An automatic arrangement and construction method for a deep learning task

Abstract

A kind of automation arrangement and construction method for deep learning task. It is related to the technical field of deep learning process arrangement. It includes successively connected functional algorithm integrated component, module intelligent recommendation component, task flow arrangement component, task execution decision component and model service generation component;The functional algorithm integrated component is used to unify and integrate management of typical algorithm and basic function module;The module intelligent recommendation component is used to realize the automatic recommendation selection of algorithm and function module when task flow arrangement;The task flow arrangement component is used to convert the arrangement result based on visualization into process configuration description file that can be parsed by task execution decision component;The present application can effectively improve system resource utilization, realize deep learning task automation operation, complete model one-key service and version rapid iterative update.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning process orchestration, and in particular to an automatic orchestration and construction method for deep learning tasks. BACKGROUND

[0002] A typical deep learning process usually includes data preprocessing, model training, hyperparameter tuning, model evaluation, model deployment, model inference and other main links. In the traditional deep learning development process, the functions are not clearly divided in data processing and model generation, and some functions are repeatedly implemented, resulting in low reusability of algorithms or modules. For different business scenarios, experienced engineers need to design special processes.

[0003] At the same time, with the rise of the AI industry, a large number of special AI chip architectures have emerged, including NPU, XPU, GPGPU, GPU, etc. Generally, deep learning training tasks have relatively high computing and network requirements. How to fully utilize the AI chip computing power and reasonably schedule deep learning tasks for efficient operation requires both extensive and professional technical requirements. SUMMARY

[0004] The present application aims to at least solve one of the technical problems existing in the prior art.

[0005] The technical solution of the present application is: an automatic orchestration and construction method for deep learning tasks, comprising a function algorithm integration component, a module intelligent recommendation component, a task flow orchestration component, a task execution decision component and a model service generation component connected in turn;

[0006] The function algorithm integration component is used to integrate and manage typical algorithms and basic function modules uniformly;

[0007] The module intelligent recommendation component is used to realize automatic recommendation and selection of algorithms and function modules during task flow orchestration;

[0008] The task flow orchestration component is used to convert the visualized orchestration results into a process configuration description file that can be parsed by the task execution decision component;

[0009] The task execution decision component is used to read and parse the exclusive configuration description file and start process execution;

[0010] The model service generation component is used to provide unified access to the trained model in a service-oriented manner;

[0011] It also includes a computing resource scheduling component, which is used to connect the task execution decision component and the model service generation component.

[0012] The computing resource scheduling component is used for deep learning task execution and model service running, and dynamically allocates computing resources.

[0013] Preferably, the functional algorithm integration component provides a unified integration and management interface for algorithms and functional modules required for rich deep learning, and provides a solid base support for accelerating deep learning task orchestration.

[0014] The typical algorithms include image binarization, image geometric transformation, color space conversion;

[0015] The basic functional modules include target labeling, model training, model evaluation, and model deployment.

[0016] Preferably, the module intelligent recommendation component is used to realize the automatic recommendation and selection of algorithms and functional modules in task flow orchestration, and improve the efficiency of deep learning task flow orchestration. This process specifically includes:

[0017] S1.1: According to a large amount of past deep learning task flow orchestration data, fuse classification, statistics, aggregation and other recommendation algorithms to generate an intelligent recommendation model;

[0018] S1.2: In the process of flow orchestration, use the intelligent recommendation model for real-time analysis, automatically predict the next link function according to the last link function, and give TOP3 for priority selection in orchestration;

[0019] S1.3: Real-time record orchestration data, when reaching the set threshold value, such as 500 orchestration data, start the recommendation model self-evolution (i.e. continuous optimization and update of the model) mode, and gradually improve the process orchestration prediction accuracy.

