Method and apparatus for training model, communication device, and computer-readable storage medium

Abstract

The present application discloses a method and apparatus for training a model, and a communication device, the method comprising: on the basis of a first data set, a network device trains a first model, the first model comprising a first part and a second part; and sends a second data set to a terminal device, the first data set comprising M first samples, the second data set comprising N second samples, the N second samples being obtained by N first samples being processed by the first part of the first model, and M and N being positive integers. The second data set is used by the terminal device to train a second model, and the trained second model and the second part of the first model are used for CSI reporting.

Description

Model training method and device, communication equipment, and computer-readable storage medium

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application is based on the Chinese patent application with application number 202311394364.8 and application date of October 25, 2023, and claims the priority of the Chinese patent application. The entire content of the Chinese patent application is hereby introduced into this application as a reference. Technical Field

[0003] The present application relates to the field of wireless communication technology, and in particular to a model training method and apparatus, communication equipment, and computer-readable storage medium. Background Art

[0004] Introducing artificial intelligence (AI) into the radio access network (RAN) can improve the performance of the wireless air interface. A typical application is AI-based channel state information (CSI) feedback.

[0005] AI-based CSI feedback can be implemented using AI models (referred to as models). In some scenarios, a pair of AI models must be deployed on both the terminal device and the network equipment side, known as a dual-sided AI model. The terminal device and network equipment each use their own AI model to implement CSI feedback. Currently, there is a lack of comprehensive technical solutions for training a pair of AI models for CSI feedback.

[0006] Summary of the Invention

[0007] The embodiments of the present application provide a model training method and apparatus, communication equipment, chip, and computer-readable storage medium.

[0008] The model training method provided in the embodiment of the present application includes:

[0009] The network device trains a first model based on a first data set, where the first model includes a first part and a second part, and sends a second data set to the terminal device, where the first data set includes M first samples, and the second data set includes N second samples, where the N second samples are obtained by processing the N first samples by the first part of the first model, where M and N are positive integers;

[0010] The second data set is used for the terminal device to train a second model, and the trained second model and the second part of the first model are used for CSI reporting.

[0011] The model training method provided in the embodiment of the present application includes:

[0012] The terminal device trains a third model based on the fifth data set, the third model including a first part and a second part, and sends a sixth data set to the network device, where the fifth data set includes M first samples, the sixth data set includes N first samples and N second samples, the N second samples being obtained by processing the N first samples through the first part of the third model, where M and N are positive integers;

[0013] The sixth data set is used by the network device to train a fourth model, and the trained fourth model and the first part of the third model are used for CSI reporting.

[0014] The model training device provided in the embodiment of the present application is applied to a network device, and the device includes:

[0015] a training unit configured to train a first model based on a first data set, wherein the first model includes a first part and a second part;

[0016] a sending unit configured to send a second data set to a terminal device, wherein the first data set includes M first samples, and the second data set includes N second samples, where the N second samples are obtained by processing the N first samples by the first part of the first model, and M and N are positive integers;

[0017] The second data set is used for the terminal device to train a second model, and the trained second model and the second part of the first model are used for CSI reporting.

[0018] The model training device provided in the embodiment of the present application is applied to a terminal device, and the device includes:

[0019] a training unit configured to train a third model based on a fifth data set, wherein the third model includes a first part and a second part;

[0020] a sending unit configured to send a sixth data set to a network device, wherein the fifth data set includes M first samples, the sixth data set includes N first samples and N second samples, the N second samples being obtained by processing the N first samples by the first part of the third model, and M and N are positive integers;

[0021] The sixth data set is used by the network device to train a fourth model, and the trained fourth model and the first part of the third model are used for CSI reporting.

[0022] The communication device provided in an embodiment of the present application includes: a processor and a memory, the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory to execute any one of the above-mentioned model training methods.

[0023] The chip provided in an embodiment of the present application includes: a processor, which is used to call and run a computer program from a memory, so that a device equipped with the chip executes any of the above methods.

[0024] The computer-readable storage medium provided in an embodiment of the present application is used to store a computer program, and the computer program enables a computer to execute any one of the above methods.

[0025] In the above technical solution of the embodiment of the present application, the network device initiates model training. Specifically, the network device trains the first model based on the first data set, and the network device sends the second data set to the terminal device, so that the terminal device trains the second model based on the second data set, and the trained second model and the second part of the first model (which can be understood as a pair of AI models) are used for CSI reporting. In this way, the training method of a pair of AI models for CSI reporting is clarified. Alternatively, the terminal device initiates model training. Specifically, the terminal device trains the third model based on the fifth data set, and the terminal device sends the sixth data set to the network device, so that the network device trains the fourth model based on the sixth data set, and the trained fourth model and the first part of the third model (which can be understood as a pair of AI models) are used for CSI reporting. In this way, the training method of a pair of AI models for CSI reporting is clarified. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0027] FIG1 is a schematic diagram of an application scenario of an embodiment of the present application;

[0028] FIG2 is a schematic diagram of AI-based CSI feedback;

[0029] Figure 3 is a schematic diagram of independently training an AI model;

[0030] FIG4 is a flow chart of a model training method according to an embodiment of the present application;

[0031] FIG5 is a second flow chart of the model training method provided in an embodiment of the present application;

[0032] FIG6 is a schematic diagram of the structure of the model training device provided in an embodiment of the present application;

[0033] FIG7 is a second schematic diagram of the structure of the model training device provided in an embodiment of the present application;

[0034] FIG8 is a schematic structural diagram of a communication device provided in an embodiment of the present application;

[0035] FIG9 is a schematic structural diagram of a chip according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] The following will describe the technical solutions in the embodiments of this application in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0037] FIG1 is a schematic diagram of an application scenario of an embodiment of the present application.

[0038] As shown in Figure 1, a communication system 100 may include a terminal device 110 and a network device 120. The network device 120 may communicate with the terminal device 110 via an air interface. The terminal device 110 and the network device 120 support multi-service transmission.

[0039] It should be understood that the embodiments of the present application are only illustrative of the communication system 100, but the embodiments of the present application are not limited thereto. In other words, the technical solutions of the embodiments of the present application can be applied to various communication systems, such as: 5G communication systems (also known as New Radio (NR) communication systems), or future communication systems.

[0040] In the communication system 100 shown in Figure 1, the network device 120 may be an access network device that communicates with the terminal device 110. The access network device may provide communication coverage for a specific geographical area and may communicate with the terminal device 110 (eg, UE) located within the coverage area.

[0041] The network device 120 may be a base station (gNB) in an NR system, or a network device in a future evolved public land mobile network (PLMN).

[0042] The terminal device 110 may be any terminal device, including but not limited to a terminal device connected to the network device 120 or other terminal devices by wire or wireless connection.

[0043] For example, the terminal device 110 may refer to an access terminal, user equipment (UE), a subscriber unit, a subscriber station, a mobile station, a mobile station, a remote station, a remote terminal, a mobile device, a user terminal, a terminal, a wireless communication device, a user agent, or a user apparatus. An access terminal may be a cellular phone, a cordless phone, a Session Initiation Protocol (SIP) phone, an IoT device, a satellite handheld terminal, a Wireless Local Loop (WLL) station, a Personal Digital Assistant (PDA), a handheld device with wireless communication capabilities, a computing device or other processing device connected to a wireless modem, an in-vehicle device, a wearable device, a terminal device in a 5G network, or a terminal device in a future evolution network, etc.

[0044] FIG1 exemplarily shows a base station and two terminal devices. Optionally, the wireless communication system 100 may include multiple base station devices and each base station may include other numbers of terminal devices within its coverage area, which is not limited in the embodiments of the present application.

[0045] It should be noted that Figure 1 is merely an example of a system applicable to this application. Of course, the methods described in the embodiments of this application can also be applied to other systems. Furthermore, the terms "system" and "network" are often used interchangeably herein. The term "and / or" herein simply describes an association relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " generally indicates that the associated objects are in an "or" relationship. It should also be understood that the "indication" mentioned in the embodiments of this application can be a direct indication, an indirect indication, or an indication of an association relationship. For example, "A indicates B" can mean that A directly indicates B, for example, B can obtain information through A; it can also mean that A indirectly indicates B, for example, A indicates C, and B can obtain information through C; or it can mean that A and B have an association relationship. It should also be understood that the "correspondence" mentioned in the embodiments of this application can mean that there is a direct or indirect correspondence between two objects, or that there is an association relationship between the two objects, or a relationship between an indicator and the indicated, a configuration and the configured, and so on. It should also be understood that the “predefined” or “predefined rules” mentioned in the embodiments of the present application can be implemented by pre-saving corresponding codes, tables, or other methods that can be used to indicate relevant information in a device (for example, including a terminal device and a network device). The present application does not limit its specific implementation method. For example, predefined may refer to a definition in a protocol. It should also be understood that in the embodiments of the present application, the “protocol” may refer to a standard protocol in the field of communications, for example, it may include an NR protocol and related protocols used in future communication systems, and the present application does not limit this.

[0046] To facilitate understanding of the technical solutions of the embodiments of the present application, the relevant technologies of the embodiments of the present application are described below. The following relevant technologies can be arbitrarily combined with the technical solutions of the embodiments of the present application as optional solutions, and they all fall within the protection scope of the embodiments of the present application.

