WO2023123429A1

WO2023123429A1 - Method and device for training channel information feedback model, apparatus, and storage medium

Info

Publication number: WO2023123429A1
Application number: PCT/CN2021/143874
Authority: WO
Inventors: 李德新; 田文强; 刘文东; 肖寒
Original assignee: Oppo广东移动通信有限公司
Priority date: 2021-12-31
Filing date: 2021-12-31
Publication date: 2023-07-06
Also published as: CN118120156A

Abstract

The present application relates to the technical field of communications. Disclosed are a method and device for training a channel information feedback model, an apparatus, and a storage medium. The method comprises: performing a mask operation on initial channel information to obtain masked channel information; inputting the masked channel information into a channel information feedback model, and outputting restored channel information; and training the channel information feedback model on the basis of a discrepancy between the restored channel information and the initial channel information.

Description

Training method, device, equipment and storage medium of channel information feedback model

technical field

The present application relates to the field of communication technology, and in particular to a training method, device, equipment and storage medium of a channel information feedback model.

Background technique

The terminal device generally generates channel information through Channel State Information (CSI) measurement, and feeds the channel information back to the network device.

In related technologies, a channel information feedback scheme based on deep learning has been introduced: the channel information is regarded as an image to be compressed, the encoder is used to compress and feed back the channel information, and the decoder is used at the sending end to reconstruct the compressed channel information , the channel information can be preserved to a greater extent.

Contents of the invention

Embodiments of the present application provide a training method, device, equipment, and storage medium for a channel information feedback model. Described technical scheme is as follows:

According to one aspect of the present application, a method for training a channel information feedback model is provided, which is applied to a source-side terminal, and the method includes:

Perform a masking operation on the initial channel information to obtain masked channel information;

Input the masked channel information into the channel information feedback model, and output the restored channel information;

The channel information feedback model is trained based on an error between the restored channel information and the initial channel information.

According to one aspect of the present application, a method for training a channel information feedback model is provided, which is applied to a target-side terminal, where the channel information feedback model includes: a second encoder and a second decoder, and the method includes:

generating said second decoder;

receiving second transfer learning information sent by the network device, where the second transfer learning information is used to assist in transfer learning, where the second transfer learning information includes: matrix size information corresponding to the second encoder and mask operation, The second encoder is obtained by training based on the mask operation;

jointly training the second encoder and the second decoder based on the matrix size information;

sending the trained second decoder to the network device.

According to one aspect of the present application, a method for training a channel information feedback model is provided, which is applied to a network device, and the method includes:

Sending second transfer learning information to the target side terminal, where the second transfer learning information is used to assist transfer learning, where the second transfer learning information includes: matrix size information corresponding to the second encoder and mask operation, the The second encoder is obtained by training based on the mask operation;

receiving the second decoder sent by the target-side terminal, where the second decoder is trained by the target-side terminal after performing transfer learning based on the second transfer learning information.

According to one aspect of the present application, a training device for a channel information feedback model is provided, the device comprising: a mask module, a model processing module and a training module;

The masking module is configured to perform a masking operation on initial channel information to obtain masked channel information;

The model processing module is configured to input the masked channel information into the channel information feedback model, and output restored channel information;

The training module is configured to train the channel information feedback model based on the error between the recovered channel information and the initial channel information.

According to one aspect of the present application, a training device for a channel information feedback model is provided, the channel information feedback model includes: a second encoder and a second decoder, and the device includes: a decoder generating module and an information receiving module , a training module and a decoder sending module;

The decoder generating module is configured to generate the second decoder;

The information receiving module is configured to receive second transfer learning information sent by a network device, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: the second encoder and mask Matrix size information corresponding to the code operation, the second encoder is obtained by training based on the mask operation;

The training module is configured to jointly train the second encoder and the second decoder based on the matrix size information;

The decoder sending module is configured to send the trained second decoder to the network device.

According to one aspect of the present application, a training device for a channel information feedback model is provided, the device comprising: an information sending module and a decoder receiving module;

The information sending module is configured to send second transfer learning information to the target terminal, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: a second encoder and a mask operation Corresponding matrix size information, the second encoder is obtained by training based on the mask operation;

The decoder receiving module is configured to receive the second decoder sent by the target terminal, the second decoder is obtained by training after the target terminal performs transfer learning based on the second transfer learning information .

According to one aspect of the present application, a terminal device is provided, and the terminal device includes: a processor; wherein,

The processor is configured to perform a masking operation on initial channel information to obtain masked channel information;

The processor is configured to input the masked channel information into a channel information feedback model, and output restored channel information;

The processor is configured to train the channel information feedback model based on an error between the recovered channel information and the initial channel information.

According to one aspect of the present application, a terminal device is provided, and the terminal device includes: a processor and a transceiver connected to the processor; wherein,

the processor, configured to generate a second decoder;

The transceiver is configured to receive second transfer learning information sent by a network device, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: the second encoder and a mask Operating the corresponding matrix size information, the second encoder is obtained by training based on the mask operation;

The processor is configured to jointly train the second encoder and the second decoder based on the matrix size information;

The transceiver is configured to send the trained second decoder to the network device.

According to one aspect of the present application, a network device is provided, and the network device includes: a transceiver; wherein,

The transceiver is configured to send second transfer learning information to the target-side terminal, the second transfer learning information is used to assist in transfer learning, and the second transfer learning information includes: a second encoder corresponding to a mask operation The matrix size information of the second encoder is obtained by training based on the mask operation;

The transceiver is configured to receive the second decoder sent by the target-side terminal, where the second decoder is trained by the target-side terminal after performing transfer learning based on the second transfer learning information.

According to one aspect of the present application, a computer-readable storage medium is provided, and executable instructions are stored in the readable storage medium, and the executable instructions are loaded and executed by a processor to implement the channel described in the above aspect Training methods for information feedback models.

According to an aspect of an embodiment of the present application, a chip is provided, the chip includes a programmable logic circuit and/or program instructions, and when the chip is run on a computer device, it is used to realize the channel information described in the above aspect Feedback model training method.

According to one aspect of the present application, a computer program product is provided. When the computer program product is run on a processor of a computer device, the computer device executes the method for training a channel information feedback model described in the above aspect.

