Algorithm online and operation control method for multi-algorithm link and program product

Abstract

According to the algorithm online and operation control method and program product for the multi-algorithm link, efficient deployment under the multi-algorithm link is achieved through an automatic data label of a data warehouse, flexible training task management of a task processing module, model algorithm version control, layered packaging and an intelligent scheduling mechanism, and the algorithm online and operation control method and program product for the multi-algorithm link are provided. Namely, automatic online and deployment under a multi-algorithm link; under the action of the multi-layer detachable algorithm package, each algorithm module belongs to the respective layer to independently run and can be dynamically updated, so that the flexibility, the maintainability and the expansibility are greatly improved; meanwhile, efficient execution of reasoning tasks and optimized utilization of resources are ensured through automatic online and hierarchical scheduling, rapid adjustment and switching are achieved for multi-algorithm links, and the method has remarkable technical advantages and can be suitable for various complex artificial intelligence application scenes.

Description

Technical Field

[0001] This application relates to the field of machine learning application technology, specifically to a method and computer program product for algorithm deployment and operation control in multi-algorithm chains. Background Technology

[0002] In many applications with inference and prediction capabilities, especially in fields such as computer vision, autonomous driving, and intelligent monitoring, multiple algorithm modules typically work together to complete the inference task. In other words, the response to the inference task required by the application often involves different algorithms, such as object detection, object tracking, behavior recognition, and temporal prediction, with each algorithm module corresponding to a different stage of the inference task processing.

[0003] However, with the continuous development of algorithms and the expansion of applications, existing algorithm implementations and deployments have many shortcomings in supporting application functionality. Inference and prediction capabilities in applications often can be achieved through multiple algorithmic pathways; however, the deployment and implementation of any algorithmic pathway requires manual configuration and adjustment, a process that is cumbersome, error-prone, and consumes significant development costs. Furthermore, it lacks the capability for rapid adjustment and switching between multiple algorithmic pathways. Summary of the Invention

[0004] One objective of this application is to enable automated deployment and online operation under multiple algorithm links, and to enable rapid adjustment and switching of inference and prediction capabilities in applications across multiple algorithm links.

[0005] According to one aspect of the embodiments of this application, a method for algorithm deployment and operation control for multi-algorithm chains is disclosed, characterized in that the method includes: The raw data is automatically labeled using a data warehouse, and the resulting labeled data is then solidified into a new version of the dataset. The training task created by the task management module specifies the dataset version, target algorithm link, and corresponding algorithm module, and executes the created training task on the specified algorithm module using the labeled data in the specified version dataset. After the training task is completed, a new version of the corresponding model algorithm is obtained. The new version of the model algorithm and other algorithm modules called for the training task are then adapted to the target algorithm link topology and encapsulated into a multi-layer decomposable algorithm package, where each layer corresponds to an algorithm module on the target algorithm link. By using a multi-layered, detachable algorithm package that can be deployed online, the algorithm sequence of the target algorithm chain is automatically deployed online, and the deployed algorithm sequence is scheduled and executed for inference tasks in a hierarchical manner.

[0006] According to one aspect of the embodiments of this application, a computer device is disclosed, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described above.

[0007] According to one aspect of the embodiments of this application, a computer program product is disclosed, including a computer program that, when executed by a processor, implements the steps of the method as described above.

[0008] According to one aspect of the embodiments of this application, a computer-readable storage medium is disclosed having a computer program stored thereon that, when executed by a processor, implements the steps of the method as described above.

[0009] The data warehouse of this application features automated data labeling, flexible training task management in the task processing module, model algorithm version control, layered encapsulation, and intelligent scheduling mechanisms. These mechanisms enable efficient deployment across multiple algorithm chains, i.e., automated online deployment and deployment across multiple algorithm chains. Furthermore, the multi-layered, detachable algorithm packages allow each algorithm module to operate independently at its own level and be dynamically updated, greatly improving flexibility, maintainability, and scalability. Simultaneously, automated online deployment and layered scheduling ensure efficient execution of inference tasks and optimized resource utilization. This allows for rapid adjustment and switching across multiple algorithm chains, demonstrating significant technical advantages and making it applicable to various complex artificial intelligence application scenarios.

[0010] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0011] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application. Attached Figure Description

[0012] The above and other objectives, features and advantages of this application will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0013] Figure 1 This is a flowchart of an algorithm deployment and operation control method for multiple algorithm paths provided in the embodiments of this application.

[0014] Figure 2 It is based on Figure 1 The corresponding embodiment shows a flowchart describing the steps of specifying the dataset version, target algorithm link, and corresponding algorithm module for the training task created by the task management module, and executing the created training task using the labeled data in the specified version dataset.

[0015] Figure 3 It is based on Figure 1The corresponding embodiment shows a flowchart describing the steps of obtaining a new version of the corresponding model algorithm after the training task is completed, and adapting the new version of the model algorithm and other algorithm modules called for the training task to the target algorithm link topology and encapsulating them layer by layer into a multi-layer decomposable algorithm package.

[0016] Figure 4 It is based on Figure 1 The corresponding embodiment shows a flowchart describing the steps of automatically uploading the algorithm sequence of the target algorithm link online through an uploadable multi-layered detachable algorithm package.

[0017] Figure 5 This is a flowchart illustrating a method for algorithm deployment and operation control for a multi-algorithm link, according to another exemplary embodiment of this application.

[0018] Figure 6 This is a flowchart illustrating another exemplary embodiment of the present application, describing a method for algorithm deployment and operation control of a multi-algorithm link.

[0019] Figure 7 This is a schematic diagram of the functional architecture of an algorithm platform, illustrated based on a specific example.

[0020] Figure 8 It is based on Figure 7 A schematic diagram of the algorithm platform interface shown in the corresponding embodiment.

[0021] Figure 9 It is based on Figure 7 The corresponding embodiment shows a visual interface diagram of the algorithm platform monitoring the operation of the inference service. Detailed Implementation

[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make the description of this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The drawings are merely illustrative of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted.

[0023] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of exemplary embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced with one or more of the specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0024] Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0025] For the inference capabilities required by an application, there are multiple different algorithm chains to complete the same or similar inference needs. However, to adapt to different needs and performance requirements, the most suitable algorithm chain needs to be selected in actual deployment and online operation.