[0047] Traditional codebook-based (such as Type-I / Type-II / e-Type-II codebook) CSI feedback has problems such as limited feedback accuracy or large feedback overhead. AI-based CSI feedback can effectively solve these problems of traditional solutions. The AI-based CSI feedback process is as follows: the terminal device inputs the channel information obtained by CSI measurement into the model on the terminal device side, and the model processes the channel information to obtain CSI quantization bits; the terminal device sends the CSI quantization bits to the network device. After receiving the CSI quantization bits, the network device inputs the CSI quantization bits into the model on the network device side, and the model recovers the corresponding channel information from the CSI quantization bits. The channel information includes eigenvectors, beam information, delay information, etc.

[0048] Figure 2 shows a schematic diagram of AI-based CSI feedback, where feature vector feedback is used as an example. The AI ​​model deployed on the terminal device side includes an encoder module and a quantization module, and the AI ​​model deployed on the network device side includes a dequantization module and a decoding module. The AI-based feature vector feedback process is as follows: the encoding module and quantization module on the terminal device side compress and quantize the feature vectors of k (k ≥ 1) subbands respectively, and feed back the quantized bits (i.e., the quantized bits of the feature vectors) to the network device. After receiving the bits fed back by the terminal device, the network device dequantizes and decompresses them through the dequantization module and the decoding module respectively, thereby recovering the feature vectors of the k subbands.

[0049] In the AI-based CSI feedback process described above, a pair of AI models must be deployed on the terminal device side and the network device side, respectively. This is referred to as a two-sided AI model. In actual model deployment, there are three ways to implement the training of a pair of AI models.

[0050] (1) The network device trains a pair of AI models (including CSI generation and CSI reconstruction) and sends the CSI generation part to the terminal device. This method can be called a network device-initiated training method. Alternatively, the terminal device trains a pair of AI models (including CSI generation and CSI reconstruction) and sends the CSI reconstruction part to the network device. This method can be called a terminal device-initiated training method.

[0051] (2) The network device side and the terminal device side jointly train a pair of AI models (including the CSI generation part and the CSI recovery part), where the CSI generation part is deployed on the terminal device side and the CSI recovery part is deployed on the network device side.

[0052] (3) The network device side and the terminal device side each train their own AI model. The terminal device side trains the CSI generation part, and the network device side trains the CSI recovery part. The CSI generation part is deployed on the terminal device side, and the CSI recovery part is deployed on the network device side.

[0053] It should be noted that the CSI generation part mentioned above represents the AI ​​model on the terminal device side of the pair of AI models, which may actually include an encoder module and a quantization module. The CSI recovery part mentioned above represents the AI ​​model on the network device side of the pair of AI models, which may actually include a dequantization module and a decoder module.

[0054] In the above method (1), it is necessary to transmit the model between the network device and the terminal device. The model transmitted by the network device to the terminal device may not match the model supported by the terminal device, or the model transmitted by the terminal device to the network device may not match the model supported by the network device.

[0055] In the above-mentioned method (2), a forward propagation process and a backward propagation process need to be performed between the network device and the terminal device, resulting in a large air interface overhead.

[0056] In the above method (3), it is necessary to ensure that the models trained on the network device side and the terminal device side can be used together and meet the performance requirements.

[0057] In some technical solutions, improvements can be made based on the above method (3). Taking the training method initiated by the network device as an example, a potential training method is: after the network device trains the CSI generation part 1 and the CSI recovery part (i.e., a pair of AI models) based on the data set A, it sends the data set B to the terminal device. The data set B includes N (N≥1) first samples and N second samples obtained after these N first samples are processed by the CSI generation part. The terminal device trains the CSI generation part 2 based on the received data set B to ensure that it can be used in conjunction with the CSI recovery part on the network device side. Referring to Figure 3, the process of training their respective AI models on the network device side and the terminal device side is as follows:

[0058] ① On the network equipment side, train the CSI generation part 1 and the CSI recovery part (i.e., a pair of AI models), where the input of the CSI generation part 1 is Vin (i.e., dataset A), the output of the CSI generation part 1 is Vq, the input of the CSI recovery part is Vq, and the output of the CSI recovery part is Vout.

[0059] ② The network device sends data set B (including N first samples and N second samples after these N first samples pass through the CSI generation part on the network device side) to the terminal device.

[0060] ③ The terminal device trains CSI generation part 2, where the input of CSI generation part 2 is Vin and the output of CSI generation part 2 is Vq. The terminal device trains CSI generation part 2 based on data set B.

[0061] ④ Combine the CSI recovery part obtained by training on the network device side and the CSI generation part 2 obtained by training on the terminal device side as a pair of AI models for CSI feedback.

[0062] Assuming that the CSI recovery portion of the network device uses a transformer model structure, the simulation results of the terminal device under different sample sizes of dataset B and using different model structures are shown in Table 1 below. The square of generalized cosine similarity (SGCS) of the channel information is used as the simulation result for illustration. Among them, Case 1 is a joint training case, that is, the network device trains a pair of AI models (including CSI generation part 1 and CSI recovery part); Case 2 is a separate training case, that is, the terminal device side trains CSI generation part 2 (using the Transformer model structure), and the network device side trains the CSI recovery part (using the Transformer model structure); Case 3 is another independent training case, that is, the terminal device side trains CSI generation part 3 (using the EVCsiNet model structure), and the network device side trains the CSI recovery part (using the Transformer model structure); Case 4 is another independent training case, that is, the terminal device side trains CSI generation part 4 (using the MLP-Mixer model structure), and the network device side trains the CSI recovery part (using the Transformer model structure).

[0063] Table 1

[0064] It should be noted that in Table 1, for simplicity, the CSI generation part is described as the encoder, and the CSI recovery part is described as the decoder. In Case 1, encoder 1 and decoder are models trained on the network device side based on training dataset A and using the transformer model structure. In Case 2, encoder 2 is a model trained on the terminal device side based on dataset B and using the transformer model structure. In Case 3, encoder 3 is a model trained on the terminal device side based on dataset B and using the EVCsiNet model structure. In Case 4, encoder 4 is a model trained on the terminal device side based on dataset B and using the MLP-Mixer model structure. Among them, Case 2, Case 3, and Case 4 all assume that the terminal device side and the network device side use the same quantization / dequantization method.

[0065] From the above simulation results, it can be seen that in order to realize model training, the terminal device needs to obtain data set B. Furthermore, the terminal device also needs to obtain information about the quantization method and basic model structure adopted by the network device.

[0066] In some technical solutions, training can also be initiated by the terminal device. The method for initiating training by the terminal device is similar to the method for initiating training by the network device mentioned above. One potential training method is: After the terminal device trains the CSI generation part and the CSI recovery part 1 (i.e., a pair of AI models) based on data set A, it sends data set B to the network device. Data set B includes M (M ≥ 1) first samples and M second samples obtained after these M second samples are processed by the CSI generation part. The network device trains the CSI recovery part 2 based on the received data set B to ensure that it can be used in conjunction with the CSI generation part on the terminal device side.

[0067] Furthermore, after the CSI generation and CSI recovery components are trained and deployed on the terminal device and network equipment, respectively, if the scenario or configuration of the terminal device or network equipment changes, such as when the terminal device switches to another cell or the number of transmit antennas or subbands configured on the network equipment changes, the originally deployed model may not be suitable for the new scenario. Therefore, the originally deployed model needs to be updated or fine-tuned.

[0068] To facilitate understanding of the technical solutions of the embodiments of the present application, the technical solutions of the present application are described in detail below through specific embodiments. The above related technologies can be combined arbitrarily with the technical solutions of the embodiments of the present application as optional solutions, and all of them fall within the scope of protection of the embodiments of the present application. The embodiments of the present application include at least part of the following contents.

[0069] It should be noted that the "CSI reporting" described in the embodiments of this application can also be described as "CSI feedback". Specifically, the "CSI feedback" refers to AI-based CSI feedback. The process of AI-based CSI feedback can refer to the relevant description above.

[0070] FIG4 is a flow chart of a model training method according to an embodiment of the present application. As shown in FIG4 , the model training method includes:

[0071] Step 401: The network device trains a first model based on a first data set, where the first model includes a first part and a second part, and sends the second data set to the terminal device. The first data set includes M first samples, and the second data set includes N second samples, where the N second samples are obtained by processing the N first samples through the first part of the first model, and M and N are positive integers. The second data set is used by the terminal device to train a second model, and the trained second model and the second part of the first model are used for CSI reporting.

[0072] In this embodiment of the present application, the network device is responsible for training the first model, while the terminal device is responsible for training the second model. The first model includes a first part and a second part, and the trained second model and the second part of the first model are used for CSI reporting. The second part of the first model and the second model can be understood as a pair of AI models for CSI reporting, i.e., a two-sided AI model.

[0073] In some implementations, the first part of the first model is a CSI generation part (CSI generation part), and the second part of the first model is a CSI reconstruction part (CSI reconstruction part).

[0074] In some embodiments, the second model is a CSI generation part. In order to distinguish the CSI generation part in the first model from the CSI generation part in the second model, the CSI generation part in the first model can be referred to as CSI generation part 1, and the CSI generation part in the second model can be referred to as CSI generation part 2.