The technical solutions provided by the embodiments of the present application at least include the following beneficial effects:

In the case of implementing the channel information feedback scheme based on deep learning, during model training, use the mask operation to shield part of the initial channel information, reduce the redundant information input during the channel information feedback model training, and reduce the resources for model training Overhead, speed up the training speed of the model, and improve the generalization ability of the training model.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

FIG. 1 is a schematic diagram of a channel information feedback system provided by an exemplary embodiment of the present application;

FIG. 2 is a schematic diagram of transfer learning based on a pre-training-fine-tuning mode provided by an exemplary embodiment of the present application;

Fig. 3 is a block diagram of a communication system provided by an exemplary embodiment of the present application;

FIG. 4 is a flow chart of a method for training a channel information feedback model provided in an exemplary embodiment of the present application;

Fig. 5 is a schematic diagram of a mask operation provided by an exemplary embodiment of the present application;

Fig. 6 is a schematic diagram of a channel information feedback model in the form of an encoder-decoder provided by an exemplary embodiment of the present application;

FIG. 7 is a flow chart of a method for training a channel information feedback model provided in an exemplary embodiment of the present application;

Fig. 8 is a schematic diagram of a mask operation provided by an exemplary embodiment of the present application;

FIG. 9 is a schematic diagram of a channel information feedback system provided by an exemplary embodiment of the present application;

FIG. 10 is a flowchart of a method for training a channel information feedback model provided in an exemplary embodiment of the present application;

Fig. 11 is a flowchart of a training method of a channel information feedback model provided by an exemplary embodiment of the present application;

FIG. 12 is a flow chart of a method for training a channel information feedback model provided in an exemplary embodiment of the present application;

FIG. 13 is a schematic diagram of a training process of a channel information feedback model provided by an exemplary embodiment of the present application;

Fig. 14 is a structural block diagram of a training device for a channel information feedback model provided by an exemplary embodiment of the present application;

FIG. 15 is a structural block diagram of a training device for a channel information feedback model provided by an exemplary embodiment of the present application;

Fig. 16 is a structural block diagram of a training device for a channel information feedback model provided by an exemplary embodiment of the present application;

Fig. 17 is a schematic structural diagram of a communication device provided by an exemplary embodiment of the present application.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application clearer, the implementation manners of the present application will be further described in detail below in conjunction with the accompanying drawings.

First, a brief introduction to the technical knowledge involved in the embodiments of this application:

Codebook-Based Eigenvector Feedback Scheme

In the current New Radio (NR) system, a codebook-based eigenvector feedback scheme is usually used to enable the base station to obtain downlink CSI. Specifically, the base station sends a downlink channel state information reference signal (Channel State Information-Reference Signal, CSI-RS) to the terminal, and the terminal uses the CSI-RS to estimate the CSI of the downlink channel, and performs eigenvalue decomposition on the estimated downlink channel, The eigenvector corresponding to the downlink channel is obtained. According to certain rules, the terminal calculates the corresponding matching codeword coefficients of the feature vector in the preset codebook and performs quantization feedback, and the terminal restores the feature vector according to the quantized CSI fed back by the user.

Channel information feedback scheme based on deep learning

In view of artificial intelligence (AI) technology, especially deep learning has achieved great success in computer vision, natural language processing, etc., the field of communication has begun to try to use deep learning to solve technical problems that are difficult to solve by traditional communication methods, such as deep learning. study. The neural network architecture commonly used in deep learning is nonlinear and data-driven. It can extract features from the actual channel matrix data, and restore the channel matrix information compressed and fed back by the terminal side as much as possible on the base station side. While ensuring the restoration of channel information, It also provides a possibility for the terminal side to reduce the CSI feedback overhead. The CSI feedback based on deep learning regards the channel information as the image to be compressed, uses the deep learning self-encoder to compress the channel information, and reconstructs the compressed channel image at the sending end, which can preserve the channel information to a greater extent .

A typical channel information feedback system is shown in FIG. 1 . The entire feedback system is divided into encoder and decoder parts, which are deployed at the sending end and receiving end respectively. After the transmitting end obtains the channel information through channel estimation, the channel information matrix is compressed and encoded through the neural network of the encoder, and the compressed bit stream is fed back to the receiving end through the air interface feedback link, and the receiving end passes the decoder according to the feedback bit stream The channel information is restored to obtain complete feedback channel information.

The encoder shown in Figure 1 uses the superposition of multiple layers of fully connected layers, and the design of the convolutional layer and residual structure is used in the decoder. Exemplarily, on the side of the encoder, the information is input into the encoder, and the information is firstly convoluted through the convolution (conv) layer, and then the dimensions of the information are changed through the reshape (Reshape) layer, and then through the full connection (dense ) layer to complete the encoding of the information; on the decoder side, the input information is first processed through the fully connected (dense) layer, and then the information is input into the semantic segmentation network RefineNet for processing. RefineNet includes: reshaping ( Reshape) layer, at least one convolution (conv) layer and the design of the residual structure, and then perform convolution (conv) on the information to complete the decoding of the information. Under the condition that the encoding and decoding framework remains unchanged, the network model structure inside the encoder and decoder can be flexibly designed.

Transfer learning based on pre-training-fine-tuning model

Migration learning can be understood as using existing knowledge, models, and structures to help achieve learning goals on target data. Transfer learning based on the pre-training-fine-tuning mode refers to: training a network in the source domain, directly using it for the data of the target domain, and fine-tuning on the data of the target domain, as shown in Figure 2. Therefore, transfer learning based on the pre-training-fine-tuning mode can make better use of limited computing resources, and can also deal with the problem of insufficient data in new scenarios.

The channel information feedback in the related art is a codebook-based eigenvector feedback scheme. However, this scheme only selects the optimal feedback matrix and corresponding feedback coefficients from the codebook according to the estimated channel, but the codebook itself is The preset finite sequence, that is, the mapping process from the estimated channel to the channel in the codebook is quantized and lossy. At the same time, the fixed codebook design cannot be dynamically adjusted according to channel changes, which reduces the accuracy of the feedback channel information, thereby reducing the performance of precoding.

Furthermore, the existing deep learning-based channel information feedback schemes use deep neural networks (Deep Neural Networks, DNN), convolutional neural networks (Convolution Neural Networks, CNN) to directly encode the channel information obtained after channel estimation. Compressed feedback, compared with the traditional codebook-based channel information feedback, significantly improves the feedback accuracy. However, the model performance of the channel information feedback scheme based on deep learning is strongly related to the diversity of data, which requires a large amount of real channel data to provide support, and the cost of real channel data collection is high. At the same time, the training process also brings a lot of computing overhead.

In addition, due to the unstable wireless environment, the data distribution will change over time. Under the limited data set, even if the model is fully trained, it is difficult to guarantee the performance of the model after the data distribution changes over time.

Therefore, how to cope with the changes in data distribution brought about by time lapse under different channel scenarios, while ensuring the accuracy of channel vector compression feedback and recovery, is an urgent model generalization problem to be solved.

In view of the above problems, the embodiment of the present application proposes a training method of a channel information feedback model. In the case of implementing a channel information feedback scheme based on deep learning, when performing model training, the mask operation is used to shield part of the initial channel information, Reduce the redundant information input during channel information feedback model training, reduce the resource overhead of model training, accelerate the training speed of the model, and improve the generalization ability of the training model.

FIG. 3 shows a block diagram of a communication system provided by an exemplary embodiment of the present application. The communication system may include: an access network 12 and a terminal device 14 .

The access network 12 includes several network devices 120 . The network device 120 may be a base station, and the base station is a device deployed in an access network to provide a wireless communication function for a terminal. The base station may include various forms of macro base stations, micro base stations, relay stations, access points and so on. In systems using different wireless access technologies, the names of devices with base station functions may be different. For example, in LTE systems, they are called eNodeB or eNB; in 5G NR-U systems, they are called gNodeB or gNB. . As communications technology evolves, the description "base station" may change. For convenience in this embodiment of the present application, the above-mentioned devices that provide the wireless communication function for the terminal device 14 are collectively referred to as network devices.

The terminal device 14 may include various handheld devices with wireless communication functions, vehicle-mounted devices, wearable devices, computing devices or other processing devices connected to wireless modems, as well as various forms of user equipment, mobile stations (Mobile Station, MS) , terminal (terminal device) and so on. For convenience of description, the devices mentioned above are collectively referred to as terminals. The network device 120 and the terminal device 14 communicate with each other through a certain air interface technology, such as a Uu interface.

