AI model integration method, platform, storage medium and electronic device

Abstract

The present application discloses an AI model integration method, platform, storage medium and electronic device, and relates to the field of artificial intelligence technology. The method includes: obtaining a model training task to be executed; determining the target training mode corresponding to the model to be trained based on the model training task and the preset training mode prediction model, and training the model to be trained according to the target training mode; determining the target deployment mode of the trained model according to the actual application conditions of the trained model and the preset deployment mode prediction model, and deploying the trained model according to the target deployment mode; classifying the trained model according to the structure, indicators and application scenarios of the trained model, and storing the trained model in the corresponding area according to the classification results; in response to the user's model recommendation request, determining the model recommendation factor corresponding to the user according to the user's identity information, and recommending the corresponding model to the user. The present application can manage AI models in a unified manner.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an AI model integration method, platform, storage medium and electronic device. Background Art

[0002] As the driving force of the new round of industrial transformation, artificial intelligence (AI) has achieved remarkable results in empowering and upgrading traditional industries and spawning emerging industries. As the types of AI models being developed continue to increase, there is currently a lack of an effective means of unified management of AI models. Summary of the invention

[0003] In view of this, the present application provides an AI model integration method, platform, storage medium and electronic device, the main purpose of which is to enable unified management of AI models.

[0004] According to a first aspect of the present application, an AI model integration method is provided, which is applied to an AI model integration platform, wherein the AI ​​model integration platform adopts a microservice architecture and cloud native technology, and the method comprises:

[0005] Get the model training task to be executed;

[0006] Based on the model training task and the preset training mode prediction model, determine the target training mode corresponding to the model to be trained, and train the model to be trained according to the target training mode to obtain a trained model;

[0007] Determine the target deployment mode of the trained model according to the actual application conditions of the trained model and the preset deployment mode prediction model, and deploy the trained model according to the target deployment mode;

[0008] Classify the trained model according to the structure, index and application scenario of the trained model, and store the trained model in a corresponding area according to the classification result;

[0009] In response to a model recommendation request from a user, a model recommendation factor corresponding to the user is determined according to the identity information of the user, and a corresponding model is recommended to the user based on the model recommendation factor.

[0010] According to a second aspect of the present application, an AI model integration platform is provided, wherein the AI ​​model integration platform adopts a microservice architecture and cloud native technology, and the platform includes:

[0011] The model training module is used to obtain the model training task to be executed; based on the model training task and the preset training mode prediction model, determine the target training mode corresponding to the model to be trained, and train the model to be trained according to the target training mode to obtain a trained model;

[0012] A model deployment module, used to determine the target deployment mode of the trained model according to the actual application conditions of the trained model and the preset deployment mode prediction model, and deploy the trained model according to the target deployment mode;

[0013] The model management module is used to classify the trained model according to the structure, indicators and application scenarios of the trained model, and store the trained model in the corresponding area according to the classification result; in response to the user's model recommendation request, determine the model recommendation factor corresponding to the user according to the user's identity information, and recommend the corresponding model to the user based on the model recommendation factor.

[0014] According to the third aspect of the present application, a storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned AI model integration method is implemented.

[0015] According to the fourth aspect of the present application, an electronic device is provided, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor implements the above-mentioned AI model integration method when executing the program.

[0016] Through the above-mentioned technical scheme, the present application provides an AI model integration method, platform, storage medium and electronic device, which can realize unified management of AI models by automatically classifying and storing trained models, and automatically recommending corresponding models to users according to user identities when users request recommended models. In addition, the present application automatically trains and deploys models through preset training mode prediction models and preset deployment mode prediction models, and can realize one-stop service for model training, deployment and management, thereby saving the time cost of model development and improving model development efficiency.

[0017] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0019] Figure 1 A schematic diagram of a process of an AI model integration method provided in an embodiment of the present application is shown;

[0020] Figure 2 A schematic diagram of the structure of an AI model integration platform provided in an embodiment of the present application is shown;

[0021] Figure 3 A structural schematic diagram of another AI model integration platform provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0022] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0023] As the types of AI models being developed continue to increase, there is currently a lack of an effective means of unified management of AI models.

[0024] In order to solve the above problems, an embodiment of the present invention provides an AI model integration method, which is applied to an AI model integration platform, including: a model training module, a model deployment module and a model management module. The AI ​​model integration platform adopts a microservice architecture and cloud native technology, such as Figure 1 As shown, the method includes:

[0025] Step 101: Obtain the model training task to be executed.

