A visual interactive artificial intelligence algorithm automatic arrangement deployment method

Abstract

The application discloses an automatic arrangement and deployment method of visual interactive artificial intelligence algorithm, and belongs to the technical field of artificial intelligence development and deployment. The method comprises the following steps: after uploading a training data set, the system automatically analyzes the format and generates a visual analysis report; importing an algorithm running environment and code, configuring parameters and required data sets through a visual interface; after a user inputs a training task, the system dynamically schedules algorithms and data nodes to automatically arrange a process and generate a directed acyclic graph; a resource intelligent scheduler allocates computing resources and monitors a training state; and after training is completed, a model API is published through a service deployment pipeline. The method can automatically arrange the training process of the algorithm, greatly shortens the training process of the algorithm, and solves the problems of the prior art, such as lack of automatic arrangement function, non-support of automatic parameter mapping, single resource scheduling strategy and inability to meet the demand of joint training of multiple institutions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence development and deployment, and specifically relates to a visual interactive artificial intelligence algorithm automatic arrangement and deployment method. BACKGROUND

[0002] The current AI technology faces the following key problems in practical application:

[0003] 1. High technical threshold:

[0004] Traditional AI development process requires professional personnel to write code for data preprocessing, model training and deployment;

[0005] Existing MLOps tools (such as Kubeflow, MLflow) still require YAML / JSON configuration, which is difficult for non-technical personnel to operate.

[0006] 2. Low process efficiency:

[0007] It takes more than 4.5 hours on average from model training to deployment;

[0008] Each link (data cleaning, parameter adjustment, service release) needs manual intervention, which is easy to produce bottleneck.

[0009] 3. Poor collaboration security:

[0010] When cross-department collaboration, sensitive data (such as medical images, military data) exist the risk of leakage;

[0011] Lack of compliance solutions that meet the requirements of "data not out of domain".

[0012] 4. Low resource utilization:

[0013] Static resource allocation results in less than 50% average utilization of GPU;

[0014] Unable to dynamically adjust resource allocation according to task priority.

[0015] The existing technology has realized containerized scheduling, but still has the following shortcomings:

[0016] Lack of automatic arrangement function;

[0017] Does not support automatic parameter mapping;

[0018] Resource scheduling strategy is single;

[0019] Unable to meet the needs of multi-institution joint training. SUMMARY

[0020] In view of the above problems, the present application provides a visual interactive artificial intelligence algorithm automatic arrangement and deployment method. The method can automatically arrange the training process of the algorithm, greatly shortening the training process of the algorithm.

[0021] To achieve the above object, the technical scheme adopted by the present application is as follows:

[0022] A visual interactive artificial intelligence algorithm automatic arrangement and deployment method, comprising the following steps:

[0023] Step 1: upload the training data set, the system automatically parses the data format and generates a visual analysis report;

[0024] Step 2: import the running environment and code of the artificial intelligence algorithm, and configure the parameters of the artificial intelligence algorithm to be deployed and the required training data set through the visual interface;

[0025] Step 3: the user inputs the algorithm task to be trained, the system dynamically schedules the algorithm nodes and data nodes to arrange the training task process, and automatically generates a directed acyclic graph;

[0026] Step 4: start the training process, the resource intelligent scheduler dynamically allocates computing resources, and monitors the training status in real time;

[0027] Step 5: after the training is completed, the model API is published through the service deployment pipeline.

[0028] Further, in step 1, for image classification artificial intelligence algorithms, the training data set is a folder named by category name, and each folder stores image files of the corresponding category; for target detection artificial intelligence algorithms, the training data set adopts COCO or VOL format; for text classification artificial intelligence algorithms, the training data set is a CSV text and corresponding label.

[0029] Further, in step 2, the imported artificial intelligence algorithm running environment is in the form of a Docker image, the system performs dependency conflict detection on the uploaded Docker image, automatically extracts the algorithm hyperparameter range through static code analysis, the parameters of the artificial intelligence algorithm include the built-in parameter list, the input parameter list and the output parameter list, the system automatically deploys the algorithm image, and registers the parameter information of the artificial intelligence algorithm in the system.