Optionally, the terminal device 14 includes: a source-side terminal and a target-side terminal. Wherein, the source-side terminal is a device for performing the pre-training phase of the model in the transfer learning, and the target-side terminal is a device for performing the fine-tuning phase of the model in the transfer learning.

The technical solutions of the embodiments of the present application can be applied to various communication systems, such as: Global System of Mobile Communication (GSM) system, Code Division Multiple Access (CDMA) system, wideband code division multiple access (Wideband Code Division Multiple Access, WCDMA) system, General Packet Radio Service (General Packet Radio Service, GPRS), Long Term Evolution (Long Term Evolution, LTE) system, LTE Frequency Division Duplex (Frequency Division Duplex, FDD) system, LTE Time Division Duplex (Time Division Duplex, TDD) system, Advanced Long Term Evolution (LTE-A) system, New Radio (NR) system, evolution system of NR system, LTE on unlicensed frequency band (LTE-based access to Unlicensed spectrum, LTE-U) system, NR-U system, Universal Mobile Telecommunication System (UMTS), Worldwide Interoperability for Microwave Access (WiMAX) communication system, Wireless local area network (Wireless Local Area Networks, WLAN), wireless fidelity (Wireless Fidelity, WiFi), 6th generation mobile communication technology (6-Generation, 6G) system, next generation communication system or other communication systems, etc.

Generally speaking, the number of connections supported by traditional communication systems is limited and easy to implement. However, with the development of communication technology, mobile communication systems will not only support traditional communication, but also support, for example, Device to Device (Device to Device, D2D) communication, Machine to Machine (M2M) communication, Machine Type Communication (MTC), Vehicle to Vehicle (V2V) communication and Vehicle to Everything (V2X) system, etc. The embodiments of the present application may also be applied to these communication systems.

Fig. 4 shows a flowchart of a method for training a channel information feedback model provided by an exemplary embodiment of the present application. The method can be applied to a terminal device in a communication system as shown in FIG. 3, and the method includes:

Step 410: Perform a masking operation on the initial channel information to obtain masked channel information.

Wherein, the initial channel information is channel information determined after the terminal device performs channel estimation.

Wherein, the masking operation refers to an operation of masking part of information to reduce redundant information.

It can be understood that channel information is characterized by high redundancy. In order to solve this situation, some channel information can be shielded by adding a mask to reduce redundant information. In the field of image processing, visual images are also characterized by high redundancy, for example, missing pixel information can be recovered from adjacent pixel blocks. In the embodiment of the present application, a method of using a mask operation to hide initial channel information is proposed, thereby reducing redundant information.

Exemplarily, a schematic diagram of a masking operation is shown in FIG. 5 . After performing the masking operation on the initial channel information H, the masked channel information H' is obtained. Compared with the masked channel information H', the initial channel information H has more redundant information. It can be understood that FIG. 5 is only an exemplary illustration, and in practice, channel information may not be presented in the same form as the image shown in FIG. 5 .

Step 420: Input the masked channel information into the channel information feedback model, and output the restored channel information.

Wherein, the channel information feedback model is a model for compressing and feeding back the input channel information, and reconstructing and recovering the compressed channel information.

In the embodiment of the present application, after the initial channel information is masked to obtain the masked channel information, the masked channel information is used as the input of the channel information feedback model, and the channel information feedback model is used to predict the masked channel information, so that the output recovery channel information.

Optionally, the channel information feedback model is in the form of an encoder-decoder.

Exemplarily, refer to FIG. 6 , which shows a schematic diagram of processing channel information using an encoder-decoder channel information feedback model. In the current feedback cycle, after the transmitting end performs channel estimation, the encoder compresses and encodes the estimated channel information H, and feeds it back to the receiving end through the feedback link of the air interface. In more detail, the feedback link of the air interface actually transmits a feedback vector, which is obtained from the output of the neural network of the encoder at the transmitting end, and is used as part of the input of the neural network at the receiving end for channel information at the receiving end. recover.

Step 430: Based on the error between the recovered channel information and the initial channel information, train the channel information feedback model.

After obtaining the restored channel information output by the channel information feedback model, compare the restored channel information with the corresponding initial channel information to judge the accuracy of the masked content in the initial channel information predicted by the channel information feedback model, and then When there is an error between the restored channel information and the corresponding initial channel information, the channel information feedback model is corrected according to the existing error, so that the generated channel information feedback model has the ability to reconstruct and restore the channel information.

To sum up, the technical solution provided by this embodiment, in the case of implementing the channel information feedback scheme based on deep learning, uses the mask operation to shield part of the initial channel information during model training, reducing the channel information feedback model training. The redundant information input at the time reduces the resource overhead of model training, accelerates the training speed of the model, and improves the generalization ability of the training model.

Next, the masking operation will be further described.

Fig. 7 shows a flowchart of a method for training a channel information feedback model provided by an exemplary embodiment of the present application. The method can be applied to a terminal device in a communication system as shown in FIG. 3, and the method includes:

Step 710: Divide the channel matrix used to represent the initial channel information into multiple non-overlapping matrix blocks.

Wherein, the matrix size information corresponding to each divided matrix block is the same.

Exemplarily, the channel matrix used to represent the initial channel information is a 25*25 matrix, and the matrix is divided into 25 5*5 matrix blocks.

Step 720: Generate position indices for the matrix blocks to form a sequence of matrix blocks.

Wherein, the position index is an index used to characterize the position of each matrix block in the matrix block sequence.

Exemplarily, 25 matrix blocks correspond to position indices of 0, 1, .

Step 730: Sampling the matrix block sequence, and masking unsampled matrix blocks in the matrix block sequence to obtain masked channel information.

That is, the sampled matrix blocks in the matrix block sequence are retained, and the unsampled matrix blocks are deleted, so as to obtain masked channel information.

Optionally, the sampling manner corresponding to the sampling includes: random sampling; or grid sampling. That is, the selection scheme of the masking operation includes a random masking strategy and a grid masking strategy. It can be understood that the selection scheme of the masking operation is not limited to the above two masking strategies, for example, using other prior knowledge to set the mask distribution is within the protection scope of the present application. Wherein, the grid sampling may be grid sampling at equal intervals.

Exemplarily, the sampling corresponding to the mask operation shown in FIG. 5 is random sampling, and the sampling corresponding to the mask operation shown in FIG. 8 is grid sampling.

Exemplarily, the terminal device randomly samples the matrix block sequence according to uniform distribution, and the sampling rate is 50%. Exemplarily, the terminal device performs grid sampling on the matrix block sequence according to uniform distribution, and the sampling rate is 25%.

It can be understood that the foregoing sampling rate is only an exemplary description, and this embodiment of the present application does not limit the numerical value of the sampling rate. Exemplarily, when the amount of channel information on the local side is large, a smaller sampling rate is adopted; when the amount of channel information on the local side is small, a larger sampling rate is adopted.

Step 740: Input the masked channel information into the channel information feedback model, and output the restored channel information.

For the specific implementation manner of this step, refer to the above-mentioned step 420, which will not be repeated here.

Step 750: Based on the error between the restored channel information and the initial channel information, train the channel information feedback model.

