Algorithm model management architecture and method

Abstract

The invention relates to an algorithm model management architecture and method, the algorithm model management architecture is applied to an algorithm warehouse platform, the algorithm model management architecture comprises various algorithm components, the algorithm model management architecture comprises an actuator, a configuration file class, a register, a file read-write module and a communication component, and the register registers an algorithm model through a model decorator; the configuration file class carries out unified management on configuration files of various algorithm components, the file read-write module provides various file rear ends, calling interfaces of the various algorithm components are unified through the various file rear ends, and the executor calls at least one algorithm component from the algorithm bin platform based on the unified calling interface. Different types of algorithm models are uniformly trained through at least one algorithm component and a configuration file, and a communication component coordinates data synchronization and sharing among different computing resources in distributed training of an actuator, so that the models are uniformly managed and deployed from registration to data synchronization and sharing of the algorithm models; iterative development is reduced, and deployment efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an algorithm model management architecture and method. Background Art

[0002] With the rapid development of autonomous driving technology, environmental perception methods such as object detection and semantic segmentation are crucial for the safe operation of autonomous vehicles. However, existing technical solutions have pain points in algorithm development, management, and deployment, which hinder the efficiency of technological iteration and deployment.

[0003] In existing technologies, algorithm models independently developed by different research teams use differentiated data interfaces, non-uniform evaluation metric implementation methods, and decentralized optimizer configuration processes. This makes it difficult to effectively test and verify the performance of different algorithms and select the optimal algorithm, resulting in wasted resources and inefficiency. Furthermore, the lack of a unified algorithm repository means developers often need to develop new algorithms from scratch when faced with new application scenarios, making it difficult to directly reuse existing algorithm resources. Furthermore, the algorithm models independently developed by different research teams use differentiated data interfaces, non-uniform evaluation metric implementation methods, and decentralized optimizer configuration processes. This makes it difficult to effectively test and verify the performance of different algorithms and select the optimal algorithm, resulting in wasted resources and inefficiency. Furthermore, the lack of a unified algorithm repository means developers often need to develop new algorithms from scratch when faced with new application scenarios, making it difficult to directly reuse existing algorithm resources. These problems have seriously affected the iterative development and deployment efficiency of autonomous driving technology, hindering its rapid development and widespread application.

[0004] Therefore, there is an urgent need to propose an algorithm model management architecture and method to solve the problem that existing technologies cannot uniformly manage and deploy different algorithm models, which seriously affects the iterative development and deployment efficiency of autonomous driving technology. Summary of the Invention

[0005] In view of this, it is necessary to provide an algorithm model management architecture and method to solve the technical problem that the existing technology cannot uniformly manage and deploy different algorithm models, which seriously affects the iterative development and deployment efficiency of autonomous driving technology and hinders the rapid development and widespread application of autonomous driving technology.

[0006] In order to solve the above problems, in a first aspect, the present invention provides an algorithm model management architecture, including: Applied to the algorithm warehouse platform, which includes various algorithm components, including executors, configuration file classes, registrars, file reading and writing modules, and communication components; The registrar is used to register the algorithm model through the model decorator; The configuration file class is used to uniformly manage the configuration files of the various algorithm components; The file reading and writing module is used to provide multiple file backends and unify the calling interfaces of the various algorithm components through the multiple file backends; The executor is configured to call at least one algorithm component from the algorithm warehouse platform based on the unified calling interface, and uniformly train different types of algorithm models through the at least one algorithm component and the configuration file; The communication component is used to coordinate data synchronization and sharing between different computing resources during distributed training performed by the executor.

[0007] In one possible implementation, the various algorithm components include a data loader component, an algorithm model component, and an evaluator component; The data loader component is used to load the training data set, the validation data set, and the test data set of the algorithm model according to the configuration information in the configuration file; The algorithm model component is used to perform a model training operation and a model verification operation on the algorithm model according to the training data set and the verification data set, respectively, and perform a model testing operation on the algorithm model based on the test data set to obtain a test result; The evaluator component is used to evaluate the test results according to multiple evaluation indicator categories to obtain the evaluation results of the algorithm model. When the evaluation results meet the expected results, a trained algorithm model is obtained.

[0008] In a possible implementation, the evaluator component includes an evaluator class, each evaluator class includes multiple evaluation indicator classes, and the evaluator class is used to provide format conversion and distributed communication support for the evaluation indicator classes.

