Channel compression feedback-oriented base model training method and device, computer equipment and storage medium

Abstract

The invention relates to a channel compression feedback-oriented base model training method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: processing sample channel state information of each user side through a model of the user side to obtain target bit stream data, processing the target bit stream data of each user side through a model of a base station side to obtain reconstructed channel state information, and obtaining a trained base model; by adopting the method, the task of channel compression feedback of different heterogeneous configurations can be flexibly solved through the trained base model; the condition that multiple users exist in a wireless communication system is comprehensively considered, and the sample channel state information estimated by the multiple users is combined for processing, so that the model training efficiency is improved, and the performance of the base model obtained through training is improved.

Description

Technical Field

[0001] The present application relates to the field of wireless communication technology, and in particular to a training method, apparatus, computer equipment, computer-readable storage medium, and computer program product for a base model for channel compression feedback. Background Art

[0002] The channel compression and feedback (CFE) task is a key means of obtaining downlink channel state information (CSI) in Multiple-Input Multiple-Output (MIMO) and Orthogonal Frequency-Division Multiplexing (OFDM) systems.

[0003] Traditional technologies typically implement channel compression feedback through vector quantization or codebook schemes. However, with the continuous advancement of technology, the number of antennas in wireless communication systems is also increasing, and the feedback volume in traditional schemes is also increasing, resulting in unacceptable feedback overhead. Compressed sensing (CS)-based schemes can fully exploit channel sparsity to achieve efficient compression, but this approach significantly reduces feedback accuracy and also introduces some practical deployment issues. Deep learning technology, due to its powerful nonlinear modeling capabilities, can reduce feedback overhead for channel compression feedback tasks, but it still has limitations. The further development of deep learning technology has led to the emergence of foundation models, which have achieved great success in fields such as natural language processing. These foundation models, which employ large-scale parameter neural networks, can demonstrate excellent few-shot learning capabilities on downstream tasks after self-supervised pre-training on large datasets. The development of large foundation models for various tasks, such as channel compression feedback, remains a pressing technical challenge. Summary of the Invention

[0004] Based on this, it is necessary to provide a training method, device, computer equipment, computer-readable storage medium and computer program product for a base model for channel compression feedback that can flexibly handle multiple heterogeneous configurations to address the above technical problems.

[0005] In a first aspect, the present application provides a training method for a base model for channel compression feedback, wherein the base model includes a base station end and at least one user end, the user end includes an encoder and a transformation module; the base station end includes an inverse transformation module and a decoder; the encoder includes at least one expert hybrid encoding module; the decoder includes at least one expert shared decoding module; the method includes:

[0006] Acquire sample channel state information of at least one user terminal, where the sample channel state information includes information of a target dimension, where the target dimension includes at least two of a time dimension, a space dimension, and a frequency dimension;

[0007] Obtaining a target token corresponding to the sample channel state information of the user end, encoding the target token using the expert model and the gating network in each of the expert hybrid coding modules corresponding to the user end to obtain a first output token; and transforming the first output token using the transformation module to obtain target bit stream data;

[0008] Performing inverse transformation processing on the target bit stream data corresponding to each of the user terminals through the inverse transformation module of the base station terminal to obtain a second output token;

[0009] The second output token is processed by at least one expert shared decoding module included in the decoder to obtain reconstructed channel state information corresponding to each user terminal, and training is performed based on each reconstructed channel state information to obtain a trained base model.

[0010] In one embodiment, the encoder further includes a first position encoding module; the transformation module includes a compression layer and a quantization encoding layer; encoding the target token using the expert model and the gating network in each of the expert hybrid encoding modules corresponding to the user end to obtain a first output token; and transforming the first output token using the transformation module to obtain target bit stream data, including:

[0011] Position-encode the target token using the first position encoding module to obtain a first position encoding token;

[0012] Encoding the first position coding token through the gating network and multiple target expert models in each of the expert hybrid coding modules connected end to end to obtain a first output token;

[0013] compressing the first output token by the compression layer to obtain a compressed feature vector;

[0014] The compressed feature vector is quantized by using the quantization coding layer and the quantization bit width to obtain target bit stream data corresponding to the user end.

[0015] In one embodiment, the hybrid expert encoding module includes a pre-processing layer, an expert processing layer, and a normalization layer, wherein the expert processing layer includes multiple expert models and multiple gating networks; encoding the first position code token through the gating networks and multiple target expert models in the hybrid expert encoding modules connected end to end to obtain the first output token includes:

[0016] For the i-th mixed-expert coding module, perform attention mechanism processing and normalization processing on the input token of the i-th mixed-expert coding module through the pre-processing layer in the i-th mixed-expert coding module to obtain a first token;

[0017] Determining a model weight corresponding to the first token based on the gating network in the i-th expert hybrid encoding module, and screening out an expert activation model from the expert models included in the i-th expert hybrid encoding module based on the model weight;

[0018] Encoding the first tokens by using the expert activation models to obtain the expert output results corresponding to the expert activation models; and obtaining the output tokens of the i-th expert hybrid coding module by using the expert output results of the expert activation models and the model weights corresponding to the expert activation models;

[0019] Determine that the output token of the expert hybrid coding module located at the last position is the first output token; the input token of the expert hybrid coding module located at the initial position is the first position coding token, and the input token of the i-th expert hybrid coding module is the output token of the i-1-th expert hybrid coding module.

[0020] In one embodiment, the inverse transform module includes a splicing layer, a dequantization layer, and an upsampling layer; the inverse transform module at the base station performs inverse transform processing on the target bit stream data corresponding to each user terminal to obtain a second output token, including:

[0021] For each user terminal, splicing the bit streams fed back by the user terminal based on the bit stream reception order by a splicing layer in the inverse transform module of the base station to obtain a spliced ​​bit stream corresponding to the user terminal;

[0022] Performing type inverse transformation on the spliced ​​bit streams corresponding to the respective user terminals through the dequantization layer to obtain feature maps of the respective user terminals;

[0023] The feature maps of each user terminal are input into the upsampling module in parallel to obtain sampling tokens corresponding to each user terminal, and the sampling tokens are spliced ​​to obtain a second output token.

[0024] In one embodiment, the decoder further includes a second position encoding module; and the processing of the second output token by at least one expert shared decoding module included in the decoder to obtain the reconstructed channel state information corresponding to each of the user terminals includes:

[0025] Position-encoding the second output token by the second position encoding module to obtain a second position-encoded token;

[0026] The second position coding token is decoded through the gating network, at least one shared expert model and multiple expert activation models in each expert shared decoding module connected end to end to obtain a third output token, and based on the third output token, the reconstructed channel state information corresponding to each user terminal is obtained.

[0027] In one embodiment, the base model further includes an output module, and obtaining the reconstructed channel state information corresponding to each of the user terminals based on the third output token includes:

[0028] The third output token is reconstructed through the fully connected layer in the output module to obtain reconstructed channel state information corresponding to the sample channel state information of each user terminal.

