Model training reasoning method and device, equipment, medium and program product

Abstract

The invention provides a model training reasoning method and device, equipment, a medium and a program product, and relates to the technical field of artificial intelligence. The method comprises the following steps: receiving a data processing task sent by a demand end; determining a target execution end from a plurality of execution ends according to the data processing task; sending the data processing task to a target execution end; receiving a data processing result sent by the target execution end, wherein the data processing result is obtained by executing the data processing task by the target execution end; and sending the data processing result to the demand end. According to the model training reasoning method provided by the invention, by dynamically distributing dispersed processor resources, social idle processor resources are effectively integrated and utilized, so that the model training reasoning task can be efficiently completed on a proper execution end. The resource utilization rate and the task execution efficiency are improved, the dependence on a traditional data center and a large-scale server cluster is reduced, and a computing power solution which is lower in cost and higher in expansibility is provided.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a model training reasoning method, device, equipment, medium and program product. Background Art

[0002] Model training and reasoning are the core processes of machine learning and deep learning. Model training is the process of learning patterns and rules from a large amount of data through algorithms, aiming to optimize model parameters to improve its performance on specific tasks; while model reasoning is to use the trained model to predict or classify new data. It is widely used in image recognition, natural language processing, speech recognition and other fields.

[0003] In the existing technology, model training and reasoning mainly rely on centralized data centers and large processor clusters. These facilities integrate a large number of high-performance computing devices and storage systems to provide powerful computing capabilities and massive data storage space to support large-scale data processing needs and the development process of complex models. In these data centers, model training and reasoning tasks can use parallel computing and distributed storage technologies to achieve efficient processing of massive data and rapid iteration of complex algorithms.

[0004] However, in the prior art, the processor resource allocation efficiency is low. On the demand side, multiple data processing tasks may compete for limited processor resources. On the execution side, the processor is idle during non-working hours. The low efficiency of processor resource allocation makes it impossible to accurately allocate data processing tasks to idle processors, which limits the scalability and flexibility of the system and makes it difficult to adapt to dynamically changing work requirements. Summary of the invention

[0005] The present application provides a model training reasoning method, device, equipment, medium and program product to solve the problem of low processor resource allocation efficiency of the model training reasoning method in the prior art.

[0006] In a first aspect, the present application provides a model training and reasoning method, which is applied to a central end of a model training and reasoning system, wherein the model training and reasoning system further includes a demand end and multiple execution ends, and the method includes:

[0007] Receiving a data processing task sent by the demand side, where the data processing task is created by the demand side according to a preset data set;

[0008] Determining a target execution end from the multiple execution ends according to the data processing task;

[0009] Sending the data processing task to the target execution end so that the target execution end executes the data processing task;

[0010] receiving a data processing result sent by the target execution end, wherein the data processing result is obtained by the target execution end executing the data processing task;

[0011] The data processing result is sent to the demand end.

[0012] In a possible design, determining a target execution end from the multiple execution ends according to the data processing task includes:

[0013] Obtaining bandwidth requirements, processor core count requirements, and memory requirements required to perform the data processing task;

[0014] Generate a first characteristic matrix of the data processing task according to the bandwidth requirement, the processor core number requirement and the memory requirement;

[0015] Calculating the similarity between the first feature matrix and each preset second feature matrix; wherein the second feature matrix refers to a matrix composed of bandwidth, number of processor cores and memory of the execution end in an idle state;

[0016] The execution end corresponding to the second feature matrix with the greatest similarity is determined as the target execution end.

[0017] In a possible design, calculating the similarity between the first feature matrix and each preset second feature matrix includes:

[0018] Acquire the utilization rate of the processor of each execution end, and determine each execution end whose utilization rate is less than a preset threshold as an execution end in an idle state;

[0019] The similarities between the first characteristic matrix and the second characteristic matrix of the execution end in the idle state are calculated.

[0020] In a second aspect, the present application provides a model training and reasoning method, which is applied to the demand side of a model training and reasoning system, wherein the model training and reasoning system further includes a central side and multiple execution sides, and the method includes:

[0021] Create data processing tasks based on preset data sets;

[0022] Sending the data processing task to the central end; wherein the central end is used to determine a target execution end from the multiple execution ends according to the data processing task, and the target execution end is used to execute the data processing task;

[0023] Receive the data processing result sent by the central end; wherein the data processing result is obtained by the target execution end executing the data processing task and sent to the central end.

[0024] In a possible design, the preset data set is a training data set, and the data processing task is a training task;

[0025] Then, creating a data processing task according to the preset data set includes:

[0026] Package the preset running image, training code, startup script, and input and output parameters to obtain a training package;

[0027] The training task is created according to the training package, the training data set, the preset training parameters and the preset training resource configuration.

[0028] In a possible design, the preset data set is an inference data set, and the data processing task is an inference task;

[0029] Then, creating a data processing task according to the preset data set includes:

[0030] Creating an inference program according to the training model and the preset inference code;

[0031] The reasoning task is created according to the reasoning program, the reasoning data set and the preset reasoning resource configuration.

[0032] In a possible design, the data processing result is a target port, and after receiving the data processing result sent by the central end, the method further includes:

[0033] Accessing the target execution end through the target port;

[0034] Receive access data sent by the target execution end.

[0035] In a third aspect, the present application provides a model training reasoning system, including: a central end, a demand end, and multiple execution ends;

[0036] The central end is used to implement the model training and reasoning method described in any one of the first aspects;

[0037] The demand side is used to implement the model training and reasoning method described in any one of the second aspects;

[0038] The multiple execution ends include a target execution end, and the target execution end is used to execute a training task or an inference task.