[0026] An algorithm chain contains several algorithm modules, such as an algorithm module for object detection and an algorithm module for time series prediction; after the algorithm chain is deployed online, the resulting algorithm sequence can realize the sequential execution of each algorithm module.

[0027] However, the process of developing an algorithm chain to achieve the required reasoning capabilities and then deploying the resulting algorithm sequence online requires a lot of manpower and time, and it cannot quickly implement reasoning capabilities for the various applications that need to be built.

[0028] Given the continuous release of numerous applications and the iteration of each application, there is an urgent need to solve the problem of enabling each application to quickly deploy inference capabilities based on the applicable algorithm chain, without incurring significant development costs, and to achieve efficient delivery of application development.

[0029] To this end, an algorithm platform is built to enable the rapid development and deployment of inference services for various applications, thereby achieving customized development of the inference capabilities required by each application and significantly shortening the deployment cycle of inference services.

[0030] In one exemplary embodiment, the algorithm platform includes at least a data warehouse, a task management module, an algorithm warehouse, and an online inference module. The data warehouse is used for automated processing and version control of the dataset, which is used for training the algorithm module. To this end, the automated processing of the data warehouse includes automated frame extraction of the original data and automated labeling of the resulting image groups.

[0031] The data warehouse will solidify the initial uploaded raw data and the labeled data obtained from the automated annotation process into new versions, while inheriting previous versions, in order to achieve data version iteration and provide selectable data versions for the training of algorithm modules.

[0032] The task management module is responsible for training the algorithm modules, and based on this, obtains new versions of the model algorithms that can be deployed online from the algorithm repository. The algorithm repository is used to encapsulate the algorithm modules in the target algorithm chain and the new versions of the currently trained model algorithms in a layered manner to obtain multi-layered decomposable algorithm packages. Through these multi-layered decomposable algorithm packages, the online inference module automatically deploys the algorithm sequences of the target algorithm chain online.

[0033] Correspondingly, the algorithm deployment and operation control method for multi-algorithm links implemented in this application will take the algorithm platform as the execution subject, and provide external applications with efficient inference capabilities through the collaborative work of various modules of the algorithm platform.

[0034] See Figure 1 , Figure 1 This is a flowchart of an algorithm deployment and operation control method for multiple algorithm paths provided in the embodiments of this application.

[0035] The algorithm deployment and operation control method for multiple algorithm paths provided in this application includes: Step S110: Perform automated annotation on the original data through the data warehouse, and solidify the resulting annotated data into a new version dataset; Step S120: Specify the dataset version, target algorithm link, and corresponding algorithm module for the training task created by the task management module, and execute the created training task on the specified algorithm module using the labeled data in the specified version dataset. Step S130: After the training task is completed, obtain the corresponding new version of the model algorithm, and adapt the new version of the model algorithm and other algorithm modules called for the execution of the training task to the target algorithm link topology structure and encapsulate them into a multi-layer decomposable algorithm package, with each layer corresponding to an algorithm module on the target algorithm link. Step S140: The algorithm sequence of the target algorithm link is automatically uploaded through the updatable multi-layered and detachable algorithm package. The uploaded algorithm sequence is scheduled to execute inference tasks in a hierarchical manner.

[0036] These steps are explained in detail below.

[0037] In step S110, the data warehouse serves as a unified data entry point and enables distributed storage of uploaded datasets. For example, the raw data uploaded by users is received through the web console provided by the algorithm platform and stored as the initial version of the dataset. Subsequent automated annotation of this raw data yields a new version of the dataset, inheriting from the initial version.

[0038] For example, raw data can be image data or video data. The raw data and / or labeled data, as datasets comprising entity data, are stored in the data warehouse by constructing data description information. For instance, a dataset containing video data would have its data description information including dataset name, version number, annotation scheme, frame extraction strategy, and annotation status to support transaction control and versioning control related to the dataset. Based on this, the entity data carried by the dataset, such as video data, is stored in a distributed manner along with the data description information.

[0039] The data warehouse, which serves as the data entry point, performs automated annotation on the datasets it holds, thereby obtaining new versions of the dataset for data version iteration and constructing and storing data description information.

[0040] Therefore, for the subsequent training tasks implemented by the algorithm platform, there is sufficient labeled data, and the specific version of the dataset can be selected through version selection to adapt to the training process.

[0041] In summary, the data warehouse in the algorithm platform solidifies multiple versions of labeled data and supports training tasks to select the dataset to use based on the version. As data continues to accumulate, the implementation of automated labeling can be continuously optimized, and higher quality data versions can be gradually built. Subsequent training tasks can select the latest or optimal version to achieve continuous iteration of algorithm performance.

[0042] Furthermore, for algorithm platforms that enable flexible deployment of algorithms, the support of multiple data versions can meet the different data requirements of business implementation in different applications. For example, one model algorithm training task requires nighttime data, another model algorithm training task requires industrial inspection data, and yet another model algorithm training task requires traffic video data. All of these data requirements can be met with the support of the algorithm platform. Versioned datasets allow the algorithm platform to customize training data for different training tasks of different applications, support the coexistence of multiple model algorithms and customized development, and ultimately improve the adaptability of the provided inference services and applications.

[0043] Furthermore, when the algorithm platform performs parallel training of multiple models and multiple tasks, the deployment of multi-version datasets in the data warehouse can avoid resource conflicts, realize the safe reuse of data across multiple tasks, and significantly improve development efficiency.

[0044] The automated annotation of the data warehouse supports various types of object detection, including rotated object detection, instance segmentation, object tracking and detection, and behavior detection. In addition, as a supplement, existing algorithms can be used for automated annotation, followed by manual correction to improve efficiency. The annotated data is then solidified into a new version dataset, providing the necessary data description information. For example, for a new version dataset formed from annotated data, if the original data is video data, the distributed storage data description information includes the version number, frame extraction strategy, annotation scheme, and label set. Correspondingly, the annotated data includes the target location, category, and confidence score. The target location includes the bounding box location information; for example, this could be the coordinates of the top-left corner of the bounding box, as well as its width and height. The category refers to the corresponding image's category label, and the confidence score is used to support subsequent object tracking.

