of explanation, refer to FIG24, which is a schematic diagram of the structure of a communication device 2400 provided in an embodiment of this application. The communication device 2400 includes a processor 2401 and a transceiver 2402. The communication device 2400 can be a transmitting device, or a chip or chip system therein; or, the communication device 2400 can be a receiving device, or a chip or module therein. FIG24 only shows the main components of the communication device 2400. In addition to the processor 2401 and transceiver 2402, the communication device may further include a memory 2403.

[0469] Optionally, the processor 2401 is mainly used to process communication protocols and communication data, control the entire communication device, execute software programs, and process the data of the software programs. The memory 2403 is mainly used to store software programs and data. The transceiver 2402 may include radio frequency (RF) circuitry and an antenna. The RF circuitry is mainly used for converting baseband signals to RF signals and processing RF signals. The antenna is mainly used for transmitting and receiving RF signals in the form of electromagnetic waves.

[0470] Optionally, the processor 2401, transceiver 2402, and memory 2403 can be connected via a communication bus.

[0471] When the communication device is powered on, the processor 2401 can read the software program in the memory 2403, interpret and execute the instructions of the software program, and process the data of the software program. When data needs to be transmitted wirelessly, the processor 2401 performs baseband processing on the data to be transmitted and outputs the baseband signal to the radio frequency (RF) circuit. The RF circuit processes the baseband signal and transmits the RF signal outward in the form of electromagnetic waves through the antenna. When data is sent to the communication device, the RF circuit receives the RF signal through the antenna, converts the RF signal into a baseband signal, and outputs the baseband signal to the processor 2401. The processor 2401 converts the baseband signal into data and processes the data.

[0472] In another implementation, the radio frequency circuitry and antenna can be set up independently of the processor performing baseband processing. For example, in a distributed scenario, the radio frequency circuitry and antenna can be arranged remotely, independent of the communication device.

[0473] In some embodiments, this application also provides a communication device, which includes a processor for implementing the methods in any of the above method embodiments. The communication device can be a transmitting device or a receiving device as described in the above method embodiments.

[0474] As one possible implementation, the communication device also includes a memory. This memory stores necessary computer programs and data. The computer program may include instructions, which a processor can invoke to instruct the communication device to execute the methods described in any of the above method embodiments. Alternatively, the memory may not be present in the communication device.

[0475] As another possible implementation, the communication device also includes an interface circuit, which is a code / data read / write interface circuit, used to receive computer execution instructions (which are stored in memory and may be read directly from memory or may be transmitted through other devices) and transmit them to the processor.

[0476] As another possible implementation, the communication device also includes a communication interface for communicating with modules outside the communication device.

[0477] It is understood that the communication device can be a chip or a chip system. When the communication device is a chip system, it can be composed of chips or may include chips and other discrete devices. This application does not specifically limit this.

[0478] This application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a computer, implements the functions of any of the above-described method embodiments.

[0479] This application also provides a computer program product that, when executed by a computer, implements the functions of any of the above method embodiments.

[0480] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0481] It is understood that the systems, apparatuses, and methods described in this application can also be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0482] The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. The components shown as units may or may not be physical units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0483] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0484] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive (SSD)). In this embodiment, the computer may include the aforementioned apparatus.

[0485] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

Claims

1. A communication method characterized by comprising: The method comprises: receiving first information, the first information being compressed information; obtaining an inference result according to the first information and a first inference model, the first inference model being obtained by model training based on a preconfigured model and a first module, the first module being used for compressing and / or decompressing data.

2. The method of claim 1, wherein, The first information is obtained by compressing intermediate features output by a second inference model; The first inference model is obtained by model training based on a preconfigured model and a first module, comprising: The first inference model and the second inference model are obtained by training based on a first sub-model, a second sub-model in the preconfigured model, and the first module, and the sum of the number of model layers of the first sub-model and the number of model layers of the second sub-model is equal to the number of model layers of the preconfigured model.

3. The method according to claim 1 or 2, characterized in that, The method is performed by a receiving end device, and the receiving first information comprises: receiving the first information from a sending end device; Before the receiving the first information from the sending end device, the method further comprises: receiving first capability information, the first capability information indicating inference capability of the sending end device; determining the first inference model and / or a second inference model according to the first capability information, the second inference model being obtained by training based on a first sub-model, a second sub-model in the preconfigured model, and the first module, and the sum of the number of model layers of the first sub-model and the number of model layers of the second sub-model being equal to the number of model layers of the preconfigured model; sending first indication information, the first indication information indicating the first inference model and / or the second inference model.

4. The method according to claim 1 or 2, characterized in that, The method is performed by a receiving end device, and before the receiving first information, the method further comprises: obtaining second capability information, the second capability information indicating inference capability of the receiving end device; determining the first inference model and / or a second inference model according to the second capability information, the second inference model being obtained by training based on a first sub-model, a second sub-model in the preconfigured model, and the first module, and the sum of the number of model layers of the first sub-model and the number of model layers of the second sub-model being equal to the number of model layers of the preconfigured model; sending second indication information, the second indication information indicating the first inference model and / or the second inference model.

5. The method according to any one of claims 1 to 4, characterized in that, The obtaining an inference result according to the first information and a first inference model comprises: performing inference on the first information according to the first inference model to obtain the inference result.

6. The method according to any one of claims 1 to 4, characterized in that, The obtaining an inference result according to the first information and a first inference model comprises: decompressing the first information to obtain to-be-inferred data; performing inference on the to-be-inferred data according to the first inference model to obtain the inference result.

7. The method of claim 6, wherein, The decompressing the first information to obtain to-be-inferred data comprises: decompressing the first information by using a first decompression module to obtain the to-be-inferred data, the first module being related to the first decompression module.

