AIGC model optimization method with adaptive learning capability

Abstract

The invention discloses an AI GC model optimization method with adaptive learning ability, and relates to the technical field of model optimization, comprising an information acquisition module, an architecture selection module, an optimization module, a verification module, an architecture change module, a data analysis module and an influence analysis module. Compared with the prior art, the system has the advantages that the appropriate architecture can be quickly screened out according to different information requirements through the architecture selection module, the adaptability of the model and the task is improved, the potential of different architectures on different information is mined, model parameters can be adjusted more meticulously through the optimization module, the model performance is improved, and the system is more practical. The accuracy of the model can be effectively evaluated through the verification module, the defects of the model can be found in time, a basis is provided for further optimization, the framework change module is used for avoiding falling into a local optimal solution, the continuous effectiveness of model optimization is kept, and the model improvement is more targeted through the arrangement of the data analysis module.

Description

Technical Field

[0001] The present invention relates to the technical field of model optimization, and in particular to an AIGC model optimization method with adaptive learning capability. Background Art

[0002] AIGC refers to a technology based on artificial intelligence (AI) methods such as generative adversarial networks and large pre-trained models. It generates relevant content with appropriate generalization capabilities through learning and identifying existing data. The core idea of ​​AIGC technology is to use artificial intelligence algorithms to generate content with certain creativity and quality. Through training models and learning from a large amount of data, AIGC can generate relevant content based on input conditions or guidance.

[0003] When using common model optimization methods, it is impossible to analyze the adaptation relationship between different AIGC model architectures and various information requirements, which limits the performance of the model and fails to fully tap the potential of different architectures in different information processing. It is easy to fall into local optimal solutions. Therefore, we propose an AIGC model optimization method with adaptive learning capabilities. Summary of the invention

[0004] The purpose of the present invention is to provide an AIGC model optimization method with adaptive learning capability.

[0005] In order to solve the problems raised in the above background technology, the present invention provides the following technical solutions: an AIGC model optimization method with adaptive learning capability, comprising an information acquisition module, wherein the information acquisition module can collect text data to provide information support for the optimization of the AIGC model, and further comprising an architecture selection module, an optimization module, a verification module, and an architecture change module;

[0006] The architecture selection module is responsible for collecting the architecture information of various AIGC models, analyzing and determining the information directions applicable to each model, and then converting these directions into directional features and recording them on the corresponding AIGC models;

[0007] The optimization module selects the known optimization information and the corresponding result information, analyzes the information characteristics of the optimization information, selects the architecture information of the AIGC model according to the information characteristics, generates a test model, inputs the optimization information and the result information into the test model, and extracts the optimization information within the X time period, where X is an unknown number, generates a staged model, analyzes the result information at all X intervals, completes an optimization, and then extracts Optimization information within a time period, generating a secondary stage model, extracting The optimization information within the time period is used to generate three stage models, analyze the result information at all intervals of time, and complete the model optimization. After the optimization is completed, a verification model will be generated, so that the verification model can automatically perform adaptive learning;

[0008] The verification module extracts all the time points with the minimum interval, collects the time period composed of five groups of adjacent time points, and extracts the time point in the middle of the time period to generate five groups of verification points, outputs the five groups of verification points to the verification model, obtains five groups of verification result information, extracts the known result information corresponding to the five groups of verification points from the known optimization information and the corresponding result information, and analyzes the gap between the verification result information and the known result information;

[0009] The architecture change module establishes the architecture threshold and extracts the number of optimizations. When the number of optimizations exceeds the architecture threshold, the architecture selection module is reused to select the architecture information of the AIGC model.

[0010] As a further solution of the present invention: it also includes a data analysis module and an impact analysis module;

[0011] The data analysis module extracts the verification model, inputs the information to be analyzed into the verification model, uses the verification model to predict the input information, and records it, generates model prediction result information, records the actual information in real time, extracts the result information corresponding to the input information, and checks the gap between the result information and the model prediction result information;

[0012] The impact analysis module analyzes the accuracy of the verification model, extracts the normal model information, calculates the ratio between the accuracy of the verification model and the accuracy of the normal model information, obtains the optimization index of the verification model, converts the verification model into an optimization model, and outputs the optimization model.