[0030] Further, in step 3, based on the agent of the large language model, the rule agent and the API calling agent, the most suitable agent is automatically selected to process the request according to the user input training task intention, the automatic arrangement module is distributed through dynamic proxy, the connection between the data node and the algorithm node is automatically constructed, and the training process is automatically arranged; the specific mode is as follows:

[0031] (1) Task reception

[0032] The user submits an artificial intelligence algorithm task request, receives this input, and begins processing;

[0033] (2) Task understanding

[0034] Analyze user input, understand the essence and goal of the task; in this stage, the memory module provides user preferences and historical training task information to help understand user intent, and uses natural language processing techniques for intent recognition and keyword extraction on user input;

[0035] (3) Task decomposition

[0036] The planning module automatically decomposes the algorithm training task into multiple subtasks, establishing task dependency relationships and execution order;

[0037] (4) Determine the agent

[0038] Determine the agent, each agent performs different types of tasks according to its specific skills and descriptions, including generative tasks based on large language models, API calls to obtain external data, local script execution of specific operations, or calls to other services;

[0039] (5) Automatic process orchestration

[0040] Automatically generate process orchestration for algorithm training tasks, including required algorithms and datasets, and finally generate a directed acyclic graph.

[0041] Further, in step 4, Kubernetes is used for resource and data scheduling, assigning a unique task ID to the directed acyclic graph workflow, and configuring disk space, memory, CPU, and GPU resources; Each algorithm node in the directed acyclic graph corresponds to an independent Docker image, and when the process runs to the current algorithm node, the corresponding Docker container will be automatically started; Data, model files, and intermediate files during the running process are transmitted through paths, and the relevant folders will be mounted to the current container.

[0042] Further, in step 5, the trained artificial intelligence algorithm model is pushed to the server through a unified interface, and the model is named at the same time, and the model path and the metadata of the current algorithm task are pushed synchronously; The metadata is transmitted to the algorithm container by the scheduling system, and then forwarded to the model server by the algorithm; Subsequently, in the visualization interface, select the target model based on the model name, enter the model information, specify or automatically assign the server port and deployment server by the system, and perform interface testing and service log viewing through the visualization interface, verify the correctness of the service, and complete the service release after the service is verified. After the release is completed, the system automatically generates API test cases to test the service.

[0043] Compared with the prior art, the present application has the following advantages:

[0044] 1) Automatic DAG arrangement and parameter automatic mapping: automatically construct the connection of data nodes and algorithm nodes according to the user input training task, generate a directed acyclic graph, improve the arrangement efficiency, and reduce human errors.

[0045] 2) Intelligent scheduling of containerized resources: combined with Kubernetes and Docker, dynamically allocate resources, improve GPU utilization to 89%, and shorten the task deployment time to 9 minutes (improve 96.7%).

[0046] 3) Standardized data management: support multiple data formats, combined with Hadoop distributed storage, realize centralized management and visual statistical analysis of data.

[0047] 4) Extensible operator integration: encapsulate custom operators through Docker images, support open extension of algorithm ecology.

[0048] 5) Low-threshold operation system: graphical interface replaces code writing, supports non-technical personnel to participate in model testing and deployment.

[0049] 6) Automatic service deployment: unified interface and metadata push, realize the efficiency and consistency of model update. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is the principle diagram of the visual artificial intelligence algorithm automatic arrangement and deployment method in the embodiment of the present application.

[0051] Figure 2 is a flowchart of training automation arrangement. DETAILED DESCRIPTION

[0052] In order to make the person skilled in the art better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the present application, so that the advantages and characteristics of the present application can be more easily understood by the person skilled in the art, and the protection scope of the present application can be more clearly and explicitly defined. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor should belong to the protection of the present application.

[0053] A visual interactive artificial intelligence algorithm automatic arrangement and deployment method, comprising the following steps:

[0054] Step 1: upload the training data set, the system automatically parses the data format and generates a visual analysis report;

[0055] In Step 1, for image classification artificial intelligence algorithms, the training dataset is a folder named by category name, and each folder stores image files of the corresponding category; for target detection artificial intelligence algorithms, the training dataset uses COCO or VOL format; for text classification artificial intelligence algorithms, the training dataset is a CSV text and corresponding label, and the system can automatically identify and process and analyze the dataset, and encrypt sensitive data according to requirements.

