A method and system for building an AI model for a private cloud computing network

Abstract

The present invention provides an AI model building method and system for a computing power network private cloud, which belongs to the field of artificial intelligence technology, and specifically includes: using the matching conditions of different AI models in a public cloud and demand feature items to determine the available training models of the AI ​​model, based on the idle computing power resources in different workspaces in the existing computing power network, and in combination with the user's training data volume and available training models, determining the historical training problem data of different workspaces under similar training conditions, determining the secondary screening model in the available training models based on the historical training problem data, determining the recommended processing order of the secondary training model based on the training prediction time of the secondary training model in different workspaces and the historical training problem data under similar training conditions and the matching coefficient, and using the user's selection result to build the AI ​​model of the user's private cloud, thereby improving the efficiency and matching degree of the AI ​​model building process.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular, relates to an AI model building method and system for a computing power network private cloud. Background Art

[0002] The private cloud of Computing Network can meet the needs of personalized AI solutions for enterprises and universities by leasing computing power, and meet the training of AI models. Specifically, in the invention patent application CN202110769541.0 "A method for generating a private cloud algorithm model based on joint learning", the initial convolutional neural network is trained offline according to the storage data on the private cloud to obtain a personalized personal convolutional neural network model, making it easier for customers to use. However, there are the following technical problems:

[0003] When building AI models in a private cloud, users may not be able to accurately obtain AI models that fully match their own data and needs due to their lack of understanding of AI models in the public cloud. As a result, the training time and training results of the AI ​​models may not meet the requirements.

[0004] In response to the above technical problems, this application specifically provides an AI model building method and system for a computing power network private cloud. Summary of the invention

[0005] To achieve the purpose of the present invention, the present invention adopts the following technical solutions:

[0006] According to one aspect of the present invention, a method for building an AI model for a computing power network private cloud is provided.

[0007] A method for building an AI model for a computing network private cloud, specifically comprising:

[0008] S1 determines the demand feature items of the training demand based on the analysis results of the user's training demand, and determines the matching historical users by combining the distribution data of different types of training data of the user with the matching results of different historical users;

[0009] S2 determines the matching conditions between different AI models in the public cloud and the demand feature items based on the training results of the AI ​​models of different matching history users, and determines the matching coefficient of the AI ​​model and the available training model using the matching conditions;

[0010] S3 determines the historical training problem data of different workspaces under similar training conditions based on the idle computing resources in different workspaces in the existing computing network and in combination with the training data volume of the user and the available training models, and determines the secondary screening model in the available training models based on the historical training problem data;

[0011] S4 determines the predicted training duration of the secondary training model in different workspaces and the historical training problem data under similar training conditions, and determines the recommended processing order of the secondary training model in combination with the matching coefficient of the secondary training model, and uses the user's selection results to build the AI ​​model of the user's private cloud.

[0012] The beneficial effects of the present invention are:

[0013] Based on the matching of different AI models in the public cloud with the training results of the AI ​​models of historical users and the matching of demand feature items, the available training models in the AI ​​models are determined, thereby realizing the matching of the training results of the AI ​​models in matching historical users and the matching of demand feature items, identifying the AI ​​models that deviate from the user's training needs, and laying the foundation for further realizing personalized recommendation processing of AI models.

[0014] The recommended processing order of the secondary training model is determined based on the training prediction time of the secondary training model in different workspaces, the historical training problem data under similar training conditions, and the matching coefficient of the secondary training model. This not only takes into account the problem data and training prediction time of the secondary training model in the existing idle workspace, and realizes the screening of secondary training models with higher reliability and shorter training prediction time, but also takes into account the matching degree between the secondary training model and the user's training needs, thereby realizing personalized AI model recommendation processing and improving the efficiency of user selection processing.

[0015] A further technical solution is that the requirement feature items of the training requirements include input features and output quantities, model recognition accuracy, and model recognition rate.

[0016] A further technical solution is that the distribution data of the training data includes the data volume of the training data corresponding to different types of output quantities.