[0009] In a possible implementation, the evaluator component includes an evaluator class, each evaluator class includes multiple evaluation indicator classes, and the evaluator class is located above the evaluation indicator class and is used to provide format conversion and distributed communication support for the evaluation indicator class.

[0010] In one possible implementation, the registrar is further used to call the model decorator to register the algorithm model in the algorithm model registration request as a new algorithm model when receiving the algorithm model registration request, and dynamically create an algorithm model instance through the algorithm model creation method.

[0011] In a possible implementation, the registrar is further used to create a new registrar by inheritance, and the new registrar inherits the functions of the original registrar.

[0012] In one possible implementation, the communication component is also used by the executor to initialize the distributed environment and obtain status information of the distributed environment when performing distributed training on the algorithm model, and to control and schedule distributed computing based on the status information, where the status information includes the current cluster size, node ranking, and backend information.

[0013] In a possible implementation, the actuator is further configured to set control points during the training of the algorithm model, and the control points are configured to control the operation flow of the actuator.

[0014] In a possible implementation, the algorithm model management architecture further includes a deployment module, and the deployment module includes a memory and a processor; The memory is used to download the algorithm model that needs to be deployed locally on the algorithm warehouse platform to obtain the algorithm model to be deployed and the deployment method; The processor is used to deploy the algorithm model to be deployed locally according to the deployment method.

[0015] In a second aspect, the present invention further provides an algorithm model management method, comprising: Register the algorithm model through the model decorator; Unified management of configuration files of various algorithm components; Providing multiple file backends, and unifying the calling interfaces of the various algorithm components through the multiple file backends; Calling at least one algorithm component from the algorithm warehouse platform based on a unified calling interface, and uniformly training different types of algorithm models through the at least one algorithm component and configuration files; Coordinate data synchronization and sharing between different computing resources during distributed training on executors.

[0016] The beneficial effects of the present invention are: applied to an algorithm warehouse platform, including various algorithm components, the algorithm model management architecture may include an executor, a configuration file class, a registrar, a file reading and writing module and a communication component, the registrar can be used to register the algorithm model through the model decorator, the configuration file class can be used to uniformly manage the configuration files of various algorithm components, the file reading and writing module provides a variety of file backends, and the calling interfaces of various algorithm components are unified through the multiple file backends. The executor calls at least one algorithm component from the algorithm warehouse platform based on the unified calling interface, and different types of algorithm models can be uniformly trained through at least one algorithm component and a configuration file. The communication component can be used to coordinate data synchronization and sharing between different computing resources during distributed training of the executor, so that the model can be uniformly managed and deployed from the registration of the algorithm model to the data synchronization and sharing, and there is no need to iteratively develop the same algorithm model on different platforms, thereby improving deployment efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of an embodiment of the algorithm model management architecture provided by the present invention; Figure 2 A schematic diagram of the overall structure of an embodiment of the algorithm model management architecture provided by the present invention; Figure 3 A schematic diagram of another embodiment of the algorithm model management architecture provided by the present invention; Figure 4 A schematic diagram of an embodiment of the algorithm model conversion and integration provided by the present invention; Figure 5 A flow chart of an embodiment of the algorithm model management method provided by the present invention. DETAILED DESCRIPTION

[0018] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0019] like Figure 1 As shown, a specific embodiment of the present invention discloses an algorithm model management architecture, the algorithm model management architecture 100 is applied to the algorithm warehouse platform, the algorithm warehouse platform includes various algorithm components, the algorithm model management architecture 100 includes an executor 101, a configuration file class 102, a register 103, a file reading and writing module 104 and a communication component 105; Register 103 can be used to register algorithm models using a model decorator. A model decorator is a method for managing functions or classes. These functions or classes can be registered in a global dictionary for easy management and invocation. A model decorator is essentially a high-order function that accepts a function as a parameter and returns a new function. Using a decorator, additional functionality can be dynamically added to an existing function without modifying the original code.