[0013] As a further solution of the present invention: after converting the applicable information direction into a direction feature, the architecture selection module will create a folder with the corresponding architecture name as the theme, insert the converted direction feature into the folder in the form of a label, and cancel the retrieval permission for the information inside the folder.

[0014] As a further solution of the present invention: when the optimization module selects the architecture information of the AIGC model, it extracts the known optimization information and the corresponding result information, analyzes the usage direction of the known optimization information, and converts the usage direction into a retrieval feature. The retrieval features are J1, J2, J3, ..., J T , let the direction features on the folder label be F1, F2, F3, ..., F Z , let the matching feature be X 符合 ;

[0015] X 符合=(J1, J2, J3, ..., J T )∩(F1, F2, F3,…,F Z )

[0016] According to the above formula, the matching characteristics can be calculated and the number of matching characteristics can be analyzed.

[0017] As a further solution of the present invention: after calculating the matching features, the optimization module will extract the number of matching features, analyze the number of search features, and set the number of matching features as F 数量 , let the number of retrieval features be J 数量 , let the compliance index be F 指数 ;

[0018]

[0019] According to the above formula, the compliance index can be calculated, and the highest compliance index and corresponding architecture information can be extracted.

[0020] As a further solution of the present invention: when optimizing the test model, the optimization module establishes a standard optimization unit, and the standard optimization unit analyzes the standard optimization times according to the time period for extracting the optimization information and the known optimization information, and optimizes the test model according to the number of standard optimization times.

[0021] As a further solution of the present invention: the formula for calculating the difference between the verification result information and the known result information in the verification module is as follows, and an optimization threshold is established synchronously. Suppose the verification result information is Y Z , let the known result information be X Z , let the number of verification results be M, let the verification average difference be C 差异 ;

[0022]

[0023] According to the above formula, the verification average difference can be calculated. Let the optimization increase threshold be Y 阈值 ;

[0024] When C 差异 ≥Y 阈值 , the standard optimization times will be increased by one, the optimization times of the test model will be increased, and the current verification model information will be deleted;

[0025] When C 差异 <Y 阈值 , the current validation model is transferred to the data analysis module.

[0026] As a further solution of the present invention: when the architecture change module reuses the architecture selection module to select the architecture information of the AIGC model, it will synchronously transmit an architecture reselection signal to the data analysis module. When the number of optimization times is lower than the architecture threshold, it will transmit a confirmed analysis signal to the data analysis module.

[0027] As a further solution of the present invention: the data analysis module can extract verification model information, and when receiving the architecture reselection signal, the extracted verification model information will be deleted. When receiving the confirmed analysis signal, the information to be analyzed will be input into the verification model. At the same time, when checking the difference between the result information and the model prediction result information, the actual difference will be obtained, and a prediction difference threshold will be established, wherein the prediction difference threshold will extract the optimization addition threshold and establish a time index, wherein the time index is controlled by the staff, and then the prediction difference threshold is analyzed according to the time index and the optimization addition threshold, and the time index is set as S 指数 , let the prediction error threshold be Y 差阈值 ;

[0028] Y 差阈值 =Y 阈值 (1+S 指数 )

[0029] According to the above formula, the prediction difference threshold can be calculated. When the actual difference is lower than the prediction difference threshold, the verification model will be transferred to the impact analysis module. When the actual difference is higher than the prediction difference threshold, the standard optimization times will be increased by one, and the optimization times of the test model will be increased, so that it has a certain adaptive learning ability.

[0030] By adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are:

[0031] 1. The present invention helps to quickly screen out suitable architectures for different information needs through the architecture selection module, improve the adaptability of the model to the task, and tap the potential of different architectures on different information. The optimization module can adjust the model parameters more carefully to improve the model performance. The verification module can effectively evaluate the accuracy of the model, timely discover the deficiencies of the model, and provide a basis for further optimization. The architecture change module can avoid falling into the local optimal solution and maintain the continuous effectiveness of the model optimization. Through the setting of the data analysis module, the model can be improved more targeted;

[0032] 2. The present invention can avoid the interference of irrelevant information through the architecture selection module, quickly locate the characteristic information related to the architecture, and use the optimization module to accurately screen out the architecture information most relevant to the optimization requirements, so as to further evaluate the degree of adaptation between the selected architecture and the optimization information. At the same time, it can also provide a standardized basis for the optimization process, avoid the blindness and arbitrariness of the optimization, and ensure the consistency and stability of the optimization process;