[0017] A further technical solution is that the method for determining the matching historical users is:

[0018] Determining the input features and output quantity of the user based on the demand feature items of the training demand, and determining the training similar users of the historical users based on the input features and output quantity;

[0019] Based on the distribution data of different types of training data of the user, determine the data volume of the user in different types of training data, and use the data volume of different types of training data to determine the deviation of the data volume of different types of training data of users similar to the training, and determine the deviation coefficient of the data volume of different types of training data according to the deviation;

[0020] The training data deviation coefficient is determined according to the average value of the deviation coefficients of the data amounts of different types of training data, and the matching historical users among the training similar users are determined using the training data deviation coefficient.

[0021] A further technical solution is that the training similar users are historical users whose input features and output quantities are consistent with those of the user.

[0022] A further technical solution is that the matching historical users are training similar users whose training data deviation coefficient is less than a preset deviation coefficient threshold.

[0023] A further technical solution is that the method for determining the recommended processing order of the secondary training model is:

[0024] Using historical training problem data under similar training conditions in different workspaces, determine the number of historical training times when the training problem occurred in different workspaces, and determine the problem weight coefficients of different workspaces using the number of historical training times when the training problem occurred and a preset proportional factor;

[0025] Determine the training matching coefficients of different workspaces according to the ratio of the training prediction duration of different workspaces to the preset duration threshold, and determine the training matching coefficient weight sum of the secondary training model in combination with the problem weight coefficients of the different workspaces;

[0026] The comprehensive matching coefficient of the secondary training model is determined based on the training matching coefficient weight and the average value of the matching coefficient, and the recommended processing order of the secondary training model is determined using the comprehensive matching coefficient.

[0027] A further technical solution is to use the comprehensive matching coefficient to determine the recommended processing order of the secondary training model, specifically including:

[0028] The recommended processing order of different secondary training models is determined from large to small based on the comprehensive matching coefficients of the secondary training models.

[0029] In a second aspect, the present invention provides a computer system comprising: a memory and a processor that are communicatively connected, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-mentioned method for building an AI model for a private cloud of a computing power network when running the computer program.

[0030] Other features and advantages will be described in the following description. The objects and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and drawings.

[0031] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The above and other features and advantages of the present invention will become more apparent by describing in detail exemplary embodiments thereof with reference to the attached drawings.

[0033] Figure 1 It is a flow chart of a method for building an AI model for a private cloud of a computing network;

[0034] Figure 2 is a flow chart of a method for matching determinations of historical users;

[0035] Figure 3 is a flow chart of a method for determining a matching coefficient of an AI model;

[0036] Figure 4 is a flow chart of a method for determining a secondary screening model from among available training models;

[0037] Figure 5 This is a flowchart of a method for determining a recommendation processing order of a secondary training model. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0039] Because users do not have enough understanding of AI models in the public cloud, they may not be able to accurately obtain AI models that fully match their own data and needs, which makes it difficult for users to meet the requirements in terms of training time and training results.

[0040] Matching historical users: Use the user's input features and output to determine the training similar users of the historical users. According to the average value of the deviation between the training similar users and the user's data volume in different types of training data, the training similar users with less than the preset deviation are selected as matching historical users.

[0041] Matching coefficient of the AI ​​model: determine the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate for different matching history users based on the deviation between the model recognition accuracy and the model recognition rate of the AI ​​model in different matching history users and the demand characteristics; take the average value of the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate for different matching history users as the average value of the deviation rate of the matching history users; and determine the matching coefficient of the AI ​​model based on the average value of the deviation rates of different matching history users.

[0042] AI models with a matching coefficient greater than 0.6 are considered available training models.

[0043] Secondary screening model among available training models: The secondary screening model is determined by the proportion of the number of working intervals with historical training problems under similar training conditions. When the proportion is less than 0.3, the available training model is determined to be the secondary screening model.

[0044] Recommended processing order of the secondary training model: determine the problem weight coefficients of different workspaces by using the historical training times of the training problems and the preset proportional factor, determine the training matching coefficients of different workspaces according to the ratio of the training prediction duration of different workspaces to the preset duration threshold, and determine the weighted sum of the training matching coefficients of the secondary training model in combination with the problem weight coefficients of the different workspaces, determine the comprehensive matching coefficient of the secondary training model based on the weighted sum of the training matching coefficients and the average value of the matching coefficients, and use the comprehensive matching coefficient to determine the recommended processing order of the secondary training model.