[0075] It should be noted that the CSI generation part 2 described above represents the AI ​​model on the terminal device side of the pair of AI models, which may actually include an encoding module and a quantization module. The CSI recovery part described above represents the AI ​​model on the network device side of the pair of AI models, which may actually include a dequantization module and a decoding module. The functions of the encoding module, quantization module, dequantization module, and decoding module can be referred to in the relevant descriptions above.

[0076] In the embodiment of the present application, the network device trains a first model based on a first data set, and the terminal device trains a second model based on a second data set. The first data set can also be described as data set A, and the second data set can also be described as data set B, which is not limited in this application.

[0077] In the embodiment of the present application, the first data set includes M first samples, where M is a positive integer, and the first samples are channel information samples (also referred to as channel samples). The terminal device performs CSI measurement to obtain channel information and reports the measured channel information to the network device, so that the network device can obtain the channel information and use the channel information as samples for training the first model (i.e., channel information samples).

[0078] In some embodiments, the network device may obtain M first samples from one terminal device, that is, the M first samples are obtained by performing CSI measurement on one terminal device. In other embodiments, the network device may obtain M first samples from multiple terminal devices, that is, the M first samples are obtained by performing CSI measurement on multiple terminal devices.

[0079] After obtaining M first samples, the network device trains the first model based on the M first samples, that is, trains the first part and the second part of the first model. For example, the network device trains the CSI generation part 1 and the CSI recovery part based on the M first samples.

[0080] In the embodiment of the present application, the second data set includes N second samples, where N is a positive integer, and the N second samples are obtained by processing the N first samples through the first part of the first model. Here, the N second samples are obtained by processing the N first samples through the first part of the first model, specifically referring to: the N second samples are obtained by processing the N first samples through the first part of the trained first model.

[0081] In some embodiments, the N first samples are identical to the M first samples. In other embodiments, the N first samples are partially identical to the M first samples. In other embodiments, the N first samples are completely different from the M first samples.

[0082] The network device inputs N first samples into the first model, which then processes the first part of the first model to obtain N second samples. The second samples are quantized bit samples corresponding to the channel information. The N second samples constitute a second data set, which the network device transmits to the terminal device for use in training the second model. The format of the second data set is described below.

[0083] Solution 1

[0084] In some embodiments, the second data set includes N first samples and N second samples, where the N second samples are obtained by processing the N first samples with the first part of the first model.

[0085] Here, the second data set has a first format; in the first format, the second data set includes N groups of data, each of the N groups of data includes a first sample and a second sample corresponding to the first sample; wherein the N groups of data are arranged according to sample indexes.

[0086] For example, the N groups of data in the second data set may be arranged in {datasample1, datasample1′, datasample2, datasample2′, datasample3, datasample3′, …, datasampleN, datasampleN′}, where datasample1 and datasample1′ are the first group of data, datasample2 and datasample2′ are the second group of data, and so on. datasample1 represents the first sample with a sample index of 1, datasample1′ represents the second sample obtained by processing datasample1 with the first part of the first model, datasample2 represents the first sample with a sample index of 2, datasample2′ represents the second sample obtained by processing datasample2 with the first part of the first model, and so on.

[0087] For example, the arrangement of the N groups of data in the second data set can be {data1, data2, data3, ..., dataN}, where data1 = {datasample1, datasample1′}, data2 = {datasample2, datasample2′}, and so on. data1 is the first group of data, data2 is the second group of data, and so on. datasample1 represents the first sample with a sample index of 1, datasample1′ represents the second sample obtained by processing datasample1 with the first part of the first model, datasample2 represents the first sample with a sample index of 2, datasample2′ represents the second sample obtained by processing datasample2 with the first part of the first model, and so on.

[0088] For Option 1, it can be applied to at least the following situations:

[0089] 1) The M first samples are measured by the terminal device and other terminal devices, and the N first samples are all or part of the M first samples.

[0090] 2) The M first samples are all measured by the terminal device, and the N first samples are all or part of the M first samples.

[0091] 3) The M first samples are measured by the terminal device and other terminal devices, and the N first samples are completely different from or partially different from the M first samples.

[0092] 4) The M first samples are all measured by the terminal device, and the N first samples are completely different from or partially different from the M first samples.

[0093] 5) N first samples are measured by the above terminal device and other terminal devices.

[0094] 6) N first samples are measured by the terminal device.

[0095] Option 2

[0096] In some embodiments, the second data set includes N second samples, where the N second samples are obtained by processing the N first samples with the first part of the first model.

[0097] Here, the second data set has a second format; in the second format, the second data set includes N second samples, and the N second samples are arranged according to sample indices.

[0098] For example, the N second samples in the second data set may be arranged in {datasample1′, datasample2′, datasample3′, …, datasampleN′}, where datasample1′ represents the second sample obtained by processing datasample1 with the first part of the first model, datasample2′ represents the second sample obtained by processing datasample2 with the first part of the first model, and so on. datasample1 represents the first sample with a sample index of 1, datasample2 represents the first sample with a sample index of 2, and so on.

[0099] For Option 2, it can be applied to at least the following situations:

[0100] 1) The M first samples are all measured by the terminal device, and the N first samples are all or part of the M first samples.

[0101] 2) N first samples are obtained by measuring the above terminal device.

[0102] In some embodiments, when all M first samples are measured by the terminal device, and the N first samples are part of the M first samples, the network device sends first information to the terminal device, where the first information indicates the sample indexes of the N second samples. Here, since the second samples correspond to the first samples, the sample indexes of the second samples can be consistent with the sample indexes of the corresponding first samples. In this way, the terminal device can determine, based on the first information, which of the M first samples the N first samples correspond to.

[0103] In some embodiments, after obtaining M first samples measured by the terminal device, the network device uses only N first samples from the M first samples, taking into account the presence of low-quality samples. The second data set then only includes the N second samples corresponding to these N first samples. Here, there may be various criteria for determining whether samples are of low quality. For example, if some statistical features of certain samples differ significantly from those of the majority of other samples, or if the statistical features of certain samples do not meet the requirements defined by the network device, these samples may be considered to be of low quality.

[0104] In some embodiments, the first information includes a first sample index list, where the first sample index list includes sample indices of N second samples.

[0105] For example, the value of the sample index in the first sample index list may be a relative index value in the M first samples, which can be obtained by bits to represent the value of the sample index, that is, the sample index occupies For example, if the N first samples are the 1st, 3rd, and 5th samples of the M first samples, then the first sample index list is {Index#1, Index#3, Index#5}, where Index#1 represents the sample index value of 1, Index#3 represents the sample index value of 3, and Index#5 represents the sample index value of 5.

[0106] In other embodiments, the first information includes a first bitmap, each bit in the first bitmap corresponds to a sample index, and the value of the bit is used to indicate whether the second sample corresponding to the sample index corresponding to the bit is included in the second data set.

[0107] Exemplarily, the first bitmap is a bitmap of length M. In the bitmap, in order from low to high, the i-th (1≤i≤M) bit corresponds to the i-th sample in the M first samples, and the value of the i-th bit is 1, indicating that the second sample corresponding to the i-th sample in the M first samples is included in the above-mentioned second data set. The value of the i-th bit is 0, indicating that the second sample corresponding to the i-th sample in the M first samples is not included in the above-mentioned second data set.

[0108] Option 3

[0109] In some embodiments, the second data set has a third format; in the third format, the second data set includes N groups of data, each group of data in the N groups of data includes a second sample, and each group of data in some of the groups of data in the N groups of data also includes a first sample corresponding to the second sample; wherein the N groups of data are arranged according to sample indexes.

[0110] Here, Solution 3 can be understood as a combination of Solution 1 and Solution 2, that is, in Solution 3, part of the data in the second data set has the first format, and the other part of the data has the second format.

[0111] Exemplarily, the arrangement of the N groups of data in the second data set can be {data1, data2, data3, …, dataN}, where data1 is the first group of data, data2 is the second group of data, and so on. Some of these N data include a first sample and a corresponding second sample. Taking data1 as an example, data1 = {datasample1, datasample1′}, datasample1 represents the first sample with a sample index of 1, and datasample1′ represents the second sample obtained by processing datasample1 with the first part of the first model. In addition, some other data in these N data include a second sample. Taking data2 as an example, data2 = {datasample2′}, and datasample2′ represents the second sample obtained by processing datasample2 with the first part of the first model.

[0112] For Solution 3, if a sample among the N first samples is not measured by the terminal device, the second dataset needs to include that sample and a second sample obtained by processing the sample using the first part of the first model. If a sample among the N first samples is measured by the terminal device, the second dataset may only include the second sample obtained by processing the sample using the first part of the first model.

[0113] For Option 3, it is applicable to at least the following situations:

[0114] The N first samples are obtained by measuring the above terminal device and other terminal devices.

[0115] In some embodiments, P first samples out of N first samples correspond to P second samples in a second data set, where P is an integer less than N. When the second data set does not include these P first samples (e.g., when the P first samples are measured by the terminal device, the second data set may not include the P first samples), the network device sends second information to the terminal device, where the second information is used to indicate the sample indexes of the P second samples. Here, since the second samples correspond to the first samples, the sample indexes of the second samples may be consistent with the sample indexes of the corresponding first samples. In this way, the terminal device can determine, based on the second information, which of the N first samples the P first samples correspond to.