[0039] In a fourth aspect, the present application provides a model training and reasoning device, which is applied to a central end of a model training and reasoning system, wherein the model training and reasoning system further includes a demand end and multiple execution ends, and the device includes:

[0040] A first receiving module, configured to receive a data processing task sent by the demand side, wherein the data processing task is created by the demand side according to a preset data set;

[0041] A determination module, configured to determine a target execution end from the multiple execution ends according to the data processing task;

[0042] A first sending module, used for sending the data processing task to the target execution end, so that the target execution end executes the data processing task;

[0043] A second receiving module is used to receive a data processing result sent by the target execution end, where the data processing result is obtained by the target execution end executing the data processing task;

[0044] The second sending module is used to send the data processing result to the demand end.

[0045] In one possible design, the determining module includes:

[0046] An acquisition unit, used to acquire bandwidth requirements, processor core number requirements and memory requirements required to perform the data processing task;

[0047] A matrix generating unit, configured to generate a first characteristic matrix of the data processing task according to the bandwidth requirement, the processor core number requirement and the memory requirement;

[0048] A calculation unit, used to calculate the similarity between the first characteristic matrix and each preset second characteristic matrix; wherein the second characteristic matrix refers to a matrix composed of the bandwidth, the number of processor cores and the memory of the execution end in an idle state;

[0049] The determination unit is used to determine the execution end corresponding to the second feature matrix with the largest similarity as the target execution end.

[0050] In a possible design, the computing unit includes:

[0051] An acquisition component, used for acquiring the utilization rate of the processor of each execution end, and determining each execution end whose utilization rate is less than a preset threshold as an execution end in an idle state;

[0052] The calculation component is used to calculate the similarity between the first characteristic matrix and the second characteristic matrix of the execution end in the idle state.

[0053] In a fifth aspect, the present application provides a model training and reasoning device, which is applied to the demand side of the model training and reasoning system, and the model training and reasoning system also includes a central end and multiple execution ends, and the device includes:

[0054] A first creation module is used to create a data processing task according to a preset data set;

[0055] A third sending module is used to send the data processing task to the central end; wherein the central end is used to determine a target execution end from the multiple execution ends according to the data processing task, and the target execution end is used to execute the data processing task;

[0056] The third receiving module is used to receive the data processing result sent by the central end; wherein the data processing result is obtained by the target execution end executing the data processing task and sent to the central end.

[0057] In a possible design, the preset data set is a training data set, and the data processing task is a training task;

[0058] Then the first creation module includes:

[0059] A packaging unit is used to package the preset running image, training code, startup script and input and output parameters to obtain a training package;

[0060] The training task creation unit is used to create the training task according to the training package, the training data set, preset training parameters and preset training resource configuration.

[0061] In a possible design, the preset data set is an inference data set, and the data processing task is an inference task;

[0062] Then the first creation module includes:

[0063] An inference program creation unit, used to create an inference program according to the training model and a preset inference code;

[0064] The reasoning task creation unit is used to create the reasoning task according to the reasoning program, the reasoning data set and the preset reasoning resource configuration.

[0065] In a possible design, the data processing result is a target port, and the model training and reasoning device further includes:

[0066] An access module, used for accessing the target execution end through the target port;

[0067] The fourth receiving module is used to receive the access data sent by the target execution end.

[0068] In a sixth aspect, the present application provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor;

[0069] The memory stores computer-executable instructions;

[0070] When the processor executes the computer-executable instructions stored in the memory, it is used to implement the model training and reasoning method as described in any one of the first aspects, or to implement the model training and reasoning method as described in any one of the second aspects.

[0071] In the seventh aspect, the present application provides a computer-readable storage medium, which stores computer execution instructions. When the computer execution instructions are executed by a processor, they are used to implement the model training and reasoning method as described in any one of the first aspect, or to implement the model training and reasoning method as described in any one of the second aspect.

[0072] In an eighth aspect, the present application provides a computer program product, comprising a computer program, which, when executed by a processor, is used to implement the model training and reasoning method as described in any one of the first aspects, or to implement the model training and reasoning method as described in any one of the second aspects.

[0073] The model training reasoning method, device, equipment, medium and program product provided by the present application include: receiving a data processing task sent by the demand end, the data processing task is created by the demand end according to a preset data set; according to the data processing task, determining a target execution end from the multiple execution ends; sending the data processing task to the target execution end so that the target execution end executes the data processing task; receiving a data processing result sent by the target execution end, the data processing result is obtained by the target execution end executing the data processing task; sending the data processing result to the demand end. The model training reasoning method of the present application improves resource utilization efficiency and reduces computing costs by dynamically allocating dispersed processor resources, and provides an efficient computing power management and execution mechanism for model training and reasoning. The method can effectively integrate and utilize idle processor resources in society, so that model training and reasoning tasks can be efficiently completed on a suitable execution end. This not only improves the utilization rate of processor resources and the efficiency of task execution, but also reduces the dependence on traditional data centers and large server clusters, providing a lower cost and more scalable computing power solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0075] Figure 1A schematic diagram of an application scenario of the model training reasoning method provided in an embodiment of the present application;

[0076] Figure 2 Schematic diagram of the process of model training and reasoning method provided in the embodiment of the present application Figure 1 ;

[0077] Figure 3 Schematic diagram of the process of model training and reasoning method provided in the embodiment of the present application Figure 2 ;

[0078] Figure 4 Schematic diagram of the process of model training and reasoning method provided in the embodiment of the present application Figure 3 ;

[0079] Figure 5 A timing diagram of the model training reasoning method provided in an embodiment of the present application;

[0080] Figure 6 A schematic diagram of the structure of the model training and reasoning device provided in the embodiment of the present application Figure 1 ;

[0081] Figure 7 A schematic diagram of the structure of the model training and reasoning device provided in the embodiment of the present application Figure 2 ;

[0082] Figure 8 A schematic diagram of the structure of the model training and reasoning device provided in the embodiment of the present application Figure 3 ;

[0083] Fig. 9 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0084] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0085] In the embodiments of the present application, words such as "first" and "second" are used to distinguish between identical or similar items with substantially the same functions and effects. Those skilled in the art will understand that words such as "first" and "second" do not limit the quantity and execution order, and words such as "first" and "second" do not necessarily limit the difference. It should be noted that in the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "exemplary" or "for example" in this application should not be interpreted as being more preferred or more advantageous than other embodiments or design. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way. In the embodiments of the present application, "at least one" refers to one or more, and "more" refers to two or more.