[0026] The model training task includes the structure of the model to be trained, sample data set, training algorithm and training requirements. The structure of the model to be trained may include convolutional neural network, recurrent neural network, deep belief network, etc. The training algorithm may include gradient descent algorithm, back propagation algorithm, stochastic gradient descent algorithm, Adam algorithm, AdaGrad algorithm, Momentum algorithm, etc. The training requirements may include accuracy, precision and recall, etc.

[0027] For the embodiment of the present invention, developers can submit model training tasks to the model training module of the platform, and the model training module can complete the training of the model based on the model training tasks.

[0028] Step 102: Based on the model training task and the preset training mode prediction model, determine the target training mode corresponding to the model to be trained, and train the model to be trained according to the target training mode to obtain a trained model.

[0029] The preset training mode prediction model may specifically be a training mode classification model, such as a support vector machine, a decision tree, a random forest, a neural network model, etc. The target training mode may specifically be one of online training, offline training, and hybrid training. Online training specifically refers to model training on a local server, offline training refers to model training on a cloud server, and hybrid training refers to assigning a portion of the training tasks to a local server and another portion of the training tasks to a cloud server.

[0030] For the embodiment of the present invention, in order to determine the target training mode corresponding to the model to be trained, step 102 specifically includes: determining the computing resources required for the model training task based on the structure, data volume, training algorithm and training requirements of the model to be trained in the model training task; determining the training features of the model to be trained based on the computing resources, as well as the data security and data portability of the model to be trained; inputting the training features into the preset training mode prediction model to perform training mode prediction, and obtaining the target training mode corresponding to the model to be trained.

[0031] The data volume refers to the data volume of the sample data set of the model to be trained. Data security refers to whether the sample data set contains sensitive data, and data portability refers to whether the data in the sample data set is portable. Computing resources include CPU resources, GPU resources, FPGA resources, etc. required for training the model.

[0032] Specifically, when determining the training features, the structure of the model to be trained can be encoded to obtain the feature components of the structure of the model to be trained. For example, when the structure of the model to be trained is a convolutional neural network, it is encoded to obtain a feature component of 100; when the structure of the model to be trained is a recurrent neural network, it is encoded to obtain a feature component of 010; when the structure of the model to be trained is a deep belief network, it is encoded to obtain a feature component of 001. For the data volume of the sample data set, the data volume of the sample data set can be divided into multiple ranges, and different ranges correspond to different codes. For example, when the data volume is in range A, the corresponding code is 100; when the data volume is in range B, the corresponding code is 010; when the data volume is in range C, the corresponding code is 001, thereby determining the feature components corresponding to the data volume of the model to be trained. For the training algorithm, it can be encoded to obtain the characteristic components of the training algorithm. For example, when the training algorithm is a gradient descent algorithm, it is encoded to obtain a characteristic component of 100; when the training algorithm is a back propagation algorithm, it is encoded to obtain a characteristic component of 010; when the training algorithm is a random gradient descent algorithm, it is encoded to obtain a characteristic component of 001. For the training requirements, the precision and recall rate can be divided into multiple ranges, and different ranges correspond to different codes. Taking the accuracy rate as an example, the accuracy rate can be divided into multiple ranges, and different ranges correspond to different codes. For example, when the accuracy rate is in range A, the corresponding code is 100; when the accuracy rate is in range B, the corresponding code is 010; when the accuracy rate is in range C, the corresponding code is 001, thereby determining the characteristic components corresponding to the accuracy rate of the model to be trained. After determining the characteristic components corresponding to the structure, data volume, training algorithm and training requirements of the model to be trained, each characteristic component is horizontally spliced ​​to obtain the training features of the model to be trained.

[0033] Furthermore, taking the neural network model as an example, the training features of the model to be trained are input into the neural network model for classification, and the probability values ​​of the model to be trained being applicable to online training, offline training and hybrid training are obtained, and then the training mode corresponding to the maximum probability value is selected as the target training mode of the model to be trained. After determining the target training mode, model training is performed based on the target training mode to obtain a trained model. The above model training process is executed by the model training module of the platform.

[0034] Step 103: determine the target deployment mode of the trained model according to the actual application conditions of the trained model and the preset deployment mode prediction model, and deploy the trained model according to the target deployment mode.