[0116] In some embodiments, the second information includes a second sample index list, where the second sample index list includes sample indices of P second samples.

[0117] For example, the value of the sample index in the second sample index list may be a relative index value in the N first samples, which can be obtained by bits to represent the value of the sample index, that is, the sample index occupies For example, if the P first samples are the 1st, 3rd, and 5th samples in the N first samples, then the second sample index list is {Index#1, Index#3, Index#5}, where Index#1 represents the sample index value of 1, Index#3 represents the sample index value of 3, and Index#5 represents the sample index value of 5.

[0118] In other embodiments, the second information includes a second bitmap, each bit in the second bitmap corresponds to a sample index, and the value of the bit is used to indicate whether the second sample corresponding to the sample index corresponding to the bit is included in the second data set.

[0119] Exemplarily, the second bitmap is a bitmap of length N. In the bitmap, in order from low to high, the i-th (1≤i≤N) bit corresponds to the i-th sample in the N first samples, and the value of the i-th bit is 1, indicating that the second sample corresponding to the i-th sample in the N first samples is included in the above-mentioned second data set. The value of the i-th bit is 0, indicating that the second sample corresponding to the i-th sample in the N first samples is not included in the above-mentioned second data set.

[0120] In some implementations, the network device sends third information to the terminal device, where the third information is used to indicate the quantization granularity of the quantization module in the first model.

[0121] In some embodiments, the network device receives fourth information sent by the terminal device, where the fourth information is used to indicate the second model supported by the terminal device and / or configuration information supported by the second model, where the configuration information includes at least one of the following: the number of antennas, the number of subbands, and the number of payloads required for CSI reporting. The network device can select an appropriate second data set based on the fourth information and send it to the terminal device. Here, the terminal device can report the second model supported by itself and / or the configuration information supported by the second model to the network device through UE capabilities.

[0122] Through the above scheme, it is possible to implement that the network device trains the first model based on the first data set, and the terminal device trains the second model based on the second data set, and the trained second model and the second part of the first model are used for CSI reporting. The second part of the first model and the second model can be understood as a pair of AI models for CSI reporting, that is, a bilateral AI model. After the bilateral AI model is deployed, when the configuration of the network device or the environment in which the terminal device is located changes, the network device and / or the terminal device detects that the performance of their respective models cannot meet the requirements, and a model update process can be initiated. The model update process is described below.

[0123] In some embodiments, the network device trains a first model based on a third data set and sends a fourth data set to the terminal device, where the third data set includes M' first samples, and the fourth data set includes N' second samples, where the N' second samples are obtained by processing the N' first samples through the first part of the first model, and M' and N' are positive integers; wherein the fourth data set is used by the terminal device to determine the second model.

[0124] Here, the third dataset can be understood as an updated dataset of the first dataset, and the fourth dataset can be understood as an updated dataset of the second dataset. The third dataset can also be described as dataset A', and the fourth dataset can also be described as dataset B', which is not limited in this application.

[0125] Here, the specific implementation method of "the network device trains the first model based on the third data set and sends the fourth data set to the terminal device" is similar to the specific implementation method of "the network device trains the first model based on the first data set and sends the second data set to the terminal device", and can refer to the above description.

[0126] After the network device trains the first model based on the first data set, upon obtaining the third data set, it can retrain the model based on the third data set based on the previously trained first model, thereby updating the first model. Similarly, after the terminal device trains the second model based on the second data set, upon obtaining the fourth data set, it can retrain the model based on the previously trained second model (or fine-tune the model or update the model parameters) based on the fourth data set, thereby updating the second model.

[0127] In some embodiments, the fourth dataset in the above solution has an association relationship with at least one of the following: the second dataset, the second model. The association relationship is characterized by at least one of the following:

[0128] The subordinate data set identifier of the fourth data set is equal to the identifier of the second data set;

[0129] The identifier of the fourth data set is equal to the identifier of the second data set;

[0130] The subordinate model identifier of the fourth data set is equal to the identifier of the second model.

[0131] In some embodiments, the quantization granularity of the quantization module in the first model associated with the fourth data set in the above solution is equal to the quantization granularity of the quantization module in the first model associated with the second data set.

[0132] In the above solution, the transmission method of the second data set can be, but is not limited to, the following: the network device sends the second data set to the terminal device through semi-persistent scheduling (SPS). In specific implementation, the network device sends a first radio resource control (RRC) signaling to the terminal device, the first RRC signaling including an SPS configuration, the SPS configuration including an identifier of the second data set; and the network device sends a first activation signaling to the terminal device, the first activation signaling being used to activate the SPS configuration, the SPS configuration being used for the terminal device to receive the second data set in accordance with the SPS method.

[0133] In some implementations, the first activation signaling may be a Media Access Control-Control Element (MAC-CE).

[0134] In some embodiments, after activating the SPS configuration through the first activation signaling (which can also be understood as activating an SPS transmission), the network device can transmit one or more data sets, and can include the identifier of the second data set (such as the index of the second data set) in the SPS configuration to establish an association relationship between the SPS configuration and the second data set, so that the terminal device can know that the SPS configuration is used to transmit the second data set. After completing the transmission of the second data set, the SPS configuration can be deactivated through the first activation signaling.

[0135] It should be noted that the SPS mode is a semi-static transmission mode. In order to distinguish the SPS transmission corresponding to the second data set from the ordinary SPS transmission, a new radio network temporary identifier (RNTI), such as a D-RNTI, can be defined for the SPS transmission corresponding to the second data set. The new RNTI is used to indicate that the SPS transmission is used to transmit the second data set.

[0136] During the model update process, the network device needs to send a fourth data set to the terminal device. The transmission method of the fourth data set is similar to the transmission method of the second data set. The transmission of the fourth data set can be implemented by referring to the transmission method of the second data set.

[0137] In the technical solution of the embodiment of the present application, the network device initiates model training. Specifically, the network device trains a first model based on a first data set, and the network device sends a second data set to the terminal device, so that the terminal device trains a second model based on the second data set. The trained second model and the second part of the first model (which can be understood as a pair of AI models) are used for CSI reporting. In this way, the training method of a pair of AI models used for CSI reporting is clarified.

[0138] FIG5 is a second flow chart of the model training method provided in an embodiment of the present application. As shown in FIG5 , the model training method includes:

[0139] Step 501: The terminal device trains a third model based on the fifth data set, where the third model includes a first part and a second part, and sends a sixth data set to the network device, where the fifth data set includes M first samples, and the sixth data set includes N first samples and N second samples, where the N second samples are obtained by processing the N first samples through the first part of the third model, and M and N are positive integers; wherein the sixth data set is used by the network device to train a fourth model, and the trained fourth model and the first part of the third model are used for CSI reporting.

[0140] In this embodiment of the present application, the terminal device is responsible for training the third model, and the network device is responsible for training the fourth model. The third model includes a first part and a second part. The trained first part of the third model and the fourth model are used for CSI reporting. The first part and the fourth model of the third model can be understood as a pair of AI models for CSI reporting, i.e., a two-sided AI model.

[0141] In some implementations, the first part of the third model is a CSI generation part (CSI generation part), and the second part of the third model is a CSI reconstruction part (CSI reconstruction part).

[0142] In some embodiments, the fourth model is a CSI recovery part. In order to distinguish the CSI recovery part in the third model from the CSI recovery part in the fourth model, the CSI recovery part in the third model can be referred to as CSI recovery part 1, and the CSI recovery part in the fourth model can be referred to as CSI recovery part 2.

[0143] It should be noted that the CSI generation section above represents the AI ​​model on the terminal device side of the pair of AI models, which may actually include an encoding module and a quantization module. The CSI recovery section 2 above represents the AI ​​model on the network device side of the pair of AI models, which may actually include a dequantization module and a decoding module. The functions of the encoding module, quantization module, dequantization module, and decoding module can be referred to in the relevant descriptions above.

[0144] In the embodiment of the present application, the terminal device trains the third model based on the fifth data set, and the network device trains the fourth model based on the sixth data set. The third data set can also be described as data set A, and the sixth data set can also be described as data set B, which is not limited in this application.

[0145] In this embodiment of the present application, the fifth data set includes M first samples, where M is a positive integer, and the first samples are channel information samples (also referred to as channel samples). The terminal device performs CSI measurement to obtain channel information, and uses the channel information as a sample for training the third model (i.e., the channel information sample).

[0146] After obtaining the M first samples, the terminal device trains the third model based on the M first samples, that is, trains the first and second parts of the third model. For example, the terminal device trains the CSI generation part and the CSI recovery part 1 based on the M first samples.

[0147] In the embodiment of the present application, the sixth data set includes N first samples and N second samples, the N second samples are obtained by processing the N first samples by the first part of the third model, and N is a positive integer.

[0148] In some embodiments, the N first samples are identical to the M first samples. In other embodiments, the N first samples are partially identical to the M first samples. In other embodiments, the N first samples are completely different from the M first samples.