[0056] Step 2: Import the running environment and code of the artificial intelligence algorithm, and configure the parameters of the artificial intelligence algorithm to be deployed and the required training dataset through the visual interface;

[0057] In Step 2, the imported artificial intelligence algorithm running environment is in the form of a Docker image, and the system performs dependency conflict detection on the uploaded Docker image, automatically extracts the algorithm hyperparameter range through static code analysis, the parameters of the artificial intelligence algorithm include the built-in parameter list, the input parameter list and the output parameter list, the system will automatically deploy the algorithm image, and register the parameters of the artificial intelligence algorithm and other information in the system.

[0058] Step 3: The user inputs the algorithm task to be trained, and the system dynamically schedules appropriate algorithm nodes and data nodes to arrange the training task flow and automatically generates a directed acyclic graph;

[0059] In Step 3, based on large language model (LLM) agents, rule agents, API call agents, etc., the most suitable agent is automatically selected to process the request according to the user's input training task intent, the automatic arrangement module is allocated through dynamic proxy, and the connection between data nodes and algorithm nodes is automatically constructed. The training process is automatically arranged. As shown in Figure 2 The specific workflow is as follows:

[0060] (1) Task reception

[0061] The user submits an artificial intelligence algorithm task request, receives this input, and begins processing.

[0062] (2) Task understanding

[0063] Analyze user input, understand the essence and goal of the task. In this stage, the memory module provides user preferences and historical training task information to help more accurately understand user intent. Advanced natural language processing techniques are used for intent recognition and keyword extraction on user input.

[0064] (3) Task decomposition

[0065] The planning module automatically decomposes complex algorithm training tasks into multiple executable subtasks, establishes task dependency relationships and execution order.

[0066] (4) Determining agents

[0067] Determining agents, each agent performs different types of tasks according to its specific skills and descriptions, including large language model (LLM) based generative tasks, API calls to obtain external data, local script execution of specific operations or other service calls.

[0068] (5) Automatic process orchestration

[0069] Automatically generating process orchestration of algorithm training tasks, including required algorithms and datasets, and finally generating a directed acyclic graph.

[0070] Step 4: Start the training process, and the resource intelligent scheduler dynamically allocates computing resources and monitors the training status in real time;

[0071] In step 4, Kubernetes is used for resource and data scheduling, and a unique task ID is assigned to the directed acyclic graph workflow, and resources such as disk space, memory, CPU, and GPU are configured; Each algorithm node in the directed acyclic graph corresponds to an independent Docker image, and when the process runs to the current algorithm node, the corresponding Docker container will be automatically started; The data, model files and intermediate files during the running process are transmitted through the path, and the relevant folders will be mounted to the current container.

[0072] Step 5: After the training is completed, the model API is quickly released through the service deployment pipeline.

[0073] In step 5, the trained artificial intelligence algorithm model is pushed to the server through a unified interface, and the model is named at the same time, and the model path and the metadata of the current algorithm task are pushed synchronously; The metadata is transmitted to the algorithm container by the scheduling system, and then forwarded to the model server by the algorithm; Subsequently, in the visual interface, the target model is selected based on the model name, the model information is input, the service port is specified or automatically allocated by the system, and the deployment server is specified or automatically allocated by the system, and the interface test and service log are checked through the visual interface, and the service is verified after the service is correct. The service is released, and after the release is completed, the system can automatically generate API test cases to test the service.

[0074] The following is a more specific example:

[0075] A visual interactive artificial intelligence algorithm automatic orchestration and deployment method, as shown in Figure 1 and Figure 2 , comprising the following steps:

[0076] (1) Import the required training dataset or load the built-in training dataset, and view the sample data and statistical data in the dataset on the page to confirm the correctness and completeness of the data.

[0077] The method limits the data format according to different task types. The image classification dataset is limited to the format of a series of images corresponding to the folder and class name, the target detection is limited to the common formats such as COCO and VOL, and the text classification is limited to the CSV text corresponding to the label format, etc. When integrating a new dataset, the data needs to be converted to a format compatible with the platform first, and then uploaded. According to the task type, the number and distribution of samples and labels are statistically analyzed. When the original data cannot be obtained, the algorithm developer can understand the dataset through the statistical analysis indicators of the data, and can view part of the data of the dataset on the visual page.

[0078] (2) Import the environment and code of the artificial intelligence algorithm, and set the parameters of the artificial intelligence algorithm and the required training dataset through the visual page. In addition, built-in artificial intelligence algorithms can also be loaded directly.

