A training method and a training system of an artificial intelligence model

Abstract

The present application relates to the technical field of artificial intelligence training, in particular to a kind of training method and training system of artificial intelligence model, comprising: S1: obtaining multiple model components from cloud platform;S2: in development tool, model components are combined to generate artificial intelligence model;S3: using training set, artificial intelligence model is trained, and training result is output;S4: according to training result, whether the performance parameter of artificial intelligence model reaches expected performance;If yes, output artificial intelligence model to end training process;If not, return S2 to adjust artificial intelligence model.The beneficial effects of the present application are that: by setting cloud platform, model components are integrated and packaged, then by development tool, model components are received and combined into artificial intelligence model, thereby improving development efficiency, so that users can quickly obtain multiple model components for testing in actual development process through cloud platform, to obtain artificial intelligence model with better performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence training, in particular to a training method and system of an artificial intelligence model. BACKGROUND

[0002] An artificial intelligence model refers to a mathematical model based on relevant mathematical algorithms to process and extract features of input objective data, so as to judge or predict things. With the development of computer technology, artificial intelligence technology has been applied in many fields.

[0003] In the prior art, there is a technical solution for combining existing network architecture and recognition algorithms to build a user-required artificial intelligence model. Such a technical solution is usually programmed by programmers according to the functions to be realized by the model, and the corresponding algorithms are selected, combined and connected to obtain the required artificial intelligence model. However, in the actual implementation process, the inventors found that the process of generating an artificial intelligence model in the prior art is relatively cumbersome, and it is necessary to rely on manual adjustment of the computer program itself to make the artificial intelligence model execute normally, thereby consuming a large amount of manpower in the model generation process. SUMMARY

[0004] In view of the above problems existing in the prior art, the present application provides a training method and system of an artificial intelligence model.

[0005] The specific technical solutions are as follows:

[0006] A training method of an artificial intelligence model, a development tool deployed locally and a cloud platform connected with the development tool are set in advance, and the training method comprises the following steps:

[0007] Step S1: obtaining a plurality of model components from the cloud platform;

[0008] Step S2: combining the model components in the development tool to generate the artificial intelligence model;

[0009] Step S3: obtaining a training set, training the artificial intelligence model using the training set, and outputting the training result;

[0010] Step S4: determining whether the performance parameter of the artificial intelligence model reaches the expected performance according to the training result;

[0011] If yes, output the artificial intelligence model to end the training process;

[0012] If not, return to step S2 to adjust the artificial intelligence model.

[0013] Preferably, a visualization interface is provided in the development tool, and the step S2 comprises:

[0014] Step S21: displaying the model components in the visualization interface;

[0015] Step S22: selecting the model components in the visualization interface;

[0016] Step S23: adjusting the order and connection relationship between the model components to generate the artificial intelligence model.

[0017] Preferably, the training set is pre-provided locally, and the step S3 comprises:

[0018] Step A31: acquiring and inputting the training set into the development tool;

[0019] Step A32: training the artificial intelligence model using the training set in the development tool, and outputting the training result.

[0020] Preferably, the training set is provided in the cloud platform, and the step S3 comprises:

[0021] Step B31: uploading the artificial intelligence model to the cloud platform;

[0022] Step B32: training the artificial intelligence model using the training set in the cloud platform, and outputting the training result.

[0023] Preferably, a user permission is provided for each user before the step S1;

[0024] The step S1 comprises:

[0025] Step S11: the cloud platform acquires the user from the development tool;

[0026] Step S12: the cloud platform selects the model components to be sent to the development tool according to the user permission corresponding to the user;

[0027] Step S13: the cloud platform sends the model components to the development tool.

[0028] A training system of an artificial intelligence model is used to implement the training method described above, comprising a development tool connected to a remote cloud platform;

[0029] A component storage submodule is provided in the cloud platform, and a plurality of model components are stored in the component storage submodule;

[0030] The development tool comprises:

[0031] The model component obtaining submodule is connected to the component storage submodule and obtains the model component from the component storage submodule;

[0032] The model generating submodule is connected to the model component obtaining submodule, obtains the model component from the model component obtaining submodule, and generates the artificial intelligence model according to the model component;

[0033] The training submodule is arranged in the cloud platform or the development tool, is connected to the model generating submodule and receives the artificial intelligence model, trains the artificial intelligence model, and outputs a training result;

[0034] The performance judging submodule is arranged in the development tool, is connected to the training submodule, and judges whether the training is completed according to the training result.

