Artificial intelligence model training method, apparatus, device, and readable storage medium

Abstract

Embodiments of the present application provide an AI model training method, device and equipment and readable storage medium, the method comprising: obtaining uplink channel state information; training an AI model according to the uplink channel state information, the AI model comprising a first AI model and / or a second AI model; and / or, obtaining downlink channel state information and a first result, the first result representing a calculation result obtained by reasoning the downlink channel state information in the first AI model; training the second AI model according to the first result and the downlink channel state information; wherein the first AI model is used for compressing the downlink channel state information, and the second AI model is used for decompressing the downlink channel state information.

Description

Technical Field

[0001] This application relates to the field of communication technology, specifically to an AI model training method, apparatus, device, and readable storage medium. Background Technology

[0002] Channel state information describes the current channel environment. In mobile communication networks, base stations transmit Channel State Information-Reference Signals (CSI-RS). Terminals evaluate the channel state information and quantify it to feed back to the base station. By introducing Channel State Information (CSI) feedback, the base station can make timely adjustments when transmitting the CSI-RS, thereby reducing the bit error rate and obtaining the optimal received signal at the terminal.

[0003] In wireless communication, channel prediction can be used to compensate for the delay between channel measurement and actual scheduling operations, thereby improving throughput. How to reduce the transmission overhead of training data while ensuring model performance is a pressing issue. Summary of the Invention

[0004] This application provides an AI model training method, apparatus, device, and readable storage medium to address the problem of reducing the transmission overhead of training data while ensuring the performance of the AI ​​model.

[0005] Firstly, an AI model training method is provided, applicable to network-side devices, including:

[0006] Obtain uplink channel status information;

[0007] The AI ​​model is trained based on the uplink channel state information, and the AI ​​model includes a first AI model and / or a second AI model.

[0008] And / or,

[0009] Acquire downlink channel state information and a first result, wherein the first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model;

[0010] Based on the first result and the downlink channel state information, the second AI model is trained;

[0011] The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

[0012] Optionally, the method further includes:

[0013] The trained first AI model is sent to the terminal.

[0014] Optionally, obtaining the uplink channel state information includes:

[0015] Receive one or more SRS or DMRS from one or more terminals;

[0016] Uplink channel state information is obtained based on one or more SRS or DMRS.

[0017] Optionally, based on the first result and the downlink channel state information, the second AI model is trained, including:

[0018] Based on the first result and the second AI model, the calculation result of the reasoning of the second AI model is obtained;

[0019] Obtain the loss value between the calculation result of the second AI model inference and the downlink channel state information;

[0020] The second AI model is trained based on the loss value.

[0021] Optionally, the second AI model is trained based on the loss value, including:

[0022] Based on the loss value, first information is determined, which includes: information on training depth and / or information on training method;

[0023] Based on the first information, the second AI model is trained.

[0024] Optionally, training the second AI model based on the first information includes:

[0025] Obtain gradient information, which is obtained by backpropagating the loss value in the second AI model;

[0026] The second AI model is trained based on the first information and the gradient information.

[0027] Optionally, the method further includes:

[0028] Based on the first information, a second information is sent to the terminal, the second information being used to indicate the end of training.

[0029] Optionally, the method further includes:

[0030] Based on the first information, the gradient information is sent to the terminal, which then propagates the gradient information back into the first AI model and updates the weights of the first AI model.

[0031] And / or,

[0032] Based on the first information, a third information is received, which is used to indicate whether the training has ended.

[0033] Optionally, the method further includes:

[0034] Configure the first parameter of the AI ​​model;

[0035] The first parameter includes one or more of the following:

[0036] Training interval;

[0037] Number of training samples;

[0038] Number of training iterations.

[0039] Optionally, the method further includes:

[0040] Send at least one of the following to the terminal: number of training iterations, training interval, and number of training samples.

[0041] Optionally, the method further includes:

[0042] Determine whether to reconfigure the training interval;

[0043] If it is determined that the training interval needs to be reconfigured, then the training interval is reconfigured.

[0044] Send the reconfigured training interval to the terminal.

[0045] Optionally, determine whether to reconfigure the training interval, including:

[0046] Obtain communication quality;

[0047] Based on the communication quality, determine whether to reconfigure the training interval.

[0048] Optionally, based on the communication quality, determining whether to reconfigure the training interval includes:

[0049] Compare the highest bit error rate obtained when the current first AI model performs channel compression feedback with the bit error rate threshold.

[0050] Obtain the highest bit error rate and bit error rate threshold obtained when the current first AI model performs channel compression feedback;

[0051] Compare the highest bit error rate with the bit error rate threshold;

[0052] If the maximum bit error rate is less than or equal to the bit error rate threshold, then increase the training interval;

[0053] or,

[0054] If the maximum bit error rate is greater than or equal to the bit error rate threshold, then the training interval is reduced.

[0055] Secondly, an AI model training method is provided for application on a terminal, including:

[0056] Uplink channel state information is sent to network-side devices. The uplink channel state information is used to train the AI ​​model, which includes a first AI model and / or a second AI model.

[0057] And / or,

[0058] Send downlink channel state information and a first result to the network-side device. The first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model. The first result and the downlink channel state information are used to train the second AI model.

[0059] The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

[0060] Optionally, the method further includes:

[0061] Receive the first AI model that has been trained.

[0062] Optionally, the method further includes:

[0063] Receive a second message, which indicates the end of training;

[0064] Wherein, the second information is sent by the network-side device based on the first information, the first information is determined based on the loss value, and the first information includes: training depth information and / or training method information; the loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information, and the calculation result of the second AI model inference is the result obtained by the second AI model inference calculation of the first result.

[0065] Optionally, the method further includes:

[0066] Receive gradient information, which is obtained by backpropagating the loss value in the second AI model. The loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information.

[0067] The gradient information is then backpropagated in the first AI model, and the weights of the first AI model are updated.

[0068] Optionally, the method further includes:

[0069] A third message is sent, which indicates whether the training has ended.

[0070] Optionally, the method further includes:

[0071] It receives at least one of the following: number of training iterations, training interval, and number of training samples.

[0072] Thirdly, an AI model training device is provided for use in network-side equipment, including:

[0073] The first acquisition module is used to acquire uplink channel status information;

[0074] The first training module is used to train the AI ​​model based on the uplink channel state information, wherein the AI ​​model includes a first AI model and / or a second AI model.

