A containerized deep learning model feedback training system and method

Abstract

The application discloses a containerized deep learning model feedback training system and method, the system comprises a management service module, a data set module and a model management module, the management service module is used for controlling each model to complete training and updating when receiving artificial confirmation annotation data generated by user operation business application; the data set module is used for storing the annotation data received by the management service module, converting the annotation data into training data, and sending the training data to each model for model training; the model management module comprises a plurality of models, and is used for training and updating the model corresponding to the received training data for the business application to call. The application provides a containerized deep learning model feedback training framework, uses a container to provide a unified deployment and training environment of a deep learning model, realizes automatic training and updating of the model based on user feedback annotation data, and solves the technical problem of reduced system processing efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information systems, and particularly relates to a containerized deep learning model feedback training system and method. BACKGROUND

[0002] At present, intelligent algorithm models represented by deep learning have carried out a large number of integrated applications in text information processing systems, and a large number of intelligent model services (such as entity extraction, automatic abstract, sentiment analysis, etc.) have been constructed, which greatly improves the intelligent level of the system.

[0003] However, in the traditional mode, the deep learning model deployed to the system cannot be automatically trained and updated as the processed business data changes, and the processing efficiency of the system will decrease after a period of time. SUMMARY

[0004] The main purpose of the present application is to provide a containerized deep learning model feedback training system and method, which aims to solve the technical problem of reduced system processing efficiency caused by the fact that the current deep learning model cannot be automatically trained and updated based on user feedback labeled data.

[0005] To achieve the above purpose, the present application provides a containerized deep learning model feedback training system, which comprises:

[0006] A management service module is configured to control each model to complete training and updating when receiving artificial confirmation labeled data generated by user operation business application.

[0007] A data set module is configured to store the labeled data received by the management service module, convert the labeled data into training data, and send the training data to each model for model training.

[0008] A model management module comprises a plurality of models, which are configured to train and update the models corresponding to the received training data for calling by business application.

[0009] Optionally, the management service module is further configured with a collection interface connected to the business application, which is configured to collect labeled data generated by calling the model by the business application.

[0010] Optionally, each model comprises an inference container and a training container.

[0011] The training container is configured to perform model training using the received training data according to the training instruction sent by the management service module.

[0012] The inference container is configured to perform model updating using the trained model according to the updating instruction sent by the management service module.

[0013] Optionally, the training container and the inference container mount the same model shared directory; wherein:

[0014] The model shared directory is used to store the trained model of the training container after the training container executes model training.

[0015] The model shared directory is also used to load the trained model from the model shared directory when the inference container executes model updating.

[0016] Optionally, the inference container is further configured with a service interface connected to the business application, for the business application to call the model trained and updated by the model management module.

[0017] In order to achieve the above purpose, the application further provides a containerized deep learning model feedback training method, which is used for the containerized deep learning model feedback training system as described above, and the method comprises:

[0018] S1: the management service module collects the labeled data of the business application, and the dataset module converts the labeled data into training data;

[0019] S2: the model management module executes model training according to the training data;

[0020] S3: the model management module executes model updating according to the trained model, for the business application to call.

[0021] Optionally, the step S1 specifically comprises:

[0022] S101: the business application calls the collection interface provided by the management service, and sends the labeled data generated by the user-confirmed model processing result to the management service module;

[0023] S102: the management service module receives the labeled data and stores the labeled data into the dataset module;

[0024] S103: the management service module checks the data amount in the dataset module, and if the data amount reaches the given data amount requirement for starting model training, the model training is started.

[0025] Optionally, the model management module comprises a plurality of models, each of the models comprises an inference container and a training container, and the training container and the inference container mount the same model shared directory.

[0026] Optionally, the step S2 specifically comprises:

[0027] S201: the management service module starts the training container corresponding to the model;

[0028] S202: The training container loads the model from the model sharing directory, downloads the training data from the dataset module, and starts the model training.

[0029] S203: After the training is completed, the training container saves the new model to the model sharing directory.