[0045] Embodiment 1 To solve the above problem, according to one aspect of the present invention, Figure 1 As shown, a method for building an AI model for a computing network private cloud is provided, which specifically includes:

[0046] S1 determines the demand feature items of the training demand based on the analysis results of the user's training demand, and determines the matching historical users by combining the distribution data of different types of training data of the user with the matching results of different historical users;

[0047] Furthermore, the requirement feature items of the training requirement include input features and output quantity, model recognition accuracy, and model recognition rate.

[0048] Specifically, the distribution data of the training data includes the data volume of the training data corresponding to different types of output volumes.

[0049] It should be noted that if Figure 2 As shown, the method for determining the matching history user is:

[0050] Determining the input features and output quantity of the user based on the demand feature items of the training demand, and determining the training similar users of the historical users based on the input features and output quantity;

[0051] Based on the distribution data of different types of training data of the user, determine the data volume of the user in different types of training data, and use the data volume of different types of training data to determine the deviation of the data volume of different types of training data of users similar to the training, and determine the deviation coefficient of the data volume of different types of training data according to the deviation;

[0052] The training data deviation coefficient is determined according to the average value of the deviation coefficients of the data amounts of different types of training data, and the matching historical users among the training similar users are determined using the training data deviation coefficient.

[0053] Furthermore, the training similar users are historical users whose input features and output quantities are consistent with those of the user.

[0054] Optionally, the matching historical users are training similar users whose training data deviation coefficient is less than a preset deviation coefficient threshold.

[0055] It should be noted that the deviation coefficient of the data volume of the training data is determined according to the ratio of the deviation of the data volume of the training data of the type to the data volume of the training data of the user.

[0056] Optionally, the method for determining the matching historical user is:

[0057] Determining the input features and output quantity of the user based on the demand feature items of the training demand, and determining the training similar users of the historical users based on the input features and output quantity;

[0058] Determine, based on the distribution data of the different types of training data of the user, a distribution ratio factor of the data amount of the user in the different types of training data, and determine, according to the distribution ratio factor, a deviation coefficient of the distribution ratio factor of the different types of training data of users similar to the training;

[0059] The distribution scale factor deviation coefficient is determined according to the average value of the deviation coefficients of the distribution scale factors in different types of training data, and the matching historical users among the training similar users are determined using the distribution scale factor deviation coefficient.

[0060] Optionally, the method for determining the matching historical user is:

[0061] S11 determines the input features and output quantity of the user based on the demand feature items of the training demand, and determines the training similar users of the historical users using the input features and output quantity;

[0062] S12: determining the amount of the user's data in different types of training data based on the distribution data of the user's data in different types of training data, and using the amount of the different types of training data to determine the deviation of the amount of the training data in different types of training data from the training-similar user, and determining the deviation coefficient of the amount of the training data in different types of training data according to the deviation;

[0063] Optionally, the above step S12 includes the following contents:

[0064] S121 determines the data volume of the user in different types of training data based on the distribution data of the user's different types of training data, and uses the data volume of the different types of training data to determine the deviation of the data volume of the training similar user in different types of training data. When the training similar user does not have training data with a deviation amount that meets the requirement, it is determined that the training similar user does not belong to the matching history user. When the training similar user has training data with a deviation amount that meets the requirement, the process proceeds to step S122.

[0065] S122 determines the deviation coefficient of the data volume of different types of training data according to the deviation situation. When the deviation coefficient of the data volume of the training data does not meet the required type, it is determined that the training similar user does not belong to the matching historical user. When the deviation coefficient of the data volume of the training data meets the required type, it proceeds to step S123.