[0035] Preferably, the development tool further comprises a visualization interface connected to the model component obtaining submodule and the model generating submodule.

[0036] The model component and the artificial intelligence model are displayed in the visualization interface, and the model generating submodule is controlled to adjust the artificial intelligence model in response to an externally input operation instruction.

[0037] Preferably, the cloud platform is provided with a permission management submodule connected to the component storage submodule, which controls the component storage submodule according to user permissions to select the model component sent to the development tool.

[0038] Preferably, the model component comprises an image algorithm component, a preprocessing component, a training framework component, a training network component, and a loss function component.

[0039] The above technical solution has the following advantages or beneficial effects: the cloud platform is arranged to integrate and encapsulate the model component, and then the development tool receives the model component and combines the model component into an artificial intelligence model, thereby avoiding the problem that the code processing in the generation process of the artificial intelligence model in the prior art is relatively complicated, improving the development efficiency, and enabling the user to quickly obtain multiple model components for testing through the cloud platform in the actual development process to obtain an artificial intelligence model with good performance. BRIEF DESCRIPTION OF DRAWINGS

[0040] Embodiments of the present application will be described more fully hereinafter with reference to the accompanying drawings. However, the accompanying drawings only serve to explain and illustrate the present application and do not limit the scope of the present application.

[0041] Figure 1 is a schematic diagram of the whole embodiment of the present application;

[0042] Figure 2 is a schematic diagram of the sub-step of step S2 in the embodiment of the present application;

[0043] Figure 3 is a schematic diagram of the visual interface in the embodiment of the present application;

[0044] Figure 4 is a schematic diagram of the sub-step of step S3 in the embodiment of the present application;

[0045] Figure 5 is a schematic diagram of the sub-step of step S3 in another embodiment of the present application;

[0046] Figure 6 is a schematic diagram of the sub-step of step S1 in the embodiment of the present application;

[0047] Figure 7 is a schematic diagram of the training system in the embodiment of the present application;

[0048] Figure 8 is a schematic diagram of the training system in another embodiment of the present application. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be described clearly and completely hereinafter with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0050] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0051] The present application will be further described hereinafter with reference to the accompanying drawings and specific embodiments, but not as a limitation of the present application.

[0052] The present application includes:

[0053] A training method of an artificial intelligence model, a development tool deployed locally and a cloud platform connected with the development tool are set in advance, and as shown in the figure, the training method comprises: Figure 1 as shown in the figure, the training method comprises:

[0054] Step S1: obtaining a plurality of model components from the cloud platform;

[0055] Step S2: combining the model components in the development tool to generate the artificial intelligence model;

[0056] Step S3: obtaining a training set, training the artificial intelligence model using the training set, and outputting the training result;

[0057] Step S4: judging whether the performance parameter of the artificial intelligence model reaches the expected performance according to the training result;

[0058] If yes, output the artificial intelligence model to end the training process;

[0059] If no, return to step S2 to adjust the artificial intelligence model.

[0060] Specifically, in view of the problem that the artificial intelligence model in the prior art needs to be manually coded by programmers during the development process, and then assembled and debugged to form the artificial intelligence model, the process is relatively cumbersome, in the embodiment, a plurality of encapsulated model components are pre-set in the cloud platform, so that the developer connects to the cloud platform through the local development tool (IDE, Integrated Development Environment) to obtain the model components required for generating the artificial intelligence model, thereby realizing the rapid generation of the artificial intelligence model by adjusting the selected model components, the order between the model components, the input-output relationship and other parameters without adjusting the code part of the specific algorithm, and performing performance evaluation to quickly obtain the ideal artificial intelligence model, thereby improving the development efficiency.