[0149] The terminal device inputs N first samples into the third model, which then processes the N second samples through the first part of the third model. The second samples are quantized bit samples corresponding to the channel information. The N first samples and the N second samples constitute a sixth data set. The terminal device sends the sixth data set to the network device for use in training the fourth model. The format of the sixth data set is described below.

[0150] In some embodiments, the sixth data set has a fourth format; in the fourth format, the sixth data set includes N groups of data, each of the N groups of data includes a first sample and a second sample corresponding to the first sample; wherein the N groups of data are arranged according to sample indexes.

[0151] For example, the arrangement of the N groups of data in the sixth data set may be {datasample1, datasample1′, datasample2, datasample2′, datasample3, datasample3′, …, datasampleN, datasampleN′}, where datasample1 and datasample1′ are the first group of data, datasample2 and datasample2′ are the second group of data, and so on. datasample1 represents the first sample with a sample index of 1, datasample1′ represents the second sample obtained by processing datasample1 with the first part of the third model, datasample2 represents the first sample with a sample index of 2, datasample2′ represents the second sample obtained by processing datasample2 with the first part of the third model, and so on.

[0152] For example, the arrangement of the N groups of data in the second data set can be {data1, data2, data3, ..., dataN}, where data1 = {datasample1, datasample1′}, data2 = {datasample2, datasample2′}, and so on. data1 is the first group of data, data2 is the second group of data, and so on. datasample1 represents the first sample with a sample index of 1, datasample1′ represents the second sample obtained by processing datasample1 with the first part of the third model, datasample2 represents the first sample with a sample index of 2, datasample2′ represents the second sample obtained by processing datasample2 with the first part of the third model, and so on.

[0153] In some implementations, the terminal device sends fifth information to the network device, where the fifth information is used to indicate a quantization granularity of the quantization module in the third model.

[0154] In some embodiments, the terminal device receives sixth information sent by the network device, where the sixth information is used to indicate the configuration information of the fourth model supported by the network device and / or the fourth model supported by the network device, and the configuration information includes at least one of the following: the number of antennas, the number of subbands, and the number of payloads required for CSI reporting.

[0155] Through the above scheme, it is possible to implement that the terminal device trains the third model based on the fifth data set, and the network device trains the fourth model based on the sixth data set, and the trained fourth model and the second part of the third model are used for CSI reporting. The second part of the third model and the fourth model can be understood as a pair of AI models for CSI reporting, that is, a bilateral AI model. After the bilateral AI model is deployed, when the configuration of the network device or the environment in which the terminal device is located changes, the network device and / or the terminal device detects that the performance of their respective models cannot meet the requirements, and a model update process can be initiated. The model update process is described below.

[0156] In some embodiments, the terminal device trains a third model based on the seventh data set and sends the eighth data set to the network device, where the seventh data set includes M' first samples, and the eighth data set includes N' first samples and N' second samples, where the N' second samples are obtained by processing the N' first samples by the first part of the third model, and M' and N' are positive integers; wherein the eighth data set is used by the network device to determine the fourth model.

[0157] Here, the seventh dataset can be understood as an updated dataset of the fifth dataset, and the eighth dataset can be understood as an updated dataset of the sixth dataset. The seventh dataset can also be described as dataset A', and the eighth dataset can also be described as dataset B', which is not limited in this application.

[0158] Here, the specific implementation method of "the terminal device trains the third model based on the seventh data set and sends the eighth data set to the network device" is similar to the specific implementation method of "the terminal device trains the third model based on the fifth data set and sends the sixth data set to the network device", and can refer to the above description.

[0159] After the terminal device trains the third model based on the fifth data set, upon obtaining the seventh data set, it may train again based on the seventh data set on the basis of the third model trained previously, thereby updating the third model. Similarly, after the network device trains the fourth model based on the sixth data set, upon obtaining the eighth data set, it may train again based on the eighth data set on the basis of the fourth model trained previously (or fine-tune the model or update the model parameters), thereby updating the fourth model.

[0160] In some embodiments, the eighth dataset in the above solution has an association relationship with at least one of the following: the sixth dataset, the fourth model. The association relationship is characterized by at least one of the following:

[0161] The subordinate dataset identifier of the eighth dataset is equal to the identifier of the sixth dataset;

[0162] The identifier of the eighth data set is equal to the identifier of the sixth data set;

[0163] The subordinate model identifier of the eighth data set is equal to the identifier of the fourth model.

[0164] In some embodiments, the quantization granularity of the quantization module in the third model associated with the eighth data set in the above solution is equal to the quantization granularity of the quantization module in the third model associated with the sixth data set.

[0165] In the above scheme, the transmission method of the sixth data set can be, but is not limited to, the following: the terminal device sends the sixth data set to the network device through a configured grant (CG). In specific implementation, the terminal device receives a second RRC signaling sent by the network device, the second RRC signaling includes a CG-PUSCH configuration, and the CG-PUSCH configuration includes an identifier of the sixth data set; the terminal device receives a second activation signaling sent by the network device, the second activation signaling is used to activate the CG-PUSCH configuration, and the CG-PUSCH configuration is used by the terminal device to send the sixth data set in a CG manner.

[0166] In some implementations, the second activation signaling may be downlink control information (DCI).

[0167] In some embodiments, after activating the CG-PUSCH configuration through the second activation signaling (which can also be understood as activating a CG-PUSCH transmission), the terminal device can transmit one or more data sets, and can include the identifier of the sixth data set (such as the index of the sixth data set) in the CG-PUSCH configuration to establish an association relationship between the CG-PUSCH configuration and the sixth data set. In this way, the terminal device can know that the CG-PUSCH configuration is used to transmit the sixth data set. After completing the transmission of the sixth data set, the CG-PUSCH configuration can be deactivated through the first activation signaling.

[0168] It should be noted that the CG-PUSCH mode is a semi-static transmission mode. In order to distinguish the CG-PUSCH transmission corresponding to the sixth data set from the ordinary CG-PUSCH transmission, a new RNTI, such as a D-RNTI, can be defined for the CG-PUSCH transmission corresponding to the sixth data set. This new RNTI is used to indicate that the CG-PUSCH transmission is used to transmit the sixth data set.

[0169] During the model update process, the terminal device needs to send the eighth data set to the network device. The transmission method of the eighth data set is similar to the transmission method of the sixth data set mentioned above, and the transmission of the eighth data set can be implemented by referring to the transmission method of the sixth data set mentioned above.

[0170] In the technical solution of the embodiment of the present application, the terminal device initiates model training. Specifically, the terminal device trains the third model based on the fifth data set, and the terminal device sends the sixth data set to the network device, so that the network device trains the fourth model based on the sixth data set. The trained fourth model and the first part of the third model (which can be understood as a pair of AI models) are used for CSI reporting. In this way, the training method of the pair of AI models used for CSI reporting is clarified.

[0171] The following describes the technical solutions of the embodiments of the present application with reference to specific application examples. The following application example 1 corresponds to the embodiment related to FIG. 4 above, and the description of the embodiment related to FIG. 4 can be referenced in application example 1. The following application example 2 corresponds to the embodiment related to FIG. 5 above. The description of the embodiment related to FIG. 5 can be referenced in application example 2.

[0172] Application Example 1

[0173] In this application example, the network device is a base station, the terminal device is UE1, the first data set is data set A for example, the second data set is data set B, the first part of the first model is CSI generation part 1 (referred to as encoder1 for short), the second part of the first model is the CSI recovery part (referred to as decoder for short), the second model includes CSI generation part 2 (referred to as encoder2 for short), the first sample is a channel sample, and the second sample is a quantized bit sample (hereinafter referred to as channel data).

[0174] Model training process

[0175] 1) The base station trains encoder 1 and decoder 1 based on dataset A. After training, it sends dataset B to UE 1. Dataset A includes M channel samples, where M is a positive integer. The formats of dataset B can be as follows:

[0176] Solution 1: Dataset B includes N channel samples and N channel data, which are obtained after the N channel samples are processed by encoder 1. The arrangement of the N channel samples and the N channel data in dataset B can be described with reference to the embodiment related to FIG4 .

[0177] As a case, if M channel samples are measured jointly by UE1 and other UEs, and N channel samples are all or part of the M channel samples, then the above solution 1 can be used to determine the format of data set B.

[0178] Solution 2: Dataset B includes only N channel data, which are obtained by processing N channel samples by encoder 1. The arrangement of the N channel data in dataset B can be referred to the description of the embodiment related to FIG4 .

[0179] As a case, if all the M channel samples are measured by UE1, and the N channel samples are all or part of the M channel samples, then the above solution 2 can be used to determine the format of the data set B.

[0180] In some implementations, if the base station considers that there are low-quality channel samples among the M channel samples and only uses N channel samples (N < M) among the M channel samples, then data set B includes N channel data obtained after these N channel samples are processed by encoder 1, and the base station needs to additionally indicate the sample indexes of these N channel samples to UE1 through the first information. Here, the specific implementation of the first information can be referred to the description of the embodiment related to Figure 4.

[0181] Solution 3: Data set B includes N groups of data, each group of data in the N groups of data includes one channel data, and each group of data in some of the N groups of data also includes a channel sample corresponding to the channel data. For example, for some channel samples among the N channel samples, since they are not measured by UE1, it is necessary to include the channel sample and the channel data obtained after the channel sample is processed by encoder1 in data set B. For other channel samples among the N channel samples, since they are measured by UE1, it is sufficient to only include the channel data obtained after the channel sample is processed by encoder1 in data set B. The arrangement of the N groups of data in data set B can refer to the description in the embodiment related to Figure 4.