[0066] S123 obtains the average value of the deviation coefficient of the data volume of different types of training data. When the average value of the deviation coefficient of the data volume of different types of training data is less than the preset deviation coefficient threshold, it is determined that the training similar user belongs to the matching historical user. When the average value of the deviation coefficient of the data volume of different types of training data is within the preset deviation coefficient interval, it proceeds to step S13. When the average value of the deviation coefficient of the data volume of different types of training data is not within the preset deviation coefficient interval, it is determined that the training similar user does not belong to the matching historical user.

[0067] S13, based on the distribution data of the different types of training data of the user, determining a distribution ratio factor of the data amount of the user in the different types of training data, and determining a deviation coefficient of the distribution ratio factor of the different types of training data of a user similar to the training according to the distribution ratio factor;

[0068] Optionally, the above step S13 includes the following contents:

[0069] S131 determines the distribution ratio factor of the data volume of the user in different types of training data based on the distribution data of the user. When there is no type of training data whose distribution ratio factor meets the requirement, it is determined that the training similar user does not belong to the matching history user. When there is a type of training data whose distribution ratio factor meets the requirement, it proceeds to step S132.

[0070] S132 determines the deviation coefficient of the distribution scale factor of the training data of different types of users similar to the training according to the distribution scale factor. When the average value of the deviation coefficient of the distribution scale factor of the training data of different types is less than the preset distribution factor deviation coefficient, the process proceeds to step S14. When the average value of the deviation coefficient of the distribution scale factor of the training data of different types is not less than the preset distribution factor deviation coefficient, the process proceeds to step S133.

[0071] S133 takes the type of training data whose distribution ratio factor meets the requirements as the distribution similarity type. When the number of the distribution similarity types is greater than the preset number of similarity types, it proceeds to step S14. When the number of the distribution similarity types is not greater than the preset number of similarity types, it determines that the training similar user does not belong to the matching historical user.

[0072] S14 determines the data deviation coefficient of different types of training data based on the deviation coefficient of the distribution ratio factor of different types of training data and the deviation coefficient of the data volume of the training data, and determines the training data deviation coefficient according to the average value of the data deviation coefficient of different types of training data, and uses the training data deviation coefficient to determine the matching historical users among the training similar users.

[0073] Optionally, the above step S14 includes the following contents:

[0074] S141 determines the data deviation coefficient of different types of training data based on the deviation coefficient of the distribution scale factor of different types of training data and the deviation coefficient of the data volume of the training data. When the average value of the data deviation coefficient of different types of training data is greater than the preset data deviation coefficient threshold, it is determined that the training similar user does not belong to the matching historical user. When the average value of the data deviation coefficient of different types of training data is not greater than the preset data deviation coefficient threshold, it proceeds to step S142.

[0075] S142 determines the training data deviation coefficient according to the average value of the data deviation coefficients of different types of training data, and uses the training data deviation coefficient to determine the matching historical users among the training similar users.

[0076] Furthermore, the matching status of the AI ​​model and the demand feature item is determined based on the deviation between the model recognition accuracy and model recognition rate of users in different matching history and the demand feature.

[0077] S2 determines the matching conditions between different AI models in the public cloud and the demand feature items based on the training results of the AI ​​models of different matching history users, and determines the matching coefficient of the AI ​​model and the available training model using the matching conditions;

[0078] Specifically, Figure 3 As shown, the method for determining the matching coefficient of the AI ​​model is:

[0079] Determine the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate of different matching history users based on the deviation amount of the model recognition accuracy and the model recognition rate of the AI ​​model in different matching history users and the demand characteristics;

[0080] Based on the average of the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate of different matching historical users, the average deviation rate of the matching historical users is used as the average deviation rate of the matching historical users;

[0081] The matching coefficient of the AI ​​model is determined according to the average value of the deviation rate average values ​​of different matching history users.

[0082] Furthermore, the matching coefficient of the AI ​​model is determined based on the difference between a preset value and an average value of the deviation rates of different matching history users.

[0083] Optionally, the matching coefficient of the AI ​​model ranges from 0 to 1, and when the matching coefficient of the AI ​​model is greater than a preset matching coefficient threshold, the AI ​​model is determined to be an available training model.