[0079] The method splits the data and model into different nodes, and when connecting, the data_path (data path), num_classes (number of classes) and other fields of the data are assigned to the same fields of the model. The training process is based on these fields to obtain training data, complete the association of data and model, and the model node will be configured by the algorithm developer according to the logic of the algorithm code. Training parameters that can be modified.

[0080] Among them, the algorithm parameters are configured by the visual page, defining the built-in parameter list, input parameter list, and output parameter list, which can set default values and be changed when arranging. After completing the production of data and operator nodes, non-algorithm programming personnel with data authority can also operate through the visual page to replace the dataset, change the training parameters, etc.

[0081] (3) Input the task target of model training, the system will automatically select the appropriate algorithm and dataset, automatically arrange the data in the training dataset and the artificial intelligence algorithm, and form a DAG (Directed Acyclic Graph).

[0082] (4) Start and monitor the entire DAG process, fill in the required CPU, memory, video memory and other resources, then intelligently distribute the process tasks to the execution server nodes and divide the resources, start the task execution process, and in the execution process, real-time feedback of occupied resources, logs, intermediate files and other contents, and in the form of state icons and colors, show the state of each node of the DAG process, including not started, executing, execution success, execution failure, etc. After the entire DAG process is executed, the result log file and result model file and other contents are saved to the file storage system, and the page provides viewing and downloading of the result file. If the process is a training process, the output model is registered in the publishable model table, and the model file, model name, output time and other metadata information are saved.

[0083] The method uses kubenets to schedule resources and data. When starting to run the visual process in the page, a task ID is assigned to the entire running process, and disk space, memory, CPU and GPU are allocated. Each algorithm node corresponds to a docker image. When running to the current algorithm node, the corresponding docker container is started. Data, model files and transfer files are passed through the path type configuration parameter. The corresponding folder is mounted to the container of the current process, and the page provides viewing and downloading of the output logs and files.

[0084] For example, in the image classification task, the output parameters of the data are data_path and num_classes. After assigning values to the same field in the input parameter information of the algorithm, the data_path corresponding directory is mounted to the algorithm container through the kubenets configuration file, and the algorithm code reads the path to obtain the data content. After the algorithm completes the training, the trained model is saved to model_path, and the log is saved to log_path. They are still mapped to the disk space, and through the task ID, algorithm ID, process ID and other information, the file location is confirmed for the user to view and download on the page.

[0085] (5) The output model is intelligently deployed as a service.

[0086] After the algorithm completes the training and outputs the model, in order to realize one-key deployment, the model needs to be pushed to the server through a unified interface, and the model needs to be named. When uploading the model, the model path and the metadata of the current task of the algorithm are pushed. The metadata is sent to the algorithm container by the scheduling system, and then forwarded to the model server by the algorithm.

[0087] After training and generating the model, the user can select the model based on metadata such as model name on the page, fill in the required resources, and complete the service release. The user can specify or automatically assign the service port and service deployment server, and provide a page for interface testing and service log viewing to verify the correctness of the service.

[0088] The service release supports the configuration of resources such as memory, CPU, and GPU, and the port is allocated by the scheduling system. The code logic for model deployment is implemented in the built-in algorithm, and the deployment process is initiated by the scheduling system, with the run_type field set to deploy. The pushed model is mounted to the container and the model path is passed in. At the same time, the configuration file sent together when the model is pushed is passed into the deployment container. The algorithm developer writes code to complete the model deployment process based on the configuration parameters. Non-algorithm personnel can select models through the page to publish services.

[0089] The method can automatically add data and algorithms to the same task space and assign the same parameters between nodes through automatic wiring. Based on the kubenets mapping file path, the method can complete the transfer of data, models, logs, and other files in the model training orchestration business process. In addition, by passing the model file, training parameters, metadata, and the image containing the training and deployment code written by the developer, the one-key service release function can be completed.

[0090] The above method can be implemented by an automatic deployment system based on large model artificial intelligence algorithms. The system includes a data management module, an algorithm containerized registration center, an automatic DAG construction engine, a resource intelligent scheduler, and a service deployment module. Specifically:

[0091] (1) Data management module

[0092] (101) Provide built-in datasets for users to use directly. In the data management page, query the dataset according to the data type (text, image, etc.) and task type (image classification, object detection, semantic segmentation, etc.), and view the contents of each data in the dataset on the page.