[0075] And / or,

[0076] The second acquisition module is used to acquire downlink channel state information and a first result, wherein the first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model;

[0077] The second training module is used to train the second AI model based on the first result and downlink channel state information.

[0078] The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

[0079] Fourthly, an AI model training method is provided for application on a terminal, including:

[0080] The sixth sending module is used to send uplink channel state information to the network-side device. The uplink channel state information is used to train the AI ​​model, which includes a first AI model and / or a second AI model.

[0081] And / or,

[0082] The seventh sending module is used to send downlink channel state information and a first result to the network-side device. The first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model. The first result and the downlink channel state information are used to train the second AI model.

[0083] The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

[0084] Fifthly, a communication device is provided, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, perform the steps as described in the first or second aspect.

[0085] A sixth aspect provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, perform the steps as described in the first or second aspect.

[0086] In one embodiment of this application, the network-side device obtains the uplink channel state information sent by the terminal and uses the reciprocity of the uplink and downlink channels to train the first AI model and / or the second AI model. This avoids the need for the network-side device to send a downlink channel reference signal to the terminal, the terminal to perform downlink channel measurement, and then the terminal to feed back the downlink channel measurement results to the network side when the network-side device uses the downlink channel state information for training. This can effectively reduce the transmission overhead of the data required for training while ensuring the performance of the AI ​​model.

[0087] In another embodiment of this application, the network-side device obtains the downlink channel state information and the first result sent by the terminal; the network-side device trains the second AI model based on the first result and the downlink channel state information. It can be understood that, in addition to the downlink channel state information sent by the terminal to the network-side device, the amount of training data sent by the terminal to the network-side device includes the amount of data related to the first result fed back in a single time during the training process. This makes the transmission overhead of the training data required in this embodiment significantly reduced compared with the transmission overhead of the terminal needing to feed back all layer parameters of the trained first AI model to the network-side device. This achieves an effective reduction in the transmission overhead of the data required to train the second AI model while ensuring the performance of the second AI model. Attached Figure Description

[0088] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0089] Figure 1 This is a schematic diagram of a communication system provided in an embodiment of this application;

[0090] Figure 2 This is one of the flowcharts of the AI ​​model training method provided in the embodiments of this application;

[0091] Figure 3This is the second flowchart of the AI ​​model training method provided in the embodiments of this application;

[0092] Figure 4 This is the third flowchart of the AI ​​model training method provided in the embodiments of this application;

[0093] Figure 5 This is the fourth flowchart of the AI ​​model training method provided in the embodiments of this application;

[0094] Figure 6 This is a diagram illustrating the selection of model retraining depth;

[0095] Figure 7 This is a schematic diagram illustrating the impact of model retraining depth on model performance;

[0096] Figure 8 This is a schematic diagram illustrating the reciprocity of uplink and downlink channels;

[0097] Figure 9 This is one of the schematic diagrams of an AI model training device provided in an embodiment of this application;

[0098] Figure 10 This is a second schematic diagram of the AI ​​model training device provided in the embodiments of this application;

[0099] Figure 11 This is a schematic diagram of a communication device provided in an embodiment of this application;

[0100] Figure 12 This is a schematic diagram of pre-training provided in an embodiment of this application. Detailed Implementation

[0101] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0102] The term "comprising," and any variations thereof, used in the specification and claims of this application, is intended to cover a non-exclusive inclusion. For example, a process, manner, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, manner, product, or apparatus. Furthermore, the use of "and / or" in the specification and claims indicates at least one of the connected objects, such as A and / or B, indicating the inclusion of A alone, B alone, or both A and B.

[0103] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0104] See Figure 1 This is a schematic diagram of the architecture of a wireless communication system provided in an embodiment of the present invention. The wireless communication system may include a network-side device 11 and a terminal 12, wherein the terminal 12 can communicate with the network-side device 11 (transmitting signaling or transmitting data). In practical applications, the connection between the above-mentioned devices can be a wireless connection. To conveniently and intuitively represent the connection relationship between the devices, Figure 1 Solid lines are used to represent the meaning.

[0105] The terminals involved in this application can be mobile phones, tablet computers, laptop computers (also known as notebook computers), personal digital assistants (PDAs), handheld computers, netbooks, ultra-mobile personal computers (UMPCs), mobile internet devices (MIDs), augmented reality (AR) / virtual reality (VR) devices, robots, wearable devices, vehicle-mounted devices (VUEs), pedestrian terminals (PUEs), smart home devices (home appliances with wireless communication capabilities, such as refrigerators, televisions, washing machines, or furniture), game consoles, personal computers (PCs), ATMs, or self-service machines, etc. Wearable devices include: smartwatches, smart bracelets, smart headphones, smart glasses, smart jewelry (smart bracelets, smart chains, smart rings, smart necklaces, smart anklets, smart anklets, etc.), smart wristbands, smart clothing, etc. It should be noted that the specific type of terminal is not limited in the embodiments of this application.

[0106] The network-side equipment involved in this application may include access network equipment, which may also be referred to as radio access network equipment, radio access network (RAN), radio access network function, or radio access network unit. Access network equipment may include base stations, WLAN access points, or WiFi nodes, etc. Base stations may be referred to as Node B, evolved Node B (eNB), access point, base transceiver station (BTS), radio base station, radio transceiver, basic service set (BSS), extended service set (ESS), home B node, home evolved B node, transmitting and receiving point (TRP), or any other suitable term in the art.

[0107] See Figure 2 This application provides an AI model training method applied to a network-side device, with specific steps including: steps 201, 202, and / or steps 203 and 204.

[0108] Step 201: Obtain uplink channel state information;

[0109] Uplink channel state information can also be called uplink channel data;

[0110] Optionally, step 201 may include: receiving one or more Sounding Reference Signals (SRS) or Demodulation Reference Signals (DMRS) from one or more terminals; and obtaining uplink channel state information based on the one or more SRS or DMRS. For example, the network-side device obtains measured uplink channel state information based on multiple SRS or DMRS from a terminal, or the network-side device performs uplink channel measurement based on one or more SRS or DMRS sent by multiple terminals to obtain measured uplink channel state information, so that the network side can obtain sufficient uplink channel data samples, thereby improving the accuracy of the trained AI model.