[0086] It should be noted that the "at..." in the embodiments of the present application can be the instant when a certain situation occurs, or it can be a period of time after a certain situation occurs, and the embodiments of the present application do not specifically limit this. In addition, the model training reasoning method and the use method provided in the embodiments of the present application are only used as examples, and the model training reasoning method and the use method can also include more or less content.

[0087] In order to clearly describe the technical solutions of the embodiments of the present application, some terms and technologies involved in the embodiments of the present application are briefly introduced below:

[0088] Model training and reasoning: In the field of artificial intelligence and machine learning, model training and model reasoning are collectively referred to as the two key processes. Model training is to adjust the model parameters by inputting a large amount of data so that it can learn and identify patterns or features in the data; while model reasoning is the process of using the trained model to predict or classify new data. The training phase focuses on the learning ability and accuracy of the model, while the reasoning phase emphasizes the execution efficiency and real-time performance of the model.

[0089] Computing power: refers to computing power or computing resources, which is used to measure the ability of computing devices, such as processors, computers, servers or data centers, to process data and perform computing tasks in a unit of time. The strength of computing power directly affects the efficiency and speed of the system in performing complex calculations, processing large-scale data and running artificial intelligence algorithms.

[0090] Running image: It is an executable software package that contains an operating system, application, dependent libraries, and configuration files. It provides an independent running environment so that the application can run on any system that supports containerization technology without reconfiguring the environment.

[0091] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0092] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0093] In order to clearly understand the technical solution of the present application, the solution of the prior art is first introduced in detail. The model is usually untrained when it is just built, which means that its parameters are randomly initialized and lack understanding and adaptability to specific tasks or data sets. Therefore, the untrained model cannot make accurate predictions or decisions when faced with actual data. Based on this problem, it is necessary to train the model. Through the training process, the model can adjust its parameters to learn patterns and features in the data. Subsequently, through the reasoning process, the trained model can be applied to new data to achieve intelligent prediction and decision-making. Therefore, it is necessary to train and reason the model.

[0094] Current model training and inference methods mainly rely on centralized data centers and large processor clusters, which support model development and deployment by providing powerful computing power. However, this centralized approach has the problem of inefficiency in processor resource allocation: on the demand side, multiple data processing tasks usually need to be performed simultaneously, causing them to compete for limited processor resources, which may cause performance bottlenecks and delays; on the execution side, processor resources are often idle during non-working hours or low-load periods, and are not fully utilized, resulting in resource waste. This unbalanced resource allocation limits the popularity and efficiency improvement of model training and inference technology, and there is an urgent need to improve the utilization of processor resources through innovative resource integration and management solutions. Therefore, the current model training and inference methods have the problem of inefficient processor resource allocation.

[0095] Therefore, in view of the problem of inefficient processor resource allocation in the model training and inference methods in the prior art, the research found that in order to solve this problem, it is necessary to adjust the processor allocation strategy according to different training and inference tasks: ① Match the training and inference tasks to the most suitable processor by matching the resources required to execute the training and inference tasks with the execution resources available at different execution ends; ② During the non-working hours of the execution end, for example, when the user is not at home, the devices equipped with processors at home are idle. A system can be built to connect the idle processors to the system. The demand end can use the idle processors at the execution end to perform model training and inference, which can reduce the pressure on the processor on the demand end and the waste of processor resources on the execution end.

[0096] Specific:

[0097] The network resources required to execute model training and reasoning tasks and the configuration of each processor on the execution end can be obtained, and the network resources required to execute model training and reasoning tasks are matched with the configuration of each processor on the execution end to obtain the most suitable execution end processor, and use this processor to execute model training and reasoning tasks.

[0098] The model training and reasoning method of the embodiment of the present application matches the most suitable execution end for each model training and reasoning task, solving the problem of inefficient processor resource allocation in the current model training and reasoning method. By dynamically selecting the most suitable execution end to process a specific task, it ensures that the data processing task on the demand side matches the resource capacity of the execution end, thereby maximizing resource utilization, reducing resource contention on the demand side and resource idleness on the execution side, and improving processor resource allocation efficiency.

[0099] Based on the above creative findings, the technical solution of the present application is proposed.

[0100] The following introduces the application scenarios of the model training and reasoning method provided by the embodiments of the present invention. Figure 1 Schematic diagram of application scenarios of the model training and reasoning method provided in the embodiment of the present application. Figure 1 As shown, the application scenario includes a demand end 101, a central end 102, a first execution end 103, and a second execution end 104. The central end 102 receives the model training reasoning task sent by the demand end 101, selects a target execution end from the first execution end 103 and the second execution end 104, and sends the model training task to the target execution end. After the target execution end executes the model training reasoning task, it sends the model training reasoning result to the demand end 101 through the central end 102.

[0101] The embodiments of the present invention are described below in conjunction with the accompanying drawings.

[0102] Figure 2Schematic diagram of the process of model training and reasoning method provided in the embodiment of the present application Figure 1 .like Figure 2 As shown, in this embodiment, the execution subject of the embodiment of the present invention is the central end. Then the model training and reasoning method provided in this embodiment includes the following steps:

[0103] S201. Receive a data processing task sent by a demand side, where the data processing task is created by the demand side based on a preset data set.

[0104] Specifically, the demand side can send the data processing tasks it creates to the central side through the network communication interface, so that the central side can coordinate and allocate appropriate execution ends to complete the tasks according to the data processing tasks, thereby achieving effective utilization of resources and efficient execution of tasks.

[0105] S202: Determine a target execution end from multiple execution ends according to the data processing task.