[0020] Preferably, the task flow orchestration component is used to convert the visual-based orchestration results into a process configuration description file that can be parsed by the task execution decision component. This process specifically includes:

[0021] S2.1: Input this time's flow orchestration data and the set strategy value of the orchestration plan, wherein the setting of the strategy value includes:

[0022] 1) Data-driven strategy setting: when the data reaches a certain threshold value, such as 1000 orchestration data, the process orchestration discovers and automatically schedules the execution in time;

[0023] 2) Event-driven strategy setting: when the trigger condition reaches the preset value, such as the user specifies the immediate execution condition, the process orchestration discovers and automatically schedules the execution in time;

[0024] 3) Strategy setting based on periodic driving: periodic strategy setting can be performed in different granularities such as year, month, week, day, hour, minute, second, etc. When the time meets the strategy requirement, the process arrangement discovers and automatically schedules the execution in time.

[0025] S2.2: The task process arrangement component first resolves and converts the visual arrangement description form into a code type arrangement configuration description form;

[0026] S2.3: The code type arrangement configuration description form is semantically analyzed and decomposed into multiple execution units;

[0027] S2.4: Analyze the dependency relationship between each execution unit to form a basic DAG (Directed Acyclic Graph);

[0028] S2.5: Analyze the logical function association of the execution unit and the unit input and output association, start the function horizontal parallel optimization and input and output cache reuse optimization, optimize the basic DAG structure, and generate a DAG that is beneficial to concurrent execution and cache reuse;

[0029] S2.6: Convert the optimized DAG into a dedicated configuration description file that the task execution decision component can understand, including execution strategy, data address, computing power resource, execution process, result output, and other task required resource conditions.

[0030] Preferably, the task execution decision component reads and analyzes the dedicated configuration description file and starts the process execution.

[0031] Preferably, the computing power resource scheduling component is used for deep learning task execution and model service running, and dynamically allocates computing resources (composed of NPU, XPU, GPGPU, etc. AI chips). The computing resources involved are deployed in a multi-node distributed mode, divided into a master management node and multiple slave resource nodes, and the resource nodes periodically report resources to the master management node, and the master management node is responsible for unified management of resources and external services.

[0032] Optionally, the allocation of computing power resources can be based on the virtual container mode to achieve finer granularity division and calling at the chip level.

[0033] Preferably, the model service generation component is used to provide unified access to the trained model in a service-oriented manner, and can trigger the execution of the task process according to the strategy setting, realize the full-automatic evolution of model upgrade deployment and service release. This process specifically includes:

[0034] S3.1: Deploy and publish the model generated based on the process arrangement training in a service-oriented manner, and return the service Restful API calling interface;

[0035] S3.2: When the process orchestration is updated or the policy-driven condition is triggered, the retraining process will be started, a new model will be generated, and the version will be marked;

[0036] S3.3: Stop the old version service, start the new version service, and keep the service Restful API call interface unchanged.

[0037] Optionally, the model service can be packaged, deployed and published in the form of a container image, realizing the isolation of computing resources between different model services.

[0038] In the working process, the existing algorithms and basic function modules can be fully reused, the intelligent recommendation technology is superimposed, the traditional manual code type orchestration is changed into a semi-automatic or automatic way of orchestration based on visualization, the efficiency of deep learning task orchestration is greatly improved, the DAG topological relationship graph between the execution module units is parsed and optimized, and the heterogeneous computing resources such as NPU, XPU and GPGPU are deeply integrated, so that the system resource utilization can be effectively improved, the data, events and periodic policy setting are used to set the customized scheduling rules for the process orchestration, the automatic operation of the deep learning task is realized, the model one-key service and the rapid iteration and update of the version are completed. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The flow component architecture of the present application is shown in the figure,

[0040] Figure 2 The function module splitting example diagram is shown in the figure,

[0041] Figure 3 The intelligent recommendation working principle example diagram is shown in the figure,

[0042] Figure 4 The process orchestration parsing and optimization principle example diagram is shown in the figure,

[0043] Figure 5 The model service automatic publishing and updating process example diagram is shown in the figure. DETAILED DESCRIPTION

[0044] In order to make the content, technical scheme and implementation effect of the present application more popular and easy to understand, the present application will be described in detail below in combination with specific embodiments and drawings.