[0111] Step 202: Train the AI ​​model based on the uplink channel state information;

[0112] The AI ​​model may include a first AI model and / or a second AI model.

[0113] The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

[0114] In step 202, the network-side device can use the received uplink channel state information as training data to train the AI ​​model of the network-side device, that is, to adjust the parameters in the AI ​​model using the training data. It is understood that existing methods can be used to train the AI ​​model in combination with the training data determined in the embodiments of this application, which will not be elaborated here.

[0115] The first AI model can also be called the encoder in the AI ​​model of the network-side device, and the second AI model can also be called the decoder in the AI ​​model of the network-side device.

[0116] Optionally, after the training process in step 202 is completed, the first AI model can be sent to the terminal.

[0117] Step 203: Obtain downlink channel state information and a first result, wherein the first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model;

[0118] For example, the network-side device receives downlink channel state information and the first result from the terminal. For instance, the network-side device sends a Channel State Information-Reference Signal (CSI-RS) to the terminal. The terminal performs downlink channel measurement based on the CSI-RS and sends the measured downlink channel state information to the network-side device.

[0119] Optionally, the first result is the intermediate result of forward propagation obtained by the terminal during the first iteration by forward propagating the measured downlink channel state information in the first AI model. Forward propagation refers to the process of progressively calculating low-level features into abstract high-level features from the input feature vector to the output and finally to the cost function until the loss is obtained. Forward propagation proceeds sequentially from the input layer to the output layer, calculating and storing the intermediate variables of the neural network.

[0120] Optionally, the downlink channel state information can be compressed and decompressed in an existing manner.

[0121] Step 204: Train the second AI model based on the first result and the downlink channel state information.

[0122] It is understandable that, in one implementation, Figure 2The method shown may include steps 201 and 202. In this embodiment, the training of the model can be understood as pre-training.

[0123] In this embodiment, during pre-training, the reciprocity between the uplink and downlink channels is utilized, and uplink channel state information is used for pre-training on the network side, which can significantly reduce training data transmission overhead. Furthermore, the pre-training data is derived from actual test data during the pre-training process, improving the model's adaptability to real-world application scenarios.

[0124] In another implementation, Figure 2 The method shown may include steps 203 and 204. In this embodiment, the training of the model can be understood as retraining.

[0125] In this embodiment, during the retraining process, the first AI model can be sent to the terminal for retraining, while the second AI model can be retrained on the network side. Considering that the number of iterations required in the retraining process is small, compared to retraining the entire model uniformly on the terminal or network side and then transmitting it, this embodiment only requires a small number of interactive training intermediate results, which can reduce the data overhead in the retraining stage.

[0126] In yet another implementation, Figure 2 The method shown may include steps 201, 202, 203 and 204. The order of steps 201 and 203 is not limited. That is, steps 201 and 202 may be executed first, and then steps 203 and 204 may be executed, or steps 201 and 203 may be executed simultaneously. In this embodiment, the training of the model includes pre-training and retraining.

[0127] In this embodiment, the transmission overhead during AI model training can be reduced while ensuring the performance of the first AI model and / or the second AI model. Based on the reciprocity of uplink and downlink channels, the first AI model and / or the second AI model are pre-trained using uplink channel state information. Subsequently, the second AI model is retrained on the network side using measured downlink channel state information with a small number of iterations. This mechanism can ensure the availability of the first AI model and / or the second AI model by performing pre-training and retraining at a very low frequency when the channel scenario does not change rapidly and drastically.

[0128] Furthermore, the retraining data during the retraining process comes from actual test data, which improves the adaptability of the second AI model to real-world application scenarios.

[0129] In one embodiment of this application, step 204 may include:

[0130] Step 2041: Based on the first result and the second AI model, obtain the calculation result of the reasoning of the second AI model;

[0131] In other words, the network-side device propagates the first result forward into the second AI model to obtain the calculation result of the second AI model's inference.

[0132] Forward propagation refers to the process by which data starts from the input layer, passes through hidden layers (if any), and finally reaches the output layer.

[0133] In this process, the information represented by the node's output value becomes more hierarchical and generalized with each layer of data propagation. The output value of a node is the result of a weighted sum of the output values ​​of all nodes in the preceding layer connected to it.

[0134] The calculation results of the second AI model's inference can also be referred to as the forward propagation intermediate layer results of the second AI model.

[0135] Step 2042: Obtain the loss value between the calculation result of the second AI model inference and the downlink channel state information;

[0136] In other words, the network-side device calculates the loss value between the two based on the calculation results inferred by the second AI model and the downlink channel state information;

[0137] Step 2043: Train the second AI model based on the loss value.

[0138] In this embodiment, the second AI model is retrained using the loss value. A good performance of the second AI model can be obtained with only a small number of retraining iterations and retraining data.

[0139] In one embodiment of this application, step 2043 includes:

[0140] Step 20431: Based on the loss value, determine the first information, which includes: information on training depth and / or information on training method;

[0141] Among them, the training depth information is used to indicate the type of model involved in training (such as retraining), or the number of layers of the model involved.

[0142] The training method information is used to indicate either method 1) the way in which the second AI model participates in training (e.g., retraining), or method 2) the way in which the first AI model and the second AI model participate in training (e.g., retraining).

[0143] During the training process in steps 203 and 204, for a second AI model with a certain depth, taking a fully connected model as an example, not all layers need to participate in the training process (e.g., retraining). It is possible to freeze the parameters of some layers and allow only a small number of layers to participate in training (e.g., retraining), such as... Figure 6 As shown.

[0144] from Figure 7 The results show that as the number of layers involved in the training (e.g., retraining) process increases, the performance gain of the second AI model decreases continuously. This means that using excessively deep training (e.g., retraining) depth has very limited impact on the performance improvement of the second AI model. Considering the different training depths, the separate training process for the first and second AI models will be divided into two cases based on whether the training (e.g., retraining) depth will affect the first AI model: the way the first AI model participates in training, and the way the first and second AI models participate in training, such as... Figure 4 and Figure 5 As shown.

[0145] The criterion for determining which training method to use can be the Loss value calculated during the first iteration. The network-side device can preset the Loss threshold according to the type of the first AI model used, as shown in Table 1.

[0146]

[0147]

[0148] Table 1: Basis for selecting the number of training (e.g., retraining) layers.