[0035] The preset deployment mode prediction model may specifically be a deployment mode classification model, such as a support vector machine, a decision tree, a random forest, a neural network model, etc. The target deployment method may specifically be one of cloud deployment, edge deployment, and hybrid deployment. Hybrid deployment is a combination of cloud deployment and edge deployment.

[0036] For the embodiment of the present invention, in order to determine the target deployment mode corresponding to the model to be trained, step 103 specifically includes: determining the real-time requirements, data volume, data security, cloud device performance, edge device performance and business elasticity of the trained model according to the actual application conditions of the trained model;

[0037] Based on the real-time requirements, data volume, data security, cloud device performance, edge device performance and business elasticity of the trained model, the deployment characteristics of the trained model are determined; the deployment characteristics are input into a preset deployment mode prediction model for deployment mode prediction to obtain the target deployment method of the trained model.

[0038] Among them, the data volume of the trained model specifically refers to the data volume of the predicted data. The real-time requirement specifically refers to the time requirement for each run of the trained model. Data security refers to whether the predicted data contains sensitive data or whether the model parameters contain sensitive parameters. Cloud device performance and edge device performance specifically include processor performance, storage performance, memory performance, network bandwidth, etc. Processor performance specifically refers to main frequency, number of cores, cache size, etc. Storage performance specifically refers to disk speed, capacity, read and write rate, etc. Memory performance specifically refers to capacity, speed, etc. Business elasticity refers to whether business expansion will be carried out in the future.

[0039] Specifically, when determining the deployment characteristics, for real-time requirements, the required running time can be divided into multiple ranges, and different ranges correspond to different codes. For example, when the required running time is in range A, the corresponding code is 100; when the required running time is in range B, the corresponding code is 010; when the required running time is in range C, the corresponding code is 001, thereby determining the characteristic components corresponding to the real-time requirements. Similarly, the characteristic components corresponding to the data volume, cloud device performance, and edge device performance can be determined. For data security, when the predicted data contains sensitive data or the model parameters include sensitive parameters, it is encoded as 01; when the predicted data does not contain sensitive data and the model parameters do not include sensitive parameters, it is encoded as 10. For business elasticity, when the model needs to be expanded in the future, it is encoded as 01; when the model does not need to be expanded in the future, it is encoded as 10. After determining the feature components corresponding to the real-time requirements, data volume, data security, cloud device performance, edge device performance, and business elasticity of the trained model, the feature components are horizontally spliced ​​to obtain the deployment characteristics of the trained model.

[0040] Furthermore, taking the neural network model as an example, the deployment features of the trained model are input into the neural network model for classification, and the probability values ​​of the trained model being suitable for cloud deployment, edge deployment, and hybrid deployment are obtained. Then, the deployment method corresponding to the maximum probability value is selected as the target deployment method of the trained model. After determining the target deployment method, the trained model is deployed based on the target deployment method. The above model deployment process is executed by the model deployment module of the platform.

[0041] Furthermore, when the model is deployed and officially runs, for cloud deployment, the platform can automatically expand or reduce computing resources according to the user's access volume and computing load to cope with changes in peak and trough periods, thereby ensuring that the model is always in the best operating state. At the same time, the platform will monitor the model's indicators and system load, and adjust resource allocation in a timely manner to ensure the stability and efficiency of the model; for edge deployment, the platform can reasonably schedule the deployment location of the model according to the computing power and network bandwidth of the edge device, and optimize the model's computing logic and communication method to reduce latency and improve response speed. The platform will also dynamically adjust the model deployment strategy according to the status of the edge device and environmental changes to cope with different scenarios and needs; for hybrid deployment, the platform will flexibly combine cloud and edge deployment methods according to the actual needs and resource conditions of the user to maximize the use of existing computing resources and optimize the model's operating efficiency. The platform will also monitor the status and performance indicators of the deployment nodes and adjust the deployment strategy in a timely manner to ensure the stable operation and efficient response of the model in different environments.

[0042] Step 104: classify the trained model according to the structure, indicators and application scenarios of the trained model, and store the trained model in a corresponding area according to the classification result.

[0043] The structure of the trained model includes convolutional neural network, recurrent neural network, deep belief network, etc. The indicators of the trained model include accuracy, recall rate, precision rate, etc. The application scenarios of the trained model include image recognition, natural language processing or recommendation system, etc. Different types of trained models are located in different storage areas of the model management module.