[0182] Solution 3 can be understood as a combination of Solution 1 and Solution 2. Similar to Solution 2, if the base station only uses a portion of the channel samples measured by UE1 (e.g., P channel samples), the base station also needs to indicate the sample indexes of these P channel samples to UE1 through the second information. The specific implementation of the second information can be found in the description of the embodiment related to FIG4 .

[0183] In addition, the base station needs to indicate to UE1 the quantization granularity K of the base station-side quantization module associated with data set B, that is, in the bit data after the channel data of each channel sample passes through encoder1 and is quantized, each K bits corresponds to a quantization unit or a dequantization unit.

[0184] The above dataset B can be identified using Dataset ID#1.

[0185] 2) After receiving data set B, UE1 trains encoder2 based on data set B, where encoder2 is called model#B or functionality#B. Encoder2 forms an association with the decoder on the base station side to form a pair of AI models for UE1's CSI reporting.

[0186] Here, UE1 can report to the base station through the UE capability reporting process that it supports model#B or functionality#B, and that this model#B or functionality#B is obtained from data set B sent by the base station. In addition, UE1 can also report to the base station the number of base station antennas, the number of subbands, the number of payloads required for CSI reporting, etc. that this model#B or functionality#B can support.

[0187] The above model#B or functionality#B can be identified using model ID#1 or functionality ID#1.

[0188] The model training process described above can be used to train the model. The trained decoder is deployed on the base station side, and the trained encoder2 is deployed on the terminal device side. The decoder and encoder2 form a pair of AI models.

[0189] The transmission method of the above-mentioned data set B can be described with reference to the embodiment related to FIG. 4 .

[0190] Model update process

[0191] 3) After the AI ​​model is deployed, if the configuration of the base station or the environment in which UE1 is located changes, the base station or UE1 detects that the performance of their respective models does not meet the requirements and can initiate a model update process. The base station trains encoder1 and decoder based on data set A′, and after training, sends data set B′ to UE1, thus initiating the model training process again. UE1 can use data set B′ to fine-tune the model or update the model parameters based on the last trained encoder2, thereby achieving the purpose of updating encoder2. The updated model of encoder2 is recorded as encoder2′.

[0192] The dataset B′ can be identified using Dataset ID#2. The dataset B′ has the following association with the dataset B, model#B, or functionality#B in the model training phase:

[0193] a) Dataset B′ and dataset B have the following relationship:

[0194] Dataset B′ has a subordinate dataset ID, and the subordinate dataset ID = Dataset ID#1 of dataset B;

[0195] Dataset ID#2 of dataset B′=Dataset ID#1 of dataset B.

[0196] b) Dataset B′ and model#B or functionality#B have the following relationship:

[0197] Data set B′ has subordinate model ID#1′ or functionality ID#1′, and model ID#1′=model ID#1 or functionality ID#1′=functionality ID#1.

[0198] In addition, the granularity N′ of the quantization module on the base station side associated with the data set B′ is equal to the granularity N associated with the data set B.

[0199] The transmission method of the above-mentioned data set B′ can be referred to the description of the embodiment related to FIG. 4 .

[0200] The technical solution of the embodiment of the present application proposes a model training initiated by a base station, including a model training process and a model update process. On the one hand, the training of the bilateral model can be completed without transmitting the model, which provides a feasible training method for the use of the bilateral model. On the other hand, when the performance of the model does not meet the requirements, the technical solution of the embodiment of the present application provides a method for quickly updating the model based on the deployed model, which ensures that the bilateral model can quickly adapt to changes in the base station configuration and the UE environment. At the same time, compared with retraining the model, the model update is based on the trained model to fine-tune the model or update the model parameters, which can significantly reduce the training data volume requirements and the computing power and other overheads of model training while ensuring better reasoning performance. In addition, the matching degree of the two side models of the bilateral model is usually higher than that of the retrained model after the updated model.

[0201] Application Example 2

[0202] Model training process

[0203] 1) The UE trains encoder and decoder 1 based on dataset A. After training, dataset B is sent to the base station. Dataset A includes M channel samples, where M is a positive integer. The format of dataset B can be as follows:

[0204] Dataset B includes N (N≤M) channel samples and N channel data, which are obtained after the N channel samples are processed by the encoder. The arrangement of the N channel samples and the N channel data in dataset B can be referred to the description of the embodiment related to FIG5 .

[0205] In addition, the UE needs to report the granularity L of the quantization module on the UE side associated with the data set B to the base station, that is, in the bit data after the channel data of each channel sample passes through the encoder and is quantized, each L bits corresponds to a quantization unit or a dequantization unit.

[0206] The above dataset B can be identified using Dataset ID#1.

[0207] 2) After receiving data set B, the base station trains decoder2 based on data set B, where decoder2 is called model#B or functionality#B. Decoder2 is associated with the encoder on the UE side to form a paired AI model for UE1's CSI reporting.

[0208] Here, the base station may indicate to the UE that it supports model#B or functionality#B, and that this model#B or functionality#B is obtained from the data set B sent by the UE. In addition, the base station may also indicate to the UE the number of base station antennas, the number of subbands, the number of payloads required for CSI reporting, etc. that this model#B or functionality#B can support.

[0209] The above model#B or functionality#B can be identified using model ID#1 or functionality ID#1.

[0210] The model training process described above can be used to train the model. The trained decoder2 is deployed on the base station side, and the trained encoder is deployed on the terminal device side. The decoder2 and encoder form a pair of AI models.

[0211] The transmission method of the above-mentioned data set B can be referred to the description in the embodiment related to Figure 5.

[0212] Model update process

[0213] 3) After the AI ​​model is deployed, if the base station configuration or the UE's environment changes, and the base station or UE detects that the performance of their respective models does not meet the requirements, they can initiate a model update process. The UE trains encoder1 and decoder1 based on dataset A', and after training, sends dataset B' to the base station, initiating the model training process again. The base station can use dataset B' to fine-tune the model or update the model parameters based on the last trained decoder2, thereby achieving the purpose of updating decoder2. The updated model of decoder2 is recorded as decoder2'.

[0214] The dataset B′ can be identified using Dataset ID#2. The dataset B′ has the following association with the dataset B, model#B, or functionality#B in the model training phase:

[0215] a) Dataset B′ and dataset B have the following relationship:

[0216] Dataset B′ has a subordinate dataset ID, and the subordinate dataset ID = Dataset ID#1 of dataset B;

[0217] Dataset ID#2 of dataset B′=Dataset ID#1 of dataset B.

[0218] b) Dataset B′ and model#B or functionality#B have the following relationship:

[0219] Data set B′ has subordinate model ID#1′ or functionality ID#1′, and model ID#1′=model ID#1 or functionality ID#1′=functionality ID#1.

[0220] In addition, the granularity N′ of the quantization module on the base station side associated with the data set B′ is equal to the granularity N associated with the data set B.

[0221] The transmission method of the above-mentioned data set B′ can refer to the description of the embodiment related to Figure 5.

[0222] The technical solution of the embodiment of the present application proposes a model training initiated by the UE, including a model training process and a model update process. On the one hand, the training of the bilateral model can be completed without transmitting the model, which provides a feasible training method for the use of the bilateral model. On the other hand, when the performance of the model does not meet the requirements, the technical solution of the embodiment of the present application provides a method for quickly updating the model based on the deployed model, which ensures that the bilateral model can quickly adapt to changes in the base station configuration and the UE environment. At the same time, compared with retraining the model, the model update is based on the trained model to fine-tune the model or update the model parameters, which can significantly reduce the training data volume requirements and the computing power and other overheads of model training while ensuring better reasoning performance. In addition, the matching degree of the two side models of the bilateral model is usually higher than that of the retrained model after the updated model.

[0223] FIG6 is a schematic diagram of the structure of a model training device provided in an embodiment of the present application, which is applied to a network device. As shown in FIG6 , the model training device includes:

[0224] A training unit 601 is configured to train a first model based on a first data set, where the first model includes a first part and a second part;

[0225] A sending unit 602 is configured to send a second data set to a terminal device, where the first data set includes M first samples, and the second data set includes N second samples, where the N second samples are obtained by processing the N first samples by the first part of the first model, and M and N are positive integers;

[0226] The second data set is used for the terminal device to train a second model, and the trained second model and the second part of the first model are used for CSI reporting.

[0227] In some embodiments, the second data set also includes the N first samples; the second data set has a first format; in the first format, the second data set includes N groups of data, each of the N groups of data includes a first sample and a second sample corresponding to the first sample; wherein the N groups of data are arranged according to sample indexes.

[0228] In some embodiments, the second data set has a second format; in the second format, the second data set includes N second samples, and the N second samples are arranged according to sample indices.

[0229] In some implementations, the sending unit 602 is further configured to send first information to the terminal device, where the first information is used to indicate sample indexes of the N second samples.

[0230] In some embodiments, the first information includes a first sample index list, which includes sample indices of the N second samples; or, the first information includes a first bit map, each bit in the first bit map corresponds to a sample index, and the value of the bit is used to indicate whether the second sample corresponding to the sample index corresponding to the bit is included in the second data set.

