Semantic communication system generation method and apparatus, electronic device, and medium

Abstract

The application discloses a generation method and device of a semantic communication system, electronic equipment and a medium. Through the technical scheme of the application, a complete task of training a semantic communication system can be divided into multiple sub-training tasks according to modules. The multiple sub-training tasks are respectively distributed to a training object group capable of training the sub-training tasks. Each client in the training object group trains the sub-training task allocated thereto based on local data, and sends the model update parameter of the corresponding module to the server after obtaining the model update parameter. The server updates each module of the semantic communication system according to the model update parameter of each module, and obtains a target semantic communication system after training. Thus, on the one hand, the local privacy data of the client is not leaked. On the other hand, the performance consumption of the server caused by training the communication system is saved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a generation method and device of a semantic communication system, an electronic device and a medium. BACKGROUND

[0002] In the related art, semantic communication systems are continuously used with the development of various businesses. Among them, the current DLSC (DNN-based End-to-End Communication System) is implemented based on deep learning, which relies on deep neural networks to extract and encode semantic information, and then uses wireless channels to achieve the purpose of signal transmission.

[0003] Further, the semantic communication system contains millions of neural parameters, that is, it has a very large demand for data. In the prior art, uploading a large amount of data collected offline by the client to the server to train the semantic communication system not only causes privacy problems, but also is a great challenge to the performance load of the server. SUMMARY

[0004] The embodiments of the present application provide a generation method and device of a semantic communication system, an electronic device and a medium. The method is used to solve the problem of excessive server load caused by uploading a large amount of data to the server to train the semantic communication system in the related art.

[0005] According to an aspect of the embodiments of the present application, a generation method of a semantic communication system is provided, wherein:

[0006] a plurality of training tasks are established for an initial semantic communication system to be trained, wherein each training task is used to train at least one module in the initial semantic communication system;

[0007] each training task is respectively assigned to a corresponding training object group, so that the training object group trains the received training task, wherein each training object group includes a plurality of clients;

[0008] model update parameters sent by each training object group are respectively received, and the initial semantic communication system is updated based on the model update parameters to obtain a target semantic communication system that has been trained.

[0009] Optionally, in another embodiment based on the above-mentioned method of the present application, the plurality of training tasks are established for the initial semantic communication system to be trained, including:

[0010] establishing, for the initial semantic communication system, a first training task for training a semantic encoder module, a second training task for training a semantic decoder module, a third training task for training a wireless transmission channel module, and a fourth training task for training a classifier.

[0011] Optionally, in another embodiment based on the above method of the present application, before the respective distribution of each training task to the corresponding training object group, further comprising:

[0012] obtaining attribute information sent by a plurality of clients, the attribute information including device type, computing hardware, and service load condition of the clients;

[0013] based on the attribute information of each client, determining the computing performance corresponding to each client, and dividing the plurality of clients into a plurality of training object groups according to the computing performance, wherein the computing performance of the clients in each training object group is the same.

[0014] Optionally, in another embodiment based on the above method of the present application, after the plurality of clients are divided into a plurality of training object groups, further comprising:

[0015] based on the training module corresponding to each training task, determining the computing performance index corresponding to each training task;

[0016] based on the computing performance corresponding to each training object group, matching a training object group capable of meeting the computing performance index for each training task.

[0017] Optionally, in another embodiment based on the above method of the present application, after the plurality of clients are divided into a plurality of training object groups, further comprising:

[0018] from the plurality of training object groups, determining a first training object group with the lowest computing performance;

[0019] distributing the first training task for extracting semantic information to the first training object group, so that the first training object group extracts semantic information in the training data.

[0020] Optionally, in another embodiment based on the above method of the present application, the respective distribution of each training task to the corresponding training object group comprises:

[0021] distributing each training task, a semantic encoder, and an OFDM transmitter to the corresponding training object group, so that the training object group trains the received training task using the semantic encoder; and sends the model update parameter to the initial semantic communication system using the OFDM transmitter.

[0022] Optionally, in another embodiment based on the above method of the present application, after the respective distribution of each training task to the corresponding training object group, it further includes:

[0023] training the received training task by each client in the training object group using local data.