[0061] In the implementation process, the above development tool is a local installed integrated development environment (IDE, Integrated Development Environment), which is pre-provided with a communication module connected with the cloud platform, an editor and a graphical interface, and then obtains model components from the cloud platform, so that the user assembles the model components in the graphical interface, thereby realizing the graphical development of the artificial intelligence model. In this process, the editor automatically adjusts the input and output of the model components according to the user's operation instructions, and provides the user with corresponding parameter setting ports to realize partial customization of the model components. Different types of model components are pre-stored in the cloud platform, such as image algorithm components, preprocessing components, training framework components, training network components and loss function components. Each model component includes a pre-encapsulated algorithm, such as Roberts operator, Sobel operator, blind convolution, grayscale, binarization, dilation algorithm, erosion algorithm, opening operation, closing operation, blur effect, and hybrid processing algorithm in image algorithm component and preprocessing component; each training framework component encapsulates Tensorflow, Pytorch, Keras training framework, etc. The development tool realizes the assembly of the artificial intelligence model by obtaining the above components and adjusting the order and connection relationship of each component.

[0062] In a preferred embodiment, the development tool is provided with a visualization interface, and as shown in Figure 2 , step S2 includes:

[0063] Step S21: displaying the model components in the visualization interface;

[0064] Step S22: selecting the model components in the visualization interface;

[0065] Step S23: adjusting the order and connection relationship between the model components to generate the artificial intelligence model.

[0066] Specifically, in the prior art, the programmer needs to manually code the recognition algorithm in the development process of the artificial intelligence model, and then assemble and debug the artificial intelligence model, which is a relatively cumbersome process. As shown in Figure 3 , in the present embodiment, the visualization interface is integrated in the development tool, so that the user can select the required model components in the visualization interface and add them to the editing area, and then adjust the component part and input and output order in the model, thereby realizing the editing of the artificial intelligence model in a codeless development environment.

[0067] In a preferred embodiment, a training set is provided locally, and as shown in Figure 4 , step S3 includes:

[0068] Step A31: obtaining and inputting the training set into the development tool;

[0069] Step A32: training the artificial intelligence model in the development tool using the training set, and outputting the training result.

[0070] Specifically, to achieve a more convenient development effect, in the embodiment, the training set is set locally, and then the user locally edits the artificial intelligence model, and then trains using the local training set. In the training process, the training environment can be instrumented to obtain the data changes in the training process, thereby facilitating the user to debug the artificial intelligence model.

[0071] As an optional implementation, the cloud platform pre-stores a training set of a specific category, and the user can retrieve and obtain the training set on the cloud platform through the development tool, so that the user can select the corresponding training set for training according to the artificial intelligence model generated as needed.

[0072] In a more preferred embodiment, the training set is set in the cloud platform, and as shown in Figure 5 Step S3 includes:

[0073] Step B31: uploading the artificial intelligence model to the cloud platform;

[0074] Step B32: training the artificial intelligence model in the cloud platform using the training set, and outputting the training result.

[0075] Specifically, to achieve a more convenient development effect, in the embodiment, the edited artificial intelligence model is sent to the cloud platform, and is hosted on the cloud platform for training. In the training process, the cloud platform returns debugging data to the development tool, so that the user can train the artificial intelligence model with the computing power of the cloud platform, and at the same time, according to the debugging data, the training of the artificial intelligence model is obtained, thereby improving the development efficiency of the model.

[0076] In a more preferred embodiment, a user permission is set for each user before step S1;

[0077] Step S1 includes: Figure 6

[0078] Step S11: the cloud platform obtains the user from the development tool;

[0079] Step S12: the cloud platform selects the model component sent to the development tool according to the user permission corresponding to the user;

[0080] Step S13: the cloud platform sends the model component to the development tool.

[0081] ​Specifically, to achieve the management and control effect on the overall development process, in the embodiment, different levels of user permissions are assigned to each user in advance on the cloud platform, and the model components that can be opened to the user are selected according to the user permissions, thereby avoiding the leakage problem caused by issuing some confidential components to users with lower permissions, or the waste of computing power caused by the improper selection of model components by the user when the user selects cloud training, thereby improving the development efficiency.