[0045] In combination with the drawings, Figure 1 The embodiment proposes an automatic orchestration and construction method for deep learning tasks, which is realized based on a function algorithm integrated component 1, a module intelligent recommendation component 2, a task process orchestration component 3, a task execution decision component 4, a computing resource scheduling component 5 and a model service generation component 6,

[0046] The functional algorithm integration component is used to integrate and manage typical algorithms and basic functional modules in a unified manner;

[0047] The intelligent recommendation component of the module is used to automatically recommend and select algorithms and functional modules when orchestrating task flows;

[0048] The task flow orchestration component is used to convert the visualization-based orchestration results into a flow configuration description file that can be parsed by the task execution decision component;

[0049] The task execution decision component is used to read and parse the dedicated configuration description file and initiate process execution;

[0050] The model service generation component is used to provide unified access to the trained model in a service-oriented manner;

[0051] The computing resource scheduling component is used to dynamically allocate computing resources for deep learning task execution and model service operation.

[0052] Functional Algorithm Integration Component 1 provides a unified integration and management interface for the algorithms and functional modules required for deep learning. Typical algorithms such as image binarization, image geometric transformation, and color space conversion are integrated and managed in a unified manner with basic functional modules such as model services and model evaluation. Unlike traditional functional algorithm implementations, this method decomposes functions and algorithms at a fine-grained level, ensuring the specificity of algorithm and module functions. This effectively improves the reusability of algorithms and functional modules and enhances the flexibility of task orchestration.

[0053] Combined with appendix Figure 2 A schematic diagram of functional decomposition and integration is given. To ensure the single functionality of each module, the output of each module must be processed as necessary and can be used as the input of the next module.

[0054] This example combines a typical image data processing procedure, which can be broken down into two modules: an image scaling module and an image rotation module. In actual operation, it can be further broken down into more granular functional modules according to business needs, such as image binarization, image geometric transformation, and image color transformation, providing rich functional module support for the reuse of functional modules and task flow orchestration.

[0055] The module intelligent recommendation component can continuously evolve the model based on user workflow orchestration behavior (by constantly learning, optimizing and updating the model), achieving more accurate personalized recommendations of algorithms and functional modules during task workflow orchestration, and providing a visual and automated module orchestration selection interface, which greatly improves the efficiency of deep learning task workflow orchestration.

[0056] Combined with appendix Figure 3, taking image data processing arrangement as an example, the module intelligent recommendation component 2 implements the process including the following steps:

[0057] S1.1: Collect user image processing history operation data, such as function module selection, module arrangement sequence, etc., integrate and form historical process arrangement data, input into the recommendation algorithm model training component;

[0058] S1.2: According to the existing historical process arrangement data, using classification, statistics, aggregation and other data analysis techniques, the user process arrangement habit model is initially generated;

[0059] S1.3: When the user needs to start a new round of image data processing, such as selecting the first image scaling module, the module intelligent recommendation component 2 will rely on the generated process arrangement habit model to automatically calculate the second module that may be needed, such as image rotation, and so on. The third, fourth, and Nth modules can be calculated to facilitate the user to quickly select the function modules required for this round of image processing and determine the execution sequence of each module;

[0060] S1.4: Continuously collect user process arrangement data this time and input into the recommendation algorithm model training component;

[0061] S1.5: According to the newly collected process arrangement data, continuously optimize and update the user process arrangement habit model, and gradually improve the accuracy of intelligent recommendation.

[0062] Compared with the traditional code type task arrangement method, the module intelligent recommendation component 2 can learn the task arrangement habits of the user in depth, and realize the personalized and accurate recommendation of the function modules and the automatic execution sequence arrangement in a visual form, which can greatly improve the efficiency of the user's deep learning task process arrangement.