[0024] According to yet another aspect of embodiments of the present application, there is provided a semantic communication system generation apparatus, comprising:

[0025] a creating module configured to establish a plurality of training tasks for an initial semantic communication system to be trained, wherein each training task is used to train at least one module in the initial semantic communication system;

[0026] a sending module configured to respectively distribute each training task to a corresponding training object group, so that the training object group trains the received training task, wherein each training object group comprises a plurality of clients;

[0027] an updating module configured to respectively receive model update parameters sent by each training object group, and update the initial semantic communication system based on the model update parameters to obtain a target semantic communication system trained.

[0028] According to yet another aspect of embodiments of the present application, there is provided an electronic device, comprising:

[0029] a memory for storing executable instructions; and

[0030] a display for executing the executable instructions with the memory to complete the operation of any of the above-mentioned semantic communication system generation methods.

[0031] According to still another aspect of embodiments of the present application, there is provided a computer readable storage medium for storing computer readable instructions, which when executed perform the operation of any of the above-mentioned semantic communication system generation methods.

[0032] In the present application, multiple training tasks can be established for the initial semantic communication system to be trained, wherein each training task is used to train at least one module in the initial semantic communication system; each training task is respectively distributed to a corresponding training object group, so that the training object group trains the received training task, wherein each training object group includes multiple clients; model update parameters sent by each training object group are respectively received, and the initial semantic communication system is updated based on the model update parameters to obtain a target semantic communication system after training. By applying the technical solution of the present application, the complete task of training a semantic communication system can be divided into multiple sub-training tasks according to each module. The multiple sub-training tasks are respectively distributed to training object groups that have the ability to process. Each client in the training object group trains the sub-training task allocated to it, so that the model update parameters of the corresponding module are sent to the server after being obtained. The server updates each module of the semantic communication system according to the model update parameters of each module, and then obtains a target semantic communication system after training. Thus, on the one hand, the local privacy data of the client is not leaked. On the other hand, the performance consumption of the server caused by training is also saved.

[0033] The technical solution of the present application will be described in further detail below by means of the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0034] The accompanying drawings, which form a part of the specification, illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application.

[0035] The present application can be more clearly understood and appreciated from the following detailed description, taken in conjunction with the accompanying drawings, in which:

[0036] Figure 1 A schematic diagram of a semantic communication system generation method provided by an embodiment of the present application is shown;

[0037] Figure 2 A schematic diagram of the architecture of a semantic communication system provided by an embodiment of the present application is shown;

[0038] Figure 3 A flowchart of a semantic communication system generation method provided by an embodiment of the present application is shown;

[0039] Figure 4 A schematic diagram of the system architecture of a semantic communication system generation method provided by an embodiment of the present application is shown;

[0040] Figure 5 A training flowchart of a semantic communication system provided by an embodiment of the present application is shown;

[0041] Figure 6 Fig. 1 shows a structural schematic diagram of an electronic device according to an embodiment of the present application;

[0042] Figure 7 Fig. 1 shows a structural schematic diagram of an electronic device according to an embodiment of the present application;

[0043] Figure 8 Fig. 1 shows a structural schematic diagram of an electronic device according to an embodiment of the present application; DETAILED DESCRIPTION

[0044] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of the components and steps set forth in the embodiments, numerical expressions, and numerical values are not limiting to the scope of the present application unless specifically stated otherwise.

[0045] It should be understood that the sizes of the various portions shown in the drawings are not necessarily drawn to scale for ease of understanding.

[0046] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way limiting to the scope of the application or its applications or uses.

[0047] Techniques, methods, and apparatus known to those of ordinary skill in the relevant art can not be discussed in detail herein, but should be considered as part of the specification, where appropriate.

[0048] It should be noted that like identifiers and letters refer to like items throughout the several views, like numbers and letters need not be discussed further where, once defined in one drawing, they are carried over into the following drawings.

[0049] In addition, the technical solutions among the various embodiments of the present application can be combined with each other, but it must be based on the fact that a person of ordinary skill in the art can realize it, and when the combination of technical solutions contradicts each other or cannot be realized, it should be considered that the combination of technical solutions does not exist, nor is it within the scope of protection required by the present application.

[0050] It should be noted that all directional indications, such as up, down, left, right, front, back, etc., are only used to explain the relative position relationship, movement condition, etc. between the components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications will also change accordingly.