For the specific implementation manner of this step, refer to the above-mentioned step 430, which will not be repeated here.

To sum up, the technical solution provided by this embodiment provides different masking strategies such as a random masking strategy and a grid masking strategy to perform a masking operation to ensure the rationality of the masking operation.

Exemplarily, in combination with the above mask operation and the encoder-decoder model structure, the overall architecture of this embodiment may be shown in FIG. 9 .

In FIG. 9 , the following operation flow is mainly shown: mask operation, encoder, codeword stuffing and decoder.

Masking operation: After the channel matrix H corresponding to the initial channel information is masked, the masked channel information H' is obtained.

Encoder: The masked channel information H' is input to the encoder, and compression coding is performed to obtain compressed coding information.

Codeword stuffing: performing codeword stuffing on the compressed coded information to obtain the filled compressed coded information, that is, to obtain complete compressed coded information.

Decoder: The filled compressed coded information is input to the decoder for decompression to obtain the restored channel information H".

Correspondingly, if the channel information feedback model on the local side includes: a first encoder and a first decoder; inputting the masked channel information into the channel information feedback model, the step of obtaining the restored channel information includes:

(1) The masked channel information is input as a model of the first coder, and the masked channel information is compressed through the first coder to obtain compressed coded information.

(2) Filling the compressed coded information with codewords to obtain the filled compressed coded information.

Codeword filling refers to filling codewords at masked positions.

Exemplarily, the compressed coding information obtained by the first encoder is the coding of the visible matrix blocks in the matrix block sequence corresponding to the channel matrix, and based on the position index, the codeword is filled in the corresponding position of the mask.

(3) The compressed coded information after filling is input as the model of the first decoder, and the compressed coded information after filling is decompressed through the first decoder to obtain the restored channel information.

In an exemplary embodiment, the training of the channel information feedback model by the terminal device corresponds to the pre-training stage in the transfer learning of the pre-training-fine-tuning mode. The terminal device is a source-side terminal, and the source-side terminal also needs to The trained encoder is uploaded, and the target terminal performs the fine-tuning stage, and the trained decoder is uploaded to the network device.

Fig. 10 shows a flowchart of a method for training a channel information feedback model provided by an exemplary embodiment of the present application. The method can be applied to a communication system as shown in FIG. 3, and the method includes:

Step 1010: After the training of the channel information feedback model is completed, the terminal at the source side sends the first transfer learning information of the channel information feedback model to the network device.

Correspondingly, the network device receives the first transfer learning information of the channel information feedback model sent by the source-side terminal. The first transfer learning information is used to perform transfer learning on the channel information feedback model.

Wherein, the channel information feedback model at the terminal at the source side includes: a first encoder, and the first transfer learning information includes:

• A first encoder.

That is, the first transfer learning information carries model parameters of the first encoder.

• The matrix size information corresponding to the mask operation.

That is, the first transfer learning information carries matrix size information corresponding to the mask operation performed by the terminal at the source side. Wherein, the matrix size information is used to indicate the size of each matrix block in the matrix block sequence of the input channel information feedback model.

Step 1020: The network device sends second transfer learning information to the target terminal.

Correspondingly, the target-side terminal receives the second migration learning information sent by the network device. The second transfer learning information is used to assist transfer learning.

Wherein, the second transfer learning information includes:

• Second encoder.

That is, the second transfer learning information carries the model parameters of the second encoder. Wherein, the second encoder is obtained by training based on the mask operation.

• The matrix size information corresponding to the mask operation.

That is, the second transfer learning information carries matrix size information corresponding to the mask operation performed by the terminal at the source side. Wherein, the matrix size information is used to indicate the size of each matrix block in the matrix block sequence of the input channel information feedback model.

It can be understood that the matrix size information corresponding to the mask operation in the second transfer learning information is the matrix size information corresponding to the mask operation in the first transfer learning information, and the second encoder in the second transfer learning information is It is obtained based on the first encoder in the first transfer learning information.

Step 1030: The target-side terminal generates a second decoder.

That is, the target side terminal generates a new decoder.

Step 1040: The target terminal performs joint training on the second encoder and the second decoder based on the matrix size information.

That is, the target terminal uses the pre-trained second encoder to jointly train the second encoder and the new second decoder under the new data set, so as to complete transfer learning.

It is understandable that the transfer learning of the pre-training-fine-tuning mode refers to the pre-training of a network, directly using it for the data of the target scene, and fine-tuning on the target scene data. The model can help other scenarios to achieve the same function. In the embodiment of the present application, the second encoder is pre-trained, and the pre-trained second encoder is used for retraining together with the new second decoder, thereby saving computing resources of the target-side terminal.

Step 1050: The target-side terminal sends the trained second decoder to the network device.

Correspondingly, the network device receives the second decoder sent by the target-side terminal, and the second decoder is trained by the target-side terminal after performing transfer learning based on the second transfer learning information.

Optionally, after step 1050, the target-side terminal uses the trained second encoder, and the network device side uses the received second decoder, with the target-side terminal as the sender of the channel information, and the network device as the source of the channel information The receiving end implements a channel information feedback scheme based on deep learning by using the second encoder of the terminal on the target side and the second decoder on the network device side.

To sum up, the technical solution provided by this embodiment is to enhance the design of the channel information feedback model in the form of encoder-decoder in the migration scenario of pre-training-fine-tuning mode, and use the mask operation to reduce the input in the pre-training stage. Redundant information accelerates the pre-training speed of the model, improves the generalization ability of the pre-trained model, and improves model performance.

Next, a manner in which the target terminal performs joint training on the second encoder and the second decoder based on the matrix size information will be described.

(1) Divide the channel matrix used to represent the initial channel information into multiple non-overlapping matrix blocks according to the matrix size information, and the multiple matrix blocks form a matrix block sequence.

Wherein, the matrix size information is used to indicate the size of each matrix block in the matrix block sequence of the input channel information feedback model.

Exemplarily, the matrix size information indicates that the size of the matrix block is 5*5, and then the terminal on the target side divides the channel matrix corresponding to the channel information on its own side into multiple matrix blocks of 5*5.

(2) The matrix block sequence is input as a model of the second encoder, and the matrix block sequence is compressed through the second encoder to obtain compressed coding information.

(3) The compressed coded information is input as a model of the second decoder, and the compressed coded information is decompressed via the second decoder to obtain restored channel information.

(4) Jointly train the second encoder and the second decoder based on the error between the recovered channel information and the initial channel information.

As shown in the above steps, after migrating to the target domain, the second encoder in the target domain supports variable-length sequence input, and the input of the second encoder is a matrix block sequence of a complete channel, do not perform mask processing, and correspondingly , after the second encoder outputs the compressed encoding information, it does not need to perform codeword padding, so as to make full use of the limited data of the current scene.

In a possible implementation manner, the second encoder is an encoder indicated to the network device by a source-side terminal.

That is, after receiving the first transfer learning information sent by the source-side terminal, the network device directly sends the first transfer-learning information as the second transfer-learning information to the target-side terminal device. The second encoder in the second transfer learning information in the above embodiment is equivalent to the first encoder in the first transfer learning information.