[0084] In another embodiment, the matching coefficient of the AI ​​model is determined by:

[0085] Determine the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate of different matching history users based on the deviation amount of the model recognition accuracy and the model recognition rate of the AI ​​model in different matching history users and the demand characteristics;

[0086] Determining users with similar needs to the matching history users based on the deviation rates of the model recognition accuracy and the deviation rates of the model recognition rates among different matching history users;

[0087] The matching coefficient of the AI ​​model is determined according to the proportion of the users with similar needs in the matching history users.

[0088] Furthermore, the demand-similar users are matching history users whose deviation rate of model recognition accuracy and deviation rate of model recognition speed are both within a preset deviation rate range.

[0089] Optionally, the matching coefficient of the AI ​​model is determined by:

[0090] S21 determines the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate of different matching history users based on the deviation amount of the model recognition accuracy and the model recognition rate of the AI ​​model in different matching history users and the demand characteristics, and determines the matching deviation coefficient of different matching history users according to the average value of the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate in different matching history users;

[0091] Optionally, the above step S21 includes the following contents:

[0092] S211 determines the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate of different matching history users based on the deviation of the model recognition accuracy and the model recognition rate of the AI ​​model in different matching history users and the demand characteristics. When the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate of different matching history users are both greater than the preset deviation rate threshold, it is determined that the AI ​​model does not belong to the available training model. When any one of the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate of different matching history users is not greater than the preset deviation rate threshold, the process proceeds to step S212.

[0093] S212 determines the matching deviation coefficients of different matching history users according to the average value of the deviation rate of the model recognition accuracy and the deviation rate of the model recognition rate of different matching history users. When the matching deviation coefficients of different matching history users are all greater than the preset matching deviation coefficient threshold, it is determined that the AI ​​model does not belong to the available training model. When there is a matching history user whose matching deviation coefficient is not greater than the preset matching deviation coefficient threshold, the process proceeds to step S213.

[0094] S213 When the average value of the matching deviation coefficients of different matching history users meets the requirements, proceed to step S22. When the average value of the matching deviation coefficients of different matching history users does not meet the requirements, it is determined that the AI ​​model does not belong to an available training model.

[0095] S22 determines users with similar needs among the matching history users according to the matching deviation coefficients of different matching history users;

[0096] Optionally, the above step S22 includes the following contents:

[0097] S221 determines the demand-similar users among the matching history users according to the matching deviation coefficients of different matching history users. When the number of demand-similar users among the matching history users is not greater than the preset number of similar users, it is determined that the AI ​​model does not belong to the available training model. When the number of demand-similar users among the matching history users is greater than the preset number of similar users, the process proceeds to step S222;

[0098] S222: when the number of users with similar needs among the matching historical users is within the preset similar user number range, the process proceeds to step S23; when the number of users with similar needs among the matching historical users is not within the preset similar user number range, the process proceeds to step S223;

[0099] S223: When the average value of the matching deviation coefficients of the users with similar needs is within the preset matching deviation coefficient range, it is determined that the AI ​​model does not belong to an available training model; when the average value of the matching deviation coefficients of the users with similar needs is not within the preset matching deviation coefficient range, the process proceeds to step S23.

[0100] S23 determines the matching coefficient of the AI ​​model based on the proportion of the number of users with similar needs and the matching deviation coefficients of different matching history users.

[0101] S3 determines the historical training problem data of different workspaces under similar training conditions based on the idle computing resources in different workspaces in the existing computing network and in combination with the training data volume of the user and the available training models, and determines the secondary screening model in the available training models based on the historical training problem data;

[0102] Furthermore, the idle computing resources in the workspace include idle GPU types in the workspace and the number of different idle GPU types.

[0103] Specifically, the similar training condition is a training condition in which, in the available training model, the deviation of the idle computing resources in the workspace and the deviation of the amount of training data of the user are both within a preset deviation range.

[0104] It should be noted that the historical training problem data includes the number of training interruptions and the number of alarms of different types of hardware devices.