[0106] Specifically, load balancing strategies can be used to determine the target executor from multiple executors. These algorithms may consider a variety of factors, including the computing requirements of the task, the current load of the executor, network latency, the processing power of the executor, and the priority of the task. By comprehensively analyzing these factors, the system can dynamically select the executor that best suits the current task. Doing so can improve the overall efficiency and responsiveness of the system, optimize resource utilization, reduce task processing time, and improve the scalability and stability of the system, thereby enhancing user experience and system reliability.

[0107] S203: Send the data processing task to the target execution end so that the target execution end executes the data processing task.

[0108] Specifically, data processing tasks can be sent to the target execution end through a network communication protocol so that the target execution end can execute the data processing tasks. The central end packages the data processing tasks into an appropriate format and sends them to the selected target execution end through a secure and reliable communication channel. This process may involve serializing task data, encrypting transmission content to ensure data security, and using confirmation mechanisms to ensure successful receipt and processing of tasks. Doing so can achieve efficient distribution and execution of tasks, ensuring that data processing tasks can reach the target execution end accurately, thereby reducing delays and errors in task transmission and improving system reliability and throughput.

[0109] S204: Receive a data processing result sent by the target execution end, where the data processing result is obtained by the target execution end executing the data processing task.

[0110] Specifically, the data processing results sent by the target execution end can be received through the network communication mechanism. After completing the task, the target execution end packages the result data and sends it back to the central end through the communication protocol. The central end needs to have the ability to process the returned data, which may involve decoding the data, verifying the integrity and correctness, and storing or further processing the results. Doing so can ensure that the data processing results can be reliably transmitted back to the central end, support real-time data updates and decision-making, enhance the responsiveness and accuracy of the system, and provide timely feedback and result delivery to the demand side.

[0111] S205: Send the data processing result to the demand side.

[0112] Specifically, after receiving the data processing results from the target execution end, the center will package the result data into an appropriate format and send it to the demand end through a secure and reliable communication channel. This process may involve serialization and encrypted transmission of the result data to ensure data security, as well as the use of a confirmation mechanism to ensure that the demand end successfully receives the data. Doing so can achieve efficient and accurate delivery of data processing results, enabling the demand end to obtain processing results in a timely manner, thereby supporting rapid decision-making and subsequent operations, and improving the overall response speed of the system and user experience.

[0113] A model training reasoning method provided in this embodiment includes: receiving a data processing task sent by a demand end, the data processing task is created by the demand end according to a preset data set; determining a target execution end from multiple execution ends according to the data processing task; sending the data processing task to the target execution end so that the target execution end executes the data processing task; receiving the data processing result sent by the target execution end, the data processing result is obtained by the target execution end executing the data processing task; sending the data processing result to the demand end. A model training reasoning method achieves the following technical effects: by receiving the data processing task of the demand end at the central end, and intelligently selecting the most suitable execution end for processing according to the characteristics and resource requirements of the task. This process involves dynamic matching of the resources required for the task and the available resources of each execution end, so as to assign the task to the most suitable processor. This method can effectively allocate processor resources for the demand end, avoid resource competition and performance bottlenecks, and make full use of idle resources at the execution end to reduce waste. Ultimately, this method improves the utilization of processor resources, optimizes the efficiency and speed of task processing, and provides a computing power solution with lower cost and stronger scalability.

[0114] In a possible design, S202, according to the data processing task, determining a target execution end from multiple execution ends includes:

[0115] S2021. Obtain the bandwidth requirements, processor core number requirements, and memory requirements required to perform data processing tasks.

[0116] Specifically, the bandwidth requirements, processor core count requirements, and memory requirements required to perform data processing tasks can be obtained by analyzing the specific requirements and characteristics of the data processing tasks. This usually involves calculating the computational complexity, data volume, real-time requirements, etc. of the task to determine the system resources required for the task to be executed. It can ensure that when selecting the target execution end, it can be matched according to the actual resource requirements of the task, thereby improving resource utilization efficiency, reducing task execution time, and ensuring that the task can be successfully completed on the appropriate execution end. By generating a first characteristic matrix corresponding to the task requirements, the most suitable execution end can be found among multiple execution ends, thereby optimizing the overall performance of the system.

[0117] S2022. Generate a first characteristic matrix of the data processing task according to bandwidth requirements, processor core number requirements, and memory requirements.

[0118] Specifically, the bandwidth requirements, processor core count requirements, and memory requirements can be used as different dimensions or elements of the matrix, and combined into a unified feature representation. This matrix representation facilitates subsequent similarity calculations. By converting the resource requirements of the data processing task into a standardized feature matrix, it can be directly compared with the resource characteristics of the execution end. This method helps to accurately select the execution end that best matches the current task requirements among multiple execution ends, thereby improving the overall efficiency and resource utilization of the system.

[0119] S2023. Calculate the similarity between the first characteristic matrix and each preset second characteristic matrix; wherein the second characteristic matrix refers to a matrix composed of the bandwidth, the number of processor cores and the memory of the execution end in the idle state.

[0120] Specifically, the similarity between the first feature matrix and each preset second feature matrix can be calculated by using a similarity measurement method. First, the first feature matrix and each second feature matrix are standardized, and then the selected similarity measurement method is applied to calculate the similarity score between them. The degree of resource matching between the data processing task and each execution end is evaluated by a quantified similarity score. The higher the similarity score, the more the resource configuration of the execution end meets the requirements of the task. In this way, the target execution end that is most suitable for executing the task can be effectively identified, thereby optimizing resource allocation, improving the efficiency of task processing and the overall performance of the system.

[0121] S2024. Determine the execution end corresponding to the second feature matrix with the greatest similarity as the target execution end.