[0029] In one embodiment, the performing training based on each of the reconstructed channel state information to obtain a trained base model includes:

[0030] Performing loss prediction based on each sample channel state information and reconstructed channel state information corresponding to each sample channel state information to obtain compression reconstruction loss;

[0031] Determining a token routing ratio based on the model weight and the number of target tokens; and obtaining a task load balancing loss based on the number of expert models to be activated, the token routing ratio, and an average activation weight corresponding to the gating network;

[0032] Calculating a quantization loss using the compressed feature vector corresponding to the user end and the target bit stream data corresponding to the user end;

[0033] Obtaining a model training loss based on the compression and reconstruction loss, the task load balancing loss, the quantization loss, and a loss weighting coefficient;

[0034] The model parameters of the base model are updated through the model training loss until the preset training completion conditions are met to obtain a trained base model.

[0035] In a second aspect, the present application further provides a training device for a base model for channel compression feedback, wherein the base model includes a base station end and at least one user end, the user end includes an encoder and a transformation module; the base station end includes an inverse transformation module and a decoder; the encoder includes at least one expert hybrid encoding module; the decoder includes at least one expert shared decoding module; and includes:

[0036] A first acquisition module is configured to acquire sample channel state information of at least one user terminal, where the sample channel state information includes information of a target dimension, where the target dimension includes at least two of a time dimension, a space dimension, and a frequency dimension;

[0037] a second acquisition module, configured to acquire a target token corresponding to the sample channel state information of the user end, encode the target token using the expert model and the gating network in each of the expert hybrid encoding modules corresponding to the user end to obtain a first output token; and transform the first output token using the transformation module to obtain target bitstream data;

[0038] an inverse transformation processing module, configured to perform inverse transformation processing on the target bit stream data corresponding to each of the user terminals through the inverse transformation module of the base station terminal, to obtain a second output token;

[0039] The first processing module is used to process the second output token through at least one expert shared decoding module included in the decoder to obtain reconstructed channel state information corresponding to each of the user terminals, and perform training based on each of the reconstructed channel state information to obtain a trained base model.

[0040] In a third aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in this embodiment when executing the computer program.

[0041] In a fourth aspect, the present application further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps in this embodiment are implemented.

[0042] In a fifth aspect, the present application also provides a computer program product, including a computer program, which implements the steps in this embodiment when executed by a processor.

[0043] The above-mentioned training method, device, computer equipment, computer-readable storage medium and computer program product for the base model for channel compression feedback, wherein the method includes: processing the sample channel state information of each user end through the model of the user end to obtain target bit stream data, and processing the target bit stream data of each user end through the model of the base station end to obtain reconstructed channel state information and a trained base model; by adopting this method, the task of channel compression feedback of different heterogeneous configurations can be flexibly solved through the trained base model; comprehensively considering the situation of multiple users in the wireless communication system, the sample channel state information estimated by multiple users is processed jointly to improve the efficiency of model training and improve the performance of the trained base model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments of the present application or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 FIG2 is a diagram illustrating an application environment of a training method for a base model for channel compression feedback in one embodiment;

[0046] Figure 2 1 is a flow chart of a method for training a base model for channel compression feedback in one embodiment;

[0047] Figure 3 Schematic diagram of a process for obtaining a first output token in one embodiment;

[0048] Figure 4 Schematic diagram of a process for obtaining a second output token in one embodiment;

[0049] Figure 5 1 is a flow chart of a method for training a base model for channel compression feedback in another embodiment;

[0050] Figure 6 is a schematic structural diagram of an expert hybrid coding module in one embodiment;

[0051] Figure 7 A schematic diagram of the structure of a shared expert decoding module in one embodiment;

[0052] Figure 8 1 is a structural block diagram of a training device for a base model for channel compression feedback in one embodiment;

[0053] Figure 9 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0055] It should be noted that the terms "first", "second", etc. used in this application may be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "including" and "having" used in this application and any variations thereof are intended to cover non-exclusive inclusions. The term "plurality" used in this application refers to two or more. The term "and / or" used in this application refers to one of the solutions or any combination of multiple solutions.

[0056] The base model in the training method of the base model for channel compression feedback provided in the embodiment of the present application can be as follows: Figure 1 As shown, the system includes a base station model 100 and multiple user-side models 200. The base station model is deployed with a base station-side network architecture, and each user-side model is deployed with a user-side network architecture. The input data of each user-side network architecture can be the estimated sample channel state information of the user corresponding to the user side. The output data of the base station can be the reconstructed channel state information corresponding to each user obtained after the base station decodes it.

[0057] Accordingly, in one embodiment, Figure 2 As shown, a training method for a base model for channel compression feedback is provided. The method is explained by taking the application of the method to a terminal device as an example. The method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The above-mentioned terminal can be, but is not limited to, various personal computers, laptops, smart phones, and tablet computers. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services. Wherein: the base model includes a base station end and at least one user end, and the user end includes an encoder and a transformation module. The base station end includes an inverse transformation module and a decoder. The encoder includes at least one expert hybrid encoding module. The decoder includes at least one expert shared decoding module. The training method for the base model for channel compression feedback includes:

[0058] Step 202: Obtain sample channel state information of at least one user terminal.

[0059] The sample channel state information includes information of target dimensions, which include at least two of the following: time dimension, spatial dimension, and frequency dimension. The time dimension includes the number of time sampling points, the spatial dimension includes the number of antennas of the base station, and the frequency dimension includes the number of subcarriers.

[0060] Optionally, the method in this embodiment can be applied to a MIMO-OFDM system, wherein the base station side is equipped with a planar array multi-antenna and the user side is equipped with a single antenna. The large base station model obtained after training in this embodiment can be a large base station model deployed on the base station side and the user side, which can handle the task of channel compression feedback of various heterogeneous configurations; the heterogeneous configuration includes one or more of the heterogeneity of the number of antennas on the base station side, the heterogeneity of the number of subcarriers in OFDM, the heterogeneity of the number of users for joint processing, the heterogeneity of the number of codewords fed back when the user feeds back the channel, and the heterogeneity of the channel distribution of the user. That is, in the wireless communication system where the large base station model obtained by training in this embodiment is deployed, this implementation does not limit the number of antennas on the base station side, the number of subcarriers in OFDM, the number of users for joint processing (that is, the number of users included in the wireless communication system), the number of codewords fed back when the user feeds back the channel, and the channel distribution of the user.

[0061] Optionally, the network weights of the deployment models of the various user terminals are the same, that is, the network parameters of the base large models deployed by the various user terminals may be the same.

[0062] Specifically, the terminal can obtain sample channel state information estimated for each user at each user terminal in the MIMO-OFDM system. For example, it can determine the users participating in training in the current MIMO-OFDM system and obtain sample channel state information estimated for each user, that is, obtain CSI data on each user side. The sample channel state information can be two-dimensional information, for example, it can include any two of the time dimension t, spatial dimension n, and frequency dimension k. For example, the terminal device needs to obtain sample channel state information estimated for each user, that is, obtain two-dimensional CSI samples estimated by the user at each user terminal. Generally, for example, it can be information in the spatial dimension and the frequency dimension, such as obtaining the number of antennas and the number of subcarriers of the base station corresponding to each user terminal.

[0063] In an example, the sample channel state information may be CSI in the space-frequency dimension by default. Optionally, the sample channel state information may also be CSI in the time-space dimension or CSI in the time-frequency dimension.

[0064] Step 204: Obtain a target token corresponding to the sample channel state information of the user end. Encode the target token using the expert models and gating networks in the expert hybrid encoding modules corresponding to the user end to obtain a first output token. Finally, transform the first output token using the transformation module to obtain target bitstream data.