[0082] A training system of an artificial intelligence model for implementing the above training method, as shown in Figure 7 and Figure 8 The development tool 1 is connected to a remote cloud platform 2.

[0083] The cloud platform 2 is provided with a component storage submodule 21, and the component storage submodule 21 stores a plurality of model components.

[0084] The development tool 1 comprises:

[0085] The model component acquisition submodule 11 is connected to the component storage submodule 21 and acquires the model components from the component storage submodule 21.

[0086] The model generation submodule 12 is connected to the model component acquisition submodule 11, acquires the model components from the model component acquisition submodule 11, and generates an artificial intelligence model according to the model components.

[0087] The training submodule 3 is provided in the cloud platform 2 or the development tool 1, is connected to the model generation submodule 12 and receives the artificial intelligence model, trains the artificial intelligence model, and outputs the training result.

[0088] The performance judgment submodule 13 is further provided in the development tool 1 and is connected to the training submodule 3. The performance judgment submodule 13 judges whether the training is completed according to the training result.

[0089] In the prior art, the programmer needs to manually code the recognition algorithm during the development of the artificial intelligence model, and then assemble and debug the artificial intelligence model, which is a relatively cumbersome process. In the embodiment, a plurality of encapsulated model components are pre-set in the cloud platform 2, so that the developer can obtain these model components through the local development tool 1, edit the model components through the model generation submodule 12 to obtain the required artificial intelligence model, train the artificial intelligence model through the training submodule 3, and then judge the training result through the performance judgment submodule 13, thereby quickly obtaining the artificial intelligence model expected by the developer.

[0090] In the implementation process, the performance judgment submodule 13 is pre-set with corresponding evaluation thresholds, including recognition accuracy, LOSS, signal-to-noise ratio, etc. The training result refers to the artificial intelligence model after training. The performance judgment submodule 13 can verify the pre-set verification set corresponding to the training set for the model, so as to obtain the overall performance of the model, or judge whether the part of the node of the artificial intelligence model meets the expected performance according to the debugging parameters in the recognition process. For example, in an artificial intelligence model used for image recognition, a monitoring point is set after the image preprocessing module to obtain whether the signal-to-noise ratio of the verification picture after processing and feature extraction exceeds the pre-set signal-to-noise ratio threshold, so as to judge whether the preprocessing part of the artificial intelligence model is qualified.

[0091] In a more preferred embodiment, the development tool 1 further comprises a visualization interface 14, and the visualization interface 14 is connected to the model component acquisition submodule 11 and the model generation submodule 12.

[0092] The model component and the artificial intelligence model are displayed in the visualization interface 14, and the model generation submodule 12 is controlled in response to an externally input operation instruction, so as to adjust the artificial intelligence model.

[0093] Specifically, in order to achieve a better user experience, in the embodiment, the visualization interface 14 is set in the development tool 1, and the model components available to the user, the current artificial intelligence model structure edited by the user, the debugging log in the editing process, and the corresponding control UI are displayed in the visualization interface 14, so that the user can achieve a better editing effect in the visualization interface.

[0094] In a more preferred embodiment, the cloud platform 2 is provided with a permission management submodule 22, and the permission management submodule 22 is connected to the component storage submodule 21. The permission management submodule 22 controls the component storage submodule 21 according to the user permission, so as to select the model components sent to the development tool 1.

[0095] Specifically, in order to realize the management and control effect of the overall development process, in the embodiment, different levels of user permissions are allocated to each user in the permission management submodule 22 in advance, and the component storage submodule 21 is controlled according to the user permission, so as to adjust the model components that can be opened to the user by the component storage submodule 21, thereby avoiding the leakage problem caused by distributing part of the confidential components to users with lower permissions, or the waste of computing power caused by the improper selection of model components by the user when the user selects cloud training, thereby improving the development efficiency.

[0096] In a more preferred embodiment, the model component includes an image algorithm component, a preprocessing component, a training framework component, a training network component, and a loss function component.

[0097] The application has the beneficial effect that by setting the cloud platform to integrate and package the model components, and then receiving the model components by the development tool and combining them into an artificial intelligence model, the problem of cumbersome code processing in the generation process of the artificial intelligence model in the prior art is avoided, thereby improving the development efficiency, so that the user can quickly obtain multiple model components for testing through the cloud platform in the actual development process to obtain an artificial intelligence model with better performance.