[0020] The configuration file class 102 can be used to uniformly manage the configuration files of various algorithm components; management can include operations such as adding, deleting, modifying, and checking the configuration files. The configuration file class can load, parse, and export configuration files in different formats, including: py files, json files, and yaml files. The configuration file class can complete the modification of the configuration file through configuration inheritance. The configuration inheritance function allows users to modify the configuration file through inheritance, which greatly improves the flexibility and maintainability of configuration management. Based on a basic configuration file, users can quickly create new configurations through inheritance and overwriting without having to repeatedly write the same configuration items. This method not only reduces the redundancy of configuration files, but also makes the configuration organization more structured and hierarchical, which facilitates the management of complex configurations in large projects.

[0021] The file read / write module 104 can be used to provide multiple file backends, unifying the call interfaces of various algorithm components and interacting with these backends. The file backend can select different storage methods based on user needs and manage files in multiple file backends. The unified interface is used for reading, writing, deleting, and copying files, thereby facilitating management. The file backend is the storage system used to actually store algorithm model files. File backends include: LocalBackend, HTTPBackend, PetrelBackend, LmdbBackend, and MemcachedBackend. File operations include read, write, delete, and copy.

[0022] The executor is a module responsible for algorithm training, testing, and verification. The executor 101 is used to integrate various algorithm components within the algorithm warehouse platform, provide a unified calling interface to the outside world, call at least one algorithm component from the algorithm warehouse platform based on the unified calling interface, and uniformly train and call different types of algorithm models by calling at least one algorithm component and configuration file to complete algorithm training, testing, and verification.

[0023] The process of training an algorithm model using an executor is as follows: the user configures the modules required for training; the training module includes defining the model structure, building a parameter scheduler, building an optimizer, and constructing a training dataset. The executor's built-in training configuration file is set to control the number of training rounds. Initializing the executor involves passing the modules required for training and the training .cfg file to the executor. Calling the executor's training method automatically allows the executor to execute the training task.

[0024] The process of testing an algorithm model using an executor is as follows: the user configures the modules required for testing; the testing module includes defining the model structure, building a test dataset, and establishing evaluation metrics. Initializing the executor involves passing the modules required for testing to the executor. Calling the executor's test method automatically executes the test task.

[0025] The process of using the executor to validate an algorithm model is as follows: the user completes the configuration of the training modules; the training module includes defining the model structure, building a parameter scheduler, building an optimizer, and constructing a training dataset. The user also completes the configuration of the validation modules; the validation module includes constructing validation metrics, building an evaluator, and constructing a validation dataset. The user then sets the executor's built-in training configuration file; the training configuration file can set the number of training rounds, validation interval, and number of validations. Initializing the executor involves passing the training modules, validation modules, and training configuration file to the executor. The executor's training method is then called, and the executor automatically performs the training and validation tasks.

[0026] The functions supported by the executor also include: configuration file-based calling, environment configuration, distributed training, and module customization; configuration file-based calling, users write all modules into the configuration file and pass the configuration file to the executor. The executor can then complete the training, testing, and verification of the algorithm model based on the contents of the configuration file; environment configuration, users can use the environment configuration file to set the randomness of the algorithm training process and adjust the underlying settings such as multi-process related settings; distributed training, the executor automatically supports distributed training; module customization, users can customize all modules in the executor.

[0027] The communication component 105 is used to coordinate data synchronization and sharing between different computing resources during distributed training on the executor. Users can perform distributed training without having to worry about the underlying communication implementation.

[0028] Compared with the prior art, the algorithm model management architecture provided in this embodiment is applied to the algorithm warehouse platform, including various algorithm components. The algorithm model management architecture may include an executor, a configuration file class, a registrar, a file reading and writing module, and a communication component. The registrar can be used to register the algorithm model through the model decorator. The configuration file class can be used to uniformly manage the configuration files of various algorithm components. The file reading and writing module provides multiple file backends, and the calling interfaces of various algorithm components are unified through multiple file backends. The executor calls at least one algorithm component from the algorithm warehouse platform based on the unified calling interface. Different types of algorithm models can be uniformly trained through at least one algorithm component and a configuration file. The communication component can be used to coordinate data synchronization and sharing between different computing resources during distributed training of the executor, so that the model can be uniformly managed and deployed from the registration of the algorithm model to the data synchronization and sharing. There is no need to iteratively develop the same algorithm model on different platforms, thereby improving deployment efficiency.