To sum up, in the technical solution provided by this embodiment, an encoder is pre-trained by a source-side terminal, and the encoder is migrated to a target-side terminal, and the redundant information input in the pre-training stage is reduced by using a mask operation. Accelerated the pre-training speed of the model.

Exemplarily, the above-mentioned implementation manner is exemplarily described in conjunction with FIG. 11 , as shown in FIG. 11 , the following steps are performed:

Step 1101: The source-side terminal obtains channel data and executes a masking strategy.

Exemplarily, the terminal at the source side divides the channel data into regular non-overlapping N small block matrices (patches). A position index 0, 1, 2, ..., N-1 is generated for each matrix block, forming a sequence of matrix blocks. The sequence of matrix blocks is then sampled and the remaining matrix blocks are masked (ie deleted).

Step 1102: The source-side terminal jointly trains the encoder and the decoder.

Exemplarily, the masked channel information is used as the input of the encoder, and a codeword filling operation is added after the encoder accordingly, and the input of the decoder is the filled codeword, including the visible matrix block codeword and The padding codeword at the corresponding position of the mask.

In this embodiment, the decoder and encoder can adopt an asymmetric design. Compared with the parameter scale of the encoder, the decoder can appropriately reduce the number of network layers and parameters, thereby reducing the pre-training time.

Step 1103: The terminal at the source side sends the matrix size information corresponding to the encoder and the mask operation to the network device.

Step 1104: The network device sends the matrix size information corresponding to the encoder and the mask operation to the target terminal.

Step 1105: The target terminal generates a new decoder.

Step 1106: The terminal on the target side processes the channel information into a matrix block sequence adapted by the encoder according to the matrix size information.

Step 1107: The target side terminal uses the pre-trained encoder to combine with the new decoder, and retrains under the new data set to complete model migration.

Step 1108: the target terminal synchronizes the decoder to the network device.

In another possible implementation manner, the second encoder is a global encoder obtained by aggregate calculation of model parameters of multiple encoders by the network device, and the multiple encoders come from multiple source-side terminals respectively.

Fig. 12 shows a flowchart of a method for training a channel information feedback model provided by an exemplary embodiment of the present application. The method can be applied to a communication system as shown in FIG. 3, and the method includes:

Step 1210: After the training of the channel information feedback model is completed, multiple source-side terminals respectively send the first transfer learning information of the channel information feedback model to the network device.

Correspondingly, the network device receives the first transfer learning information of the channel information feedback model respectively sent by multiple source-side terminals. The first transfer learning information is used to perform transfer learning on the channel information feedback model.

• A first encoder.

• The matrix size information corresponding to the mask operation.

Optionally, in order to unify the masking operation, before step 1210, the network device delivers the same masking policy parameter to multiple source-side terminals, where the masking policy parameter is a parameter related to the masking operation.

Optionally, the mask policy parameters include at least one of the following:

• The matrix size information corresponding to the mask operation.

• The sampling information corresponding to the mask operation.

Wherein, the sampling information is used to indicate the execution mode of sampling in the mask operation. Exemplarily, the sampling information includes at least one of the following: sampling mode; sampling rate.

Step 1220: The network device aggregates and calculates the model parameters of multiple trained first encoders to obtain a global encoder.

Among them, aggregation calculation refers to a calculation method that calculates a set of values and returns a single value. In the embodiment of the present application, the model parameters of multiple first encoders are aggregated and calculated, and a final model parameter of a global encoder is returned. The embodiment of the present application does not limit the specific implementation form of the aggregated calculation.

Step 1230: The network device sends the second transfer learning information to the target terminal.

Wherein, the second transfer learning information includes: the global encoder and matrix size information corresponding to the mask operation, and the global encoder is obtained by training based on the mask operation.

It can be understood that the matrix size information corresponding to the mask operation in the second transfer learning information is the matrix size information corresponding to the mask operation in the first transfer learning information, and the global encoder in the second transfer learning information is based on It is obtained by aggregate calculation of multiple first encoders in the multiple first transfer learning information.

Step 1240: The target-side terminal generates a second decoder.

That is, the target side terminal generates a new decoder.

Step 1250: The target-side terminal performs joint training on the global encoder and the second decoder.

That is, the target terminal uses the pre-trained global encoder to jointly train the global encoder and the new second decoder under the new data set, so as to complete transfer learning.

Step 1260: The target-side terminal sends the trained second decoder to the network device.

To sum up, in the technical solution provided by this embodiment, multiple source-side terminals cooperate to train to obtain a shared global encoder, and the data redundancy under multiple terminal devices is higher, and the mask operation can be used to significantly reduce Small data redundancy is conducive to enhancing the representation ability of the model to extract potential features, and at the same time speeding up the pre-training speed of the model.

Exemplarily, the above-mentioned implementation manner is exemplarily described in conjunction with FIG. 13 , as shown in FIG. 13 , the following steps are performed:

Step 1301, unify the masking strategy: the network device uniformly configures masking strategy parameters, and then distributes them to n candidate source-side terminals: source-side terminal 1, source-side terminal 2, . . . , source-side terminal n.

Step 1302, pre-training the encoder: the source-side terminals each perform a masking operation, and use the masked channel information as input to train the encoder-decoder.

For a single end device, the masking strategy-based autoencoder network architecture is consistent. The components include: mask operation, encoder, codeword filling, and decoder. Each terminal device requires the above four components, and the working structure and flow of each terminal device can refer to the embodiment shown in FIG. 9 , which will not be described in detail here.

Step 1303, uploading the encoder: each source-side terminal deletes the decoder part to save device memory resources, retains only the encoder part, and uploads the encoder to the network device for synchronization.

Step 1304, aggregation calculation: the base station server or the over-the-air computing node performs aggregation calculation on the encoder model parameters of each coordinated source-side terminal to obtain a global encoder.

Step 1305, delivering global encoder and matrix size information: network devices, such as base station servers or air computing nodes, deliver the global encoder and matrix size information corresponding to mask operations to the target terminal.

It can be understood that there may be multiple target-side terminals, and it is not limited to source-side terminals. All terminals under the network device can be used as candidate target-side terminals, depending on system policies.

Step 1306, fine-tuning stage: the terminal on the target side uses the matrix size information to process the existing channel information data into a matrix block sequence, without masking, and directly inputs the complete matrix block sequence to the encoder-decoder.

It should be noted that the encoder here is a global encoder, but the target-side terminal needs to regenerate an initialized decoder. The size of the encoder model here can appropriately increase the parameter scale in order to obtain better decoding performance.

Step 1307, upload the encoder: the encoder is a model that needs to be deployed at the receiving end, so the target terminal must also send the trained decoder to the network device to ensure that the network device can send the encoder to the target terminal The codewords are correctly analyzed and restored to complete channel information.

As shown in the above steps, in the process of implementing transfer learning, each participant does not need to share the data in the local device, which fully guarantees the data privacy and security of the participants.

It should be noted that the foregoing method embodiments may be implemented individually or in combination, which is not limited in the present application.

In each of the above-mentioned embodiments, the steps performed by the source-side terminal can independently realize the training method of the channel information feedback model on the side of the source-side terminal, and the steps performed by the target-side terminal can independently realize the channel information on the side of the target-side terminal. In the training method of the information feedback model, the steps performed by the network device can be independently implemented as the training method of the channel information feedback model on the network device side.