[0045] In addition, the labeled data also includes optional attribute fields such as the target identifier of the located target and the scene to which it belongs, which are not limited here. The labeled data obtained by automatically annotating the original data, i.e., the initial version of the dataset, serves as a new version of the dataset based on this initial version. Therefore, the corresponding images, i.e., the images in the original data, can be loaded through the path mapping of the labeled data, thereby meeting the needs of subsequent training.

[0046] Thus, for a data warehouse, automated processing can be implemented based on the original data, i.e., the initial version of the dataset, to continuously improve the performance of the subsequently obtained labeled data and iterate the version continuously.

[0047] In one exemplary embodiment, automated annotation in the data warehouse is implemented using a pre-built target detection algorithm on the algorithm platform. The data warehouse invokes the pre-built target detection algorithm to locate and classify targets in the image, obtains candidate detection boxes and category labels corresponding to the targets, and then generates annotation data based on the candidate detection boxes and category labels.

[0048] In one exemplary embodiment, when the raw data of the dataset is video data, the automated processing performed by the data warehouse includes video frame extraction and object detection.

[0049] The process of performing video frame extraction on the uploaded raw data includes: the data warehouse uses a combination of fixed time intervals and scene change constraints to perform video frame extraction on the raw data to obtain image groups; and the algorithm module is called to perform automated annotation on each frame of the image group to obtain the image annotation data.

[0050] Specifically, for video data, the algorithm platform automatically triggers frame extraction after detecting that the uploaded data is video data. Considering the temporal continuity requirements of subsequent target tracking and prediction, a combined frame extraction strategy of fixed time interval and scene change constraints is adopted. In one exemplary embodiment, the execution process of fixed time interval frame extraction performed by the data warehouse includes: setting the frame extraction interval, i.e., the frame extraction step size, according to the original video frame rate and the business-expected time resolution (expected temporal analysis resolution), which is the product of the original video frame rate and the business-expected time resolution; and then extracting frames from the video data according to the frame extraction step size to obtain effective analysis frames, i.e., image groups used for target detection.

[0051] By extracting frames at fixed time intervals to ensure uniform video sampling, and by recognizing scene changes, such as intersections or congested scenes with significant local changes, frame extraction is adapted to scene change constraints to increase the extraction of key frames and ensure the complete capture of key actions and drastically changing areas, while skipping some frames to save storage.

[0052] The image groups obtained by sampling frames at fixed time intervals and the image groups obtained by sampling frames to adapt to scene changes are combined and deduplicated to obtain a set of video data for target detection.

[0053] In an exemplary embodiment, the frame extraction process of the data warehouse performing scene change constraints includes: performing inter-frame difference on the video data, determining whether the inter-frame difference is less than zero, and if the inter-frame difference of the current frame relative to the next frame is less than zero, then the next frame can be skipped until the inter-frame difference is greater than zero and the video frame is extracted.

[0054] In one exemplary embodiment, a frame skipping constraint based on the scene histogram is superimposed on the inter-frame differential to avoid the influence of noise. Specifically, the changes in the scene histogram between frames are detected, and the scene histogram is used to evaluate whether to skip frames. If the change in the scene histogram is less than a threshold, frame skipping is performed, and the corresponding frame is not extracted.

[0055] Therefore, by using the constraints of inter-frame difference and scene histogram, we can avoid frame skipping caused by the inter-frame difference being susceptible to noise. For example, if the difference between two images is mistakenly determined to be greater than zero due to only a change in illumination, then frame skipping will be implemented, resulting in redundancy in the resulting image group.

[0056] The images obtained by frame extraction are stored in chronological order to form an image group, and corresponding data description data is constructed to record for each frame: (1) source video identifier, frame number, timestamp; (2) frame extraction strategy parameters, such as original video frame rate, expected temporal analysis resolution, frame extraction step size, whether scene change constraints are used, etc., which will not be listed one by one here.

[0057] For the image group obtained by frame extraction, the data warehouse can directly call specific algorithm modules to perform automated labeling to generate labeled data for the image group.

[0058] In an exemplary embodiment, the execution process of step S110 includes: the data warehouse iterating through the original data stored in the initial version of the dataset to obtain a new version of the dataset by triggering the data version when the original data is uploaded and / or when the automated label is completed.

[0059] In other words, the data warehouse of the algorithm platform implements version control for the data held in two processes: when the original data is uploaded and when an automated annotation is completed. That is, the uploaded original data is used as the initial version dataset for storage. Subsequent automated annotation, including but not limited to the process of re-annotating the original data after automated annotation optimization, also solidifies the labeled data output by each automated annotation into a new version dataset based on the previous version. In this way, the version of data available for training based on the initially uploaded original data is also continuously updated, which enables the subsequent training of algorithm modules to be optimized, iteratively obtaining new versions of algorithm models, and realizing adaptive iteration of the deployed algorithm models.

[0060] Based on the various versions of datasets held by the data warehouse, such as a specific version of labeled data, the task processing module will create a training task for the algorithm module corresponding to the specified target algorithm link, and specify the data version of the dataset used for training the created training task. Finally, the created training task will be executed using the labeled data in the specified version of the dataset, as in step S120.

[0061] In step S120, the task management module is used to execute the training process of the algorithm modules corresponding to the target algorithm link. It should be noted that the target algorithm link is the specified algorithm link used to implement the inference service required by the external application. The target algorithm link includes several corresponding algorithm modules and the execution order between the algorithm modules; and the specific algorithm modules included in the target algorithm link need to be trained in order to have the inference ability suitable for the external application.

[0062] Based on this, for the algorithm modules included in the target algorithm chain, a training task is created through the task management module, then a dataset version is specified for the created training task, and then the training task is executed on the algorithm module using the labeled data of the specified dataset version.

[0063] For users, the task management module of the algorithm platform enables the platform to monitor the training process and generate model weight snapshots corresponding to various indicators while calling on computing power to execute training tasks. It also displays indicators such as loss curves, mAP (mean average precision), and resource usage in real time, making the training process visible.