[0231] In some embodiments, the second data set has a third format; in the third format, the second data set includes N groups of data, each group of the N groups of data includes a second sample, and each group of data in some of the N groups of data also includes a first sample corresponding to the second sample; wherein the N groups of data are arranged according to sample indexes.

[0232] In some embodiments, P first samples out of the N first samples correspond to P second samples in the second data set, where P is an integer less than N; the sending unit 602 is further configured to send second information to the terminal device, where the second information is used to indicate the sample index of the P second samples.

[0233] In some embodiments, the second information includes a second sample index list, which includes sample indices of the P second samples; or, the second information includes a second bit map, where each bit in the second bit map corresponds to a sample index, and the value of the bit is used to indicate whether the second sample corresponding to the sample index corresponding to the bit is included in the second data set.

[0234] In some implementations, the sending unit 602 is further configured to send third information to the terminal device, where the third information is used to indicate a quantization granularity of a quantization module in the first model.

[0235] In some embodiments, the apparatus further includes: a receiving unit 603, configured to receive fourth information sent by the terminal device, the fourth information being used to indicate the second model supported by the terminal device and / or configuration information supported by the second model, the configuration information including at least one of the following: the number of antennas, the number of subbands, and the number of payloads required for CSI reporting.

[0236] In some embodiments, the training unit 601 is further configured to train the first model based on a third data set;

[0237] The sending unit 602 is further configured to send a fourth data set to the terminal device, where the third data set includes M' first samples, and the fourth data set includes N' second samples, where the N' second samples are obtained by processing the N' first samples through the first part of the first model, and M' and N' are positive integers; wherein the fourth data set is used by the terminal device to determine the second model.

[0238] In some embodiments, the fourth data set is associated with at least one of the following: the second data set, the second model.

[0239] In some embodiments, the association is characterized by at least one of the following:

[0240] The subordinate dataset identifier of the fourth dataset is equal to the identifier of the second dataset;

[0241] The identifier of the fourth data set is equal to the identifier of the second data set;

[0242] The subordinate model identifier of the fourth data set is equal to the identifier of the second model.

[0243] In some embodiments, the quantization granularity of the quantization module in the first model associated with the fourth data set is equal to the quantization granularity of the quantization module in the first model associated with the second data set.

[0244] In some implementations, the sending unit 602 is configured to send the second data set to the terminal device via SPS.

[0245] In some embodiments, the sending unit 602 is configured to send a first RRC signaling to the terminal device, wherein the first RRC signaling includes an SPS configuration, and the SPS configuration includes an identifier of the second data set; and send a first activation signaling to the terminal device, wherein the first activation signaling is used to activate the SPS configuration, and the SPS configuration is used for the terminal device to receive the second data set in an SPS manner.

[0246] Those skilled in the art will appreciate that the functions implemented by each unit in the model training device shown in FIG6 can be understood with reference to the relevant description of the aforementioned method. The functions of each unit in the model training device shown in FIG6 can be implemented by a program running on a processor or by a specific logic circuit.

[0247] FIG7 is a second schematic diagram of the structure of a model training device provided in an embodiment of the present application, which is applied to a terminal device. As shown in FIG7 , the model training device includes:

[0248] A training unit 701 is configured to train a third model based on a fifth data set, wherein the third model includes a first part and a second part;

[0249] a sending unit 702 configured to send a sixth data set to a network device, wherein the fifth data set includes M first samples, and the sixth data set includes N first samples and N second samples, where the N second samples are obtained by processing the N first samples by the first part of the third model, and M and N are positive integers;

[0250] The sixth data set is used by the network device to train a fourth model, and the trained fourth model and the first part of the third model are used for CSI reporting.

[0251] In some embodiments, the sixth data set has a fourth format; in the fourth format, the sixth data set includes N groups of data, each of the N groups of data includes a first sample and a second sample corresponding to the first sample; wherein the N groups of data are arranged according to sample indexes.

[0252] In some implementations, the sending unit 702 is further configured to send fifth information to the network device, where the fifth information is used to indicate a quantization granularity of the quantization module in the third model.

[0253] In some embodiments, the device also includes: a receiving unit 703, configured to receive sixth information sent by the network device, the sixth information is used to indicate the fourth model supported by the network device and / or the configuration information supported by the four models, and the configuration information includes at least one of the following: the number of antennas, the number of subbands, and the number of payloads required for CSI reporting.

[0254] In some embodiments, the training unit 701 is further configured to train the third model based on the seventh data set, and send the eighth data set to the network device, the seventh data set includes M' first samples, the eighth data set includes N' first samples and N' second samples, the N' second samples are obtained by processing the N' first samples by the first part of the third model, and M' and N' are positive integers; wherein, the eighth data set is used by the network device to determine the fourth model.

[0255] In some embodiments, the eighth data set is associated with at least one of the following: the sixth data set, the fourth model.

[0256] In some embodiments, the association is characterized by at least one of the following:

[0257] The subordinate dataset identifier of the eighth dataset is equal to the identifier of the sixth dataset;

[0258] The identifier of the eighth data set is equal to the identifier of the sixth data set;

[0259] The subordinate model identifier of the eighth data set is equal to the identifier of the fourth model.

[0260] In some embodiments, the quantization granularity of the quantization module in the third model associated with the eighth data set is equal to the quantization granularity of the quantization module in the third model associated with the sixth data set.

[0261] In some implementations, the sending unit 702 is configured to send the sixth data set to the network device via a CG manner.

[0262] In some embodiments, the receiving unit 703 is further configured to receive a second RRC signaling sent by the network device, the second RRC signaling including a CG-PUSCH configuration, and the CG-PUSCH configuration including an identifier of the sixth data set; and receive a second activation signaling sent by the network device, the second activation signaling being used to activate the CG-PUSCH configuration, and the CG-PUSCH configuration being used by the terminal device to send the sixth data set in a CG manner.

[0263] Those skilled in the art will appreciate that the functions implemented by each unit in the model training device shown in FIG7 can be understood with reference to the relevant description of the aforementioned method. The functions of each unit in the model training device shown in FIG7 can be implemented by a program running on a processor or by a specific logic circuit.

[0264] Figure 8 is a schematic structural diagram of a communication device 800 provided in an embodiment of the present application. The communication device can be a terminal device or a network device. The communication device 800 shown in Figure 8 includes a processor 810, which can call and run a computer program from a memory to implement the method in the embodiment of the present application.

[0265] Optionally, as shown in FIG8 , the communication device 800 may further include a memory 820. The processor 810 may call and execute a computer program from the memory 820 to implement the method in the embodiment of the present application.

[0266] The memory 820 may be a separate device independent of the processor 810 , or may be integrated into the processor 810 .

[0267] Optionally, as shown in FIG8 , the communication device 800 may further include a transceiver 830 , and the processor 810 may control the transceiver 830 to communicate with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices.

[0268] The transceiver 830 may include a transmitter and a receiver. The transceiver 830 may further include an antenna, and the number of antennas may be one or more.

[0269] Optionally, the communication device 800 may specifically be a network device in an embodiment of the present application, and the communication device 800 may implement the corresponding processes implemented by the network device in each method in the embodiment of the present application. For the sake of brevity, they will not be repeated here.

[0270] Optionally, the communication device 800 may specifically be a mobile terminal / terminal device in an embodiment of the present application, and the communication device 800 may implement the corresponding processes implemented by the mobile terminal / terminal device in each method in the embodiment of the present application. For the sake of brevity, they will not be repeated here.

[0271] Figure 9 is a schematic structural diagram of a chip according to an embodiment of the present application. The chip 900 shown in Figure 9 includes a processor 910, which can call and run a computer program from a memory to implement the method according to the embodiment of the present application.

[0272] Optionally, as shown in FIG9 , the chip 900 may further include a memory 920 , wherein the processor 910 may call and execute a computer program from the memory 920 to implement the method in the embodiment of the present application.

[0273] The memory 920 may be a separate device independent of the processor 910 , or may be integrated into the processor 910 .

[0274] Optionally, the chip 900 may further include an input interface 930. The processor 910 may control the input interface 930 to communicate with other devices or chips, and specifically, may obtain information or data sent by other devices or chips.

[0275] Optionally, the chip 900 may further include an output interface 940. The processor 910 may control the output interface 940 to communicate with other devices or chips, and specifically, may output information or data to other devices or chips.

[0276] Optionally, the chip can be applied to the network device in the embodiments of the present application, and the chip can implement the corresponding processes implemented by the network device in each method of the embodiments of the present application. For the sake of brevity, they will not be repeated here.

[0277] Optionally, the chip can be applied to the mobile terminal / terminal device in the embodiments of the present application, and the chip can implement the corresponding processes implemented by the mobile terminal / terminal device in the various methods of the embodiments of the present application. For the sake of brevity, they will not be repeated here.

[0278] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0279] It should be understood that the processor of the embodiments of the present application may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiment can be completed by hardware integrated logic circuits in the processor or software instructions. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The various methods, steps, and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly embodied as being executed by a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0280] It is understood that the memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM bus random access memory (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0281] It should be understood that the above-mentioned memories are exemplary but not restrictive. For example, the memories in the embodiments of the present application may also be static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM RAM (DR RAM), etc. In other words, the memories in the embodiments of the present application are intended to include, but are not limited to, these and any other suitable types of memories.