[0149] The network side can determine the number of model layers that need to participate in the training (e.g., retraining) process based on the type of the first AI model currently being used and the Loss value calculated in the first iteration, according to Table 1. It can then decide which training (e.g., retraining) method to use based on whether the model layers participating in the training (e.g., retraining) process involve the first AI model.

[0150] For example, if the first AI model is of type A, and the loss belongs to (A1, A2), then the number of retraining layers is determined to be 1. Similarly, if the first AI model is of type B, and the loss belongs to (B2, B3), then the number of retraining layers is determined to be 2.

[0151] Step 20432: Based on the first information, train (e.g., retrain) the second AI model. That is, the second AI model can be trained based on the training depth information and / or training method information.

[0152] In this embodiment, there is flexibility in the selection of training (e.g., retraining) mode. The training (e.g., retraining) method can be flexibly adjusted according to the loss value, so that the data transmission overhead of the training (e.g., retraining) process is minimized.

[0153] In one embodiment of this application, the 20432 includes:

[0154] Step 204321: Obtain gradient information, which is obtained by backpropagating the loss value in the second AI model;

[0155] For example, back propagation can calculate the gradient information of the loss value with respect to each parameter using the chain rule of derivatives, and update the parameters based on the gradient information.

[0156] Step 204322: Train the second AI model based on the first information and the gradient information.

[0157] It is understandable that after executing step 204322, the terminal can return to the step of obtaining downlink channel state information by performing downlink channel measurement based on the Channel State Information-Reference Signal (CSI-RS), and then execute the training steps in steps 203 and 204 until the training is completed.

[0158] In one embodiment of this application, the method further includes:

[0159] Based on the first information, a second information is sent to the terminal, the second information being used to indicate the end of training (e.g., retraining).

[0160] For example, if the training method information in the first information indicates that only the second AI model is involved in training (e.g., retraining), then the network-side device needs to indicate to the terminal that training has ended. Furthermore, if the training method information in the first information indicates that only the second AI model is involved in training, and the loss value is lower than or equal to a preset value, then the network-side device can indicate to the terminal that training has ended.

[0161] In the training (e.g., retraining) mode where only the second AI model is frozen, the network-side device notifies the terminal that the network-side device's second AI model has completed training (e.g., retraining).

[0162] The second piece of information mentioned above can also be called the retraining iteration confirmation signal.

[0163] In another embodiment of this application, the method further includes:

[0164] Based on the first information, the gradient information is sent to the terminal, which then propagates the gradient information back into the first AI model and updates the weights of the first AI model.

[0165] For example, when both the first AI model and the second AI model participate in retraining, the gradient information is sent to the terminal. In this embodiment, during the retraining process of the first AI model and the second AI model, the network-side device can determine the required retraining depth based on the Loss value calculated in the first iteration, and determine the specific retraining method based on the retraining depth. When the retraining depth does not involve the first AI model, the retraining process of the first AI model and the second AI model is completed entirely on the network side. When the retraining depth involves the first AI model and the second AI model, the network side needs to send the gradient information and the Loss value to the terminal, and the terminal will continue to backpropagate the gradient information in the first AI model and update the weights of the first AI model.

[0166] Understandably, during the retraining process of the first AI model and the second AI model, the terminal needs to feed back the intermediate forward propagation results of the first AI model to the network side. When the retraining depth does not involve the first AI model, this feedback only occurs in the first retraining iteration. When the retraining depth involves the first AI model, the terminal needs to feed back the intermediate forward propagation results of the first AI model to the network side in each retraining iteration. In addition, the terminal needs to feed back the channel sounding reference signal (CSI) to the network side in each retraining iteration.

[0167] Furthermore, optionally, the method further includes:

[0168] Based on the first information, a third information is received, which is used to indicate whether the training (e.g., retraining) has ended. For example, the third information is received when both the first AI model and the second AI model are participating in retraining.

[0169] For example, if the training method information in the first message indicates that both the first AI model and the second AI model are participating in training (e.g., retraining), then the network-side device needs to receive a third message from the terminal indicating whether the training has ended.

[0170] The aforementioned third piece of information can also be referred to as the CSI-RS retraining request signaling.

[0171] In this embodiment, if the third information is used to indicate that retraining has not ended, the network-side device can send CSI-RS to the terminal and then repeat the retraining steps.

[0172] In this embodiment, in a mode where both the first AI model and the second AI model participate in training (e.g., retraining), the terminal notifies the network-side device that it has completed backpropagation and declares whether the retraining iteration has been completed.

[0173] It is understandable that when the training (e.g., retraining) depth involves the first AI model, the terminal needs to send third information to the network side after each training (e.g., retraining) iteration, such as a retraining CSI-RS request signaling.

[0174] In another embodiment of this application, the method further includes:

[0175] Configure the first parameter of the AI ​​model; wherein the first parameter includes one or more of the following:

[0176] (1) Training interval, also known as retraining interval;

[0177] Optionally, the training interval is the number of downlink channel state information feedbacks between the time the AI ​​model completes retraining and the time before the next training.

[0178] (2) Number of training samples;

[0179] Optionally, the number of training samples includes: the number of pre-training samples and the number of retraining samples;

[0180] (3) Number of training iterations;

[0181] Optionally, the number of training iterations includes: the number of pre-training iterations and the number of retraining iterations.

[0182] The number of retraining iterations refers to the number of iterations required for the AI ​​model to be retrained using measured downlink channel state information after pre-training using uplink channel state information.

[0183] In one embodiment of this application, the method further includes:

[0184] Send at least one of the following to the terminal: number of training iterations, training interval, and number of training samples.

[0185] The network-side equipment can determine the number of uplink channel state information based on the number of training samples, and the terminal can determine the number of SRS or DMRS to be sent based on the number of training samples.

[0186] The network-side equipment can determine the number of times it needs to acquire downlink channel state information and first results based on the number of training iterations, and the terminal can determine the number of times it needs to report downlink channel state information and first results based on the number of training iterations.

[0187] Network-side devices and terminals can determine the interval between two adjacent training sessions, such as time, based on the training interval.

[0188] Optionally, pre-training on the network side requires the network side to pre-configure the number of pre-training iterations and the number of samples required for pre-training and retraining. The pre-training process itself is only completed on the network side, and the training process is independent of the terminal. Therefore, the number of pre-training iterations does not need to be synchronized with the terminal. However, the pre-training samples required by the network side need to be obtained with the assistance of the sounding reference signal (SRS) sent by the terminal. Retraining also requires the cooperation between the network side and the terminal. Therefore, the number of samples required for pre-training and retraining configured on the network side needs to be synchronized with the terminal.