[0051] The following will be described in detail Figures 1-5A method for generating a semantic communication system according to an example embodiment of the present application will be described. It should be noted that the following application scenarios are only shown for the purpose of facilitating the understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect. On the contrary, the embodiments of the present application can be applied to any applicable scenario.

[0052] The present application also proposes a method, device, electronic equipment and medium for generating a semantic communication system.

[0053] Figure 1 A flowchart of a method for generating a semantic communication system according to an example embodiment of the present application is schematically shown. As shown in the figure, the method comprises: Figure 1

[0054] S101, a plurality of training tasks are established for an initial semantic communication system to be trained, wherein each training task is used to train at least one module in the initial semantic communication system.

[0055] S102, each training task is respectively distributed to a corresponding training object group, so that the training object group trains the received training task, wherein each training object group comprises a plurality of clients.

[0056] S103, model update parameters sent by each training object group are respectively received, and the initial semantic communication system is updated based on the model update parameters to obtain a target semantic communication system trained.

[0057] In the related art, due to the increasing number of emerging applications such as Internet of Things (IoT), human-to-machine dialogue (H2M) and machine-to-machine communication (M2M) generating a large amount of data, and mobile traffic will grow at about 55% per year in the next few years. Another significant trend is that these applications need to interpret information in a more intelligent way at the receiver, such as image recognition or video analysis.

[0058] Among them, end-to-end semantic communication (ESC) as a revolutionary wireless communication paradigm has great potential in high-demand applications that meet the above requirements. The principle behind it is that semantic communication essentially relies on deep neural networks (DNNs) to transmit only the semantics of data rather than the entire data to improve communication efficiency. For this purpose, a DNN-based semantic communication system has recently been proposed, which uses a neural structure similar to an autoencoder, using a semantic encoder at the sender to extract semantic representations, and then using a decoder at the receiver to reconstruct information.

[0059] ​Despite this, a major drawback of semantic communication systems is that they require a large amount of data to train millions of neural parameters. In practice, this data can contain sensitive information about end users, such as health conditions, political stance, and service access history, so training semantic communication systems on a centralized server will lead to some serious privacy issues.

[0060] On the other hand, the recent emergence of federated learning (FL) (i.e. instead of uploading data to the server for centralized training, each client trains the model using its own data, and finally aggregates the model parameters on the server side) and split learning (SL) (i.e. split the model, the client keeps a part of the model, and the server keeps a part of the model, and train together) provides some new ideas for semantic communication systems to solve the above problems: collaborative training of neural networks without showing exchange data samples and updating parameters across multiple decentralized devices.

[0061] Such learning paradigms do not upload users' local data, which can protect users' privacy. However, directly combining semantic communication systems with FL / SL can suffer from the following problems:

[0062] 1. Limited computing power of mobile devices: In general, a semantic communication system consists of multiple components, including a semantic encoder, a channel encoder / decoder, and a semantic decoder. Jointly training such an ESC system on mobile devices and then aggregating a global model on the server can be impractical in the real world, as the computing power of these devices is very limited.

[0063] 2. High training latency: Splitting the network into encoder and decoder parts can be more flexible. However, it introduces higher communication costs, as gradients need to be passed over the air when computing forward and backward.

[0064] To address the above problems, the present application proposes a distributed learning framework in the context of wireless network semantic communication, which has a flexible learning mechanism that can be flexibly adjusted according to the computing power of user devices.

[0065] Further, the following specific description of the semantic communication system proposed by the present application:

[0066] As shown in Figure 2 The semantic communication system proposed by the present application includes five modules, namely a semantic encoder module, an OFDM transmitter module, an OFDM receiver module, a semantic decoder module, and a classifier module. Among them,

[0067] Semantic encoder module: in one way, for image data, the semantic encoder module in the embodiments of the present application can be constructed based on a convolutional neural network model. In another way, for text data, the semantic encoder module in the embodiments of the present application can be based on a Transformer model. That is, the input data is represented as X, and X is input into the semantic encoder Semantic Encoder to obtain the semantic information of the data.