[0065] The target token corresponding to the sample channel state information of the user terminal is obtained by embedding the sample channel state information. The user terminal may also include an embedding module. The encoder includes a first position encoding module and at least one hybrid expert encoding module; each hybrid expert encoding module includes a gating network and multiple expert models. The gating network can output model weights of each expert model, etc. Based on the model weights output by the gating network, an expert activation model is determined from the multiple expert models included in the hybrid expert encoding module.

[0066] Specifically, for each user end, the terminal device can, after obtaining the sample channel state information corresponding to the user end, perform block processing on the sample channel state information to obtain multiple sample channel state blocks, that is, obtain a sample channel state block sequence, and perform feature extraction processing on each sample channel state block in the sequence through the embedded module to obtain a unit token sequence, that is, determine the unit token sequence as the target token, wherein the target token is a one-dimensional token and the unit token sequence is a one-dimensional token sequence.

[0067] In this way, the terminal device can input the target token into the encoder, and perform position encoding processing on the target token through the first position encoding module in the encoder to obtain a first position encoding token, and input the first position encoding token into at least one expert hybrid encoding module, and encode the first position encoding token through the gating network and multiple expert activation models in at least one of the expert hybrid encoding modules to obtain the first output token output by the encoder.

[0068] Accordingly, the user terminal further includes a transformation module, which may include a compression layer and a quantization coding layer, wherein the compression layer may be a feature compression layer. In this way, the terminal device may compress the first output token using the compression layer to obtain a compression result, and quantize the compression result based on the quantization layer to obtain target bitstream data corresponding to the user terminal.

[0069] Step 206: Perform inverse transformation on the target bit stream data corresponding to each user terminal through the inverse transformation module of the base station to obtain a second output token.

[0070] Specifically, the base station can deploy a base station side with a large base model to be trained, i.e., a base station-side network architecture. The base station side can include an inverse transformation module; the terminal device can receive target bit stream data fed back by each user terminal via an uplink communication link through the inverse transformation module, and then, through the inverse transformation module, splice the received target bit stream data in the order in which they were received to obtain spliced ​​bit streams for each user terminal. The spliced ​​bit streams are then dequantized and upsampled to obtain a second output token.

[0071] Step 208: Process the second output token through at least one expert shared decoding module included in the decoder to obtain reconstructed channel state information corresponding to each user terminal, and perform training based on each reconstructed channel state information to obtain a trained base model.

[0072] The decoder includes a second position encoding module and at least one expert shared decoding module.

[0073] Specifically, the terminal device can position-encode the second output token using the second position encoding module to obtain a second position-coded token. In this way, the terminal device can decode the second position-coded token using multiple end-to-end expert shared decoding modules to obtain a third output token. The fully connected layer in the output module included in the base station reconstructs the third output token to obtain reconstructed channel state information corresponding to the sample channel state information of each user terminal. In this way, the terminal device can obtain a trained large base model based on the reconstructed channel state information.

[0074] In the above-mentioned training method for the base model for channel compression feedback, the user-side model processes the sample channel state information of each user end to obtain the target bit stream data, and the base station model processes the target bit stream data of each user end to obtain reconstructed channel state information and a trained base model. By adopting this method, the trained base model can flexibly solve the channel compression feedback task of different heterogeneous configurations. Taking into account the presence of multiple users in the wireless communication system, the sample channel state information estimated by multiple users is processed jointly to improve the efficiency of model training and the performance of the trained base model.

[0075] In one embodiment, the encoder further comprises a first position encoding module. The transform module comprises a compression layer and a quantization encoding layer. Specifically, the compression layer may be a feature compression layer.

[0076] The specific implementation process of the step of "encoding the target token using the expert models and the gating network in each expert hybrid encoding module corresponding to the user end to obtain a first output token. And transforming the first output token using the transformation module to obtain target bit stream data" may include:

[0077] The target token is position-encoded using the first position encoding module to obtain a first position-encoded token. The first position-encoded token is encoded using the gated network and multiple target expert models in the end-to-end expert hybrid encoding modules to obtain a first output token. The first output token is compressed using the compression layer to obtain a compressed feature vector. The compressed feature vector is quantized using the quantization encoding layer and the quantization bit width to obtain the target bitstream data corresponding to the user end.

[0078] Among them, the encoder includes a first position encoding module and at least one expert hybrid encoding module; the expert hybrid encoding module includes a pre-processing layer, an expert processing layer and a normalization layer, and the pre-processing layer includes a multi-head attention layer and a normalization layer; the expert processing layer includes a gating network and multiple expert models. The gating network can determine the model weights of each expert model based on the input data of the encoder, etc. Based on the model weights output by the gating network, one or more expert activation models (that is, one or more target expert models) are determined from the multiple expert models contained in the expert hybrid encoding module, and the expert activation model is used to process the first position encoding token.

[0079] Specifically, the terminal device can input the target token into the encoder, and perform position encoding processing on the target token through the first position encoding module in the encoder to obtain a first position encoding token, and input the first position encoding token into at least one expert hybrid encoding module, and encode the first position encoding token through the gating network and multiple expert activation models in at least one of the expert hybrid encoding modules to obtain the first output token output by the encoder.

[0080] In this way, the first output token output by the encoder can be a two-dimensional feature map. The terminal device can compress the first output token using a CNN (Convolutional Neural Network) network in the compression layer, alternating with downsampling operations, to obtain a compression result. Specifically, the length and width of the first output token are compressed to obtain a compression result. In this way, the terminal device can perform scalar quantization on the compression result using the quantization coding layer to obtain target bit stream data corresponding to the user terminal, and transmit the target bit stream data corresponding to the user terminal to the base station via the uplink communication link between the user terminal and the base station.

[0081] Optionally, the specific process of performing scalar quantization processing can be that the compression result can be a feature map, and the terminal device performs a scalar quantization operation on the feature map element by element based on a randomly selected quantization bit width through a quantization coding layer (the quantization bit width applied to each element in the feature map is the same) to obtain the target bit stream data corresponding to the user terminal. During the model training process, the value of the quantization bit width can be uniformly sampled from a quantization bit width list within a preset range. During the model application process, the value of the quantization bit width can be determined based on the requirements of the actual scenario corresponding to the model application.

[0082] In this embodiment, the channel state information estimated by different users is adaptively adapted through multiple expert activation models in the expert hybrid coding module, thereby improving the model performance of the base large model.

[0083] In one embodiment, the hybrid-of-experts coding module includes a pre-processing layer, an expert processing layer, and a first normalization layer. The expert processing layer includes a gating network and multiple expert models to be activated. The pre-processing layer includes a multi-head attention layer and a second normalization layer. The expert processing layer can be a MoE feedforward network layer.

[0084] Accordingly, if Figure 3 As shown, the specific implementation process of the step of "encoding the first position coding token through the gating network in each expert hybrid coding module connected end to end and multiple target expert models to obtain the first output token" may include:

[0085] Step 302: For the i-th expert hybrid coding module, the input token of the i-th expert hybrid coding module is processed by the attention mechanism and normalized through the pre-processing layer in the i-th expert hybrid coding module to obtain a first token.

[0086] Specifically, for the input token of the i-th expert joint encoding module, the input token is processed by the multi-head attention layer in the pre-processing layer to obtain an attention result, and the attention result is input into the second normalization layer for normalization processing to obtain a first normalized result; the terminal device can superimpose the first normalization result with the input token to obtain a superimposed result, and the terminal can determine that the superimposed result is the first token.