[0064] The task management module is used to create training tasks, select algorithms, and choose data versions, thereby enabling no-code, visual algorithm model building and training. It should be understood that for a dataset whose original data is video data, this initial version of the dataset has different versions of labeled data. The different versions of labeled data will be used to adapt the training of the time series prediction algorithm to different business needs, so as to ensure that the algorithm platform can achieve flexible development and online deployment of the adapted algorithms for different applications.

[0065] See Figure 2 , Figure 2 It is based on Figure 1 The corresponding embodiment shows a flowchart describing the steps of specifying the dataset version, target algorithm link, and corresponding algorithm module for the training task created by the task management module, and executing the created training task using the labeled data in the specified version dataset.

[0066] The step S120 provided in this application embodiment, which specifies the dataset version, target algorithm link, and corresponding algorithm module for the training task created by the task management module, and executes the created training task on the specified algorithm module using the labeled data in the specified version dataset, includes: Step S121: Specify the target algorithm link and the corresponding algorithm module on the target algorithm link through the task management module. The algorithm module is a preset model algorithm and / or an existing model. Step S122: Create training tasks for the algorithm modules in the target algorithm chain and select a dataset version for the training tasks; Step S123: Execute the training task using the dataset version, and train the algorithm modules of the target algorithm through the execution of the training task.

[0067] These steps are explained in detail below.

[0068] The selected algorithm modules are either pre-configured by the algorithm platform or imported locally, thereby providing usable algorithms directly and quickly for the application's algorithm implementation, such as target tracking algorithm modules and time series prediction algorithm modules.

[0069] A training task is created, corresponding to a specified data version and algorithm model. This algorithm model serves two purposes: firstly, it executes the training task, and secondly, it provides features and class labels for the execution of the training task. In other words, the algorithm model can be functionally divided into an algorithm module for performing annotation, an algorithm module for performing detection to construct features, and an algorithm module for implementing inference and prediction. Therefore, for the created training task, on the one hand, it is necessary to implement annotation and feature construction through the algorithm modules on the target algorithm chain; on the other hand, it is necessary to execute the created training task on the algorithm module for implementing inference and prediction using the constructed features and the obtained labeled data.

[0070] In the execution of training tasks, in order to solve the problem of computing resource contention when multiple tasks are running in parallel, each training task is run in an isolated container environment. Resources are released when the training task is completed. Under the control of the container environment, dependency conflicts and port conflicts between different training tasks are avoided. Also, the automatic release of resources after task completion prevents zombie processes from occupying computing power for a long time.

[0071] As the training task is executed, the parameter states corresponding to the training time are fully recorded to generate model weight snapshots corresponding to the checkpoints.

[0072] When the training task is completed, each algorithm module on the target algorithm link is encapsulated in step S130. It should be clear that the encapsulated algorithm modules include the new version model algorithm obtained by performing training tasks on specific algorithm modules on the target algorithm link.

[0073] In step S130, the algorithm platform's algorithm repository encapsulates the new version of the obtained model algorithm and other algorithm modules called for executing the training task, i.e. other algorithm modules on the target algorithm chain, layer by layer and packages them together to obtain a multi-layered decomposable algorithm package.

[0074] The multi-layered decomposable algorithm package refers to encapsulating the various algorithm modules involved in the algorithm chain in a hierarchical structure. Each layer represents an independent algorithm module, and these layers are loosely coupled, allowing for independent disassembly, replacement, or upgrades. Each layer of algorithm modules interacts through clearly defined interfaces and data structures, ensuring high maintainability and scalability of the system.

[0075] By encapsulating multiple algorithm modules at different levels, each level becomes an independent unit, decoupling them from one another. Each level can independently handle its own algorithmic tasks, reducing inter-layer dependencies and making the inference services provided by the algorithm platform to external applications more flexible in terms of algorithm updates, version management, and scheduling.

[0076] Furthermore, due to the acquisition of multi-layered decomposable algorithm packages and the deployment of algorithm sequences, the operation of the inference service is oriented towards each algorithm level and the layers are decoupled from each other. This allows for the selection of the most suitable inference path to adapt to the scenarios of input data from external applications, and dynamic hierarchical scheduling for inference implementation. This ensures that the implemented inference service is not constrained by rigid algorithm design and fixed architecture, supports intelligent scheduling, and greatly improves operating efficiency and adaptability.

[0077] Furthermore, the implementation of multi-layered decomposable algorithm packages and version control of model algorithms enable the inference service provided to external applications to support version control and rollback. That is, if the performance of the new version of the model algorithm is not as expected, it can be quickly rolled back to the previous version without affecting the normal operation of the entire inference chain, and other layers are not affected during the update.

[0078] In summary, the availability of multi-layered, decomposable algorithm packages provides significant flexibility and room for innovation in the iterative updates, customized scheduling, and resource optimization management of algorithms deployed on the algorithm platform. Whether in optimizing the algorithm chain, scheduling inference, or adapting to hardware resources, a high degree of intelligence and automation can be achieved, thereby greatly improving performance and adaptability.

[0079] See Figure 3 , Figure 3 It is based on Figure 1 The corresponding embodiment shows a flowchart describing the steps of obtaining a new version of the corresponding model algorithm after the training task is completed, and adapting the new version of the model algorithm and other algorithm modules called for the training task to the target algorithm link topology and encapsulating them layer by layer into a multi-layer decomposable algorithm package.

[0080] The step S130 provided in this application embodiment, which involves obtaining a new version of the corresponding model algorithm after the training task is completed, and adapting the new version of the model algorithm and other algorithm modules called for executing the training task to the target algorithm link topology and encapsulating them layer by layer into a multi-layer decomposable algorithm package, includes: Step S131: In response to the completion of the training task, the algorithm repository automatically pulls back the model weight snapshot corresponding to the best performance during the training process. The model weight snapshot includes a complete record of the parameter state of the time series prediction algorithm at a specific training time. Step S132: Load the model weight snapshot to restore and obtain a new version of the model algorithm that can be deployed online; Step S133: The model algorithm and other algorithm modules that the training task depends on are abstracted into independent and encapsulated algorithm layers, and the algorithm layers are stacked hierarchically to adapt to the target algorithm link topology to obtain a multi-layer decomposable algorithm package.