[0189] In one embodiment of this application, the method further includes:

[0190] Determine whether to reconfigure the training interval;

[0191] If it is determined that the training interval needs to be reconfigured, then the training interval is reconfigured (e.g., retraining).

[0192] Send the reconfigured training interval (e.g., retraining) to the terminal.

[0193] Further, optionally, the method for determining whether to reconfigure the training interval includes: obtaining the communication quality; and determining whether to reconfigure the training interval based on the communication quality.

[0194] Further, optionally, based on the communication quality, determining whether to reconfigure the training interval includes:

[0195] Obtain the highest bit error rate and bit error rate threshold obtained when the current first AI model performs channel compression feedback;

[0196] Compare the highest bit error rate with the bit error rate threshold;

[0197] If the highest bit error rate is less than or equal to the bit error rate threshold, then increase the training (e.g., retraining) interval.

[0198] or,

[0199] If the highest bit error rate is greater than or equal to the bit error rate threshold, then reduce the training (e.g., retraining) interval.

[0200] In this embodiment, the network-side device can determine whether the training (e.g., retraining) interval needs to be adjusted based on information such as the transmission bit error rate. If adjustment is required, the network-side device can reconfigure the training (e.g., retraining) interval based on higher-layer information and send the new training (e.g., retraining) interval to the terminal.

[0201] In this embodiment, the frequency of training (such as retraining) is flexibly adjusted by leveraging the network side's assessment of transmission quality, thereby minimizing the waste of computing and data transmission resources while ensuring communication quality.

[0202] In one embodiment of this application, the network-side device obtains the uplink channel state information sent by the terminal and uses the reciprocity of the uplink and downlink channels to train the first AI model and / or the second AI model. In this way, during the training process, the terminal only needs to send the uplink channel reference signal (such as SRS or DMRS) to the network-side device. This avoids the need for the network-side device to send the downlink channel reference signal to the terminal, the terminal to perform downlink channel measurement, and then the terminal to feed back the downlink channel measurement results to the network side when the network-side device uses the downlink channel state information for training. This can effectively reduce the transmission overhead of the data required for training while ensuring the performance of the AI ​​model.

[0203] In another embodiment of this application, the network-side device obtains the downlink channel state information and the first result sent by the terminal; the network-side device trains the second AI model based on the first result and the downlink channel state information. It can be understood that, in addition to the downlink channel state information sent by the terminal to the network-side device, the amount of training data sent by the terminal to the network-side device includes: the amount of data of the first result fed back in a single training process * the number of training iterations. In this embodiment, the transmission overhead of the data required for training is significantly reduced compared with the transmission overhead of the terminal needing to feed back all layer parameters of the trained first AI model to the network-side device, thereby effectively reducing the transmission overhead of the data required for training the second AI model while ensuring the performance of the second AI model.

[0204] See Figure 3 This application provides an AI model training method for use on a terminal, and the specific steps include: step 301 and / or step 302.

[0205] Step 301: Send uplink channel state information to the network-side device. The uplink channel state information is used to train the AI ​​model. The AI ​​model includes a first AI model and / or a second AI model.

[0206] Step 302: Send downlink channel state information and a first result to the network-side device. The first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model. The first result and the downlink channel state information are used to train the second AI model.

[0207] The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

[0208] In one embodiment of this application, the method further includes:

[0209] Receive the first AI model that has been trained.

[0210] In one embodiment of this application, the method further includes:

[0211] Receive a second message, which indicates the end of training;

[0212] Wherein, the second information is sent by the network-side device based on the first information, the first information is determined based on the loss value, and the first information includes: training depth information and / or training method information; the loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information, and the calculation result of the second AI model inference is the result obtained by the second AI model inference calculation of the first result.

[0213] In one embodiment of this application, the method further includes:

[0214] Receive gradient information, which is obtained by backpropagating the loss value in the second AI model. The loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information.

[0215] The gradient information is then backpropagated in the first AI model, and the weights of the first AI model are updated.

[0216] In one embodiment of this application, the method further includes:

[0217] A third message is sent, which indicates whether the training has ended.

[0218] In one embodiment of this application, the method further includes:

[0219] It receives at least one of the following: number of training iterations, training interval, and number of training samples.

[0220] In one embodiment of this application, during the training process, the terminal only needs to send an uplink channel reference signal (such as SRS or DMRS) to the network-side device. This avoids the need for the network-side device to send a downlink channel reference signal to the terminal, the terminal to perform downlink channel measurement, and then the terminal to feed back the downlink channel measurement results to the network side when the network-side device uses downlink channel state information for training. This can effectively reduce the transmission overhead of the data required for training while ensuring the performance of the AI ​​model.

[0221] In another embodiment of this application, in addition to the terminal sending downlink channel state information to the network-side device, the amount of training data sent by the terminal to the network-side device includes: the amount of data of the first result fed back in a single training process * the number of training iterations. In this embodiment, the transmission overhead of the training data is significantly reduced compared with the transmission overhead of the terminal needing to feed back all layer parameters of the first AI model trained to the network-side device. This achieves an effective reduction in the transmission overhead of the data required to train the second AI model while ensuring the performance of the second AI model.

[0222] For ease of understanding, the implementation methods of this application are described below with reference to Embodiment 1 and Embodiment 2, wherein the first AI model is an encoder and the second AI model is a decoder.

[0223] Example 1: Mode in which only the decoder participates in retraining

[0224] See Figure 4 The specific steps are as follows:

[0225] Step 0: The terminal synchronizes the encoder and decoder required for channel compression with the network side;

[0226] Step 1: Configure the retraining interval N_d on the network side, which is the number of feedbacks to wait for after training is completed, and the number of samples required for pretraining and / or retraining;

[0227] Step 2: Configure the number of pre-training and / or retraining iterations on the network side at higher levels;

[0228] Step 3: The terminal and network sides synchronize the number of retraining iterations, the retraining interval N_d, and the number of samples required for pre-training and / or retraining;

[0229] Step 4: On the network side, the initial weight parameters of the encoder and decoder are randomly set based on the synchronized encoder and decoder framework;

[0230] Step 5: The terminal sends SRS or DMRS to the network side;

[0231] Step 6: The network side processes and parses the SRS or DMRS to obtain uplink channel state information;

[0232] Step 7: The network side utilizes uplink channel state information and the reciprocity between uplink and downlink channels to pre-train the initial weights W1 of the encoder and decoder on the network side;

[0233] Considering that the uplink and downlink channel center frequencies are similar, their channel characteristics are very similar, and they have a certain degree of reciprocity, such as Figure 8 As shown. From Figure 8As can be seen, although the uplink and downlink channels differ significantly in value (mainly due to the independent phases of the uplink and downlink channels), the distribution of each component in the time delay domain and angle domain is highly similar. Therefore, it is possible to use the uplink channel state information to provide coarser-grained measured raw data for the pre-training of the encoder and decoder.