[0087] Step 304 : determining the model weight corresponding to the first token based on the gating network in the i-th expert hybrid coding module, and screening out the expert activation model from the expert models included in the i-th expert hybrid coding module based on the model weight.

[0088] Specifically, the terminal can input a first token into an expert processing layer; the expert processing layer includes a gating network (threshold network) and multiple expert models to be activated. The terminal device can process the first token through the gating network to obtain a model weight corresponding to the first token. The model weight corresponding to the first token can be the weight of each expert model corresponding to the first token. The expert model to be activated corresponding to the highest weight of the top target number can be selected through a Top-k sampling method to be the expert activation model, or the target expert model, i.e., the expert model used to process the first token. This embodiment does not limit the specific value of the target number, and those skilled in the art can determine it based on actual application scenarios.

[0089] Optionally, Nr routing expert models, i.e., expert models to be activated, are pre-configured in the expert hybrid coding model. The terminal can screen multiple routing expert models configured in the expert hybrid coding module based on the model weights output by the gating network and the Top-k sampling method to determine the K expert activation models corresponding to the first token.

[0090] In step 306, the first token is encoded using each expert activation model to obtain the expert output corresponding to each expert activation model. The output token of the i-th expert hybrid encoding module is obtained by combining the expert output of each expert activation model and the model weight corresponding to each expert activation model.

[0091] Specifically, the terminal may encode the first token based on each expert activation model and input the first token into each expert activation model to obtain an expert output result. The terminal device may perform a weighted calculation on the expert output results of each expert activation model and the model weights corresponding to each expert activation model to obtain a weighted result, perform a normalization process based on the weighted result, and determine the obtained normalized result as the output token of the i-th expert joint encoding module.

[0092] Step 308: Determine the output token of the expert hybrid encoding module at the last position as the first output token.

[0093] The input token of the expert hybrid coding module at the initial position is the first position coding token, and the input token of the i-th expert hybrid coding module is the output token of the i-1-th expert hybrid coding module.

[0094] Specifically, the terminal may determine the output token of the expert hybrid coding module located at the end of the encoder as the first output token of the encoder. That is, if there are n expert hybrid coding modules in the encoder, the terminal may determine the output token of the nth expert hybrid coding module as the first output token of the encoder.

[0095] Among them, the input token of the first expert hybrid coding module located at the initial position in the encoder is the first position coding token, that is, the terminal can input the first position coding token into the first expert hybrid coding module; the input token of the i-th expert hybrid coding module is the output token of the i-1-th expert hybrid coding module, that is, the input token of the i-th expert hybrid coding module is the output token of the previous expert hybrid coding module (i-1-th) of this module.

[0096] In this embodiment, an expert processing layer including a gating network and multiple routing expert models is used to determine an expert activation model that matches the user for processing. This allows the expert activation characteristics corresponding to different users to be learned, and the adaptability of the routing expert model allows for better processing of the specificity of channels between different users.

[0097] In one embodiment, the inverse transform module includes a splicing layer, a dequantization layer, and an upsampling layer. Specifically, the splicing layer can be a codeword splicing layer, which receives the bit stream sent by the user via the uplink communication link and splices it in sequence; the dequantization layer can be a conversion process for converting the bit stream into floating point form. The upsampling layer is used to restore the length and width of the feature map.

[0098] like Figure 4 As shown, the specific implementation process of the step of "performing an inverse transformation process on the target bit stream data corresponding to each user terminal through the inverse transformation module on the base station side to obtain the second output token" may include:

[0099] Step 402 : For each user end, the splicing layer in the inverse transform module of the base station splices the bit streams fed back by the user end based on the bit stream reception order to obtain a spliced ​​bit stream corresponding to the user end.

[0100] The bit stream receiving order may be the order in which the base station receives the target bit stream data sent by the user terminal via the uplink communication link; the bit stream fed back by the user terminal is the target bit stream data sent by the user terminal via the uplink communication link.

[0101] Specifically, the terminal device can perform splicing processing on the bit streams fed back by the received user terminals based on the receiving order of the bit streams through the splicing layer to obtain user-sent bit streams corresponding to each user, namely, spliced ​​bit streams.

[0102] Step 404 : Perform type inverse transformation on the spliced ​​bit stream corresponding to each user terminal through the dequantization layer to obtain a feature map of each user terminal.

[0103] Specifically, the terminal device can convert the representation of the spliced ​​bit stream of each user into a floating-point representation in the dequantization layer, and obtain the feature map of the user based on the floating-point numbers corresponding to each user.

[0104] Optionally, for each user terminal, the terminal can obtain a spliced ​​bit stream corresponding to the user terminal; based on the communication protocol and / or fixed length of the uplink communication link, the spliced ​​bit stream is grouped, and each group of binary bits is converted into an integer value to obtain an integer value corresponding to the spliced ​​bit stream, and the type of the integer value is converted and normalized to obtain a normalized floating-point number, and the floating-point number is rearranged based on the width, height, and number of channels of the original image to obtain a multidimensional array, and the multidimensional array is determined as a feature map, wherein the width, height, and number of channels of the original image can be determined randomly or based on the actual application scenario.

[0105] In step 406, the feature maps of each user terminal are input to the upsampling layer in parallel to obtain sampling tokens corresponding to each user terminal, and the sampling tokens are concatenated to obtain a second output token.

[0106] Specifically, the terminal device can input the feature maps of each user in parallel into the upsampling module. The upsampling model restores the size of the feature map to its pre-compression size, thereby obtaining a restored feature map. The size of the restored feature map is consistent with the size of the feature map input to the compression layer. The terminal device can convert the restored feature map corresponding to each user terminal into a token form to obtain a sampling token corresponding to each user terminal. The terminal device can then concatenate the sampling tokens corresponding to each user terminal to obtain a second output token.

[0107] In this embodiment, by performing splicing processing, dequantization type conversion processing and upsampling processing, the feature map can be accurately and comprehensively restored, providing a reliable data basis for the reconstruction process of the signal state information of the subsequent model.

[0108] In one embodiment, the decoder further includes a second position encoding module. The specific implementation process of the step of "processing the second output token by at least one expert shared decoding module included in the decoder to obtain reconstructed channel state information corresponding to each user terminal" may include:

[0109] The second output token is position-encoded by the second position encoding module to obtain a second position-coded token. The second position-coded token is decoded by the gating network, at least one shared expert model, and multiple expert activation models in the end-to-end expert shared decoding modules to obtain a third output token. Based on the third output token, reconstructed channel state information corresponding to each user terminal is obtained.

[0110] The expert shared decoding module includes a pre-processing layer, an expert processing layer, and a first normalization layer. The expert processing layer includes a gating network, at least one shared expert model, and multiple expert models to be activated. The pre-processing layer includes a multi-head attention layer and a second normalization layer. The expert processing layer can be an SR MoE feedforward network layer.

[0111] Specifically, the second output token is position-encoded by the second position encoding module to obtain a second position-encoded token. The second position-encoded token is encoded by the gating network, at least one shared expert model, and multiple expert activation models in the end-to-end expert shared decoding module blocks to obtain a third output token. The first output token is compressed by the compression layer to obtain a compressed feature vector. The compressed feature vector is quantized by the quantization coding layer and the quantization bit width to obtain the target bit stream data corresponding to the user end. The third output token is reconstructed by the fully connected layer in the output module to obtain the reconstructed channel state information corresponding to the sample channel state information of each user end.