[0068] OFDM transmitter module: in one way, the embodiments of the present application can realize data transmission between each client and the server through an OFDM transmitter. Further, the OFDM transmitter module utilizes the fast Fourier transform (IFFT) of the semantic information and adds a cyclic prefix to reduce the peak-to-average power ratio (PAPR) through signal clipping. Here, a pilot is also added for channel estimation.

[0069] Wireless channel module: in one way, the wireless transmission channel in the embodiments of the present application adopts a Rayleigh fading channel. For example, by giving an input X o , the output of the wireless transmission channel X ′ o .

[0070] OFDM receiver module: in one way, the embodiments of the present application can realize data transmission between each client and the server through an OFDM receiver. Further, the OFDM receiver module utilizes the OFDM receiver to perform the inverse process used by the OFDM transmitter, removes the cyclic prefix, and converts the signal with fast Fourier transform (FFT) to produce frequency domain pilot and data symbols. Here, further combined with channel estimation and channel equalization to improve transmission performance.

[0071] Semantic decoder module: the semantic decoder corresponds to the semantic encoder one-to-one, and the purpose is to decode the semantic information transmitted on the channel into the original image or text data.

[0072] Classifier module: the classifier will accept the data decoded by the semantic decoder and classify it. In one way, the classifier in the embodiments of the present application can measure the semantic information involved in the input data (i.e. image or text).

[0073] Further, as Figure 3 shown, the flowchart of the generation method of the semantic communication system proposed in the present application includes:

[0074] Step 1, establishing, for an initial semantic communication system, a first training task for training a semantic encoder module, a second training task for training a semantic decoder module, a third training task for training a wireless transmission channel module, and a fourth training task for training a classifier.

[0075] Further, first of all, the present application needs to construct an end-to-end virtual semantic communication system model. In one way, the embodiments of the present application can use joint source channel coding (JSCC) for text or image data transmission over a multipath fading channel.

[0076] Step 2, obtaining attribute information sent by a plurality of clients, the attribute information including device type, computing hardware and service load condition of the clients.

[0077] Step 3, determining the computing performance corresponding to each client based on the attribute information of each client, and dividing the plurality of clients into a plurality of training object groups according to the computing performance.

[0078] Among them, the computing performance of the clients in each training object group is the same.

[0079] In one way, the present application can determine the computing performance corresponding to each client according to the attribute information fed back by each client to the server, and divide the clients belonging to the same computing performance into the same training object group.

[0080] In one way, the present application does not limit the number of clients included in each training object group, which can be the same or different.

[0081] Step 4, determining the computing performance index corresponding to each training task based on the training module corresponding to each training task.

[0082] Among them, since the training module can include a semantic encoder module, a semantic decoder module, a wireless transmission channel module and a classifier module, and for training each training module, the computing performance index required is not the same.

[0083] For example, for the semantic encoding and decoding module, it needs more training parameters relative to the wireless transmission channel module. Therefore, in one way, the computing performance index corresponding to the first training task for training the semantic encoder module and the second training task for training the semantic decoder module is higher than the computing performance index corresponding to the third training task for training the wireless transmission channel module.

[0084] In another way, the embodiments of the present application can also determine the corresponding computing performance index for each training module according to the training of each module of the current initial semantic communication system.

[0085] Step 5, match each training task with a training object group capable of meeting the computing performance index based on the computing performance corresponding to each training object group.

[0086] In one way, if the computing performance corresponding to a certain training object group is higher, then it can be assigned a larger number of training tasks, and / or training tasks that require higher computing performance indicators.

[0087] For example, if the computing performance of training object group A is higher, it can be assigned 2 or 3 training tasks out of a total of 4 training tasks. And the 2 or 3 training tasks also include at least one training task that requires higher computing performance indicators to process.

[0088] Step 6, determine the first training object group with the lowest computing performance from the plurality of training object groups.

[0089] Step 7, assign the first training task for extracting semantic information to the first training object group, so that the first training object group extracts semantic information from the training data.

[0090] In one way, if the computing performance corresponding to a certain training object group is lower, then it can be assigned a smaller number of training tasks, and / or training tasks that only extract semantic information from training data (i.e. image data or text data).

[0091] Step 8, assign each training task, semantic encoder, and OFDM transmitter to the corresponding training object group.

[0092] Step 9, the training object group trains the received training task using the semantic encoder and the local data.