[0044] For the embodiment of the present invention, after the model training module completes the model training, the model management module can classify and store the trained model. For this process, step 104 specifically includes: determining the attribute characteristics of the trained model according to the structure, indicators and application scenarios of the trained model; inputting the attribute characteristics into a preset classification model for classification, and obtaining the category information of the trained model.

[0045] Among them, the preset classification model can specifically be a support vector machine, a decision tree, a random forest, a neural network model, etc.

[0046] Specifically, when determining the attribute features, the structure of the trained model can be encoded to obtain the characteristic components of the trained model structure. For example, when the structure of the trained model is a convolutional neural network, it is encoded to obtain a characteristic component of 100; when the structure of the trained model is a recurrent neural network, it is encoded to obtain a characteristic component of 010; when the structure of the trained model is a deep belief network, it is encoded to obtain a characteristic component of 001. For the indicators of the trained model, the indicator size of the trained model can be divided into multiple ranges, and different ranges correspond to different codes. For example, when the accuracy rate is in range A, the corresponding code is 100; when the accuracy rate is in range B, the corresponding code is 010; when the accuracy rate is in range C, the corresponding code is 001, thereby determining the characteristic components corresponding to the indicators of the trained model. For application scenarios, they can be encoded to obtain the characteristic components of the application scenarios. For example, when the application scenario of the trained model is image recognition, it is encoded to obtain the characteristic component 100; when the application scenario of the trained model is natural language processing, it is encoded to obtain the characteristic component 010; when the application scenario of the trained model is a recommendation system, it is encoded to obtain the characteristic component 001. After determining the characteristic components corresponding to the structure, indicators and application scenarios of the trained model, the characteristic components are horizontally spliced ​​to obtain the attribute characteristics of the trained model;

[0047] Furthermore, taking the neural network model as an example, the attribute features of the trained model are input into the neural network model for classification, and the probability values ​​of the trained model belonging to different categories are obtained, and then the category corresponding to the maximum probability value is selected as the category information of the trained model. After the category information of the trained model is determined, it is stored in the corresponding storage area.

[0048] Step 105: In response to the user's model recommendation request, determine the model recommendation factor corresponding to the user according to the user's identity information, and recommend a corresponding model to the user based on the model recommendation factor.

[0049] For the embodiment of the present invention, the model management module can not only classify and store the trained models, but also recommend models that meet the user's requirements. For this process, step 105 specifically includes: if the user is determined to be a developer according to the user's identity information, then determine whether the developer provides model description information when applying for recommendation; if the developer provides the model description information, then directly determine the structure, index and application scenario of the model required by the developer according to the model description, and determine the structure, index and application scenario of the model required by the developer as the model recommendation factor, and recommend the corresponding model to the developer based on the model recommendation factor; if the developer does not provide the model description information, then determine the target project group to which the developer belongs according to the developer's identity information; based on the target project group, infer the first application scenario of the model required by the developer, and determine the inferred first application scenario as the model recommendation factor; determine multiple first target models that meet the inferred first application scenario; sort the multiple first target models according to the index level to obtain a first sorting result, and recommend the corresponding model to the developer from the multiple first target models according to the first sorting result.

[0050] Specifically, when the user who requests model recommendation is a developer, if the developer directly provides description information, such as the structure, indicators, and application scenarios of the required model, the corresponding model will be recommended directly according to the developer's description, because the developer is a professional and his description information is relatively professional, so the structure, indicators, and application scenarios of the required model can be directly determined; if the developer does not directly provide description information, the department and project group to which the developer belongs can be determined based on the developer's identity information, such as the lung image segmentation project group, so that the model application scenario (first application scenario) required by the developer can be determined. After that, the model management module can determine all models (multiple first target models) in the application scenario, and sort each first target model according to the indicator level (such as accuracy), and finally recommend the model with the best indicator to the developer, such as recommending the model with the highest accuracy to the developer.

[0051] In some embodiments, if the user is determined to be an ordinary user based on the user's identity information, keywords in the model description information provided by the ordinary user are extracted; based on the keywords, a second application scenario of the model required by the ordinary user is inferred, and the inferred second application scenario is determined as the model recommendation factor; multiple second target models that meet the inferred second application scenario are determined; and based on the multiple second target models, corresponding models are recommended to the user.

[0052] Specifically, when the user who requests model recommendation is an ordinary user, since the description information provided by ordinary people is not professional enough, it is necessary to perform word segmentation on the description information of ordinary people, then select keywords from each word segmentation, and determine the second application scenario corresponding to the keyword from the preset scenario library. Then the model management module can determine all models (multiple second target models) under the application scenario, and recommend corresponding models to users based on the multiple second target models.