[0112] In one example, the decoder may include n expert shared decoding modules, each of which may be connected end to end. The expert processing layer of each expert shared decoding model includes multiple expert routing models, a gating network, and at least one expert shared model.

[0113] For the i-th expert shared decoding module among the n expert shared decoding modules, the terminal can perform attention mechanism processing and normalization processing on the input token of the i-th expert shared decoding module through the pre-processing layer in the i-th expert shared decoding module to obtain a first normalization result, and superimpose the first normalization result with the input token to obtain a second token.

[0114] Based on the gating network in the i-th expert shared decoding module, the model weight corresponding to the second token is determined, and based on the model weight, the expert activation model is screened out from the expert models (routing expert models) included in the i-th expert shared decoding module.

[0115] The second token is encoded using each expert activation model and the expert shared model in the expert shared decoding module to obtain multiple expert output results. The expert output results of each expert activation model and the model weights corresponding to each expert activation model are weighted to obtain a weighted result, and the weighted result is superimposed with the expert output results of each expert shared model to obtain a first superimposed result. The first superimposed result is normalized using a first normalization layer to obtain a normalized result. The normalized result is superimposed with the second token to obtain a second superimposed result, and the second superimposed result is determined to be the output token of the i-th expert shared decoding module.

[0116] The terminal may determine that the output token of the expert shared decoding module at the end position of the decoder is the third output token. The input token of the expert shared decoding module at the initial position is the second position coding token, and the input token of the i-th expert shared decoding module is the output token of the i-1-th expert shared decoding module.

[0117] This embodiment utilizes multiple routing expert models to determine the expert activation model that matches the user for processing. This allows the expert activation characteristics corresponding to different users to be learned, and the adaptability of the routing expert model allows for better handling of the specificity of channels between different users. Furthermore, the shared expert model in the expert layer extracts relevant information about the channels between multiple users, further improving decoder performance and providing accurate data for the output model to reconstruct channel states.

[0118] In one embodiment, the base model further includes an output module. The specific implementation process of the step "obtaining the reconstructed channel state information corresponding to each user terminal based on the third output token" may include:

[0119] The third output token is reconstructed through the fully connected layer in the output module to obtain reconstructed channel state information corresponding to the sample channel state information of each user terminal.

[0120] Among them, the output module can include a fully connected layer.

[0121] Specifically, the terminal can reconstruct the third output token output by the decoder through the fully connected layer in the output module. For example, the fully connected layer includes a network layer composed of fully connected neurons. The terminal outputs the third output token to the fully connected layer. The features of the third output token can be linearly transformed through the weight matrix and bias term of the fully connected layer, and then a nonlinear mapping is introduced through the activation function. The third output token is transformed by the nonlinear mapping through the activation function, and finally a reconstruction result that matches the dimension of the original token (i.e., the sample channel state information) of each user terminal is output, that is, the reconstructed channel state information corresponding to the sample channel state information of each user terminal is obtained.

[0122] In this embodiment, the base model can learn how to recover the original information from the compressed or abstract token representation, thereby optimizing the feature extraction capability of the model.

[0123] In one embodiment, the specific implementation process of the step of "performing training based on each reconstructed channel state information to obtain a trained base model" may include:

[0124] Based on the channel state information of each sample and the reconstructed channel state information corresponding to each sample channel state information, loss prediction is performed to obtain the compression reconstruction loss. Based on the model weight and the number of target tokens, the token routing ratio is determined. And based on the number of expert models to be activated, the token routing ratio and the average activation weight corresponding to the gating network, the task load balancing loss is obtained. The quantization loss is calculated using the compressed feature vector corresponding to the user end and the target bit stream data corresponding to the user end. Based on the compression reconstruction loss, task load balancing loss, quantization loss and loss weighting coefficient, the model training loss is obtained. The model parameters of the base model are updated through the model training loss until the preset training completion conditions are met to obtain a trained base model.

[0125] Among them, the preset training completion condition can be that the number of training times reaches the target number of times, or that the loss function obtained during the training process has satisfied the convergence condition, the convergence condition can be that the loss function has remained unchanged, or the convergence condition can be that the loss function has been close to the target value. Among them, this embodiment does not limit the specific numerical value of the number of training times and the specific numerical value of the target value, and those skilled in the art can determine it based on the needs of the actual application scenario. The preset weighting coefficient includes a first coefficient of quantization loss and a second coefficient of task load balancing loss.

[0126] Specifically, the compression reconstruction loss L1 can be the loss of channel prediction, which represents the difference between the user's estimated 2D CSI and the reconstructed CSI output by the base station; the quantization loss L2 represents the difference between the features output by the quantization layer and the feature vector input to the quantization layer; the task load balancing loss L3 is the load balancing loss for the moe architecture.

[0127] The terminal can determine the compression reconstruction loss through the mean squared error (MSE). For example, the compression reconstruction loss L1 can be calculated using the following formula:

[0128]

[0129] Among them, H is the original 2D CSI estimated by the user, that is, the sample channel state information, It is to reconstruct the channel state information.

[0130] In this way, the terminal can calculate the quantization loss L2 through the mean square error:

[0131]

[0132] Among them, s q represents the feature vector output by the quantization module (quantization layer), and s represents the feature vector input to the quantization module.

[0133] The terminal can also obtain the task load balancing loss L3 based on the number of expert models to be activated, the token routing ratio, and the average activation weight corresponding to the gating network. For example, the terminal may be executing the i-th pre-training task, and there are T target tokens in the sample channel state data Ω of a training batch. For example, the token routing ratio may be D n , n represents the number of expert models, the token routing ratio represents the ratio of all target tokens routed to expert model n, and the average activation weight of the gating network for each expert activation model n is P n , then the terminal can calculate the load balancing loss L3 using the following formula:

[0134] ;

[0135] in, is an indicator function, which is 1 when the condition in the brackets is met, otherwise it is 0; G(x)[n] represents the model weight of the nth expert model corresponding to the target token x output by the gating network G.

[0136] In this way, the terminal can perform weighted calculations on the compression reconstruction loss, the quantization loss, the task load balancing loss, the first coefficient of the quantization loss, and the second coefficient of the task load balancing loss to obtain the model training loss corresponding to the base model. After the terminal obtains the model training loss, it can determine whether the preset training completion condition is currently met. If the preset training completion condition is not met, the terminal can update the model parameters of the base model based on the model training loss to obtain an updated base model, and re-execute the method in the above embodiment with the updated base model until the preset training completion condition is met to obtain a trained base model.

[0137] In this embodiment, by comprehensively considering the loss function in terms of task load balancing, the loss function in terms of quantization, and the loss function in terms of channel information, it is possible to balance multi-objective optimization, avoid single-objective deviation, improve training efficiency, and improve the performance of the model obtained after training.

[0138] The following describes in detail the specific implementation steps of the above-mentioned training method for the base model for channel compression feedback in conjunction with a specific embodiment:

[0139] The training method of the base model for channel compression feedback provided in this embodiment is a universal network that can flexibly realize compression feedback of channel data of various heterogeneous configurations. Specifically, it is a MoE network architecture. By including multiple MoE architectures in the encoder and multiple SR MoE layers in the decoder, the correlation and specificity of channels between users are simultaneously utilized during multi-user joint decoding, thereby significantly improving the accuracy of channel acquisition on the base station side.