[0093] (102) If a custom dataset is used, first browse the data format specification document, organize the data according to the specified format according to the different task types, package it as a compressed file, and upload it on the data upload page.

[0094] (103) After uploading the dataset compressed file, the system verifies the dataset file and confirms that the format is correct. The dataset is then parsed, and information such as the number of labels and the distribution of each label sample is counted and displayed on a visual page for easy user understanding of the dataset.

[0095] (104) Using hadoop as a distributed file storage system, after uploading and parsing the dataset, save the dataset to the file storage platform, and save the data path, label statistics, and other metadata to the database.

[0096] (105) Use SM4 algorithm to encrypt sensitive data for storage;

[0097] (106) Show the newly added dataset in the data management page, and provide functions such as viewing, modifying, deleting, etc.

[0098] (2) Algorithm containerization registration center

[0099] (201) Provide built-in algorithms for users to use directly, in the operator management page, according to the task type (image classification, target detection, semantic segmentation, etc.) to query the operator, and view the operator parameters, function introduction, etc. on the page.

[0100] (202) If you use custom operators, you need to first browse the operator integration specification document, write code for operator construction, data reading, operator training, evaluation, etc. and build a docker image to integrate the code running environment, and after integration, upload to the server on the operator integration page.

[0101] (203) After uploading the operator image, add the operator in the operator management page and define the operator name, parameters, function introduction, etc.

[0102] (204) After uploading the operator image, the image verification unit detects the uploaded Docker image for dependency conflicts, and can automatically extract the algorithm hyperparameter range through static code analysis;

[0103] (205) The platform automatically uploads the operator image file to the file storage system, registers the image to the image platform, and saves the operator metadata to the database.

[0104] (206) Show the newly added operator in the operator management page, and provide functions such as viewing, modifying, deleting, etc.

[0105] (3) Automatic DAG construction engine

[0106] (301) Enter the task target of model training in the automatic arrangement page, the system will automatically select the appropriate algorithm and dataset, and automatically arrange the training data set and artificial intelligence algorithm, forming a DAG (Directed Acyclic Graph).

[0107] (302) Fill in the required parameters of each node in the current process, and you can modify optional parameters, default parameters, etc.

[0108] (303) The platform automatically assigns each ID corresponding to the flow and converts the DAG flow into an adjacency list structure, which is saved to the database.

[0109] (4) Resource intelligent scheduler

[0110] (401) After building the DAG flow, fill in the required resources (CPU, memory, GPU, etc.), and publish to the interactive execution module for execution.

[0111] (402) The platform uses kubenets for resource scheduling, intelligently selects server nodes for execution, and divides resources. When starting, it automatically pulls the required code and image for the node.

[0112] (403) After starting, the interactive execution module calls the kubenets interface to real-time feedback on resource occupation and running logs of each node, and controls the node execution state to collect intermediate files produced during node operation.

[0113] (404) The resource occupation and node state are displayed in real-time by the front-end page.

[0114] (405) After the DAG flow execution is completed, the interactive execution module pushes the intermediate files, result files, etc. to the file storage system, and provides a page for viewing and downloading the files, etc.

[0115] (406) If the current flow includes a model training node, the output model file will be pushed to the unified model storage space, and the model metadata will be saved to the database.

[0116] (5) Service deployment module

[0117] (501) Select the model output and registered in the DAG flow, fill in the required resources, and perform service publishing operations.

[0118] (502) The service publishing module automatically allocates service ports and servers based on kubenets functions (or specifies during service publishing), divides resources, pulls the required code and image for the service, and executes the startup script to start the service.

[0119] (503) Provide a page to test the interface effect, fill in the interface request parameters to send a request, and display the interface feedback content and service container logs in the page.

[0120] Through actual testing, as shown in Table 1, the system has significant advantages compared to traditional solutions:

[0121] Table 1 Performance comparison data

[0122]

[0123] The method realizes automatic DAG arrangement and parameter mapping, realizes intelligent scheduling of containerized resources by combining Kubernetes and Docker, shortens the deployment time from 4.5 hours to 9 minutes, improves the GPU utilization rate from less than 50% to 89%; reduce the technical threshold through the visual interface, support non-technical personnel operation; sensitive data encryption and distributed storage are used to ensure that the data does not go out of the domain, improve the security of cross-department cooperation; at the same time, support scalable operator integration and automatic service deployment, realize efficient management of the whole life cycle of model training, and greatly improve the algorithm development efficiency and resource utilization.