[0033] 3. The present invention can make the optimization process more flexible and intelligent through the verification module, improve the pertinence and effectiveness of model optimization, realize the collaborative work between different modules through the architecture change module, ensure the consistency and orderliness of the entire optimization process, and effectively improve the model performance. The data analysis module can make the model optimization iteration process more accurate and controllable, which helps to avoid over-optimization or under-optimization while ensuring the model optimization effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0036] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0037] Embodiment 1:

[0038] Therefore, in order to effectively solve the above problems, the present application proposes an AIGC model optimization method with adaptive learning capability, as shown in the accompanying drawings of the specification. Figure 1 As shown, it includes an information collection module, which can collect text data to provide information support for the optimization of the AIGC model, and also includes an architecture selection module, an optimization module, a verification module, and an architecture change module;

[0039] The architecture selection module is responsible for collecting the architecture information of various AIGC models, analyzing and determining the information directions applicable to each model, and then converting these directions into directional features and recording them on the corresponding AIGC models;

[0040] Combined with the characteristics of text data after dynamic screening, the compatibility of the model architecture and data can be judged more accurately. By real-time monitoring of data distribution changes, semantic features, etc., the evaluation criteria for the applicability of the architecture can be adjusted to ensure that the selected architecture can better process the text data at the current stage and improve the efficiency and effect of model optimization.

[0041] The optimization module selects the known optimization information and the corresponding result information, analyzes the information characteristics of the optimization information, selects the architecture information of the AIGC model according to the information characteristics, generates a test model, inputs the optimization information and the result information into the test model, and extracts the optimization information within the X time period, where X is an unknown number, generates a staged model, analyzes the result information at all X intervals, completes an optimization, and then extracts Optimization information within a time period, generating a secondary stage model, extracting The optimization information within the time period is used to generate three stage models, analyze the result information at all intervals of time, and complete the model optimization. After the optimization is completed, a verification model will be generated, so that the verification model can automatically perform adaptive learning;

[0042] Combine the features of different types of information organically to generate more representative feature vectors;

[0043] The verification module extracts all the time points with the minimum interval, collects the time period composed of five groups of adjacent time points, and extracts the time point in the middle of the time period to generate five groups of verification points, outputs the five groups of verification points to the verification model, obtains five groups of verification result information, extracts the known result information corresponding to the five groups of verification points from the known optimization information and the corresponding result information, and analyzes the gap between the verification result information and the known result information;

[0044] The architecture change module establishes an architecture threshold and extracts the number of optimizations. When the number of optimizations exceeds the architecture threshold, the architecture selection module is used again to select the architecture information of the AIGC model.

[0045] It also includes a data analysis module and an impact analysis module;

[0046] The data analysis module extracts the verification model, inputs the information to be analyzed into the verification model, uses the verification model to predict the input information, and records it, generates model prediction result information, records the actual information in real time, extracts the result information corresponding to the input information, and checks the gap between the result information and the model prediction result information;

[0047] The impact analysis module analyzes the accuracy of the verification model, extracts the normal model information, calculates the ratio between the accuracy of the verification model and the accuracy of the normal model information, obtains the optimization index of the verification model, converts the verification model into an optimization model, and outputs the optimization model;

[0048] Record the model's optimization index changes, architecture adjustments, and corresponding input data characteristics during each optimization iteration;

[0049] Specific workflow: collect text data, collect architecture information of different AIGC models, convert applicable information directions into directional features, record them on the corresponding AIGC model, select known optimization information and corresponding result information, input the optimization information and result information into the test model, optimize the test model, obtain five sets of verification result information, analyze the gap between the verification result information and the known result information, and when the number of optimizations exceeds the architecture threshold, re-use the architecture selection module to select the architecture information of the AIGC model, use the verification model to predict the input information, check the gap between the result information and the model prediction result information, calculate the ratio between the accuracy of the verification model and the accuracy of the normal model information, and output the optimization model;

[0050] Furthermore, the architecture selection module helps to quickly screen out suitable architectures for different information needs, improve the adaptability of the model to the task, and tap the potential of different architectures on different information. The optimization module can more finely adjust the model parameters and improve model performance. The verification module can effectively evaluate the accuracy of the model and promptly discover the deficiencies of the model to provide a basis for further optimization. The architecture change module can avoid falling into the local optimal solution and maintain the continuous effectiveness of model optimization. The setting of the data analysis module can make the improvement of the model more targeted.