Fig. 14 shows a structural block diagram of an apparatus for training a channel information feedback model provided by an exemplary embodiment of the present application. The apparatus can be implemented as a source-side terminal, or can be implemented as a part of a source-side terminal. The apparatus includes: code module 1402, model processing module 1404 and training module 1406;

The masking module 1402 is configured to perform a masking operation on initial channel information to obtain masked channel information;

The model processing module 1404 is configured to input the masked channel information into the channel information feedback model, and output restored channel information;

The training module 1406 is configured to train the channel information feedback model based on the error between the recovered channel information and the initial channel information.

In an optional embodiment, the masking module 1402 is configured to:

dividing the channel matrix used to represent the initial channel information into a plurality of non-overlapping matrix blocks;

generating a position index for the matrix block to form a sequence of matrix blocks;

Sampling the matrix block sequence and masking unsampled matrix blocks in the matrix block sequence to obtain the masked channel information.

In an optional embodiment, the sampling method corresponding to the sampling includes:

random sampling;

or,

Raster sampling.

In an optional embodiment, the channel information feedback model includes: a first encoder and a first decoder;

The model processing module 1404 is configured to:

inputting the masked channel information as a model of the first encoder, and compressing the masked channel information via the first encoder to obtain compressed encoded information;

performing codeword padding on the compressed coded information to obtain the filled compressed coded information;

inputting the filled compressed coded information as a model of the first decoder, decompressing the filled compressed coded information via the first decoder, to obtain the restored channel information.

In an optional embodiment, the device further includes: an information reporting module;

The information reporting module is configured to send the first migration learning information of the channel information feedback model to the network device after the training of the channel information feedback model is completed, and the first migration learning information is used for the channel Information feedback model for transfer learning.

In an optional embodiment, the channel information feedback model includes: a first encoder, and the first transfer learning information includes:

said first encoder;

The matrix size information corresponding to the mask operation.

In an optional embodiment, the device further includes: a parameter receiving module;

The parameter receiving module is configured to receive a masking policy parameter issued by a network device, and the masking policy parameter is a parameter related to the masking operation.

In an optional embodiment, the mask policy parameters include at least one of the following:

Matrix size information corresponding to the mask operation;

The sampling information corresponding to the mask operation.

Fig. 15 shows a structural block diagram of a training device for a channel information feedback model provided by an exemplary embodiment of the present application. The device can be implemented as a target terminal, or can be implemented as a part of the target terminal. The device includes: decoding A decoder generating module 1502, an information receiving module 1504, a training module 1506 and a decoder sending module 1508;

The decoder generation module 1502, configured to generate the second decoder;

The information receiving module 1504 is configured to receive second transfer learning information sent by a network device, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: the second encoder and Matrix size information corresponding to the mask operation, the second encoder is obtained by training based on the mask operation;

The training module 1506 is configured to jointly train the second encoder and the second decoder based on the matrix size information;

The decoder sending module 1508 is configured to send the trained second decoder to the network device.

In an optional embodiment, the training module 1506 is used for:

Divide the channel matrix used to represent the initial channel information into a plurality of non-overlapping matrix blocks according to the matrix size information, and the plurality of matrix blocks form a matrix block sequence;

The matrix block sequence is input as a model of the second encoder, and the matrix block sequence is compressed through the second encoder to obtain compressed encoding information;

inputting the compressed coded information as a model of the second decoder, and decompressing the compressed coded information via the second decoder to obtain restored channel information;

The second encoder and the second decoder are jointly trained based on an error between the recovered channel information and the initial channel information.

In an optional embodiment, the second encoder is an encoder indicated to the network device by a source-side terminal.

In an optional embodiment, the second encoder is a global encoder obtained by aggregate calculation of model parameters of multiple encoders by the network device, and the multiple encoders come from multiple one of the source-side terminals.

Fig. 16 shows a structural block diagram of a training device for a channel information feedback model provided by an exemplary embodiment of the present application. The device can be implemented as a network device, or can be implemented as a part of the network device. The device includes: an information sending module 1602 and a decoder receiving module 1604;

The information sending module 1602 is configured to send second transfer learning information to the target terminal, the second transfer learning information is used to assist in transfer learning, and the second transfer learning information includes: a second encoder and a mask Operating the corresponding matrix size information, the second encoder is obtained by training based on the mask operation;

The decoder receiving module 1604 is configured to receive the second decoder sent by the target terminal, where the second decoder is obtained by training after the target terminal performs transfer learning based on the second transfer learning information. of.

In an optional embodiment, the second encoder is an encoder indicated to the network device by a source-side terminal;

The device also includes: an information receiving module;

The information receiving module is configured to receive a piece of first transfer learning information sent by the source-side terminal, the first transfer learning information is used to assist transfer learning, and the first transfer learning information includes: a first encoder Matrix size information corresponding to the mask operation.

In an optional embodiment, the second encoder is a global encoder obtained by aggregate calculation of model parameters of multiple encoders by the network device;

The device also includes: an information receiving module and an aggregation calculation module;

The information receiving module is configured to receive a plurality of first transfer learning information respectively sent by a plurality of the source-side terminals, the first transfer learning information is used to assist transfer learning, and the first transfer learning information includes: Matrix size information corresponding to the first encoder and the mask operation;

The aggregation calculation module is configured to perform aggregation calculation on model parameters of multiple trained first encoders to obtain the global encoder.

In an optional embodiment, the device further includes: a parameter configuration module;

The parameter configuration module is configured to deliver the same masking policy parameter to multiple terminals at the source side, where the masking policy parameter is a parameter related to the masking operation.

Matrix size information corresponding to the mask operation;

The sampling information corresponding to the mask operation.

FIG. 17 shows a schematic structural diagram of a communication device (terminal device or network device) provided by an exemplary embodiment of the present application. The communication device 1700 includes: a processor 1701 , a transceiver 1702 and a memory 1703 .

The processor 1701 includes one or more processing cores, and the processor 1701 executes various functional applications by running software programs and modules.

The transceiver 1702 can be used to receive and send information, and the transceiver 1702 can be a communication chip.

The memory 1703 may be used to store a computer program, and the processor 1701 is used to execute the computer program, so as to implement various steps performed by the communication device in the foregoing method embodiments.

In addition, the memory 1703 can be realized by any type of volatile or non-volatile storage device or their combination, and the volatile or non-volatile storage device includes but not limited to: random access memory (Random-Access Memory, RAM) And read-only memory (Read-Only Memory, ROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, EPROM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, EEPROM), flash memory or other solid-state storage technologies, compact disc read-only memory (CD-ROM), high-density digital video disc (Digital Video Disc, DVD) or other optical storage, tape cartridges, tapes, disks storage or other magnetic storage devices.

Wherein, when the communication device is implemented as a source-side terminal, the processor 1701 and the transceiver 1702 involved in the embodiment of the present application can perform the steps performed by the source-side terminal in any of the methods shown in the above-mentioned embodiments, where I won't repeat them here.

In a possible implementation manner, when the communication device is implemented as a source-side terminal,

The processor 1701 is configured to perform a masking operation on initial channel information to obtain masked channel information;

The processor 1701 is configured to input the masked channel information into a channel information feedback model, and output restored channel information;

The processor 1701 is configured to train the channel information feedback model based on an error between the recovered channel information and the initial channel information.