[0122] Specifically, the target execution end can be determined by traversing all calculated similarity scores and identifying the execution end corresponding to the score with the largest value. For each execution end, the similarity score between it and the data processing task is calculated, and these scores are compared to find the execution end with the highest score, and mark it as the target execution end. By selecting the execution end with the highest similarity score, it is ensured that the selected execution end best meets the needs of the current data processing task in terms of resource allocation. This selection method can maximize the use of the available resources of the execution end, reduce resource waste, and improve the efficiency and speed of task execution. Ultimately, this optimization selection helps to improve the performance and responsiveness of the entire model training and inference system.

[0123] The technical effect of the scheme in this embodiment is: by introducing the feature matrix and similarity calculation method, intelligent matching and selection of execution end resources are achieved. This method can generate a standardized feature matrix based on the specific resource requirements of the data processing task, and compare the similarity with the resource feature matrix of each execution end, so as to identify the most suitable execution end. This method effectively improves the accuracy and efficiency of resource allocation, reduces performance bottlenecks and task delays caused by resource mismatch, and ultimately improves the overall processing capacity and response speed of the model training and reasoning system.

[0124] In a possible design, S2023, calculating the similarity between the first feature matrix and each preset second feature matrix, includes:

[0125] S20231. Obtain the utilization rate of the processor of each execution end, and determine each execution end whose utilization rate is less than a preset threshold as an execution end in an idle state.

[0126] Specifically, the processor utilization of each execution end can be obtained by collecting the processor usage of each execution end in real time through the monitoring system. The system can periodically or continuously obtain the CPU utilization data of each execution end and compare these data with the preset utilization threshold. If the processor utilization of a certain execution end is lower than the threshold, the execution end is marked as idle. By identifying and utilizing the execution ends in the idle state, computing resources can be effectively allocated to ensure that data processing tasks are assigned to execution ends with sufficient resources for processing, thereby improving the efficiency and response speed of the entire system.

[0127] S20232. Calculate the similarity between the first characteristic matrix and the second characteristic matrix of the execution end in the idle state.

[0128] Specifically, the similarity between the first feature matrix and the second feature matrix of the idle execution end can be calculated by using a vector similarity calculation method, such as cosine similarity or Euclidean distance. The first feature matrix represents the resource requirements of the data processing task, while the second feature matrix represents the available resources of the idle execution end. By calculating the similarity between the two matrices, the degree of match between the task requirements and the available resources of the execution end can be quantified. The higher the similarity, the more suitable the execution end is for executing the task. By selecting the execution end with the highest similarity as the target execution end, resource allocation can be optimized to ensure that the task runs on the most suitable execution end, thereby improving the overall performance and resource utilization efficiency of the system.

[0129] The technical effect of the scheme in this embodiment is: by obtaining the processor utilization of each execution end and marking the execution end below the preset threshold as idle, the system can effectively screen out the execution end with available resources. Then, by calculating the similarity between the feature matrix of the data processing task and the feature matrix of these idle execution ends, the system can select the most suitable execution end to process the task. This method not only improves resource utilization, but also reduces task waiting time, thereby enhancing the overall efficiency and responsiveness of the system.

[0130] Figure 3 Schematic diagram of the process of model training and reasoning method provided in the embodiment of the present application Figure 2 .like Figure 3 As shown, in this embodiment, the execution subject of the embodiment of the present invention is the demand side. The model training and reasoning method provided in this embodiment includes the following steps:

[0131] S301. Create a data processing task according to a preset data set.

[0132] Specifically, first analyze the structure and content of the preset data set to determine the type, format, and scale of the data. Then, define the goals and requirements of data processing, such as data cleaning, feature extraction, or data enhancement. Secondly, based on these goals and requirements, design specific data processing processes and algorithms to form a description of the data processing tasks. Finally, convert these descriptions into executable task formats so that they can be delivered and executed in the system. Doing so ensures that the data processing tasks can accurately reflect the needs of model training and reasoning, thereby improving the efficiency and effectiveness of data processing. Through clear data processing task descriptions, the automatic distribution and execution of tasks can be achieved in the system, reducing human intervention and errors, ensuring the consistency and reliability of data processing, and helping to optimize resource utilization and improve overall system performance.

[0133] S302, sending a data processing task to a central end; wherein the central end is used to determine a target execution end from a plurality of execution ends according to the data processing task, and the target execution end is used to execute the data processing task.

[0134] Specifically, the data processing task can be formatted into a message or request that conforms to the communication protocol, and then the message is sent to the designated receiving address or interface of the central end through the network communication interface. Make sure that the message contains the necessary task description and metadata so that the central end can correctly parse and process the task request. This can achieve centralized management of tasks. The central end is responsible for receiving and parsing task requests and selecting the appropriate execution end for task processing based on system resources and policies. This architecture helps optimize resource allocation, improve system flexibility and scalability, and simplify direct interaction between the demand end and the execution end, reducing complexity.

[0135] S303, receiving the data processing result sent by the central end; wherein the data processing result is obtained by the target execution end executing the data processing task and sent to the central end.

[0136] Specifically, the demand side can set a callback function to wait for the result message from the center side. Then, after completing the data processing task, the center side sends the result to the specified interface of the demand side through the network communication protocol. After receiving the result message, the demand side parses the data processing result and performs subsequent processing or storage. This can achieve timely acquisition and utilization of data processing results, ensuring that the demand side can quickly obtain the required information after the task is completed, thereby supporting subsequent decisions or operations. This mechanism improves the response speed and efficiency of the system, and at the same time ensures the reliability and consistency of result transmission through coordination at the center side.

[0137] The technical effect of the solution in this embodiment is that this method can effectively allocate processor resources to the demand side, avoid resource competition and performance bottlenecks, and make full use of idle resources on the execution side to reduce waste. Ultimately, this method improves the utilization of processor resources, optimizes the efficiency and speed of task processing, and provides a computing power solution with lower cost and greater scalability.

[0138] In a possible design, the preset data set is a training data set, and the data processing task is a training task. Then, S301, creating a data processing task according to the preset data set includes:

[0139] S3011. Package the preset running image, training code, startup script and input and output parameters to obtain a training package.