[0234] The process of pre-training the encoder and decoder for downlink channel compression and decompression using uplink channel state information corresponds to steps 5-7 in the flowchart. Some steps in the flowchart are simplified; the actual process... Figure 12 As shown.

[0235] During the pre-training process, the terminal needs to send a large number of SRS or DMRS to the network side to obtain sufficient uplink channel data samples. The number of samples has been configured by the network side.

[0236] Step 8: The network side transmits the pre-trained encoder to the terminal;

[0237] Step 9: The network side sends CSI-RS to the terminal;

[0238] Step 10: The terminal performs downlink channel measurements based on CSI-RS;

[0239] Step 11: The terminal forward propagates the measured downlink channel state information in the encoder (executed only in the first iteration);

[0240] Step 12: The terminal feeds back the intermediate results of the forward propagation (executed only in the first iteration) and the measured downlink channel state information (compressed according to the existing standard scheme) to the network side;

[0241] Step 13: The network side forwards the received intermediate forward propagation results in the decoder;

[0242] Step 14: The network side calculates the loss value between the intermediate forward propagation result of the decoder and the received measured downlink channel state information;

[0243] Step 15: Determine the retraining depth and retraining method to be used based on the loss value;

[0244] Step 16: The decoder on the network side performs backpropagation based on the calculated Loss value and updates the model weights, then returns to step 10 until retraining is complete, and then executes step 17.

[0245] Step 17: The network side sends a retraining iteration confirmation signal to the terminal, informing the terminal that the retraining has ended;

[0246] Step 18: The network side sends CSI-RS to the terminal;

[0247] Step 19: The terminal completes downlink channel measurements based on CSI-RS;

[0248] Step 20: The terminal uses the encoder to compress and quantize the downlink channel state information;

[0249] Step 21: The terminal feeds back the quantized bitstream to the network side;

[0250] Step 22: The network side uses a decoder to dequantize and decompress the downlink channel state information;

[0251] Step 23: The network side performs precoding based on the downlink channel state information;

[0252] Step 24: The network side determines whether a reconfiguration and retraining interval is needed based on communication quality;

[0253] Step 25: The terminal and network side synchronize and retrain at intervals N. After completion, return to step 9.

[0254] Example 2: Mode in which both encoder and decoder participate in retraining.

[0255] See Figure 5 The specific steps are as follows:

[0256] Step 0: The terminal synchronizes the encoder and decoder required for channel compression with the network side;

[0257] Step 1: Configure the retraining interval N_d on the network side, which is the number of feedbacks to wait for after training is completed, and the number of samples required for pretraining and / or retraining;

[0258] Step 2: Configure the number of pre-training and / or retraining iterations on the network side at higher levels;

[0259] Step 3: The terminal and network sides synchronize the number of retraining iterations, the retraining interval N_d, and the number of samples required for pre-training and / or retraining;

[0260] Step 4: On the network side, the initial weight parameters of the encoder and decoder are randomly set based on the synchronized encoder and decoder framework;

[0261] Step 5: The terminal sends an uplink channel reference signal (SRS) or DMRS to the network side;

[0262] Step 6: Process and parse the SRS or DMRS to obtain uplink channel data;

[0263] Step 7: Utilize uplink channel state information and the reciprocity between uplink and downlink channels to pre-train the initial weights W1 of the encoder and decoder on the network side;

[0264] Step 8: The network side only transmits the pre-trained encoder to the terminal;

[0265] Step 9: The network side sends CSI-RS to the terminal;

[0266] Step 10: The terminal performs downlink channel measurements based on CSI-RS;

[0267] Step 11: The terminal forward propagates the measured downlink channel state information in the encoder (executed only in the first iteration);

[0268] Step 12: The terminal feeds back the intermediate results of the forward propagation (executed only in the first iteration) and the measured downlink channel state information (compressed according to the existing standard scheme) to the network side;

[0269] Step 13: The network side forwards the received intermediate forward propagation results in the decoder;

[0270] Step 14: The network side calculates the loss value between the intermediate forward propagation result of the decoder and the received measured downlink channel state information;

[0271] Step 15: Determine the retraining depth and retraining method to be used based on the loss value;

[0272] Step 16: The decoder on the network side performs backpropagation based on the calculated Loss value and updates the model weights;

[0273] Step 17: The network side sends the intermediate information (gradient information) of the backpropagation to the terminal;

[0274] Step 18: The terminal encoder continues backpropagation based on the intermediate information from backpropagation and updates the encoder weights;

[0275] Step 19: The terminal sends a retraining CSI-RS request signaling to the network side to inform the network side whether the retraining has ended. If it has not ended, return to step 9; if it has ended, continue to step 20.

[0276] Step 20: The network side sends CSI-RS to the terminal;

[0277] Step 21: The terminal completes downlink channel measurements based on CSI-RS;

[0278] Step 22: Compress and quantize the downlink channel state information using an encoder;

[0279] Step 23: The terminal feeds back the quantized bitstream to the network side;

[0280] Step 24: The network side uses a decoder to dequantize and decompress the downlink channel state information;

[0281] Step 25: The network side performs precoding based on the downlink channel state information;

[0282] Step 26: The network side determines whether a reconfiguration and retraining interval is needed based on communication quality;

[0283] Step 27: The terminal and network side synchronize and retrain at intervals N. After completion, return to step 9.

[0284] See Figure 9 This application provides an AI model training device for use in network-side equipment. The device 900 includes:

[0285] The first acquisition module 901 is used to acquire uplink channel status information;

[0286] The first training module 902 is used to train the AI ​​model based on the uplink channel state information, wherein the AI ​​model includes a first AI model and / or a second AI model.