[0081] These steps are explained in detail below.

[0082] Based on the parameter states recorded at each time point during the training process, the system automatically retrieves the corresponding snapshot of the best-performing model weights to restore the new version of the algorithm model with the best performance. This process is implemented through the automated execution of the algorithm repository, without the need for manual log viewing or selection.

[0083] After restoring the new version of the algorithm model by loading the model weight snapshot, the new version of the algorithm model is renamed according to a fixed specification, and packaged with the model weights, inference configuration, training logs, environment dependency information, etc. into an algorithm package to obtain the inference prediction layer for deployment.

[0084] Correspondingly, for example, the target detection algorithm will be encapsulated as a detection layer, the target tracking algorithm will be encapsulated as a tracking layer, and finally the detection layer, tracking layer, and inference prediction layer will be formed into a multi-layered decomposable algorithm package.

[0085] The obtained multi-layered detachable algorithm package can be deployed online with one click under the action of the online inference module of the algorithm platform, so as to provide inference services for the inference implementation of external applications, as in step S140.

[0086] In one exemplary embodiment, the multi-layered decomposable algorithm package, as an algorithm version, will run independently, that is, independently in the created inference service container, and expose the service through a unified inference interface. Thus, for the application, the obtained inference service is decoupled from its own business logic, which will greatly enhance the reliability of operation.

[0087] Furthermore, in an exemplary embodiment, the execution process of step S133 includes: generating algorithm metadata to describe the new version of the online model algorithm obtained from the recovery; and performing an algorithm repository entry operation on the new version of the online model algorithm based on the algorithm metadata and the algorithm file set of the online model algorithm, wherein the new version of the model algorithm is used to replace or iterate the specified algorithm modules to form its new algorithm version.

[0088] Specifically, after restoring and obtaining a new, deployable version of the model algorithm, the first step is to generate algorithm metadata. This metadata may include the model algorithm version number, the corresponding algorithm module type, algorithm performance metrics, algorithm dependencies, and input and output formats.

[0089] Metadata descriptions of new versions of model algorithms can clearly record the algorithm's basic information, performance characteristics, and application requirements, providing accurate data support for subsequent algorithm management, version control, and updates.

[0090] The generated algorithm metadata, along with the set of new version model algorithm files that can be deployed online, i.e., the set of algorithm files (such as model weight files and training configuration files), are all stored in the algorithm repository of the algorithm platform. The new version model algorithm stored in the repository will be used to replace or iterate the specified algorithm module, forming a new version of the algorithm module, thereby automatically iterating and upgrading the implemented inference service to ensure that the inference service can always be in the optimal state.

[0091] After obtaining the multi-layered detachable algorithm package through step S130, the algorithm sequence of the target algorithm link can be automatically uploaded through the execution of step S140.

[0092] In step S140, for the multi-layered detachable algorithm package that can be put online, the multi-layered detachable algorithm package is automatically pushed to the online inference module through the algorithm repository. This triggers the online inference module to parse the multi-layered detachable algorithm package and create corresponding inference instances for each algorithm level, thereby building an inference service and enabling the online deployment of the algorithm sequence corresponding to the target algorithm link.

[0093] It should be understood that the algorithm sequence corresponding to the target algorithm link is the algorithm layer encapsulated in the multi-layer decomposable algorithm package, as well as the execution order between the algorithm layers. Therefore, the deployment of the multi-layer decomposable algorithm package in the online inference module and the implementation of the corresponding inference service realize the online operation of its corresponding algorithm sequence.

[0094] See Figure 4 , Figure 4 It is based on Figure 1 The corresponding embodiment shows a flowchart describing the steps of automatically uploading the algorithm sequence of the target algorithm link online through an uploadable multi-layered detachable algorithm package.

[0095] The step S140 of automatically uploading the algorithm sequence of the target algorithm link through the uploadable multi-layered and detachable algorithm package provided in this application embodiment includes: Step S141: Unpack the multi-layer detachable algorithm package and load the parsed algorithm layers in the inference service container in a hierarchical manner, creating inference instances for each layer; Step S142: Based on the inference service construction performed by the created inference instance, the algorithm sequence of the target algorithm link is automatically launched online.

[0096] Furthermore, the execution process of step 142 includes: First, after creating inference instances corresponding to each algorithm layer, the inference service is started and the inference instances are registered to the routing and scheduling module of the inference service. The routing and scheduling module is used to establish a schedulable execution mapping relationship between the algorithm layer and the inference service. Secondly, as the inference service enters the ready state and completes the automatic online deployment of the target algorithm link algorithm sequence, the deployed algorithm sequence is subject to hierarchical scheduling of the inference service through the routing and scheduling module.

[0097] This exemplary embodiment will now be described in detail.

[0098] In this exemplary embodiment, the multi-layered detachable algorithm package is an algorithm version that has been packaged by the algorithm platform and is ready to be deployed online. It can be stored in the algorithm repository for deployment when needed, or it can be directly deployed to the online inference module, thereby providing the inference services required by external applications in real time.

[0099] The algorithm repository automatically pushes the obtained multi-layered, decomposable algorithm packages to the online inference module. Correspondingly, the online inference module, in response to the automatic push of the multi-layered, decomposable algorithm packages, starts and runs the inference service container. The started and running inference service provides an independent runtime environment for the multi-layered, decomposable algorithm packages.

[0100] The online inference module of the inference service container is started to parse the multi-layered decomposable algorithm package to obtain several encapsulated algorithm layers, which are then loaded layer by layer into the running inference service container. Within the inference service container, corresponding inference instances are created for each layer. An inference instance is a runtime unit that carries out the execution of its corresponding algorithm layer's tasks, thereby ensuring the independent operation and flexible scheduling of each algorithm layer.

[0101] Therefore, inference services accessible to external applications can be implemented based on inference instances, and application access control can be implemented based on the published inference interface and key.

[0102] After loading the multi-layered decomposable algorithm and building inference instances for each algorithm layer, the inference service can be started to handle external inference calls.