[0063] In combination with the attached Figure 4 , the task process arrangement component 3 is used to convert the arrangement results based on visualization into process configuration description files that can be parsed by the task execution decision component 4. Specifically, it includes the following steps:

[0064] S2.1: Record this task arrangement data and task execution strategy, such as data total amount or specified time;

[0065] S2.2: Analyze and convert the visual arrangement description form into a code type arrangement configuration description form;

[0066] S2.3: Semantically analyze the code type arrangement configuration description form, and decompose it into multiple functional units such as target labeling, image scaling, model training, model evaluation, model deployment, and service publishing;

[0067] S2.4: Analyze the dependency relationship between each of the above execution units and form a DAG (Directed Acyclic Graph).

[0068] S2.5: Further analyze the logical association of each functional unit in the DAG graph and the input-output association between functional units. On the one hand, optimize the horizontal parallelism of Y1 logical functions, such as expanding the image scaling unit into multiple and concurrent execution units. On the other hand, optimize the input-output cache multiplexing of Y2, such as sharing a cache area between the model evaluation unit and the model deployment unit. Finally, automatically generate an optimized DAG directed acyclic graph;

[0069] S2.6: Output the optimized DAG directed acyclic graph as a dedicated configuration description file that can be parsed by the task execution decision component 4. The content includes the required conditions for task execution, and the format can be selected from Yaml / Json / XML file formats, etc.

[0070] The task flow arrangement component is used to convert the visual-based arrangement results into a flow configuration description file that can be parsed by the task execution decision component. The basic dependency topology relationship between multiple execution units is described in the form of a DAG directed acyclic graph. During the conversion process, the execution unit is automatically implemented with logic function horizontal parallelism, input-output cache multiplexing, and other flow optimization algorithms to generate a DAG directed acyclic graph that is beneficial for efficient execution.

[0071] In combination with the attached Figure 5 , the task execution decision component 4 and the computing resource scheduling component 5 provide necessary support for the model service generation component 6 at runtime. Taking service publishing and updating as an example, the cooperative working method of the above components is described in detail. The specific steps are as follows:

[0072] S3.1: Parse the flow arrangement dedicated configuration description file, match the most suitable node with the current resources, and schedule the execution unit to run on it;

[0073] S3.2: Deploy and run the model generated based on the above flow arrangement training, mark it as V1, and return the Restful API call interface;

[0074] S3.3: When the data volume reaches the set threshold or reaches the specified time, restart the above training process to generate a new model and mark it as version V2;

[0075] S3.4: Stop the model V1 service and start the model V2 service, keeping the Restful API call interface unchanged.

[0076] To sum up, by using the automatic arrangement and construction method for deep learning task, based on the functional algorithm integrated component 1, the module intelligent recommendation component 2, the task flow arrangement component 3, the task execution decision component 4, the computing resource scheduling component 5 and the model service generation component 6, the existing algorithms and basic function modules are reused, the traditional deep learning flow arrangement mode based on manual code writing is changed into a semi-automatic or automatic mode based on function modules, and the deep learning flow arrangement efficiency is effectively improved. By adopting the self-learning method, the model is continuously optimized and iterated, the recommendation accuracy is improved, the deep learning task automatic arrangement and efficient operation are realized, and the model self-evolution and automatic service deployment operation based on data, events and periodic strategy driving are realized.

[0077] By refining typical deep learning algorithms, the application provides basic algorithm function modules, ensures the singleness of each module function, so that the output of each function module in the deep learning process can be used as the input of the next function module after necessary processing, and at the same time, the application has the ability of intelligent selection of process arrangement and rearrangement of process, etc. Meanwhile, the application can fully utilize heterogeneous computing resources, accurately submit tasks to the best computing nodes according to different types of task resource requirements, and effectively improve the overall utilization rate of computing resources.

[0078] The above described embodiments are only the preferred embodiments of the present application and detailed description of the applied technical principles. Those skilled in the art should understand that the scope of the application involved in the present application is not limited to the technical solutions formed by the specific combination of the above technical features, and also covers other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the above features and the technical features disclosed in the present application (but not limited to) with similar functions are replaced with each other to form a technical solution, which should fall within the scope of patent protection of the present application.