[0140] In practice, this embodiment is applied to a MIMO-OFDM system, where the base station is equipped with multiple planar array antennas and the user side is equipped with a single antenna. The proposed large-scale base station model is deployed simultaneously on the user side and the base station side, and can handle channel compression feedback from various heterogeneous configurations. These heterogeneous configurations include one or more of heterogeneous number of antennas on the base station side, heterogeneous number of OFDM subcarriers, heterogeneous number of users for joint processing, heterogeneous number of codewords fed back by users during channel feedback, and heterogeneous channel distribution among users.

[0141] like Figure 5 As shown, the network architecture of this embodiment may include the network architecture of user 1, ..., the network architecture of user k, and the base station-side network architecture deployed on the base station side; wherein; the network is divided into two parts: the user-side network and the base station-side network. The user side includes multiple users, numbered from 1 to k, each user is equipped with the same network architecture and shares the same network weights. The user-side network architecture includes an embedding module, an encoder module, a feature compression module, and a quantization encoding module. The base station-side network architecture includes a codeword splicing module, a dequantization module, an upsampling module, a decoder module, and an output module.

[0142] Embedding Module: The input is the 2D CSI samples (sample channel state information) estimated by the user side. The embedding module processes the sample channel state information into a series of tokens. Specifically, the 2D CSI samples are first divided into blocks to obtain a series of 2D CSI blocks. These blocks are then embedded into a 1D token sequence using a convolutional network to obtain the target token.

[0143] Encoder module: The input token is first positionally encoded and then processed through a series of transformer blocks (multiple end-to-end mixed-expert encoding modules), each of which adopts the proposed MoE architecture. The encoder module includes an encoder positional encoding, which is the first positional encoding module, and a MoE architecture, which includes multiple end-to-end blocks (e.g., block 1, block 2, ..., block N), each of which is a mixed-expert encoding module.

[0144] Feature compression module: The token sequence (the first output token) is determined as a two-dimensional feature map. By using a CNN network (Convolutional Neural Networks) and alternating downsampling operations, the length and width of the feature map are reduced to achieve feature compression and obtain a compressed feature vector.

[0145] Quantization Coding Module: This module performs element-by-element scalar quantization on the feature map (compressed feature vector) to generate a bitstream (target bitstream data). This bitstream is then fed back to the base station via the uplink communication link. The quantization bitwidth is randomly sampled from a list of quantization bitwidths within a certain range. The specific value of this quantization bitwidth during model application can be determined based on the actual application scenario.

[0146] Codeword splicing module: After receiving the bit streams fed back by each user, it splices them in sequence to obtain the spliced ​​bit streams corresponding to each user.

[0147] Dequantization module: restores the bitstream sent by each user (joined bitstream) to floating-point form to obtain the feature map of each user.

[0148] Upsampling module: Each user's feature map is input into the upsampling module in parallel to restore it to its original size. The feature maps of each user are then converted into tokens and concatenated in sequence to obtain the second output token, which is input into the decoder module.

[0149] Decoder module: The input token is first positionally encoded and then processed through a series of transformer blocks (end-to-end expert shared decoding modules), where each transformer block adopts the SR MoE architecture. The decoder module includes the decoder position encoding and the SR MoE architecture. The encoder position encoding is the second position encoding module; the SR MoE architecture consists of multiple blocks (e.g., block 1, block 2, ..., block N) connected end-to-end, each of which is an expert shared decoding module.

[0150] Output module: The tokens output by the decoder are passed through the fully connected layer to obtain the reconstructed CSI of each user, that is, the reconstructed CSI corresponding to each user end, that is, the CSI (channel state information) of each user decoded by the base station.

[0151] like Figure 6 As shown, it can be a structural diagram of the expert hybrid coding module, which includes a pre-processing layer, an expert processing layer and a normalization layer (i.e., the first normalization layer); wherein the pre-processing layer includes a multi-head attention layer and a normalization layer (i.e., the second normalization layer), and the expert processing layer is a MoE feedforward network layer; the MoE feedforward network layer includes a threshold network and multiple routing expert models, for example, routing expert 1, routing expert 2, routing expert 3 and routing expert 4; token1, ..., token n, ..., token N can represent N input tokens; the threshold network can process each input token to obtain the model weight corresponding to the input token, and the terminal can activate the K experts with the highest weight according to the Top-k sampling method to process the token to obtain K expert activation models; for example, the model weight output by the threshold network based on the nth input token can be routing n, which represents the weight of each routing expert for the nth input token, K=2, and the expert activation models selected by the terminal device for the nth input token can be routing expert 1 and routing expert 3. Then, the terminal device can process the data input to the expert processing layer corresponding to the nth input token through routing expert 1 and routing expert 3 respectively to obtain the output results of each expert model, perform weighted calculation based on the output results of each expert model and the model weight to obtain a weighted result, and determine the weighted result as the output data of the expert processing layer corresponding to token n.

[0152] In one example, there are N r Expert networks (expert routing models, i.e., expert models to be activated) are activated each time. Each token input to the expert processing layer is passed through a threshold network to obtain its corresponding N r Expert weights are activated based on the Top-k sampling method. Experts (expert activation model) process the token and The output of the MoE feedforward layer is weighted summed up. Assume that the token input to the MoE feedforward layer is x∈R (1×D) , the corresponding threshold network is G(·), its weight distribution to the nth expert is G(x)[n], and the nth expert network is E n (·), the output of the MoE feedforward layer is y, then the following relationship is satisfied:

[0153] ;

[0154] Each expert network is a FFN architecture, namely:

[0155]

[0156] Among them, R is the set of real numbers, R (1×D) represents a D-dimensional row vector in the real number field. For example, E n (x) is a D-dimensional row vector, that is, the output vector obtained after the FFN model processes the input data x is a D-dimensional row vector.

[0157] The gating network first processes the input x through the fully connected layer to obtain N r weights, and then normalized through the Softmax layer, and finally retain the top K largest weights and reset the rest to zero. The output of the gating network is:

[0158]

[0159] Where FC(x) represents the processing of input x by the fully connected layer. (1×Nr ) Indicates that the output of the gating network G is an N r dimensional row vectors, representing the weights of each expert network for x.

[0160] like Figure 7As shown, it can be a structural diagram of the expert shared decoding module, which includes a pre-processing layer, an expert processing layer and a normalization layer (i.e., the first normalization layer); wherein, the pre-processing layer includes a multi-head attention layer, a normalization layer (i.e., the second normalization layer), and the expert processing layer is an MoE feedforward network layer; the MoE feedforward network layer includes a threshold network and at least one shared expert model, multiple routing expert models, for example, it can include routing expert 1, routing expert 2, routing expert 3, routing expert 4 and shared expert 1; token1, ..., token n, ..., token N can represent N input tokens; the threshold network can process each input token separately to obtain the model weight corresponding to each input token (i.e., the model weight of each routing expert model for the input token), and the terminal can activate the K experts with the highest weight to process the token according to the Top-k sampling method to obtain K expert activation models; for example, the model weight output by the threshold network based on the nth input token (token n) can be routing n, which means for token The weights of each routing expert of n, K=2, the expert activation model selected by the terminal device for the nth input token can be routing expert 1 and routing expert 3. Then, the terminal device can process token n respectively through shared expert 1, routing expert 1 and routing expert 3 to obtain the output results of each model, perform weighted calculation based on the output results of each model and the model weight to obtain the weighted result, and process the weighted result with the output result of the shared expert model to obtain the output result of the expert processing layer corresponding to token n.