[0051] Embodiment 2:

[0052] Based on the first embodiment, as shown in the accompanying drawings of the specification Figure 1 As shown, after converting the applicable information direction into a direction feature, the architecture selection module will create a folder with the corresponding architecture name as the theme, insert the converted direction feature into the folder in the form of a label, and cancel the retrieval permission for the information inside the folder;

[0053] When the optimization module selects the architecture information of the AIGC model, it extracts the known optimization information and the corresponding result information, analyzes the usage direction of the known optimization information, and converts the usage direction into retrieval features. The retrieval features are set as J1, J2, J3, ..., J T , let the direction features on the folder label be F1, F2, F3, ..., F Z , let the matching feature be X 符合 ;

[0054] X 符合 =(J1, J2, J3, ..., J T )∩(F1, F2, F3,…,F Z )

[0055] According to the above formula, the matching characteristics can be calculated and the number of matching characteristics can be analyzed;

[0056] After calculating the matching features, the optimization module will extract the number of matching features, analyze the number of search features, and set the number of matching features as F 数量 , let the number of retrieval features be J 数量 , let the compliance index be F 指数 ;

[0057]

[0058] According to the above formula, the compliance index can be calculated, and the highest compliance index and corresponding architecture information can be extracted;

[0059] When optimizing the test model, the optimization module establishes a standard optimization unit, and the standard optimization unit analyzes the standard optimization times according to the time period for extracting the optimization information and the known optimization information, and optimizes the test model according to the number of standard optimization times;

[0060] When the model encounters data scarcity or imbalance during optimization, appropriately reduce the weight of the number of matching features in calculating the matching index to avoid architecture selection bias due to data problems;

[0061] Specific workflow: insert the converted direction features into the folder in the form of tags, cancel the retrieval permission for the information inside the folder, analyze the usage direction of the known optimization information, convert the usage direction into retrieval features, calculate the compliance features and compliance index, extract the architecture information, analyze the standard optimization times according to the time period for extracting the optimization information and the known optimization information, and optimize the test model according to the number of standard optimization times;

[0062] Furthermore, the architecture selection module can avoid interference from irrelevant information and quickly locate feature information related to the architecture. The optimization module can accurately screen out the architecture information most relevant to the optimization requirements and further evaluate the degree of compatibility between the selected architecture and the optimization information. This provides a standardized basis for the optimization process, avoids the blindness and arbitrariness of the optimization, ensures the consistency and stability of the optimization process, and improves the predictability of the optimization effect.

[0063] Embodiment three:

[0064] Based on the second embodiment, as shown in the accompanying drawings of the specification Figure 1 As shown, the formula for calculating the difference between the verification result information and the known result information in the verification module is as follows, and the optimization threshold will be established simultaneously. Suppose the verification result information is Y Z , let the known result information be X Z , let the number of verification results be M, let the verification average difference be C 差异 ;

[0065]

[0066] According to the above formula, the verification average difference can be calculated. Let the optimization increase threshold be Y 阈值 ;

[0067] When C 差异 ≥Y 阈值 , the standard optimization times will be increased by one, the optimization times of the test model will be increased, and the current verification model information will be deleted;

[0068] When C 差异 <Y 阈值 When the current verification model is transferred to the data analysis module;

[0069] Analyze the distribution of the average validation difference in the past multiple validations. If it is found that the average validation difference has been small and the model performance has improved slowly, the optimization increase threshold can be appropriately lowered to encourage the model to make more optimization attempts to break through the performance bottleneck.

[0070] When the architecture change module reuses the architecture selection module to select the architecture information of the AIGC model, it will simultaneously transmit an architecture reselection signal to the data analysis module. When the number of optimizations is lower than the architecture threshold, it will transmit a confirmation analysis signal to the data analysis module.