[0287] And / or,

[0288] The second acquisition module 903 is used to acquire downlink channel state information and a first result, wherein the first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model.

[0289] The second training module 904 is used to train the second AI model based on the first result and the downlink channel state information.

[0290] The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

[0291] In one embodiment of this application, the apparatus further includes:

[0292] The first sending module is used to send the trained first AI model to the terminal.

[0293] In one embodiment of this application, the first acquisition module 901 includes:

[0294] A receiving unit is used to receive one or more SRS or DMRS from one or more terminals;

[0295] The acquisition unit is used to acquire uplink channel state information based on one or more SRS or DMRS.

[0296] In one embodiment of this application, the second training module 904 includes:

[0297] The first processing unit is used to obtain the calculation result of the reasoning of the second AI model based on the first result and the second AI model;

[0298] The first acquisition unit is used to acquire the loss value between the calculation result of the second AI model inference and the downlink channel state information;

[0299] The first training unit is used to train the second AI model based on the loss value.

[0300] In one embodiment of this application, the first training unit is further configured to:

[0301] Based on the loss value, first information is determined, which includes: information on training depth and / or information on training method;

[0302] Based on the first information, the second AI model is trained.

[0303] In one embodiment of this application, the first training unit is further configured to:

[0304] Obtain gradient information, which is obtained by backpropagating the loss value in the second AI model;

[0305] The second AI model is trained based on the first information and the gradient information.

[0306] In one embodiment of this application, the apparatus further includes:

[0307] The second sending module is used to send second information to the terminal based on the first information, the second information being used to indicate the end of training.

[0308] In one embodiment of this application, the apparatus further includes:

[0309] The third sending module is used to send the gradient information to the terminal according to the first information, and the terminal continues to backpropagate the gradient information in the first AI model and updates the weights of the first AI model.

[0310] In one embodiment of this application, the apparatus further includes:

[0311] The first receiving module is configured to receive third information based on the first information, wherein the third information is used to indicate whether the training has ended.

[0312] In one embodiment of this application, the apparatus further includes:

[0313] The configuration module is used to configure the first parameter of the AI ​​model;

[0314] The first parameter includes one or more of the following:

[0315] Training interval;

[0316] Number of training samples;

[0317] Number of training iterations.

[0318] In one embodiment of this application, the apparatus further includes:

[0319] The fourth sending module is used to send at least one of the following to the terminal: the number of training iterations, the training interval, and the number of training samples.

[0320] In one embodiment of this application, the apparatus further includes:

[0321] The decision module is used to determine whether the training interval needs to be reconfigured.

[0322] In one embodiment of this application, the determination module is further configured to: obtain communication quality; and determine whether to reconfigure the training interval based on the communication quality.

[0323] In one embodiment of this application, the judgment module is further configured to: obtain and compare the highest bit error rate obtained by the current first AI model when performing channel compression feedback with a bit error rate threshold; obtain the highest bit error rate obtained by the current first AI model when performing channel compression feedback with a bit error rate threshold; compare the highest bit error rate with the bit error rate threshold; if the highest bit error rate is less than or equal to the bit error rate threshold, increase the training interval; or, if the highest bit error rate is greater than or equal to the bit error rate threshold, decrease the training interval.

[0324] In one embodiment of this application, the apparatus further includes:

[0325] The reconfiguration module is used to reconfigure the training interval if it is determined that reconfiguration is needed.

[0326] The fifth sending module is used to send the reconfigured training interval to the terminal.

[0327] The apparatus provided in this application embodiment can achieve... Figure 2 The various processes implemented in the illustrated embodiments achieve the same technical effect, and will not be described again here to avoid repetition.

[0328] See Figure 10 This application provides an AI model training method for use on a terminal, comprising:

[0329] The sixth sending module 1001 is used to send uplink channel status information to the network side device. The uplink channel status information is used to train the AI ​​model, and the AI ​​model includes a first AI model and / or a second AI model.

[0330] And / or,

[0331] The seventh sending module 1002 is used to send downlink channel state information and a first result to the network-side device. The first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model. The first result and the downlink channel state information are used to train the second AI model.

[0332] The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

[0333] In one embodiment of this application, the apparatus further includes:

[0334] The second receiving module is used to receive the first AI model after it has been trained.

[0335] In one embodiment of this application, the apparatus further includes:

[0336] The third receiving module is used to receive the second information, which is used to indicate the end of training.

[0337] Wherein, the second information is sent by the network-side device based on the first information, the first information is determined based on the loss value, and the first information includes: training depth information and / or training method information; the loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information, and the calculation result of the second AI model inference is the result obtained by the second AI model inference calculation of the first result.

[0338] In one embodiment of this application, the apparatus further includes:

[0339] The fourth receiving module is used to receive gradient information, which is obtained by backpropagating the loss value in the second AI model. The loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information.

[0340] The second processing module is used to continue backpropagating the gradient information in the first AI model and update the weights of the first AI model.

[0341] In one embodiment of this application, the apparatus further includes:

[0342] The eighth sending module is used to send third information, which is used to indicate whether the training has ended.

[0343] In one embodiment of this application, the apparatus further includes:

[0344] The fifth receiving module is used to receive at least one of the following: the number of training iterations, the training interval, and the number of training samples.

[0345] The apparatus provided in this application embodiment can achieve... Figure 3 The various processes implemented in the illustrated embodiments achieve the same technical effect, and will not be described again here to avoid repetition.

[0346] like Figure 11 As shown, this application embodiment also provides a communication device 1100, including a processor 1101, a memory 1102, and a program or instructions stored in the memory 1102 and executable on the processor 1101. When the program or instructions are executed by the processor 1101, they implement the above-mentioned... Figure 2 or Figure 3 The various processes in the illustrated embodiments achieve the same technical effect. To avoid repetition, they will not be described again here.

[0347] This application embodiment also provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the above-described functionality. Figure 2 or Figure 3 The various processes in the illustrated embodiments can achieve the same technical effect, and will not be described again here to avoid repetition.

[0348] The processor mentioned above is the processor in the terminal described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0349] The steps of the methods or algorithms described in this application can be implemented in hardware or by executing software instructions on a processor. The software instructions can consist of corresponding software modules, which can be stored in RAM, flash memory, ROM, EPROM, EEPROM, registers, hard disk, portable hard disk, read-only optical disk, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can be housed in an ASIC. Alternatively, the ASIC can be housed in a core network interface device. Of course, the processor and storage medium can also exist as discrete components in the core network interface device.