[0098] The above is only the preferred embodiment of the application, and does not limit the implementation and protection scope of the application. Those skilled in the art should be able to realize that any equivalent replacement and obvious changes made by applying the content of the specification and drawings should be included in the protection scope of the application.

Claims

1. A training method for an artificial intelligence model, characterized in that: A development tool deployed locally and a cloud platform connected to the development tool are pre-set, and the training method includes: Step S1: obtaining multiple model components from the cloud platform; Step S2: combining the model components in the development tool to generate the artificial intelligence model; Step S3: Obtain a training set, use the training set to train the artificial intelligence model, and output the training results; Step S4: judging whether the performance parameters of the artificial intelligence model meet the expected performance according to the training results; If so, output the artificial intelligence model to end the training process; If not, return to step S2 to adjust the artificial intelligence model; When the training set is pre-set locally, step S3 includes: Step A31: Obtain and input the training set into the development tool; Step A32: training the artificial intelligence model using the training set in the development tool, and outputting the training results; Furthermore, during the training process of step A32, by burying points in the training environment, data changes during the training process are obtained, and the artificial intelligence model is debugged; The development tool includes a plurality of submodules, and a monitoring point corresponding to the artificial intelligence model to be trained is set at the rear stage of the submodule to determine whether the part of the artificial intelligence model at the front stage of the monitoring point is qualified; The development tool further includes a performance judgment submodule, which is configured with corresponding evaluation thresholds for different indicators, thereby judging whether some nodes of the model meet the expected performance.

2. The training method according to claim 1, characterized in that If the development tool is provided with a visual interface, step S2 includes: Step S21: displaying the model component in the visualization interface; Step S22: selecting the model component in the visualization interface; Step S23: Adjust the order and connection relationship between the model components to generate the artificial intelligence model.

3. The training method according to claim 1, characterized in that If the training set is set in the cloud platform, step S3 includes: Step B31: Uploading the artificial intelligence model to the cloud platform; Step B32: Use the training set to train the artificial intelligence model in the cloud platform and output the training results.

4. The training method according to claim 1, characterized in that Before step S1, user permissions are set for each user; Then the step S1 includes: Step S11: the cloud platform obtains the user from the development tool; Step S12: the cloud platform selects the model component to be sent to the development tool according to the user authority corresponding to the user; Step S13: The cloud platform sends the model component to the development tool.

5. A training system for an artificial intelligence model, characterized in that: Used to implement the training method according to any one of claims 1 to 4, comprising a development tool, wherein the development tool is connected to a remote cloud platform; The cloud platform is provided with a component storage submodule, wherein the component storage submodule stores a plurality of model components; The development tools include: A model component acquisition submodule, the model component acquisition submodule is connected to the component storage submodule, and acquires the model component from the component storage submodule; A model generation submodule, the model generation submodule being connected to the model component acquisition submodule, the model generation submodule acquiring the model components from the model component acquisition submodule, and generating the artificial intelligence model based on the model components; It also includes a training submodule, which is set in the cloud platform or the development tool, the training submodule is connected to the model generation submodule and receives the artificial intelligence model, and the training submodule trains the artificial intelligence model and outputs the training results; The development tool is further provided with a performance judgment submodule, which is connected to the training submodule and determines whether the training is completed based on the training result.

6. The training system according to claim 5, characterized in that The development tool further includes a visual interface, wherein the visual interface connects the model component acquisition submodule and the model generation submodule; The model components and the artificial intelligence model are displayed in the visualization interface, and the model generation submodule is controlled in response to an external input operation instruction, thereby adjusting the artificial intelligence model.

7. The training system according to claim 5, characterized in that The cloud platform is provided with a rights management submodule, which is connected to the component storage submodule. The rights management submodule controls the component storage submodule according to user rights to select the model component to be sent to the development tool.

8. The training system according to claim 5, characterized in that The model components include: image algorithm component, preprocessing component, training framework component, training network component and loss function component.

Images