[0105] Optionally, if Figure 4 As shown, the method for determining the secondary screening model in the available training model is:

[0106] The historical training times of the available training model under similar training conditions are used as the matching training times, and the working interval with historical training problems is determined according to the historical training problem data in the matching training times in different working spaces, and is used as the problem working interval;

[0107] The training problem coefficients of different problem working intervals are determined by using the proportion of matching training times with historical training problems in different problem working intervals;

[0108] Based on the average values ​​of the training problem coefficients of different problem working intervals and the proportion of the problem working intervals in the working intervals, the problem coefficient evaluation value of the available training model is determined, and the problem coefficient evaluation value is used to determine whether the available training model is a secondary screening model.

[0109] Furthermore, the problem coefficient evaluation amount of the available training model is determined based on the product of the average value of the training problem coefficients of different problem working intervals and the proportion of the number of the problem working intervals in the working interval.

[0110] In another embodiment, when the problem coefficient evaluation value of the available training model is greater than a preset evaluation value threshold, it is determined that the available training model does not belong to the secondary screening model.

[0111] S4 determines the predicted training duration of the secondary training model in different workspaces and the historical training problem data under similar training conditions, and determines the recommended processing order of the secondary training model in combination with the matching coefficient of the secondary training model, and uses the user's selection results to build the AI ​​model of the user's private cloud.

[0112] Furthermore, the training prediction duration is determined based on an average of historical training durations of the secondary training model under similar training conditions in the workspace.

[0113] Specifically, Figure 5 As shown, the method for determining the recommended processing order of the secondary training model is:

[0114] Using historical training problem data under similar training conditions in different workspaces, determine the number of historical training times when the training problem occurred in different workspaces, and determine the problem weight coefficients of different workspaces using the number of historical training times when the training problem occurred and a preset proportional factor;

[0115] Determine the training matching coefficients of different workspaces according to the ratio of the training prediction duration of different workspaces to the preset duration threshold, and determine the training matching coefficient weight sum of the secondary training model in combination with the problem weight coefficients of the different workspaces;

[0116] The comprehensive matching coefficient of the secondary training model is determined based on the training matching coefficient weight and the average value of the matching coefficient, and the recommended processing order of the secondary training model is determined using the comprehensive matching coefficient.

[0117] Further, the recommended processing order of the secondary training model is determined by using the comprehensive matching coefficient, specifically including:

[0118] The recommended processing order of different secondary training models is determined from large to small based on the comprehensive matching coefficients of the secondary training models.

[0119] In the second aspect of Example 2, the present invention provides a computer system, comprising: a memory and a processor that are communicatively connected, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-mentioned method for building an AI model for a private cloud of a computing power network when running the computer program.

[0120] Optionally, the method for determining the secondary screening model in the available training model is:

[0121] The number of historical trainings of the available training model under similar training conditions is used as the number of matching trainings, and when no historical training problem data appears in the matching trainings in different workspaces, it is determined that the available training model belongs to a secondary screening model;

[0122] When there is a working interval with historical training problem data in the matching training times:

[0123] The working interval in which the historical training problems occur in the matching training times is regarded as the problematic working interval, and when the proportion of the problematic working interval in the working interval does not meet the requirement, it is determined that the available training model does not belong to the secondary screening model;

[0124] When the proportion of the problematic working intervals in the working intervals meets the requirement:

[0125] Determine the basic problem coefficient of the problem working interval according to the number of the problem working intervals and the proportion of the problem working intervals in the working intervals. When the basic problem coefficient of the problem working interval is not greater than a preset problem coefficient threshold, determine that the available training model belongs to a secondary screening model.

[0126] When the basic problem coefficient of the problem working interval is greater than the preset problem coefficient threshold:

[0127] The training problem coefficients of different problem working intervals are determined by using the proportion of the matching training times of historical training problems in different problem working intervals, and when the average value of the training problem coefficients of different problem working intervals is greater than a preset threshold value of the problem coefficient, it is determined that the available training model does not belong to the secondary screening model;

[0128] When the average value of the training problem coefficients of different problem working intervals is not greater than the preset threshold of the problem coefficient:

[0129] Based on the average values ​​of the training problem coefficients of different problem working intervals and the proportion of the problem working intervals in the working intervals, the problem coefficient evaluation value of the available training model is determined, and the problem coefficient evaluation value is used to determine whether the available training model is a secondary screening model.