[0030] Optionally, the step S3 specifically comprises:

[0031] S301: The management service module restarts the inference container of the model one by one;

[0032] S302: After the inference container is restarted, the newly trained model is loaded from the model sharing directory, and the service update is completed;

[0033] S303: The business application calls the service interface provided by the inference container, invokes the model to process the business data, generates a processing structure, and generates the labeled data after the user confirms.

[0034] The embodiment of the application provides a containerized deep learning model feedback training system and method, the system comprises a management service module, a dataset module and a model management module, the management service module is used for controlling each model to complete training and updating when receiving the artificial confirmation labeled data generated by the user operation business application; the dataset module is used for storing and collecting the labeled data received by the management service module, and converting the labeled data into training data and sending the training data to each model for model training; the model management module comprises a plurality of models, and is used for training and updating the model corresponding to the received training data for the business application to call. The application provides a containerized deep learning model feedback training framework, and utilizes the container to provide a unified deployment and training environment of the deep learning model, so that the model automatic training and updating based on the user feedback labeled data are realized. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 Fig. 1 is a schematic diagram of the principle of the containerized deep learning model feedback training system of the application;

[0036] Figure 2 Fig. 2 is a schematic diagram of the principle of the containerized deep learning model feedback training method of the application.

[0037] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0038] It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0039] At present, intelligent algorithm models represented by deep learning have carried out a large number of integrated applications in text information processing systems, and a large number of intelligent model services (such as entity extraction, automatic abstract, sentiment analysis, etc.) have been built, which greatly improves the intelligent level of the system. In the traditional mode, after the deep learning model is deployed to the system, it cannot be automatically trained and updated with the change of the processed business data, and the processing efficiency of the system will decrease after a period of time. However, there are many manual confirmation links in the system, which can produce a large amount of high-quality labeled data that fits the actual business scenario. A feedback mechanism can be built to automatically collect labeled data in the manual confirmation link, update and release the model after incremental training, so as to keep the freshness of the model.

[0040] The embodiment is based on the above principle, and proposes a containerized deep learning model feedback training system, which uses a container to provide a unified deployment and training environment for deep learning models, and realizes automatic training and updating of the model based on user feedback labeled data.

[0041] The embodiment of the application provides a containerized deep learning model feedback training system, which is described with reference to Figure 1 , Figure 1 FIG. 1 is a schematic diagram of the principle of the containerized deep learning model feedback training system of the application.

[0042] The containerized deep learning model feedback training system of the embodiment comprises:

[0043] The management service module is configured to control each model to complete training and updating when receiving the manual confirmation labeled data generated by the user operation business application.

[0044] The data set module is configured to store the labeled data received by the management service module, convert the labeled data into training data, and send the training data to each model for model training.

[0045] The model management module comprises a plurality of models, and is configured to train and update the models corresponding to the received training data, and provide the business application with the trained models.

[0046] Specifically, the framework of the containerized deep learning model feedback training system comprises a management service, a data set, and a plurality of models.

[0047] In the preferred embodiment, the management service provides a labeled data collection interface, collects the manual confirmation labeled data generated by the user operation business application, and controls each model to complete training and service updating.

[0048] In the preferred embodiment, the data set stores the collected labeled data, and provides the labeled data to the training container of each model for model training.

[0049] In a preferred embodiment, one model comprises one training container and at least one inference container. The training container and the inference container of the same model mount the same model shared directory, and the trained model file is stored in the model shared directory. The inference container provides a service interface, and the business application completes business data processing by calling the interface.

[0050] The embodiment of the present application also provides a containerized deep learning model feedback training method, which refers to Figure 2 , Figure 2 The embodiment of the present application also provides a containerized deep learning model feedback training method, which refers to

[0051] The embodiment of the present application also provides a containerized deep learning model feedback training method, which refers to

[0052] S1: The management service module collects the labeled data of the business application, and the data set module converts the labeled data into training data;

[0053] S2: The model management module performs model training according to the training data;

[0054] S3: The model management module performs model updating according to the trained model, for calling by the business application. Specifically, the working process of the containerized deep learning model feedback training method of the embodiment is as follows:

[0055] (I) Labeled data collection

[0056] 1. The business application calls the service interface provided by the model inference container, processes the business data, and generates a processing result;

[0057] 2. The user confirms the processing result, and generates labeled data;

[0058] 3. The business application calls the collection interface provided by the management service, and sends the labeled data to the management service;

[0059] 4. The management service receives the labeled data, and stores the labeled data in the data set;

[0060] 5. The management service checks the data amount in the data set, and if the data amount meets the given data amount requirement for starting model training, the model training is started, otherwise, the model training is skipped.