[0029] In some embodiments of the present invention, the executor 101 is also used to set control points during the training of the algorithm model. The control points are used to control the operating process of the executor. The operating process may include general functions in controlling the execution process, such as logging, counting iteration time, saving weights at specified intervals, and clearing the PyTorch CUDA cache. The control point function enhances the flexibility and controllability of the executor in the training, testing, and verification processes. It allows specific control points to be set in the execution process to implement general functions such as logging, counting iteration time, saving weights at specified intervals, and clearing the PyTorch CUDA cache. This design makes the algorithm execution process more controllable, facilitates monitoring and optimization, and improves the efficiency of algorithm development and debugging.

[0030] In some embodiments of the present invention, the various algorithm components include a data loader component, an algorithm model component, and an evaluator component; The data loader component is used to load the training dataset, validation dataset, and test dataset of the algorithm model according to the configuration information in the configuration file. The executor provides the dataset base class. In addition to supporting the native PyTorch dataset class, the dataset base class supports Lazyinitialize and annotation serialization, and provides a unified interface to the outside world. The Lazyinitialize function allows the dataset to be initialized only when actually needed, which can reduce memory usage and improve system resource utilization efficiency. The annotation serialization function can speed up the data reading and processing process. These two functions together improve the efficiency and flexibility of data loading, especially when processing large-scale datasets, which can significantly improve the training and testing speed of the algorithm.

[0031] The algorithm model component is used to calculate the loss and return the test results. The calculated loss is used to calculate the gradient and parameter update. The algorithm model is trained and verified based on the training data set and the validation data set respectively. The algorithm model is tested based on the test data set to obtain the test results. The prediction results are used to cooperate with the evaluator to calculate the corresponding evaluation indicators. The evaluator component is used for model evaluation. It evaluates the test results according to multiple evaluation indicators to obtain the evaluation results of the algorithm model. If the evaluation results meet the expected results, the trained algorithm model is obtained.

[0032] In some embodiments of the present invention, the evaluator component includes an evaluation indicator class and an evaluator class. Each evaluator class may contain multiple evaluation indicator classes and can evaluate model performance at multiple scales. The evaluation indicator class is used to calculate model-specific accuracy indicators based on test data and model prediction results. The evaluator class is located at the upper layer of the evaluation indicator class and is used to provide format conversion and distributed communication support for the evaluation indicator class.

[0033] The multi-scale evaluation functionality of this embodiment allows for evaluating model performance at different scales, which is particularly important for object detection and semantic segmentation tasks in autonomous driving scenarios. By evaluating models at multiple scales, a more comprehensive understanding of their performance on objects of varying sizes is achieved, helping developers optimize their models for complex real-world scenarios. This evaluation approach improves the comprehensiveness and accuracy of model evaluation, contributing to the development of more robust and versatile autonomous driving algorithms.

[0034] In some embodiments of the present invention, the various algorithm components further include an optimizer component and a parameter scheduler component; The optimizer in the algorithm component is uniformly encapsulated and used to be directly called in the data loader algorithm model, centralizing the algorithm optimization process in the model algorithm. The optimizer component can be used to optimize the model parameters of the algorithm model during the training process of the algorithm model; The parameter scheduler component can be used to adjust the hyperparameters of the optimizer component during the algorithm model training process.

[0035] In some embodiments of the present invention, the registrar 103 is also used to provide a unified registration mechanism for different types of algorithm models. When an algorithm model registration request is received, the registration mechanism calls the model decorator to register the algorithm model in the algorithm model registration request as a new algorithm model, and dynamically creates an algorithm model instance through the algorithm model creation method. The user calls the registrar method to define the MODEL type, and registers the algorithm model class to be registered as a MODEL type model through the @MODEL.register_module decorator. The algorithm model instance can then be dynamically created through the MODEL.build(model_cfg) method.

[0036] In some embodiments of the present invention, the registrar 103 is also used to create a new registrar by inheritance, and the new registrar inherits the functions of the original registrar. The registrar can also create a new registrar by inheritance under the defined MODEL type; the new registrar inherits the functions of the MODEL registrar and allows the registration of new algorithm models based on it.

[0037] In some embodiments of the present invention, the communication component 105 is also used by the executor 101 to initialize the distributed environment when performing distributed training on the algorithm model, to ensure effective communication between different nodes and GPUs, and to obtain status information of the distributed environment, to control and schedule distributed computing based on the status information, where the status information includes the current cluster size, node ranking, and backend information.