[0350] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this application can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of a computer program from one place to another. Storage media can be any available medium accessible to a general-purpose or special-purpose computer.

[0351] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this application.

[0352] Those skilled in the art will understand that embodiments of this application can be provided as a method, system, or computer program product. Therefore, embodiments of this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, embodiments of this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0353] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of embodiments of the present application, apparatus (systems), and computer program products. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0354] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0355] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0356] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. An artificial intelligence (AI) model training method, applied to network-side devices, characterized in that, include: Obtain uplink channel status information; The AI ​​model is trained based on the uplink channel state information, and the AI ​​model includes a first AI model and a second AI model. The trained first AI model is sent to the terminal; Obtain downlink channel state information and a first result, wherein the first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model; Based on the first result and the downlink channel state information, the second AI model is trained; The second AI model is trained based on the first result and the downlink channel state information, including: Based on the first result and the second AI model, the calculation result of the reasoning of the second AI model is obtained; Obtain the loss value between the calculation result of the second AI model inference and the downlink channel state information; Based on the loss value, first information is determined, which includes: information on training depth and / or information on training method; Based on the first information, the second AI model is trained; The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

2. The method according to claim 1, characterized in that, The acquisition of uplink channel state information includes: Receive one or more probe reference signals (SRS) or demodulation reference signals (DMRS) from one or more terminals; Uplink channel state information is obtained based on one or more SRS or DMRS.

3. The method according to claim 1, characterized in that, The step of training the second AI model based on the first information includes: Obtain gradient information, which is obtained by backpropagating the loss value in the second AI model; The second AI model is trained based on the first information and the gradient information.

4. The method according to claim 1, characterized in that, The method further includes: Based on the first information, a second information is sent to the terminal, the second information being used to indicate the end of training.

5. The method according to claim 3, characterized in that, The method further includes: Based on the first information, the gradient information is sent to the terminal, which then propagates the gradient information back into the first AI model and updates the weights of the first AI model. And / or, Based on the first information, a third information is received, which is used to indicate whether the training has ended.

6. The method according to claim 1, characterized in that, The method further includes: Configure the first parameter of the AI ​​model; The first parameter includes one or more of the following: Training interval; Number of training samples; Number of training iterations.

7. The method according to claim 6, characterized in that, The method further includes: Send at least one of the following to the terminal: number of training iterations, training interval, and number of training samples.

8. The method according to claim 6, characterized in that, The method further includes: Determine whether to reconfigure the training interval; If it is determined that the training interval needs to be reconfigured, then the training interval is reconfigured. Send the reconfigured training interval to the terminal.

9. The method according to claim 8, characterized in that, Determining whether to reconfigure the training interval includes: Obtain communication quality; Based on the communication quality, determine whether to reconfigure the training interval.

10. The method according to claim 9, characterized in that, Based on the communication quality, determine whether to reconfigure the training interval, including: Compare the highest bit error rate obtained when the current first AI model performs channel compression feedback with the bit error rate threshold. Obtain the highest bit error rate and bit error rate threshold obtained when the current first AI model performs channel compression feedback; Compare the highest bit error rate with the bit error rate threshold; If the maximum bit error rate is less than or equal to the bit error rate threshold, then increase the training interval; or, If the maximum bit error rate is greater than or equal to the bit error rate threshold, then the training interval is reduced.

11. An AI model training method, applied to a terminal, characterized in that, include: Uplink channel state information is sent to network-side devices. The uplink channel state information is used to train the AI ​​model, which includes a first AI model and a second AI model. Receive the first AI model that has been trained; Send downlink channel state information and a first result to the network-side device. The first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model. The first result and the downlink channel state information are used to train the second AI model. The method further includes: Receive a second message, which indicates the end of training; Wherein, the second information is sent by the network-side device based on the first information, the first information is determined based on the loss value, and the first information includes: training depth information and / or training method information; the loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information, and the calculation result of the second AI model inference is the result obtained by the second AI model inference calculation of the first result; The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

12. The method according to claim 11, characterized in that, The method further includes: Receive gradient information, which is obtained by backpropagating the loss value in the second AI model. The loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information. The gradient information is then backpropagated in the first AI model, and the weights of the first AI model are updated.

13. The method according to claim 12, characterized in that, The method further includes: A third message is sent, which indicates whether the training has ended.

14. The method according to claim 11, characterized in that, The method further includes: It receives at least one of the following: number of training iterations, training interval, and number of training samples.

15. An AI model training device, applied to network-side equipment, characterized in that, include: The first acquisition module is used to acquire uplink channel status information; The first training module is used to train the AI ​​model based on the uplink channel state information. The AI ​​model includes a first AI model and a second AI model. The trained first AI model is sent to the terminal; The second acquisition module is used to acquire downlink channel state information and a first result, wherein the first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model; The second training module is used to train the second AI model based on the first result and downlink channel state information. The second training module is further configured to: obtain the calculation result of the second AI model inference based on the first result and the second AI model; and obtain the loss value between the calculation result of the second AI model inference and the downlink channel state information; Based on the loss value, first information is determined, which includes: information on training depth and / or information on training method; based on the first information, the second AI model is trained. The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

16. An AI model training method, applied to a terminal, characterized in that, include: The sixth sending module is used to send uplink channel state information to the network-side device. The uplink channel state information is used to train the AI ​​model, which includes a first AI model and a second AI model. Receive the first AI model that has been trained; The seventh sending module is used to send downlink channel state information and a first result to the network-side device. The first result represents the calculation result obtained by reasoning the downlink channel state information in the first AI model. The first result and the downlink channel state information are used to train the second AI model. Receive a second message, which indicates the end of training; Wherein, the second information is sent by the network-side device based on the first information, the first information is determined based on the loss value, and the first information includes: training depth information and / or training method information; the loss value represents the loss between the calculation result of the second AI model inference and the downlink channel state information, and the calculation result of the second AI model inference is the result obtained by the second AI model inference calculation of the first result; The first AI model is used to compress the downlink channel state information, and the second AI model is used to decompress the downlink channel state information.

17. A communication device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 14.

18. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 14.

Images