[0282] An embodiment of the present application also provides a computer-readable storage medium for storing a computer program.

[0283] Optionally, the computer-readable storage medium can be applied to the network device in the embodiments of the present application, and the computer program enables the computer to execute the corresponding processes implemented by the network device in the various methods of the embodiments of the present application. For the sake of brevity, they are not repeated here.

[0284] Optionally, the computer-readable storage medium can be applied to the mobile terminal / terminal device in the embodiments of the present application, and the computer program enables the computer to execute the corresponding processes implemented by the mobile terminal / terminal device in the various methods of the embodiments of the present application. For the sake of brevity, they will not be repeated here.

[0285] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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.

[0286] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0287] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0288] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0289] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0290] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0291] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A model training method, the method comprising: The network device trains a first model based on a first data set, the first model includes a first part and a second part, and sends a second data set to a terminal device, the first data set includes M first samples, the second data set includes N second samples, the N second samples are obtained by processing the N first samples by the first part of the first model, and M and N are positive integers; The second data set is used for the terminal device to train a second model, and the trained second model and the second part of the first model are used for channel state information CSI reporting.

2. The method according to claim 1, wherein: The second data set also includes the N first samples; the second data set has a first format; In the first format, the second data set includes N groups of data, each of the N groups of data includes a first sample and a second sample corresponding to the first sample; wherein the N groups of data are arranged according to sample indexes.

3. The method according to claim 1, wherein: The second data set has a second format; In the second format, the second data set includes N second samples, and the N second samples are arranged according to sample indices.

4. The method according to claim 3, wherein: The method further comprises: The network device sends first information to the terminal device, where the first information is used to indicate sample indexes of the N second samples.

5. The method according to claim 4, wherein: The first information includes a first sample index list, and the first sample index list includes sample indexes of the N second samples; or, The first information includes a first bitmap, each bit in the first bitmap corresponds to a sample index, and the value of the bit is used to indicate whether a second sample corresponding to the sample index corresponding to the bit is included in the second data set.

6. The method according to claim 1, wherein: The second data set has a third format; In the third format, the second data set includes N groups of data, each group of the N groups of data includes a second sample, and each group of data in some of the N groups of data also includes a first sample corresponding to the second sample; wherein the N groups of data are arranged according to sample indexes.

7. The method according to claim 6, wherein: P first samples of the N first samples correspond to P second samples in the second data set, where P is an integer less than N; the method further includes: The network device sends second information to the terminal device, where the second information is used to indicate sample indexes of the P second samples.

8. The method according to claim 7, wherein: The second information includes a second sample index list, and the second sample index list includes sample indices of the P second samples; or, The second information includes a second bitmap, each bit in the second bitmap corresponds to a sample index, and the value of the bit is used to indicate whether the second sample corresponding to the sample index corresponding to the bit is included in the second data set.

9. The method according to any one of claims 1 to 8, wherein: The method further comprises: The network device sends third information to the terminal device, where the third information is used to indicate a quantization granularity of a quantization module in the first model.

10. The method according to any one of claims 1 to 8, wherein: The method further comprises: The network device receives fourth information sent by the terminal device, and the fourth information is used to indicate the second model supported by the terminal device and / or configuration information supported by the second model, and the configuration information includes at least one of the following: the number of antennas, the number of subbands, and the number of payloads required for CSI reporting.

11. The method according to any one of claims 1 to 8, wherein: The method further comprises: The network device trains the first model based on the third data set, and sends a fourth data set to the terminal device, wherein the third data set includes M' first samples, and the fourth data set includes N' second samples, wherein the N' second samples are obtained by processing the first part of the N' first samples by the first model, and M' and N' are positive integers; The fourth data set is used by the terminal device to determine the second model.

12. The method according to claim 11, wherein: The fourth data set is associated with at least one of the following: the second data set, the second model.

13. The method according to claim 12, wherein: The association relationship is characterized by at least one of the following: The subordinate data set identifier of the fourth data set is equal to the identifier of the second data set; The identifier of the fourth data set is equal to the identifier of the second data set; The subordinate model identifier of the fourth data set is equal to the identifier of the second model.

14. The method according to claim 11, wherein: The quantization granularity of the quantization module in the first model associated with the fourth data set is equal to the quantization granularity of the quantization module in the first model associated with the second data set.

15. The method according to any one of claims 1 to 8, wherein: The sending the second data set to the terminal device includes: The second data set is sent to the terminal device in a semi-persistent scheduling (SPS) manner.

16. The method according to claim 15, wherein: The method further comprises: The network device sends a first radio resource control RRC signaling to the terminal device, where the first RRC signaling includes an SPS configuration, and the SPS configuration includes an identifier of the second data set; The network device sends a first activation signaling to the terminal device, where the first activation signaling is used to activate the SPS configuration, and the SPS configuration is used by the terminal device to receive the second data set in an SPS manner.

17. A model training method, the method comprising: The terminal device trains a third model based on the fifth data set, the third model includes a first part and a second part, and sends a sixth data set to the network device, the fifth data set includes M first samples, the sixth data set includes N first samples and N second samples, the N second samples are obtained by processing the N first samples by the first part of the third model, and M and N are positive integers; The sixth data set is used by the network device to train a fourth model, and the trained fourth model and the first part of the third model are used for CSI reporting.

18. The method according to claim 17, wherein: The sixth data set has a fourth format; In the fourth format, the sixth data set includes N groups of data, each of the N groups of data includes a first sample and a second sample corresponding to the first sample; wherein the N groups of data are arranged according to sample indexes.

19. The method according to claim 17 or 18, wherein: The method further comprises: The terminal device sends fifth information to the network device, where the fifth information is used to indicate the quantization granularity of the quantization module in the third model.

20. The method according to claim 17 or 18, wherein: The method further comprises: The terminal device receives the sixth information sent by the network device, and the sixth information is used to indicate the fourth model supported by the network device and / or the configuration information supported by the fourth model, and the configuration information includes at least one of the following: the number of antennas, the number of subbands, and the number of payloads required for CSI reporting.

21. The method according to claim 17 or 18, wherein: The method further comprises: The terminal device trains the third model based on the seventh data set, and sends the eighth data set to the network device, the seventh data set includes M' first samples, the eighth data set includes N' first samples and N' second samples, the N' second samples are obtained by processing the N' first samples by the first part of the third model, and M' and N' are positive integers; The eighth data set is used by the network device to determine the fourth model.

22. The method according to claim 21, wherein: The eighth data set is associated with at least one of the following: the sixth data set, the fourth model.

23. The method according to claim 22, wherein: The association relationship is characterized by at least one of the following: The subordinate data set identifier of the eighth data set is equal to the identifier of the sixth data set; The identifier of the eighth data set is equal to the identifier of the sixth data set; The subordinate model identifier of the eighth data set is equal to the identifier of the fourth model.

24. The method according to claim 21, wherein: The quantization granularity of the quantization module in the third model associated with the eighth data set is equal to the quantization granularity of the quantization module in the third model associated with the sixth data set.

25. The method according to claim 17 or 18, wherein: The sending the sixth data set to the network device includes: The sixth data set is sent to the network device by configuring the authorized CG mode.

26. The method according to claim 25, wherein: The method further comprises: The terminal device receives a second RRC signaling sent by the network device, where the second RRC signaling includes a CG-PUSCH configuration, and the CG-PUSCH configuration includes an identifier of the sixth data set; The terminal device receives a second activation signaling sent by the network device, where the second activation signaling is used to activate the CG-PUSCH configuration, and the CG-PUSCH configuration is used by the terminal device to send the sixth data set in a CG manner.

27. A model training device, applied to a network device, comprising: A training unit, configured to train a first model based on a first data set, wherein the first model includes a first part and a second part; a sending unit configured to send a second data set to a terminal device, wherein the first data set includes M first samples, and the second data set includes N second samples, wherein the N second samples are obtained by processing the N first samples by the first part of the first model, and M and N are positive integers; The second data set is used for training a second model for the terminal device, and the trained second model and the second part of the first model are used for CSI reporting.

28. A model training device, applied to a terminal device, comprising: a training unit, configured to train a third model based on a fifth data set, wherein the third model includes a first part and a second part; a sending unit configured to send a sixth data set to a network device, wherein the fifth data set includes M first samples, and the sixth data set includes N first samples and N second samples, wherein the N second samples are obtained by processing the N first samples by the first part of the third model, and M and N are positive integers; The sixth data set is used by the network device to train a fourth model, and the trained fourth model and the first part of the third model are used for CSI reporting.

29. A communication device comprising: A processor and a memory, the memory being used to store a computer program, the processor being used to call and run the computer program stored in the memory to execute the method as claimed in any one of claims 1 to 26.

30. A chip, comprising: A processor, configured to call and run a computer program from a memory, so that a device equipped with the chip executes a method as claimed in any one of claims 1 to 26.

31. A computer-readable storage medium for storing a computer program, wherein the computer program causes a computer to execute the method according to any one of claims 1 to 26.

Images