[0161] Optionally, there is N s Shared expert network, N r Routing expert networks, where all shared expert networks and K routing expert networks are activated each time. Each input token obtains its corresponding N through a threshold network. r The weights of routing experts are used to activate the K experts with the highest weights to process the token based on the Top-k sampling method, and the outputs of the K experts and N s The outputs of the shared expert networks are weighted summed to obtain the output of the SR MoE feedforward layer.

[0162] Assume that the token input to the SR MoE feedforward layer is x∈R (1×D) , the corresponding threshold network is G(·), its weight distribution to the nth routing expert is G(x)[n], and the nth routing expert network is E (r,n) (·), the nth shared expert network is E (s,n) (·), the output of the SR MoE feedforward layer is y, then the following relationship is satisfied:

[0163] ;

[0164] Each expert network is a FFN architecture, namely:

[0165] ;

[0166] The gated network first processes the input x through the fully connected layer to obtain N r weights, and then normalized through the Softmax layer, and finally retain the top K largest weights and reset the rest to zero. The output of the gating network is:

[0167] .

[0168] The method provided in this embodiment is a multi-user, multi-bitrate pre-training scheme. This base model solution is capable of processing multi-user, multi-bitrate channel compression feedback. Therefore, during training, the number of users and feedback bitrate for each sample are randomly selected, allowing the model to learn more general and powerful channel compression feedback capabilities. This enables it to simultaneously handle a variety of user numbers and feedback bitrates, resolving the difficulty of simultaneously utilizing both inter-user channel correlation and specificity during multi-user joint decoding. This significantly improves the accuracy of channel acquisition at the base station, leverages the system gain brought about by multi-user joint processing, and further expands the application scenarios of the resulting large base model.

[0169] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily performed in sequence in the order indicated by the arrows. Unless clearly stated herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of the steps or stages in other steps or other steps. It is understandable that the various steps in different embodiments can be freely combined as needed, and the various non-contradictory schemes formed by the combination all fall within the scope of protection of this application.

[0170] Based on the same inventive concept, an embodiment of the present application also provides a training device for a base model for channel compression feedback for implementing the training method for the base model for channel compression feedback involved above. The implementation solution provided by the device is similar to the implementation solution described in the above method, so the specific limitations in the embodiments of one or more training devices for base models for channel compression feedback provided below can be found in the above limitations on the training method for base models for channel compression feedback, and will not be repeated here.

[0171] In an exemplary embodiment, Figure 8 As shown, a training device 800 for a base model for channel compression feedback is provided. The base model includes a base station and at least one user terminal. The user terminal includes an encoder and a transform module; the base station includes an inverse transform module and a decoder; the encoder includes at least one expert hybrid encoding module; the decoder includes at least one expert shared decoding module; the device includes:

[0172] A first acquisition module 802 is configured to acquire sample channel state information of at least one user terminal, where the sample channel state information includes information of a target dimension, where the target dimension includes at least two of a time dimension, a space dimension, and a frequency dimension;

[0173] The second acquisition module 804 is configured to obtain a target token corresponding to the sample channel state information of the user end, encode the target token using the expert model and the gating network in each expert hybrid encoding module corresponding to the user end to obtain a first output token; and transform the first output token using the transformation module to obtain target bit stream data;

[0174] An inverse transformation processing module 806 is configured to perform inverse transformation processing on the target bit stream data corresponding to each user terminal through an inverse transformation module of the base station to obtain a second output token;

[0175] The first processing module 808 is used to process the second output token through at least one expert shared decoding module included in the decoder to obtain reconstructed channel state information corresponding to each user terminal, and perform training based on each reconstructed channel state information to obtain a trained base model.

[0176] In one embodiment, the encoder further includes a first position encoding module; the transform module includes a compression layer and a quantization encoding layer; and the second acquisition module is specifically configured to:

[0177] Position-encode the target token using the first position encoding module to obtain a first position encoding token;

[0178] Encoding the first position coding token through the gating network and multiple target expert models in each of the expert hybrid coding modules connected end to end to obtain a first output token;

[0179] compressing the first output token by the compression layer to obtain a compressed feature vector;

[0180] The compressed feature vector is quantized by using the quantization coding layer and the quantization bit width to obtain target bit stream data corresponding to the user end.

[0181] In one embodiment, the expert hybrid coding module includes a pre-processing layer, an expert processing layer, and a normalization layer, wherein the expert processing layer includes multiple expert models and multiple gating networks; the second acquisition module is specifically configured to:

[0182] For the i-th mixed-expert coding module, perform attention mechanism processing and normalization processing on the input token of the i-th mixed-expert coding module through the pre-processing layer in the i-th mixed-expert coding module to obtain a first token;

[0183] Determining a model weight corresponding to the first token based on the gating network in the i-th expert hybrid encoding module, and screening out an expert activation model from the expert models included in the i-th expert hybrid encoding module based on the model weight;

[0184] Encoding the first tokens by using the expert activation models to obtain the expert output results corresponding to the expert activation models; and obtaining the output tokens of the i-th expert hybrid coding module by using the expert output results of the expert activation models and the model weights corresponding to the expert activation models;

[0185] Determine that the output token of the expert hybrid coding module located at the last position is the first output token; the input token of the expert hybrid coding module located at the initial position is the first position coding token, and the input token of the i-th expert hybrid coding module is the output token of the i-1-th expert hybrid coding module.

[0186] In one embodiment, the inverse transform module includes a splicing layer, a dequantization layer, and an upsampling layer; the inverse transform processing module is specifically configured to:

[0187] For each user terminal, splicing the bit streams fed back by the user terminal based on the bit stream reception order by a splicing layer in the inverse transform module of the base station to obtain a spliced ​​bit stream corresponding to the user terminal;

[0188] Performing type inverse transformation on the spliced ​​bit streams corresponding to the respective user terminals through the dequantization layer to obtain feature maps of the respective user terminals;

[0189] The feature maps of each user terminal are input into the upsampling module in parallel to obtain sampling tokens corresponding to each user terminal, and the sampling tokens are spliced ​​to obtain a second output token.

[0190] In one embodiment, the decoder further includes a second position encoding module; and the first processing module is specifically configured to:

[0191] Position-encoding the second output token by the second position encoding module to obtain a second position-encoded token;

[0192] The second position coding token is decoded through the gating network, at least one shared expert model and multiple expert activation models in each expert shared decoding module connected end to end to obtain a third output token, and based on the third output token, the reconstructed channel state information corresponding to each user terminal is obtained.

[0193] In one embodiment, the base model further includes an output module, and the first processing module is further specifically configured to:

[0194] The third output token is reconstructed through the fully connected layer in the output module to obtain reconstructed channel state information corresponding to the sample channel state information of each user terminal.