[0103] When the inference service starts, it first creates an internal routing and scheduling module, which is used to maintain the identifier, version information and entry function of each inference instance. Then, it registers the created inference instances at each layer to the routing and scheduling module in the form of "inference nodes" to form a schedulable inference structure.

[0104] After registration, the inference service initializes its communication capabilities, automatically generating and exposing the inference interfaces corresponding to each algorithm layer, and simultaneously generating corresponding security keys or access tokens. At this point, the inference service enters a ready state, capable of parsing and scheduling the required inference instances based on inference requests from external applications, and returning inference results. It should be understood that the configuration of the inference interfaces corresponding to each algorithm layer allows for flexible scheduling of inference request responses across different algorithm layers through inference interface calls.

[0105] In other words, see Figure 5 , Figure 5 This is a flowchart illustrating a method for algorithm deployment and operation control for a multi-algorithm link, according to another exemplary embodiment of this application.

[0106] The algorithm deployment and operation control method for multi-algorithm chains provided in this application includes: Step S210: The inference service receives an external inference request and determines the inference path by adapting the input data in the external inference request. Step S220: Perform hierarchical scheduling of the online algorithm sequence according to the inference path, and execute the inference task through the algorithm layer mapped by the scheduled level.

[0107] The following is a detailed explanation of these two steps.

[0108] After completing the layered loading of multi-layered, decomposable algorithm packages and constructing the corresponding inference instances, the online inference module can then run the inference service. The inference service provides combined inference capabilities across several algorithm layers for inference requests initiated by external applications.

[0109] In step S210, after receiving the external inference from the application, the inference service first performs structured parsing on the input data carried in the request to generate an input frame semantic embedding vector, so as to provide a basis for subsequent scene perception and scene-aware inference path selection.

[0110] In one exemplary embodiment, the type of input data includes, but is not limited to, image data, video data, sensor data, etc. Regardless of the type of input data, it corresponds to a time series, and thus the corresponding sequence, such as an image sequence, can be obtained through data parsing.

[0111] The parsed sequence is used to roughly perceive the target location, and the scene state parameters corresponding to each image are estimated based on the roughly perceived target location. In an exemplary embodiment, the estimated scene state parameters include the number of targets in the frame, target density, motion change parameters, trajectory stability parameters, and trajectory variability, which will not be listed one by one here.

[0112] Among them, target density is the number and proportion of targets roughly perceived in the image; motion change parameter is the displacement change of targets roughly perceived in the image; trajectory stability parameter can be the acceleration variance of the current trajectory, and trajectory fluctuation is the fluctuation index of the most recent few frames of images, such as velocity variance.

[0113] The inference path is selected based on the estimated scenario state parameters to obtain a dynamic inference path selection strategy adapted to external inference requests. The dynamic inference path selection strategy indicates the inference path of the inference service. The layers loaded into the inference service container, namely the detection layer, tracking layer, and timing prediction layer as mentioned above, can serve as scheduling nodes on the inference path.

[0114] For example, if the algorithm layers loaded by the running inference service container include an object detection layer, an object tracking layer, and a temporal prediction layer, then the resulting inference path includes, but is not limited to, running only the object detection layer, running from the object detection layer to the object tracking layer, running only the object tracking layer, and running from the object detection layer to the object tracking layer to the temporal prediction layer.

[0115] Therefore, specific inference paths can be mapped based on scene state parameters, serving as the current dynamic inference path selection strategy for the inference service. If the number of targets within a frame is less than a set threshold, it indicates that the targets within the frame are sparse. The distribution of the detected targets across images is the target trajectory to be tracked, meaning only the target detection layer is run. The number of targets within a frame less than the set threshold is mapped to an inference path that only runs the detection layer. This inference path is used as the current dynamic inference path selection strategy, thus eliminating the need to execute the entire inference path for hierarchical scheduling of inference. This significantly saves resources and time while ensuring the accuracy of the target trajectory tracking.

[0116] If the number of targets within the frame is dense (the number of targets within the frame is greater than a set threshold), the inference path is mapped to the target detection layer - target tracking layer, so that the target trajectory can be obtained.

[0117] For example, if the trajectory stability parameter indicates a smooth trajectory, it maps to an inference path that ignores the target detection layer and only runs the tracking layer. Balanced target trajectories mean that the target's appearance features and geometric position across frames are predictable. This significantly reduces the incremental value of implementing detection, and the execution of target tracking already has enough information to maintain target perception. The target tracking layer can predict the target position for the next frame, and errors will not accumulate. The resulting speed advantage is significant without sacrificing accuracy.

[0118] In summary, under stable trajectory conditions, the target tracking layer can achieve high-precision position inference based on the target's continuous historical state. At this time, the information gain generated by the detection layer is limited, and it brings additional computational burden and false detection risk. Therefore, the implementation of dynamic inference path can skip the target detection layer and only run the tracking path, i.e., the target tracking layer, to achieve low latency and high stability inference effect.

[0119] For example, if the trajectory fluctuation indicates large trajectory fluctuations, that is, strong changes in target behavior, then it is mapped to the inference path of target detection layer - target tracking layer - time series prediction layer to ensure the accuracy of the inference effect when the target line changes greatly.

[0120] The target detection layer, target tracking layer, and time prediction layer loaded into the inference service container each create various inference instances, thus enabling them to run independently. Specifically, step S210 is executed to perform hierarchical scheduling on the inference paths mapped by the dynamic inference path selection strategy, so as to run the corresponding inference instances according to the scheduling nodes on the inference path.

[0121] In summary, based on the scheduling nodes on the inference path, the corresponding inference instances are arranged into an execution chain with a linear or branch structure. Then, the inference results are obtained by executing each layer in the order of implementation and returned to the external application that initiated the inference request, thus realizing real-time response to inference requirements.

[0122] Thus, by using the layer-by-layer online algorithm, and the layer-by-layer loading and the inference instance created for each layer, the algorithm is deployed flexibly, and the resulting inference service is made to achieve decomposable model-level collaboration, thus building a cross-layer optimization mechanism that can automatically adapt to lightweight or full inference paths based on input data.