[0093] Step 10, send the model update parameters to the initial semantic communication system using the OFDM transmitter.

[0094] Step 11, receive the model update parameters sent by each training object group respectively, and update the initial semantic communication system based on the model update parameters to obtain the trained target semantic communication system.

[0095] In this application, multiple training tasks can be established for the initial semantic communication system to be trained, wherein each training task is used to train at least one module in the initial semantic communication system; each training task is assigned to a corresponding training object group to train the received training task by the training object group, wherein each training object group includes multiple clients; receive the model update parameters sent by each training object group respectively, and update the initial semantic communication system based on the model update parameters to obtain the trained target semantic communication system.

[0096] By applying the technical solution of the present application, the complete task of training the semantic communication system can be divided into multiple sub-training tasks according to the modules. The multiple sub-training tasks are respectively distributed to the training object groups that have the ability to process. Each client in the training object group trains the sub-training task allocated to it, and sends the model update parameters of the corresponding module to the server after obtaining the model update parameters. The server updates each module of the semantic communication system according to the model update parameters of each module, and then obtains a trained target semantic communication system. Thus, on the one hand, the local privacy data of the client is not leaked. On the other hand, the performance consumption of the server caused by training is also saved.

[0097] Optionally, in another embodiment based on the above method of the present application, multiple training tasks are established for the initial semantic communication system to be trained, including:

[0098] The first training task for training the semantic encoder module, the second training task for training the semantic decoder module, the third training task for training the wireless transmission channel module, and the fourth training task for training the classifier are established for the initial semantic communication system.

[0099] Optionally, in another embodiment based on the above method of the present application, before the respective distribution of each training task to the corresponding training object group, it further includes:

[0100] Obtain attribute information sent by multiple clients, the attribute information including device type, computing hardware and service load condition of the client;

[0101] Based on the attribute information of each client, determine the computing performance corresponding to each client, and divide the multiple clients into multiple training object groups according to the computing performance, wherein the computing performance of the clients in each training object group is the same.

[0102] Optionally, in another embodiment based on the above method of the present application, after the multiple clients are divided into multiple training object groups, it further includes:

[0103] Based on the training module corresponding to each training task, determine the computing performance index corresponding to each training task;

[0104] Based on the computing performance corresponding to each training object group, match the training object group that can meet the computing performance index for each training task.

[0105] Optionally, in another embodiment based on the above method of the present application, after the multiple clients are divided into multiple training object groups, it further includes:

[0106] From the plurality of training object groups, determine the first training object group with the lowest computational performance;

[0107] A first training task for extracting semantic information is assigned to the first training object group so that the first training object group can extract semantic information from the training data.

[0108] Optionally, in another embodiment based on the method described above in this application, the step of distributing each training task to the corresponding training object group includes:

[0109] Each training task, semantic encoder, and OFDM transmitter are distributed to the corresponding training object group so that the training object group can use the semantic encoder to train the received training task; and the model update parameters are sent to the initial semantic communication system using the OFDM transmitter.

[0110] Optionally, in another embodiment based on the method described above, after distributing each training task to the corresponding training object group, the method further includes:

[0111] Each client in the training group uses local data to train the received training task.

[0112] Furthermore, such as Figure 4 The diagram shown is a system architecture diagram of the semantic communication system generation method proposed in this application, which includes multiple clients, namely M = {M1, M2, ..., M}. N Each client's local storage contains multiple sets of training data, i.e., D = {D1, D2, ..., D}. N Each client is equipped with a semantic encoder and an OFDM transmitter, which can generate semantic representations for the central server to interpret.

[0113] It should be noted that each client node can connect to a base station, which collaborates with a central base station in the cloud to perform large-scale computations via servers. Such a central server S coordinates the different steps of the model and all participating devices during the learning process.

[0114] In one embodiment of this application, N clients can be divided into three training object groups based on their computing capabilities. This includes:

[0115] Training object group 1: The application assumes that these devices are equipped with the most powerful computing platform, so the application deploys the ESC system on these devices and trains the end-to-end system under Rayleigh fading channel using local data instances. The application also sends semantic data output by the OFDM transmitter to the central server, which is provided to the OFDM receiver, decoder and classifier deployed on the central server. It should be noted that these semantic data do not need to be forward calculated and gradient related to reduce communication cost.