8. The method of claim 7, wherein, The first module is of the same decompression mode as the first decompression module, and / or the type of the decompression parameter corresponding to the decompression mode used by the first module is the same as the type of the first decompression parameter.

9. The method according to claim 7 or 8, characterized in that, The method further comprises: obtaining configuration information, the configuration information indicating a first decompression mode and / or a first decompression parameter, the first decompression mode being a decompression mode used by the first decompression module, and the first decompression parameter being a decompression parameter corresponding to the first decompression mode.

10. The method of claim 9, wherein, The method further comprises: receiving the configuration information.

11. The method of claim 9, wherein, The method further comprises: determining the configuration information. The method further comprises: sending the configuration information.

12. The method according to any one of claims 9-11, characterized in that, The configuration information indicates a first decompression mode and / or a first decompression parameter. The configuration information indicates a first compression mode and / or a first compression parameter, the first compression mode corresponding to the first decompression mode, and the first compression parameter corresponding to the first decompression parameter.

13. A method of communication, comprising: The method comprises: performing inference on the to-be-sent data according to a second inference model to obtain intermediate features, the second inference model being obtained through model training based on a first sub-model, a second sub-model, and a first module in a pre-configured model, the sum of the number of model layers of the first sub-model and the number of model layers of the second sub-model being equal to the number of model layers of the pre-configured model, and the first module being used for compressing and / or decompressing data; compressing the intermediate data to obtain first information; sending the first information.

14. The method of claim 13, wherein, The method is performed by a sending end device, and before the performing inference on the to-be-sent data according to the second inference model to obtain intermediate features, the method further comprises: obtaining first capability information, the first capability information indicating an inference capability of the sending end device; determining the first inference model and / or the second inference model according to the first capability information; sending third indication information, the third indication information indicating the first inference model and / or the second inference model.

15. The method of claim 14, wherein, The sending of the first information comprises sending the first information to a receiving end device. Before the performing inference on the to-be-sent data according to the second inference model to obtain intermediate features, the method further comprises: receiving second capability information, the second capability information indicating an inference capability of the receiving end device; determining the first inference model and / or the second inference model according to the second capability information; sending fourth indication information, the fourth indication information indicating the first inference model and / or the second inference model.

16. The method according to any one of claims 13-15, characterized in that, The compressing of the intermediate data to obtain first information comprises: compressing the intermediate data by using a first compression module to obtain the first information, the first module being related to the first compression module.

17. The method of claim 16, wherein, The first module is of the same compression mode as the first compression module, and / or the type of the compression parameter corresponding to the compression mode used by the first module is the same as the type of the first compression parameter.

18. The method according to claim 16 or 17, characterized in that The method further comprises: Obtaining configuration information, the configuration information indicating a first compression mode and / or a first compression parameter, the first compression mode being a compression mode adopted by the first compression module, the first compression parameter being a compression parameter corresponding to the first compression mode.

19. The method of claim 18, wherein, The obtaining configuration information comprises: Determining the configuration information; The method further comprises: Sending the configuration information.

20. The method of claim 19, wherein, The obtaining configuration information comprises: Receiving the configuration information.

21. The method of any one of claims 9-12, 18-20, wherein, When the configuration information indicates the first compression mode, the configuration information comprises a first index, the first compression mode being a compression mode corresponding to the first index in a first correspondence relationship, the first correspondence relationship comprising a correspondence relationship among a plurality of compression modes and a plurality of indexes.

22. The method of any one of claims 9-12, 18-20, wherein, When the configuration information indicates the first compression mode, the configuration information comprises a second index, the first compression mode being a compression mode corresponding to the second index and a first data type in a second correspondence relationship, the first data type being a data type of the to-be-inferred data, the second correspondence relationship comprising a correspondence relationship among a plurality of data types, a plurality of compression modes and a plurality of indexes.

23. The method of claim 22, wherein, The configuration information further comprises the first data type.

24. The method of any one of claims 21-23, wherein, The plurality of compression modes comprises one or more of fixed quantization compression, 2-dimensional (2D) projection compression, dictionary compression, or sampling compression.

25. The method of claim 24, wherein, When the compression mode is the fixed quantization compression or the 2D projection compression, the type of the compression parameter is a quantization bit number; When the compression mode is the dictionary compression, the type of the compression parameter comprises a weighting parameter corresponding to the dictionary compression and / or a dimension of a dictionary matrix; When the compression mode is the sampling compression, the type of the compression parameter is a sampling rate and / or a sampling position.

26. A communications device, characterized by The communication device comprises a transceiver module and a processing module, The transceiver module is configured to perform a receiving action or a sending action in the method of any one of claims 1-12, 21-25, or perform a receiving action or a sending action in the method of any one of claims 13-25; The processing module is configured to perform a processing action in the method of any one of claims 1-12, 21-25, or perform a processing action in the method of any one of claims 13-25.

27. A communications device, characterized by The communication device comprises a processor, and the processor is configured to run a computer program or instructions to cause the communication device to perform the method of any one of claims 1-12, 21-25, or to cause the communication device to perform the method of any one of claims 13-25.

28. A computer-readable storage medium, characterized in that, A computer readable storage medium stores computer instructions or programs, which, when run on a computer, cause the method of any one of claims 1-12, 21-25 to be performed, or cause the method of any one of claims 13-25 to be performed.

29. A computer program product, characterised in that, The computer program product comprises computer instructions; when part or all of the computer instructions are run on a computer, the method as claimed in any one of claims 1-12, 21-25 is caused to be performed, the method as claimed in any one of claims 13-25 is caused to be performed.

Images