[0038] Further, such as Figure 2As shown, the algorithm model management architecture of the embodiment of the present invention may include an executor, a configuration file class, a register, a file reading and writing module, and a communication component; the executor may call various algorithm components in the configuration file class, and the various algorithm components include a data loader, an algorithm model, and an evaluator; the data loader component loads the training data set, the verification data set, and the test data set of the algorithm model according to the configuration information in the configuration file; the parameter scheduler is used to adjust the hyperparameters of the optimizer component during the algorithm model training process. The optimizer component is used to optimize the model parameters of the algorithm model during the algorithm model training process, and encapsulate and update the algorithm model to obtain the algorithm model; the algorithm model component is used to perform model training operations and model verification operations on the algorithm model according to the training data set and the verification data set, and perform model testing operations on the algorithm model based on the test data set to obtain test results; the evaluator is used to evaluate the test results according to multiple evaluation index classes, and obtain the trained algorithm model when the evaluation results meet the expected results. The registrar is used to register the algorithm model through the model decorator; the configuration file class is used to uniformly manage the configuration files of various algorithm components; the file reading and writing module is used to provide multiple file backends, unify the interface through multiple file backends, and manage files in multiple file backends; the communication component is used to coordinate data synchronization and sharing between different computing resources during distributed training of the executor.

[0039] In some embodiments of the present invention, Figure 3 As shown, the algorithm model management architecture further includes a deployment module 106, and the deployment module 106 includes a memory 301 and a processor 302; Memory 301, used to store the algorithm model to be deployed and the deployment method, and store the program that implements the algorithm warehouse platform and the deployment method; Processor 302 is used to deploy the algorithm model to be deployed according to the deployment method, execute the program to implement the algorithm warehouse platform and the deployment method, so as to implement the steps of the algorithm warehouse platform and the deployment method.

[0040] Furthermore, the process of deploying the algorithm model is as follows: the user downloads the algorithm deployment tool to the local computer. The user uses the script command to install the algorithm deployment tool and ONNX. The command calls the tool's model conversion script to convert the input model format to the model format required by the target device inference engine, such as Figure 4As shown, this includes converting PyTorch models to torch2onnx, quant, and onnx2backend, converting ONNX models to inference backend models, and converting conversion collections to ONNX, backend models, Meta Info, TorchScript, and device-independent IR models. The model conversion results are integrated, including the inference backend model, model metadata, and inference metadata. The integrated results are saved in a designated folder. Instructions are used to call the SDK inference portion of the tool, select the desired inference language, and perform inference on the integrated results. The SDK inference portion encapsulates the model's pre-processing, inference, and post-processing processes, providing a multi-language model inference interface. These languages ​​include C / C++, C#, Python, and Java.

[0041] In a specific embodiment of the present invention, specific object detection and semantic segmentation algorithms are implemented based on the infrastructure and deployment methods. These algorithms include: Swin Transformer, YOLOv5, DiffusionDet, VoxelNet, CenterPoint, PV-RCNN, UNet, ViT, Fast-RCNN, and others.

[0042] This embodiment of the present invention integrates a variety of object detection and semantic segmentation algorithms for autonomous driving. Users can call required algorithms from the algorithm warehouse through functional components provided by the platform infrastructure, eliminating the need for repetitive development. This significantly reduces development workload, improves algorithm reusability, and enhances development efficiency and costs. Furthermore, the dataset base class provided by the executor within the infrastructure is compatible with the data structures of multiple mainstream datasets and custom datasets, facilitating the selection of optimal algorithms in simulations and real-world applications. Furthermore, by iterating the algorithm warehouse platform version or allowing users to locally register algorithms through the registrar, built-in algorithms can be added, removed, and optimized, enhancing the applicability of the autonomous driving software algorithm warehouse platform in various application scenarios. Furthermore, using the algorithm warehouse platform's model deployment tools, users can quickly generate SDKs adapted for specific hardware platforms directly from trained models, significantly reducing the workload of manual adjustments and adaptations. Therefore, this embodiment of the present invention offers significant technical advantages and promising applications, providing strong support for the rapid development and widespread application of autonomous driving technology.