[0116] Training object group 2: The application assumes that these devices are not as powerful as those in group 1, and in group 1 the application only uses their local data to train the classifier of the ESC system.

[0117] Training object group 3: These devices have very limited computing power, so the application uses a pre-trained semantic encoder to generate semantic representations and sends this information to the central server to train the decoder and classifier.

[0118] Further, the specific training process is as shown in Figure 5 , which includes:

[0119] Step a, initialization:

[0120] The data collection for the initial semantic communication system training is at the local node and will not be globally shared. The central server also collects information about the devices, such as device type, computing hardware, etc., and then decides to deploy the training object group for each training task.

[0121] To initialize the global model, the central server can pre-train the initial semantic communication system, and then divide the initial semantic communication system into four modules, namely the semantic encoder, channel encoder / decoder, semantic decoder and classifier. These training tasks will be respectively assigned to the corresponding multiple training object groups.

[0122] Step b, training:

[0123] The clients of training object group 1 and training object group 2 train the neural model according to their local data, and then update the parameters respectively. Among them, the training of training object group 1 helps the initial semantic communication system to generate meaningful semantic representations, and then reconstruct the information transmitted on the Rayleigh fading channel. The input is a sentence or an image X, and the output is the reconstructed X'.

[0124] The clients of training object group 2 only train the classifier based on local data.

[0125] The clients in training object group 3 will only extract semantic representations and send them to the central server without updating the parameters. The central server can therefore also use these data to train the decoder.

[0126] Step c, aggregation:

[0127] The client of each training object group uploads the updated model weight parameters to the center server after K epochs. Then the center server updates the parameters of each module of the initial semantic communication system respectively to generate a new global model, i.e., a target semantic communication system.

[0128] Step d, client update:

[0129] The center server distributes the weight of each module to all clients and updates the neural parameters.

[0130] By applying the technical solution of the present application, the complete task of training a semantic communication system can be divided into multiple sub-training tasks according to each module. The multiple sub-training tasks are respectively distributed to training object groups that have the ability to train them. Each client in the training object group trains the sub-training task assigned to it based on its local data, so as to send the model update parameters of the corresponding module to the server after obtaining them. The server updates each module of the semantic communication system according to the model update parameters of each module, thereby obtaining a trained target semantic communication system. Thus, on the one hand, the local privacy data of the client is not leaked. On the other hand, the performance consumption of the server caused by training the communication system is also saved.

[0131] It can be understood that the unified distributed learning framework proposed in the present application can transmit and classify images and texts, and can collaboratively train an end-to-end semantic communication system at a lower communication cost. In addition, a new training scheme is introduced, which divides the semantic communication system into multiple parts and then trains these modules to greatly reduce the communication cost. The aggregation of client parameters can also be dynamically configured to improve training efficiency.

[0132] Optionally, in another embodiment of the present application, as shown in Figure 6 The present application also provides a semantic communication system generation device. It includes:

[0133] The creation module 201 is configured to establish multiple training tasks for the initial semantic communication system to be trained, wherein each training task is used to train at least one module in the initial semantic communication system;

[0134] The sending module 202 is configured to respectively distribute each training task to the corresponding training object group, so that the training object group trains the received training task, wherein each training object group includes multiple clients;

[0135] The updating module 203 is configured to receive the model updating parameters sent by each training object group respectively, and update the initial semantic communication system based on the model updating parameters to obtain a target semantic communication system after training.

[0136] By applying the technical solution of the present application, the complete task of training a semantic communication system can be divided into multiple sub-training tasks according to the modules. The multiple sub-training tasks are respectively distributed to training object groups that have the ability to train them. Each client in the training object group trains the sub-training task allocated to it based on its local data, so that the model updating parameters of the corresponding module are obtained and sent to the server. The server updates each module of the semantic communication system according to the model updating parameters of the module, thereby obtaining a target semantic communication system after training. Thus, on the one hand, the local privacy data of the client is not leaked. On the other hand, the performance consumption of the server caused by training the communication system is also saved.