[0195] In one embodiment, the performing training based on each of the reconstructed channel state information to obtain a trained base model includes:

[0196] Performing loss prediction based on each sample channel state information and reconstructed channel state information corresponding to each sample channel state information to obtain compression reconstruction loss;

[0197] Determining a token routing ratio based on the model weight and the number of target tokens; and obtaining a task load balancing loss based on the number of expert models to be activated, the token routing ratio, and an average activation weight corresponding to the gating network;

[0198] Calculating a quantization loss using the compressed feature vector corresponding to the user end and the target bit stream data corresponding to the user end;

[0199] Obtaining a model training loss based on the compression and reconstruction loss, the task load balancing loss, the quantization loss, and a loss weighting coefficient;

[0200] The model parameters of the base model are updated through the model training loss until the preset training completion conditions are met to obtain a trained base model.

[0201] Each module in the aforementioned training device for a base model with channel compression feedback can be implemented in whole or in part via software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0202] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 9 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store training data of the base model. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a training method for a base model for channel compression feedback is implemented.

[0203] Those skilled in the art will understand that Figure 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0204] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0205] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0206] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0207] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0208] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), quantum computing-based data processing logic devices, artificial intelligence (AI) processors, and the like.

[0209] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0210] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A training method for a base model for channel compression feedback, characterized in that: The base model includes a base station end and at least one user end, the user end includes an encoder and a transformation module; the base station end includes an inverse transformation module and a decoder; the encoder includes at least one expert hybrid coding module; The decoder includes at least one expert shared decoding module; the method includes: Acquire sample channel state information of at least one user terminal, where the sample channel state information includes information of a target dimension, where the target dimension includes at least two of a time dimension, a space dimension, and a frequency dimension; Obtaining a target token corresponding to the sample channel state information of the user end, encoding the target token using the expert model and the gating network in each of the expert hybrid coding modules corresponding to the user end to obtain a first output token; and transforming the first output token using the transformation module to obtain target bit stream data; Performing inverse transformation processing on the target bit stream data corresponding to each of the user terminals through the inverse transformation module of the base station terminal to obtain a second output token; The second output token is processed by at least one expert shared decoding module included in the decoder to obtain reconstructed channel state information corresponding to each user terminal, and training is performed based on each reconstructed channel state information to obtain a trained base model.

2. The method according to claim 1, characterized in that The encoder further includes a first position encoding module; the transformation module includes a compression layer and a quantization encoding layer; the target token is encoded by the expert model and the gating network in each expert hybrid encoding module corresponding to the user end to obtain a first output token; and transforming the first output token by the transformation module to obtain target bit stream data, including: Position-encode the target token using the first position encoding module to obtain a first position encoding token; Encoding the first position coding token through the gating network and multiple target expert models in each of the expert hybrid coding modules connected end to end to obtain a first output token; compressing the first output token by the compression layer to obtain a compressed feature vector; The compressed feature vector is quantized by using the quantization coding layer and the quantization bit width to obtain target bit stream data corresponding to the user end.

3. The method according to claim 2, characterized in that The hybrid expert encoding module includes a pre-processing layer, an expert processing layer, and a normalization layer. The expert processing layer includes multiple expert models and multiple gating networks. The first position coding token is encoded by the gating networks and multiple target expert models in the hybrid expert encoding modules connected end to end to obtain a first output token, including: For the i-th mixed-expert coding module, perform attention mechanism processing and normalization processing on the input token of the i-th mixed-expert coding module through the pre-processing layer in the i-th mixed-expert coding module to obtain a first token; Determining a model weight corresponding to the first token based on the gating network in the i-th expert hybrid encoding module, and screening out an expert activation model from the expert models included in the i-th expert hybrid encoding module based on the model weight; Encoding the first tokens by using the expert activation models to obtain the expert output results corresponding to the expert activation models; and obtaining the output tokens of the i-th expert hybrid coding module by using the expert output results of the expert activation models and the model weights corresponding to the expert activation models; Determine that the output token of the expert hybrid coding module located at the last position is the first output token; the input token of the expert hybrid coding module located at the initial position is the first position coding token, and the input token of the i-th expert hybrid coding module is the output token of the i-1-th expert hybrid coding module.

4. The method according to claim 1, wherein The inverse transform module includes a splicing layer, a dequantization layer, and an upsampling layer; the inverse transform module on the base station side performs inverse transform processing on the target bit stream data corresponding to each user terminal to obtain a second output token, including: For each user terminal, splicing the bit streams fed back by the user terminal based on the bit stream reception order by a splicing layer in the inverse transform module of the base station to obtain a spliced ​​bit stream corresponding to the user terminal; Performing type inverse transformation on the spliced ​​bit streams corresponding to the respective user terminals through the dequantization layer to obtain feature maps of the respective user terminals; The feature maps of each user terminal are input into the upsampling layer in parallel to obtain sampling tokens corresponding to each user terminal, and the sampling tokens are spliced ​​to obtain a second output token.

5. The method according to claim 1, wherein The decoder further includes a second position encoding module; the at least one expert shared decoding module included in the decoder processes the second output token to obtain reconstructed channel state information corresponding to each of the user terminals, including: Position-encoding the second output token by the second position encoding module to obtain a second position-encoded token; The second position coding token is decoded through the gating network, at least one shared expert model and multiple expert activation models in each expert shared decoding module connected end to end to obtain a third output token, and based on the third output token, the reconstructed channel state information corresponding to each user terminal is obtained.

6. The method according to claim 5, characterized in that The base model further includes an output module, and obtaining the reconstructed channel state information corresponding to each of the user terminals based on the third output token includes: The third output token is reconstructed through the fully connected layer in the output module to obtain reconstructed channel state information corresponding to the sample channel state information of each user terminal.

7. The method according to claim 3, characterized in that The performing training based on each of the reconstructed channel state information to obtain a trained base model includes: Performing loss prediction based on each sample channel state information and reconstructed channel state information corresponding to each sample channel state information to obtain compression reconstruction loss; Determining a token routing ratio based on the model weight and the number of target tokens; and obtaining a task load balancing loss based on the number of expert models to be activated, the token routing ratio, and an average activation weight corresponding to the gating network; Calculating a quantization loss using the compressed feature vector corresponding to the user end and the target bit stream data corresponding to the user end; Obtaining a model training loss based on the compression and reconstruction loss, the task load balancing loss, the quantization loss, and a loss weighting coefficient; The model parameters of the base model are updated through the model training loss until the preset training completion conditions are met to obtain a trained base model.

8. A training device for a base model with channel compression feedback, characterized in that: The base model includes a base station end and at least one user end, the user end includes an encoder and a transformation module; the base station end includes an inverse transformation module and a decoder; the encoder includes at least one expert hybrid coding module; The decoder includes at least one expert shared decoding module; the apparatus includes: A first acquisition module is configured to acquire sample channel state information of at least one user terminal, where the sample channel state information includes information of a target dimension, where the target dimension includes at least two of a time dimension, a space dimension, and a frequency dimension; a second acquisition module, configured to acquire a target token corresponding to the sample channel state information of the user end, encode the target token using the expert model and the gating network in each of the expert hybrid encoding modules corresponding to the user end to obtain a first output token; and transform the first output token using the transformation module to obtain target bitstream data; an inverse transformation processing module, configured to perform inverse transformation processing on the target bit stream data corresponding to each of the user terminals through the inverse transformation module of the base station terminal, to obtain a second output token; The first processing module is used to process the second output token through at least one expert shared decoding module included in the decoder to obtain reconstructed channel state information corresponding to each of the user terminals, and perform training based on each of the reconstructed channel state information to obtain a trained base model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

11. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images