[0053] When performing model recommendation, the multiple second target models are scored respectively based on the indicators of the multiple second target models to obtain model scores corresponding to the multiple second target models respectively; the multiple second target models are sorted based on the model scores to obtain a second sorting result;

[0054] Determine user characteristics based on the activity, contribution and feedback behavior of the ordinary user; input the user characteristics into a user level prediction model to perform level prediction to obtain level information of the ordinary user; and recommend a corresponding model to the user from the multiple second target models based on the level information and the second sorting result.

[0055] The user level prediction model can be a level classification model, such as support vector machine, decision tree, random forest, neural network model, etc. Activity can refer to the number of times a user browses the platform, contribution can refer to the number of times a user suggests the platform, and feedback behavior can be reflected in the number of complaints and feedback on the platform.

[0056] Specifically, the model score can be determined according to the indicators of multiple second target models. For example, if the model accuracy is below 50%, the corresponding score is 1 point, if the model accuracy is between 50% and 80%, the corresponding score is 2 points, and if the model accuracy is above 80%, the corresponding score is 3 points. Then, according to the interval range to which the accuracy of each second target model belongs, the accuracy score of each second target model is determined. Similarly, the recall score and precision score of each second target model can be determined. Finally, the accuracy score, recall score and precision score of each second target model are added to obtain the score of each second target model. Further, according to the score, the multiple second target models are sorted from high to low to obtain the second sorting result. Then, the activity, contribution and feedback behavior of ordinary users are used as user features and input into the user level classification model for level classification. For users with higher levels, models with better indicators can be recommended for them. For example, according to the second sorting result, the model ranked first is recommended to high-level users, the model ranked in the middle is recommended to medium-level users, and the model ranked at the end is recommended to low-level users. In this way, automatic recommendation and sharing of models for users can be achieved.

[0057] Furthermore, the AI ​​model integration platform in the embodiment of the present invention also includes a user management module, which is used to manage and authenticate platform users and supports username and password authentication and OAuth authentication. At the same time, the user management module also provides user authority management and user behavior auditing functions to ensure the security of the platform.

[0058] An AI model integration method provided by an embodiment of the present invention can achieve unified management of AI models by automatically classifying and storing trained models and automatically recommending corresponding models to users based on their identities when users request recommended models. In addition, an embodiment of the present invention can automatically train and deploy models through preset training mode prediction models and preset deployment mode prediction models, thereby achieving one-stop service for model training, deployment and management, thereby saving the time cost of model development and improving model development efficiency.

[0059] Furthermore, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the implementation of this embodiment, this embodiment also provides an AI model integration platform, which adopts a microservice architecture and cloud native technology, such as Figure 2As shown, the platform includes: a model training module 1, which is used to obtain a model training task to be executed; based on the model training task and a preset training mode prediction model, a target training mode corresponding to the model to be trained is determined, and according to the target training mode, the model to be trained is trained to obtain a trained model; a model deployment module 2, which is used to determine the target deployment mode of the trained model according to the actual application conditions of the trained model and a preset deployment mode prediction model, and deploy the trained model according to the target deployment mode; a model management module 3, which is used to classify the trained model according to the structure, indicators and application scenarios of the trained model, and store the trained model in a corresponding area according to the classification result; in response to a user's model recommendation request, a model recommendation factor corresponding to the user is determined according to the user's identity information, and a corresponding model is recommended to the user based on the model recommendation factor.

[0060] Among them, the specific execution process of the model training module 1, the model deployment module 2 and the model management module 3 can refer to the method content in Example 1, and the embodiment of the present invention will not be repeated here.

[0061] In addition, the model management module 3 also provides a visual comparison and evaluation function based on model versions, so that users can easily view the improvements and differences between different versions and compare and select model performance. Users can also use the model sharing function to share their models with other users or teams on the platform to promote model sharing and collaboration. The embodiment of the present invention introduces an intelligent classification and recommendation mechanism so that the models uploaded by users can be automatically classified and recommended, thereby improving the discoverability and utilization of the model.

[0062] Furthermore, the AI ​​model integration platform in the embodiment of the present invention also includes a user management module 4, such as Figure 3 As shown in the figure, the user management module is used to manage and authenticate platform users, and supports username and password authentication and OAuth authentication. At the same time, the user management module also provides user authority management and user behavior auditing functions to ensure the security of the platform.