[0137] In another embodiment of the present application, the creating module 201 is configured to perform the following steps:

[0138] The first training task for training the semantic encoder module, the second training task for training the semantic decoder module, the third training task for training the wireless transmission channel module, and the fourth training task for training the classifier are established for the initial semantic communication system.

[0139] In another embodiment of the present application, the creating module 201 is configured to perform the following steps:

[0140] Attribute information sent by multiple clients is obtained, and the attribute information includes the device type, computing hardware, and service load condition of the client.

[0141] Based on the attribute information of each client, the computing performance corresponding to each client is determined, and the multiple clients are divided into multiple training object groups according to the computing performance, wherein the computing performance of the clients in each training object group is the same.

[0142] In another embodiment of the present application, the creating module 201 is configured to perform the following steps:

[0143] Based on the training module corresponding to each training task, the computing performance index corresponding to each training task is determined.

[0144] Based on the computing performance corresponding to each training object group, each training task is matched with a training object group that can meet the computing performance index.

[0145] In another embodiment of this application, the steps of creating module 201, which is configured to be executed, include:

[0146] From the plurality of training object groups, determine the first training object group with the lowest computational performance;

[0147] A first training task for extracting semantic information is assigned to the first training object group so that the first training object group can extract semantic information from the training data.

[0148] In another embodiment of this application, the steps of creating module 201, which is configured to be executed, include:

[0149] Each training task, semantic encoder, and OFDM transmitter are distributed to the corresponding training object group so that the training object group can use the semantic encoder to train the received training task; and the model update parameters are sent to the initial semantic communication system using the OFDM transmitter.

[0150] In another embodiment of this application, the steps of creating module 201, which is configured to be executed, include:

[0151] Each client in the training group uses local data to train the received training task.

[0152] This application also provides an electronic device for executing the above-described method for generating a semantic communication system. Please refer to... Figure 7 This illustrates a schematic diagram of an electronic device provided by some embodiments of this application. For example... Figure 7 As shown, the electronic device 3 includes: a processor 300, a memory 301, a bus 302, and a communication interface 303. The processor 300, the communication interface 303, and the memory 301 are connected through the bus 302. The memory 301 stores a computer program that can run on the processor 300. When the processor 300 runs the computer program, it executes the semantic communication system generation method provided in any of the foregoing embodiments of this application.

[0153] The memory 301 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this device network element and at least one other network element is achieved through at least one communication interface 303 (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0154] The bus 302 can be an ISA bus, a PCI bus, an EISA bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, and the like. The memory 301 is configured to store programs, and the processor 300 executes the programs after receiving execution instructions. The data recognition method disclosed in any of the embodiments of the present application can be applied to the processor 300 or implemented by the processor 300.

[0155] The processor 300 can be an integrated circuit chip with processing capability. In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in the processor 300 or an instruction in the form of software. The processor 300 described above can be a general processor, including a processor (CPU), a network processor (NP), and the like; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a ready programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. Each method, step and logic block diagram disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware coding processor for execution, or a combination of hardware and software modules in the coding processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, or other mature storage media in the art. The storage medium is located in the memory 301, and the processor 300 reads the information in the memory 301 and combines the hardware to complete the steps of the above method.

[0156] The electronic device provided by the embodiments of the present application and the method for generating a semantic communication system provided by the embodiments of the present application have the same beneficial effects as the method they adopt, run or implement.

[0157] The embodiments of the present application also provide a computer readable storage medium corresponding to the method for generating a semantic communication system provided in the foregoing embodiments. Please refer to Figure 8 The computer readable storage medium shown in the figure is an optical disc 40, and a computer program (i.e. a program product) is stored on the optical disc 40. When the computer program is run by a processor, the method for generating a semantic communication system provided in any of the foregoing embodiments is executed.

[0158] It is to be noted that examples of the computer-readable storage medium can also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical, magnetic storage mediums, and the like, which are not listed one by one here.

[0159] The computer-readable storage medium provided by the above embodiments of the present application has the same beneficial effects as the method adopted, run or implemented by the application program stored therein.

[0160] It should be noted that:

[0161] In the specification provided herein, a large number of specific details are described. However, it can be understood that the embodiments of the present application can be practiced without these specific details. In some examples, well-known structures and techniques are not shown in detail in order not to obscure the understanding of the present specification.