[0043] In order to better implement the algorithm model management architecture in the embodiment of the present invention, based on the algorithm model management architecture, the embodiment of the present invention also provides an algorithm model management method, such as Figure 5 As shown, including: S501. Register the algorithm model through the model decorator; S502. Perform unified management on configuration files of various algorithm components; S503. Provide multiple file backends to unify the calling interfaces of various algorithm components through the multiple file backends; S504: Call at least one algorithm component from the algorithm warehouse platform based on a unified calling interface, and uniformly train different types of algorithm models using the at least one algorithm component and a configuration file; S505. Coordinate data synchronization and sharing between different computing resources during distributed training on the executor.

[0044] The algorithm model management method provided in the above embodiment can implement the technical solution described in the above algorithm model management architecture embodiment. The specific implementation principles of the above modules or units can be found in the corresponding content in the above algorithm model management architecture embodiment, which will not be repeated here.

[0045] The above is a detailed introduction to the algorithm model management architecture and method provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for technical personnel in this field, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. An algorithm model management architecture, characterized by: Applied to the algorithm warehouse platform, which includes various algorithm components, including executors, configuration file classes, registrars, file reading and writing modules, and communication components; The registrar is used to register the algorithm model through the model decorator; The configuration file class is used to uniformly manage the configuration files of the various algorithm components; The file reading and writing module is used to provide multiple file backends and unify the calling interfaces of the various algorithm components through the multiple file backends; The executor is configured to call at least one algorithm component from the algorithm warehouse platform based on the unified calling interface, and uniformly train different types of algorithm models through the at least one algorithm component and the configuration file; The communication component is used to coordinate data synchronization and sharing between different computing resources during distributed training performed by the executor.

2. The algorithm model management architecture according to claim 1, characterized in that: The various algorithm components include a data loader component, an algorithm model component, and an evaluator component; The data loader component is used to load the training data set, the validation data set, and the test data set of the algorithm model according to the configuration information in the configuration file; The algorithm model component is used to perform a model training operation and a model verification operation on the algorithm model according to the training data set and the verification data set, respectively, and perform a model testing operation on the algorithm model based on the test data set to obtain a test result; The evaluator component is used to evaluate the test results according to multiple evaluation indicator categories to obtain the evaluation results of the algorithm model. When the evaluation results meet the expected results, a trained algorithm model is obtained.

3. The algorithm model management architecture according to claim 2, characterized in that: The evaluator component includes an evaluator class, each evaluator class includes multiple evaluation indicator classes, and the evaluator class is used to provide format conversion and distributed communication support for the evaluation indicator class.

4. The algorithm model management architecture according to claim 2, characterized in that: The various algorithm components also include an optimizer component and a parameter scheduler component; The optimizer component is used to optimize the model parameters of the algorithm model during the training process of the algorithm model; The parameter scheduler component is used to adjust the hyperparameters of the optimizer component during the algorithm model training process.

5. The algorithm model management architecture according to claim 1, characterized in that: The registrar is also used to call the model decorator to register the algorithm model in the algorithm model registration request as a new algorithm model when receiving the algorithm model registration request, and dynamically create an algorithm model instance through the algorithm model creation method.

6. The algorithm model management architecture according to claim 1, characterized in that: The registrar is also used to create a new registrar by inheritance, and the new registrar inherits the functions of the original registrar.

7. The algorithm model management architecture according to claim 1, characterized in that: The communication component is also used by the executor to initialize the distributed environment and obtain status information of the distributed environment when performing distributed training on the algorithm model, and to control and schedule distributed computing based on the status information, wherein the status information includes the current cluster size, node ranking, and backend information.

8. The algorithm model management architecture according to claim 1, characterized in that: The actuator is also used to set control points during the training of the algorithm model, and the control points are used to control the operation process of the actuator.

9. The algorithm model management architecture according to claim 1, characterized in that: The algorithm model management architecture further includes a deployment module, and the deployment module includes a memory and a processor; The memory is used to download the algorithm model that needs to be deployed locally on the algorithm warehouse platform to obtain the algorithm model to be deployed and the deployment method; The processor is used to deploy the algorithm model to be deployed locally according to the deployment method.

10. An algorithm model management method, characterized in that: include: Register the algorithm model through the model decorator; Unified management of configuration files of various algorithm components; Providing multiple file backends, and unifying the calling interfaces of the various algorithm components through the multiple file backends; Calling at least one algorithm component from the algorithm warehouse platform based on a unified calling interface, and uniformly training different types of algorithm models through the at least one algorithm component and the configuration file; Coordinate data synchronization and sharing between different computing resources during distributed training on executors.

Images