[0130] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, equipment, and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0131] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0132] The above description is only one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, one or more embodiments of this specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of this specification shall be included in the scope of the claims of this specification.

Claims

1. A method for building an AI model for a private cloud of a computing network, characterized in that: Specifically include: Determine the demand feature items of the training demand based on the analysis results of the user's training demand, and determine the matching historical users in combination with the distribution data of different types of training data of the user and the matching results of different historical users; Based on the training results of the AI ​​models of different matching history users, the matching conditions between different AI models in the public cloud and the demand feature items are determined, and the matching coefficients of the AI ​​models and the available training models are determined using the matching conditions; Based on the idle computing resources in different workspaces in the existing computing network, and in combination with the amount of training data of the user and the available training models, historical training problem data of different workspaces under similar training conditions are determined, and a secondary screening model among the available training models is determined based on the historical training problem data; Determine the training prediction duration of the secondary screening model in different workspaces and the historical training problem data under similar training conditions, and determine the recommended processing order of the secondary screening model in combination with the matching coefficient of the secondary screening model, and use the user's selection result to build the AI ​​model of the user's private cloud; The distribution data of the training data includes the data volume of the training data corresponding to different types of output volumes.

2. The AI ​​model building method for a computing power network private cloud according to claim 1, characterized in that: The requirement feature items of the training requirement include input features and output quantity, model recognition accuracy, and model recognition rate.

3. The AI ​​model building method for a computing power network private cloud according to claim 1, characterized in that: The method for determining the matching history user is: Determining the input features and output quantity of the user based on the demand feature items of the training demand, and determining the training similar users of the historical users based on the input features and output quantity; Based on the distribution data of different types of training data of the user, determine the data volume of the user in different types of training data, and use the data volume of different types of training data to determine the deviation of the data volume of different types of training data of users similar to the training, and determine the deviation coefficient of the data volume of different types of training data according to the deviation; The training data deviation coefficient is determined according to the average value of the deviation coefficients of the data amounts of different types of training data, and the matching historical users among the training similar users are determined using the training data deviation coefficient.

4. The AI ​​model building method for a computing power network private cloud according to claim 3, characterized in that: The training similar users are historical users whose input features and output quantities are consistent with those of the user.

5. The AI ​​model building method for a computing power network private cloud according to claim 3, characterized in that: The deviation coefficient of the data volume of the training data is determined according to the ratio of the deviation of the data volume of the training data of the type to the data volume of the training data of the user.

6. The AI ​​model building method for a computing power network private cloud according to claim 1, characterized in that: The matching status of the AI ​​model and the demand feature item is determined based on the deviation between the model recognition accuracy and model recognition rate of users in different matching history and the demand feature item.

7. The AI ​​model building method for a computing power network private cloud according to claim 1, characterized in that: The method for determining the recommended processing order of the secondary screening model is: Using historical training problem data under similar training conditions in different workspaces, determine the number of historical training times when the training problem occurred in different workspaces, and determine the problem weight coefficients of different workspaces using the number of historical training times when the training problem occurred and a preset proportional factor; Determine the training matching coefficients of different workspaces according to the ratio of the training prediction duration of different workspaces to the preset duration threshold, and determine the training matching coefficient weight sum of the secondary screening model in combination with the problem weight coefficients of the different workspaces; The comprehensive matching coefficient of the secondary screening model is determined based on the training matching coefficient weight and the average value of the matching coefficient, and the recommended processing order of the secondary screening model is determined using the comprehensive matching coefficient.

8. The AI ​​model building method for a computing power network private cloud according to claim 7, characterized in that: Determining the recommended processing order of the secondary screening model using the comprehensive matching coefficient specifically includes: The recommended processing order of different secondary screening models is determined from large to small based on the comprehensive matching coefficients of the secondary screening models.

9. A computer system comprising: A memory and a processor that are communicatively connected, and a computer program stored in the memory and capable of running on the processor, characterized in that when the processor runs the computer program, an AI model building method for a computing power network private cloud as described in any one of claims 1-8 is executed.

Images