[0140] Specifically, the preset running image, training code, startup script, and input and output parameters can be packaged using containerization technology to obtain a training package. All necessary components required for training, such as the running image, training algorithm code, startup script, and input and output parameters, can be packaged into an independent container image. This ensures that the training tasks are highly portable and consistent between different execution ends. Through containerization, the training package can be executed in any environment that supports container operation without worrying about the differences in the underlying environment, thereby improving the deployment efficiency and reliability of the training tasks.

[0141] S3012: Create a training task according to the training package, the training data set, the preset training parameters and the preset training resource configuration.

[0142] Specifically, it can receive inputs of training packages, training data sets, training parameters, and resource configurations, and combine this information to generate a complete training task description. This description includes all the information required for execution, such as the container image used, the data set path, hyperparameter settings, and computing resource requirements. The central end then assigns the task to the appropriate execution end for processing. This implementation method can automate and standardize the creation and allocation process of training tasks to ensure efficient execution of tasks on different computing resources. At the same time, it allows flexible adjustment of training parameters and resource configuration to optimize the performance and cost of model training. In this way, users can manage and execute complex training tasks more conveniently, improving the scalability and responsiveness of the overall system.

[0143] The technical effect of the solution in this embodiment is: by packaging the operating environment, code, scripts and parameters required for training into a training package, and combining the training data set, training parameters and resource configuration to create a training task, the standardization and automated deployment of training tasks are achieved. This method ensures the consistency and portability of training tasks in different execution environments, avoids environmental configuration errors, and improves the efficiency of resource utilization.

[0144] In a possible design, the preset data set is an inference data set, and the data processing task is an inference task. Then, S301, creating a data processing task according to the preset data set includes:

[0145] S3011. Create an inference program based on the training model and the preset inference code.

[0146] Specifically, you can load a trained model. Then, combine the preset reasoning code with the model. These reasoning codes usually include logic such as data preprocessing, model calling, and post-processing. Finally, through compilation or packaging tools, the model and reasoning code are integrated into an executable program or service interface. By combining the model with the reasoning code, the reasoning program can be quickly deployed and run in different execution environments. In addition, this approach modularizes the reasoning process, facilitates the maintenance and updating of reasoning logic or model versions, and thus improves the scalability and adaptability of the system.

[0147] S3012: Create an inference task according to the inference program, the inference data set and the preset inference resource configuration.

[0148] Specifically, you can load the trained model and the corresponding inference code, which contains the logic required to perform inference. Then, associate the inference dataset with the inference program to ensure that the data can be correctly input into the model for inference. Finally, allocate and optimize resources according to the preset inference resource configuration to efficiently perform inference tasks. This implementation can improve the execution efficiency and accuracy of inference tasks, ensure that the model can process large amounts of data in a reasonable time, and effectively utilize system resources to avoid resource waste or bottleneck problems.

[0149] The technical effect of the scheme in this embodiment is: by systematically creating reasoning tasks, an efficient and flexible model reasoning process is achieved. The method can automatically generate a reasoning program based on a trained model and a preset reasoning code, and optimize the execution of reasoning tasks in combination with reasoning data sets and resource configuration. This method not only improves the speed and accuracy of reasoning, but also dynamically adjusts the execution strategy of reasoning tasks according to different resource conditions, thereby maintaining high performance and resource utilization in various computing environments.

[0150] Figure 4 Schematic diagram of the process of model training and reasoning method provided in the embodiment of the present application Figure 3 In this embodiment, Figure 3 Based on the provided embodiment, the model training reasoning method is further explained. The model training reasoning method includes:

[0151] S401. Create a data processing task according to a preset data set.

[0152] S402, sending a data processing task to a central end; wherein the central end is used to determine a target execution end from a plurality of execution ends according to the data processing task, and the target execution end is used to execute the data processing task.

[0153] S403, receiving the data processing result sent by the central end; wherein the data processing result is obtained by the target execution end executing the data processing task and sent to the central end.

[0154] S401 - S403 are similar to S301 - S303 and will not be described in detail in this embodiment.

[0155] S404. Access the target execution end through the target port.

[0156] Specifically, the target execution end can be accessed through the target port through the network communication protocol. The demand end sends a request to the target execution end through the target port to establish a communication channel. After receiving the request, the target execution end returns the access data through the port. This method allows the demand end to directly interact with the target execution end without the need for transfer through the central end, thereby improving the efficiency and real-time performance of data transmission. By directly accessing the target execution end, the delay in data transmission can be reduced and the response speed of the system can be improved. At the same time, this architecture can also reduce the load on the central end because the data does not need to be transferred through the central end. In addition, this method can also enhance the flexibility of the system, allowing the demand end to dynamically interact with different execution ends according to actual needs.

[0157] S405: Receive access data sent by the target execution end.

[0158] Specifically, the access data sent by the target execution end can be received through the established network connection. After the target execution end completes the task or generates data according to the request, it sends the data back to the demand end through the connection. The demand end obtains the access data by reading the data stream in the connection. This direct data reception method can reduce the interference and delay of the intermediate links and improve the efficiency and reliability of data transmission. In addition, since the data is transmitted directly from the target execution end to the demand end, this method also enhances the security of data transmission because it reduces the chance of data exposure to other intermediate nodes.

[0159] The technical effect of the scheme in this embodiment is: by directly accessing the target execution end and receiving the access data sent by it, the high efficiency and real-time performance of data transmission are achieved. The direct communication channel established through the target port reduces the intermediate links and potential delays in data transmission and improves the response speed of the system. In addition, this architecture reduces the load on the central end, optimizes resource utilization, and enhances the flexibility and scalability of the system, so that the demand end can quickly interact with different execution ends according to actual needs and adapt to dynamically changing task requirements.

[0160] An embodiment of the present application also provides a model training and reasoning system, including: a central end, a demand end, and multiple execution ends.

[0161] The central end is used to implement Figure 2 Model training inference method.