[0063] Furthermore, the AI ​​model integration platform uses Kubernetes for containerized deployment and management, ensuring the high availability and scalability of the platform. On this basis, the platform also uses Prometheus and Grafana for monitoring and log analysis. First, as a monitoring system, Prometheus can collect the operating data of each component of the system in real time and store it as time series data. It has a flexible query language and powerful alarm function, which can detect system anomalies and performance problems in a timely manner and take corresponding measures to deal with them. Through Prometheus, users can monitor various indicators of the system, such as CPU utilization, memory consumption, network traffic, etc., so as to fully understand the operating status of the system. Secondly, Grafana, as an open source data visualization tool, can graphically display the data collected by Prometheus and provide rich charts and panels to help users intuitively analyze the performance and trends of the system. Users can easily create customized monitoring dashboards through Grafana and set alarm rules to promptly discover and solve problems in the system. In addition, Grafana also supports a variety of data sources and plug-ins, and can be integrated with other monitoring systems to provide users with a more comprehensive and flexible monitoring solution. Therefore, by introducing Prometheus and Grafana, the cloud-native AI model integration platform not only ensures the high availability and scalability of the system, but also provides powerful monitoring and log analysis functions, which makes it easier for users to understand the system's operating status and performance indicators, discover and solve potential problems in a timely manner, and further improve the stability and reliability of the platform.

[0064] The AI ​​model integration platform of the embodiment of the present invention adopts a microservice architecture, so that resources can be dynamically allocated according to demand. This means that resources are allocated to specific tasks only when needed, thereby improving resource utilization. The platform supports multiple deployment methods, including cloud deployment, edge deployment, and hybrid deployment. This flexibility allows users to choose the most appropriate deployment strategy according to actual needs and further optimize resource utilization. The use of Kubernetes realizes containerized deployment and management, which can make more efficient use of computing resources because containers are lighter than traditional virtual machines, start faster, and can run more densely on the same hardware.

[0065] The user management module provides multiple authentication methods and supports user rights management and user behavior auditing functions. These functions help ensure that only authorized users can access the system, thereby reducing the risk of information leakage. Cloud native technology includes secure isolation of container networks to ensure that communication between services is secure. Container orchestration tools such as Kubernetes provide network security policies that can limit and isolate traffic between containers.

[0066] An AI model integration platform provided by an embodiment of the present invention can achieve unified management of AI models by automatically classifying and storing trained models and automatically recommending corresponding models to users based on their identities when users request recommended models. In addition, an embodiment of the present invention can automatically train and deploy models through preset training mode prediction models and preset deployment mode prediction models, thereby achieving one-stop service for model training, deployment and management, thereby saving the time cost of model development and improving model development efficiency.

[0067] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0068] The above serial numbers of this application are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of this application, but this application is not limited to them, and any changes that can be thought of by technicians in this field should fall within the scope of protection of this application.

Claims

1. An AI model integration method, characterized in that: Applied to the AI ​​model integration platform, which adopts microservice architecture and cloud native technology, including: Get the model training task to be executed; Based on the model training task and the preset training mode prediction model, determine the target training mode corresponding to the model to be trained, and train the model to be trained according to the target training mode to obtain a trained model; Determine the target deployment mode of the trained model according to the actual application conditions of the trained model and the preset deployment mode prediction model, and deploy the trained model according to the target deployment mode; Classify the trained model according to the structure, index and application scenario of the trained model, and store the trained model in a corresponding area according to the classification result; In response to a model recommendation request from a user, determining a model recommendation factor corresponding to the user according to the identity information of the user, and recommending a corresponding model to the user based on the model recommendation factor; Wherein, determining a model recommendation factor corresponding to the user according to the identity information of the user, and recommending a corresponding model to the user based on the model recommendation factor, includes: If the user is determined to be a developer based on the user's identity information, determining whether the developer provides model description information when applying for recommendation; If the R&D personnel provide the model description information, the structure, indicators and application scenarios of the model required by the R&D personnel are determined directly according to the model description, and the structure, indicators and application scenarios of the model required by the R&D personnel are determined as the model recommendation factors, and a corresponding model is recommended to the R&D personnel based on the model recommendation factors; If the R&D personnel does not provide the model description information, determining the target project group to which the R&D personnel belongs according to the identity information of the R&D personnel; Based on the target project group, inferring a first application scenario of the model required by the R&D personnel, and determining the inferred first application scenario as the model recommendation factor; Determine a plurality of first target models that conform to the first application scenario that is speculated; The multiple first target models are sorted according to the indicators to obtain a first sorting result, and a corresponding model is recommended to the R&D personnel from the multiple first target models according to the first sorting result.