[0124] In summary, the present application takes automatic arrangement as the core, the system can quickly arrange algorithms and required data sets according to the input task of the user, manage the training, testing, deployment, service release and other processes in a visual way, and realize the whole life cycle management of model training. This method greatly improves the algorithm development efficiency, reduces the technical threshold, and promotes the rapid application of artificial intelligence.

[0125] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application, any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, are included in the protection scope of the present application.

Claims

1. A method for automated orchestration and deployment of artificial intelligence algorithms with visual interaction, characterized in that, Includes the following steps: Step 1: Upload the training dataset. The system will automatically parse the data format and generate a visual analysis report. Step 2: Import the runtime environment and code of the artificial intelligence algorithm. Configure the parameters and required training dataset of the artificial intelligence algorithm to be deployed through a visual interface. The imported artificial intelligence algorithm runtime environment is in the form of a Docker image. The system performs dependency conflict detection on the uploaded Docker image and automatically extracts the hyperparameter range of the algorithm through static code analysis. The parameters of the artificial intelligence algorithm include a built-in parameter list, an input parameter list, and an output parameter list. The system automatically deploys the algorithm image and registers the parameter information of the artificial intelligence algorithm in the system. Step 3: The user inputs the algorithm task to be trained, and the system dynamically schedules algorithm nodes and data nodes to arrange the training task flow and automatically generates a directed acyclic graph. Based on large language models, proxies, rule proxies, and API call proxies automatically select the most suitable proxy to handle requests according to the user's input of the training task intent. The automatic orchestration module automatically constructs connections between data nodes and algorithm nodes through dynamic proxy allocation, automating the orchestration of the training process; the specific method is as follows: (1) Task reception The user submits an AI algorithm task request, receives this input, and begins processing it; (2) Task comprehension Analyze user input to understand the nature and objectives of the task. In this stage, the memory module provides user preferences and historical training task information to help understand user intent. Natural language processing technology is used to perform intent recognition and keyword extraction on user input. (3) Task decomposition The planning module automatically decomposes the algorithm training task into multiple sub-tasks, establishes task dependencies and execution order; (4) Determine the agent Determine the agents, each of which performs different types of tasks based on its specific skills and description, including generative tasks based on large language models, API calls to obtain external data, local scripts to perform specific operations, or calls to other services; (5) Automatic process arrangement The process of automatically generating algorithm training tasks is arranged, including the required algorithms and datasets, and finally a directed acyclic graph is generated. Step 4: Start the training process. The intelligent resource scheduler dynamically allocates computing resources and monitors the training status in real time. Kubernetes is used for resource and data scheduling. A unique task ID is assigned to the workflow of the directed acyclic graph, and disk space, memory, CPU, and graphics card resources are configured. Each algorithm node in the directed acyclic graph corresponds to an independent Docker image. When the process reaches the current algorithm node, the corresponding Docker container will be started automatically. Data, model files, and intermediate files during the process are transferred via paths, and related folders will be mounted inside the current container. Step 5: After training is complete, deploy the model API through the service-oriented deployment pipeline.

2. The method for automated orchestration and deployment of a visually interactive artificial intelligence algorithm according to claim 1, characterized in that, In step 1, for image classification AI algorithms, the training dataset consists of folders named according to category names, with each folder containing image files of the corresponding category; for object detection AI algorithms, the training dataset uses COCO or VOL format; and for text classification AI algorithms, the training dataset consists of CSV text and corresponding labels.

3. The method for automated orchestration and deployment of a visually interactive artificial intelligence algorithm according to claim 1, characterized in that, In step 5, the trained AI algorithm model is pushed to the server through a unified interface. While naming the model, the model path and metadata of the current algorithm task are pushed simultaneously. The metadata is transmitted to the algorithm container through the scheduling system, and then forwarded to the model server by the algorithm. Subsequently, the target model is selected based on the model name in the visual interface, the model information is entered, and the service port and deployment server are specified or automatically assigned by the system. Interface testing and service log viewing are performed through the visual interface. After verifying the correctness of the service, the service is published. After the publication is completed, the system automatically generates API test cases to test the service.

Images