[0162] Demand side is used to achieve Figure 3 or Figure 4 Model training inference method.

[0163] The multiple execution ends include a target execution end, and the target execution end is used to execute a training task or an inference task.

[0164] Figure 5 This is a timing diagram of the model training inference method provided in the embodiment of the present application. Figure 5 As shown, the method includes:

[0165] S501: The demand side creates a data processing task based on a preset data set.

[0166] S502: The demand side sends the data processing task to the center side.

[0167] S503: The central end determines a target execution end from multiple execution ends according to the data processing task.

[0168] S504: The central end sends the data processing task to the target execution end.

[0169] S505: The target execution end executes the data processing task and obtains the data processing result.

[0170] S506: The target execution end sends the data processing result to the central end.

[0171] S507: The center sends the data processing result to the demand side.

[0172] Figure 6 A schematic diagram of the structure of the model training and reasoning device provided in the embodiment of the present application Figure 1 The model training and reasoning device is applied to the central end of the model training and reasoning system. The model training and reasoning system also includes a demand end and multiple execution ends, such as Figure 6 As shown, the model training inference device includes:

[0173] The first receiving module 601 is used to receive a data processing task sent by a demand side, where the data processing task is created by the demand side according to a preset data set.

[0174] The determination module 602 is used to determine a target execution end from multiple execution ends according to the data processing task.

[0175] The first sending module 603 is used to send the data processing task to the target execution end so that the target execution end executes the data processing task.

[0176] The second receiving module 604 is used to receive the data processing result sent by the target execution end, where the data processing result is obtained by the target execution end executing the data processing task.

[0177] The second sending module 605 is used to send the data processing result to the demand end.

[0178] In one possible design, the determining module 602 includes:

[0179] The acquisition unit is used to obtain bandwidth requirements, processor core number requirements, and memory requirements required for executing data processing tasks.

[0180] The matrix generation unit is used to generate a first characteristic matrix of the data processing task according to bandwidth requirements, processor core number requirements and memory requirements.

[0181] The calculation unit is used to calculate the similarity between the first characteristic matrix and each preset second characteristic matrix; wherein the second characteristic matrix refers to a matrix composed of the bandwidth, the number of processor cores and the memory of the execution end in the idle state.

[0182] The determination unit is used to determine the execution end corresponding to the second feature matrix with the largest similarity as the target execution end.

[0183] In one possible design, the computing unit includes:

[0184] The acquisition component is used to acquire the utilization rate of the processor of each execution end, and determine each execution end whose utilization rate is less than a preset threshold as an execution end in an idle state.

[0185] The calculation component is used to calculate the similarity between the first characteristic matrix and the second characteristic matrix of the execution end in the idle state.

[0186] The model training and reasoning device provided in this embodiment can be executed Figure 2 The technical solution of the method embodiment shown in the figure has the same implementation principle and technical effect as Figure 2 The method embodiments shown are similar and will not be described in detail here.

[0187] Figure 7 A schematic diagram of the structure of the model training and reasoning device provided in the embodiment of the present application Figure 2 The model training and reasoning device is applied to the demand side of the model training and reasoning system. The model training and reasoning system also includes a central side and multiple execution sides, such as Figure 7 As shown, the model training inference device includes:

[0188] The first creation module 701 is used to create a data processing task according to a preset data set.

[0189] The third sending module 702 is used to send the data processing task to the central end; wherein the central end is used to determine a target execution end from multiple execution ends according to the data processing task, and the target execution end is used to execute the data processing task.

[0190] The third receiving module 703 is used to receive the data processing result sent by the central end; wherein the data processing result is obtained by the target execution end executing the data processing task and sent to the central end.

[0191] In a possible design, the preset data set is a training data set, and the data processing task is a training task;

[0192] The first creation module 701 includes:

[0193] The packaging unit is used to package the preset running image, training code, startup script and input and output parameters to obtain a training package.

[0194] The training task creation unit is used to create a training task according to a training package, a training data set, preset training parameters and preset training resource configuration.

[0195] In a possible design, the preset data set is an inference data set, and the data processing task is an inference task;

[0196] The first creation module 701 includes:

[0197] The reasoning program creation unit is used to create a reasoning program according to the training model and the preset reasoning code.

[0198] The reasoning task creation unit is used to create a reasoning task according to the reasoning program, the reasoning data set and the preset reasoning resource configuration.

[0199] Figure 8 A schematic diagram of the structure of the model training and reasoning device provided in the embodiment of the present application Figure 3 The model training and reasoning device is applied to the demand side of the model training and reasoning system. The model training and reasoning system also includes a central side and multiple execution sides, such as Figure 8 As shown, the model training reasoning device also includes:

[0200] The access module 804 is used to access the target execution end through the target port.

[0201] The fourth receiving module 805 is used to receive access data sent by the target execution end.

[0202] The model training and reasoning device provided in this embodiment can be executed Figure 3-Figure 4 The technical solution of the method embodiment shown in the figure has the same implementation principle and technical effect as Figure 3-Figure 4 The method embodiments shown are similar and will not be described in detail here.

[0203] Fig. 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Fig. 9 As shown, the electronic device includes: at least one processor 910 and a memory 920. The electronic device also includes a communication component 930. The processor 910, the memory 920 and the communication component 930 are connected via a bus 940.

[0204] During the specific implementation process, at least one processor 910 executes the computer execution instructions stored in the memory 920, so that at least one processor 910 is used to implement the model training inference method of the above embodiment.

[0205] The specific implementation process of the processor 910 can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0206] In the above embodiments, it should be understood that the processor 910 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps of the method disclosed in the invention may be directly embodied as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0207] The memory 920 may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk storage.

[0208] The bus 940 may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus 940 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus 940 in the drawings of the present application is not limited to only one bus or one type of bus.