[0162] Similarly, it is to be understood that, in the above description of exemplary embodiments of the present application, various features of the present application are sometimes grouped together in a single embodiment, figure, or description of a related aspect. This method of disclosure, however, is not to be interpreted as reflecting an intention that the claimed application requires more features than are explicitly recited in each claim. Rather, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following, in this application are hereby expressly incorporated into this detailed description, with each claim acting as a separate embodiment of the present application. The claims are not to be construed as reflecting an intention that the application requires more features than are explicitly recited in each claim.

[0163] Furthermore, those skilled in the art will recognize that references in the specification to "one embodiment", "an embodiment", "an example embodiment", means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment.

[0164] The above descriptions are only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any changes or substitutions easily conceived by those skilled in the art within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method of generating a semantic communication system, characterized by, The method comprises the following steps: establishing a plurality of training tasks for an initial semantic communication system to be trained, wherein each training task is used for training at least one module in the initial semantic communication system; respectively distributing each training task to a corresponding training object group, so that the training object group trains the received training task, wherein each training object group comprises a plurality of clients; the step of respectively distributing each training task to a corresponding training object group comprises: distributing each training task, a semantic encoder and an OFDM transmitter to a corresponding training object group, so that the training object group trains the received training task by using the semantic encoder, and sends model update parameters to the initial semantic communication system by using the OFDM transmitter; respectively receiving the model update parameters sent by each training object group, and updating the initial semantic communication system based on the model update parameters to obtain a target semantic communication system after training.

2. The method of claim 1, wherein, The step of establishing a plurality of training tasks for the initial semantic communication system to be trained comprises: establishing a first training task for training a semantic encoder module, a second training task for training a semantic decoder module, a third training task for training a wireless transmission channel module, and a fourth training task for training a classifier for the initial semantic communication system.

3. The method of claim 1, wherein, Before the step of respectively distributing each training task to a corresponding training object group, the method further comprises: obtaining attribute information sent by a plurality of clients, wherein the attribute information comprises device types, computing hardware and service load conditions of the clients; determining computing performance corresponding to each client based on the attribute information of each client, and dividing the plurality of clients into a plurality of training object groups according to the computing performance, wherein the computing performance of the clients in each training object group is the same.

4. The method of claim 3, wherein, After the step of dividing the plurality of clients into a plurality of training object groups, the method further comprises: determining a computing performance index corresponding to each training task based on the training module corresponding to each training task; matching a training object group capable of meeting the computing performance index for each training task based on the computing performance corresponding to each training object group.

5. The method of claim 3, wherein, After the step of dividing the plurality of clients into a plurality of training object groups, the method further comprises: determining a first training object group with the lowest computing performance from the plurality of training object groups; distributing a first training task for extracting semantic information to the first training object group, so that the first training object group extracts semantic information in training data.

6. The method of claim 1, wherein, After the step of respectively distributing each training task to a corresponding training object group, the method further comprises: training the received training task by each client in the training object group by using local data.

7. A generating apparatus of a semantic communication system implemented using the method of any one of claims 1 to 6, characterized by, The method comprises the following steps: a creating module is configured to establish a plurality of training tasks for an initial semantic communication system to be trained, wherein each training task is used for training at least one module in the initial semantic communication system; a sending module is configured to respectively distribute each training task to a corresponding training object group, so that the training object group trains the received training task, wherein each training object group comprises a plurality of clients; the sending module is configured to distribute each training task, a semantic encoder and an OFDM transmitter to a corresponding training object group, so that the training object group trains the received training task by using the semantic encoder, and sends model update parameters to the initial semantic communication system by using the OFDM transmitter; An updating module configured to receive model updating parameters sent by each training object group respectively, and update the initial semantic communication system based on the model updating parameters to obtain a target semantic communication system after training.

8. An electronic device, comprising: Comprising: a memory for storing executable instructions; and, a processor for executing the executable instructions with the memory to complete the operation of the generation method of the semantic communication system in any one of claims 1-6.

9. A computer readable storage medium for storing computer readable instructions, characterized in that, The instructions are executed to perform the operation of the generation method of the semantic communication system in any one of claims 1-6. The instructions are executed to perform the operation of the generation method of the semantic communication system in any one of claims 1-6.

Images