2. The method according to claim 1, characterized in that Based on the model training task and the preset training mode prediction model, determining a target training mode corresponding to the model to be trained includes: Determine the computing resources required for the model training task based on the structure, data volume, training algorithm and training requirements of the model to be trained in the model training task; Determining training features of the model to be trained based on the computing resources and data security and data portability of the model to be trained; The training features are input into the preset training mode prediction model to perform training mode prediction, and obtain the target training mode corresponding to the model to be trained.

3. The method according to claim 1, characterized in that The trained model is classified according to the structure, index and application scenario of the trained model, including: Determining the attribute characteristics of the trained model according to the structure, indicators and application scenarios of the trained model; The attribute features are input into a preset classification model for classification to obtain category information of the trained model.

4. The method according to claim 1, characterized in that: Determining a target deployment mode of the trained model according to the actual application conditions of the trained model and a preset deployment mode prediction model includes: Determine the real-time requirements, data volume, data security, cloud device performance, edge device performance, and business elasticity of the trained model based on the actual application conditions of the trained model; Determine deployment characteristics of the trained model based on real-time requirements, data volume, data security, cloud device performance, edge device performance, and business elasticity of the trained model; The deployment features are input into a preset deployment mode prediction model to perform deployment mode prediction, and obtain the target deployment mode of the trained model.

5. The method according to claim 1, characterized in that Determining a model recommendation factor corresponding to the user according to the identity information of the user, and recommending a corresponding model to the user based on the model recommendation factor, including: If the user is determined to be a common user according to the identity information of the user, extracting keywords from the model description information provided by the common user; Based on the keyword, inferring a second application scenario of the model required by the ordinary user, and determining the inferred second application scenario as the model recommendation factor; Determine a plurality of second target models that conform to the second application scenario that is speculated; Based on the multiple second target models, a corresponding model is recommended to the user.

6. The method according to claim 5, characterized in that Based on the multiple second target models, recommending a corresponding model to the user includes: Based on the indicators of the multiple second target models, scoring the multiple second target models respectively to obtain model scores corresponding to the multiple second target models respectively; Based on the model scores, sorting the multiple second target models to obtain a second sorting result; Determine user characteristics based on the activity, contribution and feedback behavior of the ordinary user; Inputting the user characteristics into a user level prediction model to perform level prediction, thereby obtaining level information of the common user; Based on the level information and the second ranking result, a corresponding model is recommended to the user from the multiple second target models.

7. An AI model integration platform, characterized in that: The AI ​​model integration platform adopts microservice architecture and cloud native technology, including: The model training module is used to obtain the model training task to be executed; based on the model training task and the preset training mode prediction model, determine the target training mode corresponding to the model to be trained, and train the model to be trained according to the target training mode to obtain a trained model; A model deployment module, used to determine the target deployment mode of the trained model according to the actual application conditions of the trained model and the preset deployment mode prediction model, and deploy the trained model according to the target deployment mode; A model management module, configured to classify the trained model according to the structure, index and application scenario of the trained model, and store the trained model in a corresponding area according to the classification result; in response to a model recommendation request from a user, determine a model recommendation factor corresponding to the user according to the identity information of the user, and recommend a corresponding model to the user based on the model recommendation factor; The model management module is specifically used to determine whether the developer provides model description information when applying for recommendation if the developer is determined to be a developer based on the user's identity information; if the developer provides the model description information, determine the structure, indicators and application scenarios of the model required by the developer directly based on the model description, and determine the structure, indicators and application scenarios of the model required by the developer as the model recommendation factors, and recommend a corresponding model to the developer based on the model recommendation factors; if the developer does not provide the model description information, determine the target project group to which the developer belongs based on the developer's identity information; Based on the target project group, the first application scenario of the model required by the R&D personnel is inferred, and the inferred first application scenario is determined as the model recommendation factor; multiple first target models that meet the inferred first application scenario are determined; the multiple first target models are sorted according to the indicators to obtain a first sorting result, and a corresponding model is recommended to the R&D personnel from the multiple first target models according to the first sorting result.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. An electronic device comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

Images