Wherein, when the communication device is implemented as a target-side terminal, the processor 1701 and the transceiver 1702 involved in the embodiment of the present application may perform the steps performed by the target-side terminal in any of the methods shown in the above-mentioned embodiments, where I won't repeat them here.

In a possible implementation manner, when the communication device is implemented as a target-side terminal,

The processor 1701 is configured to generate a second decoder;

The transceiver 1702 is configured to receive second transfer learning information sent by a network device, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: the second encoder and mask Matrix size information corresponding to the code operation, the second encoder is obtained by training based on the mask operation;

The processor 1701 is configured to jointly train the second encoder and the second decoder based on the matrix size information;

The transceiver 1702 is configured to send the trained second decoder to the network device.

Wherein, when the communication device is implemented as a network device, the processor 1701 and the transceiver 1702 involved in the embodiment of the present application can execute the steps performed by the network device in any of the methods shown in the above embodiments, which are not described here Let me repeat.

In a possible implementation manner, when the communication device is implemented as a network device,

The transceiver 1702 is configured to send second transfer learning information to the target terminal, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: a second encoder and a mask operation Corresponding matrix size information, the second encoder is obtained by training based on the mask operation;

The transceiver 1702 is configured to receive the second decoder sent by the target-side terminal, where the second decoder is trained by the target-side terminal after performing transfer learning based on the second transfer learning information.

In an exemplary embodiment, a computer-readable storage medium is also provided, the computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, the at least one instruction, the At least one section of program, the code set or instruction set is loaded and executed by the processor to implement the training method of the channel information feedback model executed by the communication device provided in the above method embodiments.

In an exemplary embodiment, a chip is also provided, the chip includes a programmable logic circuit and/or program instructions, and when the chip is run on a computer device, it is used to realize the channel information feedback described in the above aspect The training method of the model.

In an exemplary embodiment, a computer program product is also provided. When the computer program product runs on a processor of a computer device, the computer device executes the method for training a channel information feedback model described in the above aspects.

Those of ordinary skill in the art can understand that all or part of the steps for implementing the above embodiments can be completed by hardware, and can also be completed by instructing related hardware through a program. The program can be stored in a computer-readable storage medium. The above-mentioned The storage medium mentioned may be a read-only memory, a magnetic disk or an optical disk, and the like.

The above are only optional embodiments of the application, and are not intended to limit the application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the application shall be included in the protection of the application. within range.

Claims

A method for training a channel information feedback model, characterized in that the method comprises:

Perform a masking operation on the initial channel information to obtain masked channel information;

Input the masked channel information into the channel information feedback model, and output the restored channel information;

The channel information feedback model is trained based on an error between the restored channel information and the initial channel information.
The method according to claim 1, wherein the masking operation is performed on the initial channel information to obtain the masked channel information, comprising:

dividing the channel matrix used to represent the initial channel information into a plurality of non-overlapping matrix blocks;

generating a position index for the matrix block to form a sequence of matrix blocks;

Sampling the matrix block sequence and masking unsampled matrix blocks in the matrix block sequence to obtain the masked channel information.
The method according to claim 2, wherein the sampling method corresponding to the sampling comprises:

random sampling;

or,

Raster sampling.
The method according to any one of claims 1 to 3, wherein the channel information feedback model comprises: a first encoder and a first decoder;

The inputting the masked channel information into the channel information feedback model to obtain the restored channel information includes:

inputting the masked channel information as a model of the first encoder, and compressing the masked channel information via the first encoder to obtain compressed encoded information;

performing codeword padding on the compressed coded information to obtain the filled compressed coded information;

inputting the filled compressed coded information as a model of the first decoder, decompressing the filled compressed coded information via the first decoder, to obtain the restored channel information.
The method according to any one of claims 1 to 3, wherein the method further comprises:

After the training of the channel information feedback model is completed, the first transfer learning information of the channel information feedback model is sent to the network device, where the first transfer learning information is used to perform transfer learning on the channel information feedback model.
The method according to claim 5, wherein the channel information feedback model includes: a first encoder, and the first transfer learning information includes:

said first encoder;

The matrix size information corresponding to the mask operation.
The method according to any one of claims 1 to 3, wherein the method further comprises:

A masking policy parameter delivered by the network device is received, where the masking policy parameter is a parameter related to the masking operation.
The method according to claim 7, wherein the mask policy parameters include at least one of the following:

Matrix size information corresponding to the mask operation;

The sampling information corresponding to the mask operation.
A method for training a channel information feedback model, wherein the channel information feedback model includes: a second encoder and a second decoder, and the method includes:

generating said second decoder;

receiving second transfer learning information sent by the network device, where the second transfer learning information is used to assist in transfer learning, where the second transfer learning information includes: matrix size information corresponding to the second encoder and mask operation, The second encoder is obtained by training based on the mask operation;

jointly training the second encoder and the second decoder based on the matrix size information;

sending the trained second decoder to the network device.
The method according to claim 9, wherein the joint training of the second encoder and the second decoder based on the matrix size information includes:

Divide the channel matrix used to represent the initial channel information into a plurality of non-overlapping matrix blocks according to the matrix size information, and the plurality of matrix blocks form a matrix block sequence;

The matrix block sequence is input as a model of the second encoder, and the matrix block sequence is compressed through the second encoder to obtain compressed encoding information;

inputting the compressed coded information as a model of the second decoder, and decompressing the compressed coded information via the second decoder to obtain restored channel information;

The second encoder and the second decoder are jointly trained based on an error between the recovered channel information and the initial channel information.
The method according to claim 9 or 10, characterized in that,

The second encoder is an encoder indicated to the network device by a source-side terminal.
The method according to claim 9 or 10, characterized in that,

The second encoder is a global encoder obtained after the network device aggregates and calculates model parameters of multiple encoders, and the multiple encoders come from multiple source-side terminals respectively.
A method for training a channel information feedback model, characterized in that the method comprises:

Sending second transfer learning information to the target side terminal, where the second transfer learning information is used to assist transfer learning, where the second transfer learning information includes: matrix size information corresponding to the second encoder and mask operation, the The second encoder is obtained by training based on the mask operation;

receiving the second decoder sent by the target-side terminal, where the second decoder is trained by the target-side terminal after performing transfer learning based on the second transfer learning information.
The method according to claim 13, wherein the second encoder is an encoder indicated to the network device by a source-side terminal;

The method also includes:

Receive a piece of first transfer learning information sent by the source-side terminal, where the first transfer learning information is used to assist transfer learning, where the first transfer learning information includes: a first encoder corresponding to the mask operation Matrix dimension information.
The method according to claim 13, wherein the second encoder is a global encoder obtained after the network device aggregates and calculates the model parameters of multiple encoders;

The method also includes:

receiving a plurality of first transfer learning information respectively sent by a plurality of source-side terminals, the first transfer learning information is used to assist transfer learning, and the first transfer learning information includes: the first encoder and the mask The matrix size information corresponding to the code operation;

Aggregate calculation is performed on the model parameters of multiple trained first encoders to obtain the global encoder.
The method according to claim 15, further comprising:

delivering the same masking policy parameter to multiple source-side terminals, where the masking policy parameter is a parameter related to the masking operation.
The method according to claim 16, wherein the mask policy parameters include at least one of the following:

Matrix size information corresponding to the mask operation;

The sampling information corresponding to the mask operation.
A training device for a channel information feedback model, characterized in that the device includes: a mask module, a model processing module and a training module;

The masking module is configured to perform a masking operation on initial channel information to obtain masked channel information;

The model processing module is configured to input the masked channel information into the channel information feedback model, and output restored channel information;

The training module is configured to train the channel information feedback model based on the error between the recovered channel information and the initial channel information.
The device according to claim 18, wherein the mask module is configured to:

dividing the channel matrix used to represent the initial channel information into a plurality of non-overlapping matrix blocks;

generating a position index for the matrix block to form a sequence of matrix blocks;

Sampling the matrix block sequence and masking unsampled matrix blocks in the matrix block sequence to obtain the masked channel information.
The device according to claim 19, wherein the sampling mode corresponding to the sampling comprises:

random sampling;

or,

Raster sampling.
The device according to any one of claims 18 to 20, wherein the channel information feedback model comprises: a first encoder and a first decoder;

The model processing module is used for:

inputting the masked channel information as a model of the first encoder, and compressing the masked channel information via the first encoder to obtain compressed encoded information;

performing codeword padding on the compressed coded information to obtain the filled compressed coded information;

inputting the filled compressed coded information as a model of the first decoder, decompressing the filled compressed coded information via the first decoder, to obtain the restored channel information.
The device according to any one of claims 18 to 20, wherein the device further comprises: an information reporting module;

The information reporting module is configured to send the first migration learning information of the channel information feedback model to the network device after the training of the channel information feedback model is completed, and the first migration learning information is used for the channel Information feedback model for transfer learning.
The device according to claim 22, wherein the channel information feedback model includes: a first encoder, and the first transfer learning information includes:

said first encoder;

The matrix size information corresponding to the mask operation.
The device according to any one of claims 18 to 20, wherein the device further comprises: a parameter receiving module;

The parameter receiving module is configured to receive a masking policy parameter issued by a network device, and the masking policy parameter is a parameter related to the masking operation.
The device according to claim 24, wherein the mask policy parameters include at least one of the following:

Matrix size information corresponding to the mask operation;

The sampling information corresponding to the mask operation.
A training device for a channel information feedback model, characterized in that the channel information feedback model includes: a second encoder and a second decoder, and the device includes: a decoder generation module, an information receiving module, a training module, and a decoder Transmitter module;

The decoder generating module is configured to generate the second decoder;

The information receiving module is configured to receive second transfer learning information sent by a network device, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: the second encoder and mask Matrix size information corresponding to the code operation, the second encoder is obtained by training based on the mask operation;

The training module is configured to jointly train the second encoder and the second decoder based on the matrix size information;

The decoder sending module is configured to send the trained second decoder to the network device.
The device according to claim 26, wherein the training module is used for:

Divide the channel matrix used to represent the initial channel information into a plurality of non-overlapping matrix blocks according to the matrix size information, and the plurality of matrix blocks form a matrix block sequence;

The matrix block sequence is input as a model of the second encoder, and the matrix block sequence is compressed through the second encoder to obtain compressed encoding information;

inputting the compressed coded information as a model of the second decoder, and decompressing the compressed coded information via the second decoder to obtain restored channel information;

The second encoder and the second decoder are jointly trained based on an error between the recovered channel information and the initial channel information.
Apparatus according to claim 26 or 27, characterized in that,

The second encoder is an encoder indicated to the network device by a source-side terminal.
Apparatus according to claim 26 or 27, characterized in that,

The second encoder is a global encoder obtained after the network device aggregates and calculates model parameters of multiple encoders, and the multiple encoders come from multiple source-side terminals respectively.
A training device for a channel information feedback model, characterized in that the device includes: an information sending module and a decoder receiving module;

The information sending module is configured to send second transfer learning information to the target terminal, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: a second encoder and a mask operation Corresponding matrix size information, the second encoder is obtained by training based on the mask operation;

The decoder receiving module is configured to receive the second decoder sent by the target terminal, the second decoder is obtained by training after the target terminal performs transfer learning based on the second transfer learning information .
The device according to claim 30, wherein the second encoder is an encoder indicated to the network device by a source-side terminal;

The device also includes: an information receiving module;

The information receiving module is configured to receive a piece of first transfer learning information sent by the source-side terminal, the first transfer learning information is used to assist transfer learning, and the first transfer learning information includes: a first encoder Matrix size information corresponding to the mask operation.
The device according to claim 30, wherein the second encoder is a global encoder obtained by aggregate calculation of model parameters of multiple encoders by network equipment;

The device also includes: an information receiving module and an aggregation calculation module;

The information receiving module is configured to receive a plurality of first transfer learning information respectively sent by a plurality of the source-side terminals, the first transfer learning information is used to assist transfer learning, and the first transfer learning information includes: Matrix size information corresponding to the first encoder and the mask operation;

The aggregation calculation module is configured to perform aggregation calculation on model parameters of multiple trained first encoders to obtain the global encoder.
The device according to claim 32, further comprising: a parameter configuration module;

The parameter configuration module is configured to deliver the same masking policy parameter to multiple terminals at the source side, where the masking policy parameter is a parameter related to the masking operation.
The device according to claim 33, wherein the mask policy parameters include at least one of the following:

Matrix size information corresponding to the mask operation;

The sampling information corresponding to the mask operation.
A terminal device, characterized in that the terminal device includes: a processor; wherein,

The processor is configured to perform a masking operation on initial channel information to obtain masked channel information;

The processor is configured to input the masked channel information into a channel information feedback model, and output restored channel information;

The processor is configured to train the channel information feedback model based on an error between the recovered channel information and the initial channel information.
A terminal device, characterized in that the terminal device includes: a processor and a transceiver connected to the processor; wherein,

the processor, configured to generate a second decoder;

The transceiver is configured to receive second transfer learning information sent by a network device, the second transfer learning information is used to assist transfer learning, and the second transfer learning information includes: the second encoder and a mask Operating the corresponding matrix size information, the second encoder is obtained by training based on the mask operation;

The processor is configured to jointly train the second encoder and the second decoder based on the matrix size information;

The transceiver is configured to send the trained second decoder to the network device.
A network device, characterized in that the network device includes: a transceiver; wherein,

The transceiver is configured to send second transfer learning information to the target terminal, the second transfer learning information is used to assist in transfer learning, and the second transfer learning information includes: a second encoder corresponding to a mask operation The matrix size information of the second encoder is obtained by training based on the mask operation;

The transceiver is configured to receive the second decoder sent by the target-side terminal, where the second decoder is trained by the target-side terminal after performing transfer learning based on the second transfer learning information.
A computer-readable storage medium, characterized in that executable instructions are stored in the readable storage medium, and the executable instructions are loaded and executed by a processor to implement the channel according to any one of claims 1 to 17 Training methods for information feedback models.
A kind of chip, it is characterized in that, described chip comprises programmable logic circuit and/or program instruction, when described chip runs, is used for realizing the training method of channel information feedback model as described in any one of claims 1 to 17 .
A computer program product or computer program, characterized in that the computer program product or computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium, and the processor reads the computer-readable storage medium from the computer-readable storage medium And execute the computer instructions to realize the training method of the channel information feedback model according to any one of claims 1 to 17.