[0123] In another exemplary embodiment, the hierarchical scheduling execution for inference, for inference paths that do not run the timing prediction layer, will trigger the activation of the timing prediction layer when the velocity variance in the features constructed by the target trajectory exceeds a set threshold, or when the target trajectory is abnormal (such as a sudden acceleration change). This reduces a large amount of unnecessary timing inference during the operation of the inference service, thereby saving 40% to 70% of computing resources and reducing latency to 20% to 50%.

[0124] See Figure 6 , Figure 6 This is a flowchart illustrating another exemplary embodiment of the present application, describing a method for algorithm deployment and operation control of a multi-algorithm link.

[0125] The algorithm deployment and operation control method for multi-algorithm chains provided in this application includes: Step S310: Monitor and provide feedback on the operation of the inference service, and publish indicators and issue automatic alarms for the online multi-layer detachable algorithm packages. Step S320: Adapt to the published metrics and / or alarms, and make adaptive adjustments to the running inference service, including adjustments to the inference service corresponding to each layer of inference instances.

[0126] The following will explain these two steps.

[0127] In step S310, the algorithm platform continuously monitors and provides feedback on the entire operational status of the inference service. Specifically, during actual operation, the inference service collects key metrics in real time, including but not limited to: inference time, call frequency, inference success rate, cache hit rate, memory and GPU memory usage, trajectory stability metrics, detection confidence distribution, trajectory drift rate, and time-series prediction deviation for each layer (target detection layer, target tracking layer, and time-series prediction layer). Based on these metrics, a multi-dimensional operational profile is constructed, and combined with preset threshold rules, sliding window statistics, and scene change judgment strategies, the inference performance, model health, and scene status are automatically evaluated.

[0128] When situations such as exceeding limits, performance degradation, abnormal trajectory fluctuations, inference request backlog, or abnormal algorithm-level output are detected, corresponding operational metrics and alarm events will be automatically released. These metrics and alarms are pushed to the alarm center of the algorithm platform and can trigger subsequent actions such as log recording, event tracking, and inference link heatmap analysis to provide reliable data support for subsequent adaptive adjustment decisions.

[0129] In step S320, the algorithm platform automatically triggers an adaptive adjustment strategy based on the metrics and / or alarms released in step S330, thereby intelligently optimizing the currently running inference service.

[0130] Specifically, the system analyzes alarms or changes in indicators, including situations such as decreased output quality of the target detection layer, increased drift rate of the target tracking layer, increased error of the prediction layer, sudden increase in inference time, or abnormal hardware resource usage, and dynamically selects appropriate adjustment schemes based on the analysis results.

[0131] The adjustments to the scheme include, but are not limited to: (1) Adjustment at the inference instance level, i.e., dynamically scaling up and down the service threads, GPU session quantity, cache size, priority queue configuration, etc. of each inference instance to adapt to changes in real-time inference load; (2) Dynamic inference path selection strategy scheduling, i.e., when a decrease in detection quality or an increase in trajectory drift rate is detected, the algorithm platform automatically increases the frequency of the detection layer or forces the execution of the detection layer in the next round of inference to perform trajectory recalibration; when the load is high or the scene is stable, the detection frequency is automatically reduced and the tracking layer is executed first to reduce latency; (3) Hot replacement of model layers, i.e., when a decrease in inference performance or an increase in error is detected at a certain layer, it supports automatic switching from the algorithm repository to the alternative model, lightweight model, or the latest version of the model layer at the same level to ensure the stability of inference output quality.

[0132] Through an adaptive adjustment mechanism that adapts to the entire inference service process, the algorithm platform can respond to fluctuations and anomalies in real time without human intervention, ensuring the high availability, stability and multi-scenario adaptability of the inference service.

[0133] The algorithm deployment and inference service operation implemented layer by layer in this application embodiment enable the algorithms used by the inference service to be updated locally without the need for overall recompilation as in existing implementations, or the need for overall redeployment of the entire algorithm package.

[0134] Specifically, as a new version of the algorithm corresponding to a model layer is added to the algorithm repository, a version update check will be performed on the new version algorithm. If the versions are compatible and the configurations are consistent, the inference service container will stop running the corresponding inference instance, and the model layer encapsulated by the new version algorithm will be loaded into the inference service container to bind to the currently stopped inference instance to resume inference.

[0135] Thus, the entire update process does not affect the operation of the corresponding inference instances in other layers, and the inference service does not need to interrupt the current process to implement the update.

[0136] The following section will illustrate the implementation of algorithm deployment and operation control for multi-algorithm chains in this application, using the specific algorithm platform implemented as an example.

[0137] See Figure 7 , Figure 7 This is a schematic diagram illustrating the functional architecture of an algorithm platform based on a specific example. The algorithm platform comprises a data warehouse, a task management module, an algorithm warehouse, an online inference module, and a recognition ledger module.

[0138] The data warehouse is oriented towards the uploaded raw data. Under the supported automatic annotation, it realizes the version control of the dataset. That is, the obtained raw data is the initial version dataset, the labeled data obtained by automatic annotation is the new version dataset that continues the initial version dataset, and the original data is automatically re-annotated again as the annotation optimization is implemented, thus carrying out the version iteration of the labeled data.

[0139] The task management module is mainly used for algorithm training, that is, to execute training tasks for the algorithm modules that need to be trained in the online algorithm sequence, and the required labeled data is provided by the data warehouse.

[0140] For users who need to deploy a sequence of algorithms for an application, the task management module enables them to create training tasks for the algorithm model through a visual parameter configuration interface, and monitor the training process. This makes the training process visible, and the new version of the model algorithm with the best performance is automatically obtained based on the monitored training process.

[0141] At this point, given the obtained new version of the model algorithm, the algorithm repository can be used to execute the new version of the model algorithm, and to add other algorithm modules corresponding to the deployed algorithm sequence to the repository, and manage algorithm versions, such as... Figure 8 As shown, Figure 8 It is based on Figure 7 A schematic diagram of the algorithm platform interface shown in the corresponding embodiment.

[0142] The algorithm repository encapsulates algorithms, and each algorithm model is packaged as an independent algorithm layer to obtain a multi-layered, decomposable algorithm package, which is then pushed to the online inference module to launch the algorithm sequence.