Claims

1. An automated orchestration and construction method for deep learning tasks, characterized in that, It includes, in sequence, a functional algorithm integration component, a module intelligent recommendation component, a task flow orchestration component, a task execution decision component, and a model service generation component; The functional algorithm integration component is used to integrate and manage typical algorithms and basic functional modules in a unified manner; The intelligent recommendation component of the module is used to automatically recommend and select algorithms and functional modules when orchestrating task flows; The task flow orchestration component is used to convert the visualization-based orchestration results into a flow configuration description file that can be parsed by the task execution decision component; The task execution decision component is used to read and parse the dedicated configuration description file and initiate process execution; The model service generation component is used to provide unified access to the trained model in a service-oriented manner; It also includes a computing resource scheduling component, which is used to connect the task execution decision component and the model service generation component; The computing resource scheduling component is used to dynamically allocate computing resources for deep learning task execution and model service operation; the allocation of resources is based on virtualized containers to achieve chip-level partitioning and calling. The computing resource scheduling component includes multiple types of heterogeneous resources, specifically: NPU and GPGPU; The intelligent recommendation component of the module includes the following steps: S1.1: Based on a large amount of data from previous deep learning task processes, a recommendation algorithm is integrated to generate an intelligent recommendation model; S1.2: During the process orchestration, an intelligent recommendation model is used for real-time analysis to automatically predict the functions of the next step based on the functions of the previous step, and to provide a TOP3 list for priority selection during orchestration. S1.3: Record orchestration data in real time, and activate the recommendation model's self-evolution mode when the set threshold is reached; The task flow orchestration component includes the following steps: S2.1: Enter the data for this process orchestration and the set strategy values ​​for the orchestration plan; S2.2: The task flow orchestration component prioritizes parsing and converting the visualization-based orchestration results into a code-based orchestration configuration description. S2.3: Semantically parse the code-based orchestration configuration description and decompose it into multiple execution units; S2.4: Analyze the dependencies between the various execution units to form a basic DAG (Directed Acyclic Graph); S2.5: Analyze the logical function associations and input / output associations of the execution units, enable horizontal parallel optimization and input / output cache reuse optimization, optimize the basic DAG directed acyclic graph structure, and generate a DAG directed acyclic graph that is conducive to concurrent execution and cache reuse; among them, the horizontal parallel optimization specifically means: one unit is expanded into multiple units that can be executed concurrently. The input / output cache reuse optimization specifically involves different execution units sharing a single cache area. S2.6: Convert the optimized DAG (Directed Acyclic Graph) into a dedicated configuration description file that can be understood by the task execution decision component.

2. The automated orchestration and construction method for deep learning tasks according to claim 1, characterized in that, The typical algorithms include image binarization, image geometric transformation, and color space conversion; The basic functional modules include target annotation, model training, model evaluation, and model deployment.

3. The automated orchestration and construction method for deep learning tasks according to claim 1, characterized in that, In step S2.1, the setting of the strategy value includes: 1) Data-driven strategy setting: When the data reaches a predetermined threshold, the process orchestration will detect it in a timely manner and automatically schedule its execution; 2) Event-driven strategy setting: When the triggering conditions reach the preset value, the process orchestration will detect and automatically schedule the execution in a timely manner; 3) Period-driven strategy setting: When the time meets the strategy requirements, the process orchestration will promptly detect and automatically schedule the execution.

4. The automated orchestration and construction method for deep learning tasks according to claim 1, characterized in that, The model service generation component includes the following steps: S3.1: Deploy and publish the model generated based on the process orchestration training, and return the service RESTful API call interface; S3.2: When the process orchestration is updated or the policy-driven conditions are triggered, the retraining process will be started, a new model will be generated, and the version will be marked. S3.3: Stop the old version of the service, start the new version of the service, and keep the service's RESTful API call interface unchanged.

Images