[0209] The above-mentioned functions implemented by the electronic device and the main control device introduce the scheme provided by the embodiment of the present invention. It can be understood that in order to implement the above-mentioned functions, the electronic device or the main control device includes a hardware structure and / or software module corresponding to the execution of each function. In combination with the units and algorithm steps of each example described in the embodiment disclosed in the embodiment of the present invention, the embodiment of the present invention can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the technical solution of the embodiment of the present invention.

[0210] The embodiment of the present application also provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, they are used to implement the model training inference method of the above embodiment. In the specific implementation of the aforementioned multi-protocol label switching message transmission method, each module can be implemented as a processor.

[0211] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0212] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the readable storage medium can also exist as discrete components in an electronic device or a main control device.

[0213] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, is used to implement the model training inference method of the above embodiment.

[0214] The computer program is stored in a readable storage medium. At least one processor can read the computer program from the readable storage medium. At least one processor executes the computer program to execute the solution provided in any of the above embodiments.

[0215] Those skilled in the art can understand that all or part of the steps of the above method embodiments can be completed by hardware related to program instructions. The above program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above method embodiments are executed; and the above storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0216] So far, the technical solution of the present application has been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments, and the above embodiments are only used to illustrate the technical solution of the present application rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A model training inference method, characterized in that: The method is applied to the central end of the model training and reasoning system, and the model training and reasoning system also includes a demand end and multiple execution ends, and the method includes: Receiving a data processing task sent by the demand side, where the data processing task is created by the demand side according to a preset data set; Determining a target execution end from the multiple execution ends according to the data processing task; Sending the data processing task to the target execution end so that the target execution end executes the data processing task; receiving a data processing result sent by the target execution end, wherein the data processing result is obtained by the target execution end executing the data processing task; The data processing result is sent to the demand end.

2. The method according to claim 1, characterized in that Determining a target execution end from the multiple execution ends according to the data processing task includes: Obtaining bandwidth requirements, processor core count requirements, and memory requirements required to perform the data processing task; Generate a first characteristic matrix of the data processing task according to the bandwidth requirement, the processor core number requirement and the memory requirement; Calculating the similarity between the first feature matrix and each preset second feature matrix; wherein the second feature matrix refers to a matrix composed of bandwidth, number of processor cores and memory of the execution end in an idle state; The execution end corresponding to the second feature matrix with the greatest similarity is determined as the target execution end.

3. The method according to claim 2, characterized in that The calculating the similarity between the first feature matrix and each preset second feature matrix includes: Acquire the utilization rate of the processor of each execution end, and determine each execution end whose utilization rate is less than a preset threshold as an execution end in an idle state; The similarities between the first characteristic matrix and the second characteristic matrix of the execution end in the idle state are calculated.

4. A model training inference method, characterized in that: The method is applied to the demand side of the model training and reasoning system, and the model training and reasoning system also includes a central side and multiple execution sides, and the method includes: Create data processing tasks based on preset data sets; Sending the data processing task to the central end; wherein the central end is used to determine a target execution end from the multiple execution ends according to the data processing task, and the target execution end is used to execute the data processing task; Receive the data processing result sent by the central end; wherein the data processing result is obtained by the target execution end executing the data processing task and sent to the central end.

5. The method according to claim 4, characterized in that The preset data set is a training data set, and the data processing task is a training task; Then, creating a data processing task according to the preset data set includes: Package the preset running image, training code, startup script, and input and output parameters to obtain a training package; The training task is created according to the training package, the training data set, the preset training parameters and the preset training resource configuration.

6. The method according to claim 4, characterized in that The preset data set is an inference data set, and the data processing task is an inference task; Then, creating a data processing task according to the preset data set includes: Creating an inference program according to the training model and the preset inference code; The reasoning task is created according to the reasoning program, the reasoning data set and the preset reasoning resource configuration.

7. The method according to claim 4, characterized in that If the data processing result is a target port, after receiving the data processing result sent by the central end, the method further includes: Accessing the target execution end through the target port; Receive access data sent by the target execution end.

8. A model training inference system, comprising: The center side, the demand side and multiple execution sides; The central end is used to implement the model training inference method according to any one of claims 1 to 3; The demand side is used to implement the model training inference method according to any one of claims 3 to 7; The multiple execution ends include a target execution end, and the target execution end is used to execute a training task or an inference task.

9. A model training and reasoning device, characterized in that: The device is applied to the central end of the model training and reasoning system, and the model training and reasoning system also includes a demand end and multiple execution ends, and the device includes: A first receiving module, configured to receive a data processing task sent by the demand side, wherein the data processing task is created by the demand side according to a preset data set; A determination module, configured to determine a target execution end from the multiple execution ends according to the data processing task; A first sending module, used for sending the data processing task to the target execution end, so that the target execution end executes the data processing task; A second receiving module is used to receive a data processing result sent by the target execution end, where the data processing result is obtained by the target execution end executing the data processing task; The second sending module is used to send the data processing result to the demand end.

10. A model training and reasoning device, characterized in that: The device is applied to the demand side of the model training and reasoning system, and the model training and reasoning system also includes a central side and multiple execution sides, and the device includes: A first creation module is used to create a data processing task according to a preset data set; A third sending module is used to send the data processing task to the central end; wherein the central end is used to determine a target execution end from the multiple execution ends according to the data processing task, and the target execution end is used to execute the data processing task; The third receiving module is used to receive the data processing result sent by the central end; wherein the data processing result is obtained by the target execution end executing the data processing task and sent to the central end.

11. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; When the processor executes the computer-executable instructions stored in the memory, it is used to implement the model training and reasoning method as described in any one of claims 1 to 3, or to implement the model training and reasoning method as described in any one of claims 4 to 7.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the model training and reasoning method as described in any one of claims 1 to 3, or to implement the model training and reasoning method as described in any one of claims 4 to 7.

13. A computer program product, comprising a computer program, which, when executed by a processor, is used to implement the model training and reasoning method as described in any one of claims 1 to 3, or to implement the model training and reasoning method as described in any one of claims 4 to 7.

Images