[0143] Once deployed, the algorithm sequence can provide inference services to external applications, monitor their operation, and output corresponding logs, such as published metrics and / or alerts. Figure 9 As shown, Figure 9 It is based on Figure 7 The corresponding embodiment shows a visual interface diagram of the algorithm platform monitoring the operation of the inference service.

[0144] In one exemplary embodiment, this application also provides a computer device including a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the steps of the method as described above.

[0145] In one exemplary embodiment, this application also provides a computer program product including a computer program that, when executed by a processor, implements the steps of the method as described above.

[0146] In one exemplary embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method as described above.

[0147] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this application.

[0148] In an exemplary embodiment of this application, a computer program medium is also provided, on which computer-readable instructions are stored, which, when executed by a computer's processor, cause the computer to perform the methods described in the above method embodiments.

[0149] According to one embodiment of this application, a program product for implementing the methods in the above-described method embodiments is also provided. This product may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of this invention is not limited thereto. In this document, a readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0150] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0151] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0152] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0153] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0154] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0155] Furthermore, although the steps of the method in this application are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0156] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0157] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the appended claims.

Claims

1. A method for algorithm deployment and operation control in multi-algorithm chains, characterized in that, The method includes: The raw data is automatically labeled using a data warehouse, and the resulting labeled data is then solidified into a new version of the dataset. The training task created by the task management module specifies the dataset version, target algorithm link, and corresponding algorithm module, and executes the created training task on the specified algorithm module using the labeled data in the specified version dataset. After the training task is completed, a new version of the corresponding model algorithm is obtained. The new version of the model algorithm and other algorithm modules called for the training task are then adapted to the target algorithm link topology and encapsulated into a multi-layer decomposable algorithm package, where each layer corresponds to an algorithm module on the target algorithm link. By using a multi-layered, detachable algorithm package that can be deployed online, the algorithm sequence of the target algorithm chain is automatically deployed online, and the deployed algorithm sequence is scheduled and executed for inference tasks in a hierarchical manner.

2. The method according to claim 1, characterized in that, The step of performing automated annotation on the original data through a data warehouse, and then solidifying the resulting annotated data into a new version dataset, includes: The data warehouse iterates through the data version of the original data stored in the initial version of the dataset to obtain a new version dataset, triggered when the original data is uploaded and / or when the automated standard optimization is completed.

3. The method according to claim 1, characterized in that, The raw data is video data, and the automated annotation performed includes: The data warehouse uses a combination of fixed time intervals and scene change constraints to perform video frame extraction to obtain image groups from the original data. The algorithm module is invoked to perform automated annotation on each frame of the image group to obtain the annotation data of the image.

4. The method according to claim 1, characterized in that, The step of specifying the dataset version, target algorithm link, and corresponding algorithm module for the training task created by the task management module, and executing the created training task on the specified algorithm module using the labeled data in the specified version dataset, includes: The target algorithm chain and the corresponding algorithm module on the target algorithm chain are specified through the task management module. The algorithm module is a preset model algorithm and / or an existing model. Create training tasks for the algorithm modules in the target algorithm chain, and select a dataset version for the training tasks; The training task is performed using the dataset version, and the algorithm modules on the target algorithm chain are trained through the execution of the training task.

5. The method according to claim 1, characterized in that, After the training task is completed, a new version of the corresponding model algorithm is obtained. This new version of the model algorithm, along with other algorithm modules called for the training task, is then encapsulated layer by layer into a multi-layered, decomposable algorithm package, adapted to the target algorithm link topology. This package includes: In response to the completion of the training task, the algorithm repository automatically retrieves the model weight snapshot corresponding to the best performance during the training process. The model weight snapshot includes a complete record of the parameter state of the time series prediction algorithm at a specific training moment. Loading the model weight snapshot restores a new, deployable version of the model algorithm; The model algorithm and other algorithm modules that the training task depends on are abstracted into independent and encapsulated algorithm layers, and the algorithm layers are stacked hierarchically to adapt to the target algorithm link topology to obtain a multi-layer decomposable algorithm package.

6. The method according to claim 5, characterized in that, The process of obtaining a new version of the corresponding model algorithm after the training task is completed, and adapting the new version of the model algorithm and other algorithm modules called for executing the training task to the target algorithm link topology and encapsulating them layer by layer into a multi-layer decomposable algorithm package, further includes: For the recovered new version of the online model algorithm, generate algorithm metadata to describe the new version of the online model algorithm; Based on the algorithm metadata and the set of algorithm files for the new version of the online model algorithm, an algorithm repository entry operation is performed on the new version of the online model algorithm. The new version of the model algorithm is used to replace or iterate the specified algorithm modules to form its new algorithm version.

7. The method according to claim 1, characterized in that, The automatic uploading of the algorithm sequence of the target algorithm link through the uploadable multi-layered and detachable algorithm package includes: In response to the automatic push of multi-layered detachable algorithm packages to the online inference module, the inference terminal starts and runs the inference service container; The multi-layered decomposable algorithm package is unpacked, and the unpacked algorithm layers are loaded hierarchically within the inference service container, creating inference instances for each layer. The algorithm sequence of the target algorithm link is automatically deployed to the inference service built based on the inference instance that has been created.

8. The method according to claim 1, characterized in that, The method includes: The operation of the inference service is monitored and feedback is provided, and the indicators and automatic alarms are released for the online multi-layer detachable algorithm packages. Adapt to the published metrics and / or alerts, and make adaptive adjustments to the running inference service, including adjustments to the inference service corresponding to the inference instances at each layer.

9. A computer program product, characterized in that, The computer program product is used to implement the algorithm platform; The algorithm platform includes a data warehouse, a task management module, an algorithm warehouse, and an online inference module; The data warehouse is used to automatically obtain labeled data from labeled datasets and to implement version control on the labeled data to allow selection of specific versions of labeled data; The task management module is used to create training tasks for the algorithm module and execute the training tasks based on the selected version of labeled data. The algorithm repository is used to encapsulate algorithm modules into multi-layered, decomposable algorithm packages when the training task is completed. The online inference module is used to perform one-click deployment of the multi-layered decomposable algorithm package and to provide inference services to the application by responding to inference requests.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.

Images