[0061] (II) Model training

[0062] 1. The management service starts the training container corresponding to the model;

[0063] 2. The training container loads the model from the model shared directory, downloads the training data from the data set, and starts the model training;

[0064] 3. After the training is completed, the training container saves the new model into the model shared directory.

[0065] (III) service update

[0066] 1. A reasoning container for managing service restart model by model;

[0067] 2. After the reasoning container is started, a newly trained model is loaded from the model sharing directory, and the service update is completed.

[0068] 3. A service module for managing, when receiving artificial confirmation annotation data generated by user operation business application, each model to complete training and update.

[0069] The embodiment provides a containerized deep learning model feedback training system and method, by proposing a containerized deep learning model feedback training framework, using containers to provide a unified deployment and training environment for deep learning models, and realizing automatic training and updating of the models based on user feedback annotation data. The technical problem of low system processing efficiency caused by the fact that the current deep learning model cannot be automatically trained and updated based on user feedback annotation data is solved.

[0070] The above is only the preferred embodiment of the application, and does not limit the patent range of the application, and any equivalent structure or equivalent process transformation using the content of the application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection range of the application.

Claims

1. A containerized deep learning model feedback training system, characterized in that, The system comprises: a management service module, configured to control each model to complete training and updating when receiving artificial confirmation annotation data generated by a user operation service application; a data set module, configured to store the annotation data received by the management service module, convert the annotation data into training data, and send the training data to each model for model training; a model management module, comprising a plurality of models, configured to train and update the models corresponding to the received training data, and provide the models for calling by the service application; the management service module is further configured with a collection interface connected to the service application, configured to collect annotation data generated by the service application calling the models; each model comprises an inference container and a training container; wherein: the training container is configured to perform model training using the received training data according to a training instruction sent by the management service module; the inference container is configured to perform model updating using the trained model according to an updating instruction sent by the management service module; the training container and the inference container are mounted with the same model shared directory; wherein: the model shared directory is configured to store the trained model after the training container performs model training; the model shared directory is further configured to load the trained model from the model shared directory when the inference container performs model updating; the inference container is further configured with a service interface connected to the service application, configured to call the model trained and updated by the model management module by the service application.

2. A containerized deep learning model feedback training method, characterized in that, The method for the containerized deep learning model feedback training system of claim 1, comprising: S1: the management service module collects annotation data of the service application, and the data set module converts the annotation data into training data; S2: the model management module performs model training according to the training data; S3: the model management module performs model updating according to the trained model, and provides the model for calling by the service application. 3.The containerized deep learning model feedback training method of claim 2, wherein, Step S1 specifically comprises: S101: the service application calls the collection interface provided by the management service, and sends the annotation data generated by the user-confirmed model processing result to the management service module; S102: the management service module receives the annotation data, and stores the annotation data in the data set module; S103: the management service module checks the data amount in the data set module, and starts model training if the data amount reaches the given data amount requirement for starting model training. 4.The containerized deep learning model feedback training method of claim 2, wherein, The model management module comprises a plurality of models, each model comprising an inference container and a training container, and the training container and the inference container are mounted with the same model shared directory. 5.The containerized deep learning model feedback training method of claim 3, wherein, Step S2 specifically comprises: S201: the management service module starts the training container of the corresponding model; S202: the training container loads the model from the model shared directory, downloads the training data from the data set module, and starts model training; S203: after the training is completed, the training container saves the new model into the model shared directory. 6.The containerized deep learning model feedback training method of claim 3, wherein, Step S3 specifically comprises: S301: the management service module restarts the inference container of each model; S302: the inference container loads the newly trained model from the model shared directory after being restarted, and completes service updating; S303: The business application calls the service interface provided by the inference container, invokes the model to process the business data, generates a processing structure, and generates the labeled data after the user confirms.

Images