[0071] The data analysis module can extract the verification model information. When the architecture reselection signal is received, the extracted verification model information will be deleted. When the confirmation analysis signal is received, the information to be analyzed will be input into the verification model. At the same time, when checking the gap between the result information and the model prediction result information, the actual difference will be obtained, and a prediction difference threshold will be established. The prediction difference threshold will extract the optimization increase threshold and establish a time index. The time index is controlled by the staff, and then the prediction difference threshold is analyzed according to the time index and the optimization increase threshold. Let the time index be S 指数 , let the prediction error threshold be Y 差阈值 ;

[0072] Y 差阈值 =Y 阈值 (1+S 指数 )

[0073] According to the above formula, the prediction difference threshold can be calculated. When the actual difference is lower than the prediction difference threshold, the verification model will be transferred to the impact analysis module. When the actual difference is higher than the prediction difference threshold, the standard optimization times will be increased by one, and the optimization times of the test model will be increased, so that it has a certain adaptive learning ability.

[0074] Specific workflow: Determine the processing method for the current verification model, and simultaneously transmit the architecture reselection signal to the data analysis module. When the number of optimizations is lower than the architecture threshold, it will transmit a determination analysis signal to the data analysis module, and input the information to be analyzed into the verification model. When the architecture reselection signal is received, the extracted verification model information will be deleted, and a prediction difference threshold will be established. The prediction difference threshold will be analyzed based on the time index and the optimization addition threshold to determine whether to increase the number of optimizations of the test model.

[0075] Furthermore, the verification module can make the optimization process more flexible and intelligent, improve the pertinence and effectiveness of model optimization, and the architecture change module can achieve collaborative work between different modules, ensure the consistency and orderliness of the entire optimization process, and effectively improve model performance. The data analysis module can make the model optimization iteration process more accurate and controllable, which helps to avoid over-optimization or under-optimization while ensuring the model optimization effect, so that the model can reach a better performance state more quickly.

[0076] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An AIGC model optimization method with adaptive learning capability, comprising an information acquisition module, wherein the information acquisition module can collect text data to provide information support for the optimization of the AIGC model, and wherein: It also includes an architecture selection module, an optimization module, a verification module, and an architecture change module; The architecture selection module is responsible for collecting the architecture information of various AIGC models, analyzing and determining the information directions applicable to each model, and then converting these directions into directional features and recording them on the corresponding AIGC models; The optimization module selects the known optimization information and the corresponding result information, analyzes the information characteristics of the optimization information, selects the architecture information of the AIGC model according to the information characteristics, generates a test model, inputs the optimization information and the result information into the test model, and extracts the optimization information within the X time period, where X is an unknown number, generates a staged model, analyzes the result information at all X intervals, completes an optimization, and then extracts Optimization information within a time period, generating a secondary stage model, extracting The optimization information within the time period is used to generate three stage models, analyze the result information at all intervals of time, and complete the model optimization. After the optimization is completed, a verification model will be generated, so that the verification model can automatically perform adaptive learning; The verification module extracts all the time points with the minimum interval, collects the time period composed of five groups of adjacent time points, and extracts the time point in the middle of the time period to generate five groups of verification points, outputs the five groups of verification points to the verification model, obtains five groups of verification result information, extracts the known result information corresponding to the five groups of verification points from the known optimization information and the corresponding result information, and analyzes the gap between the verification result information and the known result information; The architecture change module establishes the architecture threshold and extracts the number of optimizations. When the number of optimizations exceeds the architecture threshold, the architecture selection module is reused to select the architecture information of the AIGC model.

2. The AIGC model optimization method with adaptive learning capability according to claim 1, characterized in that: It also includes a data analysis module and an impact analysis module; The data analysis module extracts the verification model, inputs the information to be analyzed into the verification model, uses the verification model to predict the input information, and records it, generates model prediction result information, records the actual information in real time, extracts the result information corresponding to the input information, and checks the gap between the result information and the model prediction result information; The impact analysis module analyzes the accuracy of the verification model, extracts the normal model information, calculates the ratio between the accuracy of the verification model and the accuracy of the normal model information, obtains the optimization index of the verification model, converts the verification model into an optimization model, and outputs the optimization model.

3. The AIGC model optimization method with adaptive learning capability according to claim 1, characterized in that: After converting the applicable information direction into a direction feature, the architecture selection module will create a folder with the corresponding architecture name as the theme, insert the converted direction feature into the folder in the form of a label, and cancel the retrieval permission for the information inside the folder.

4. The AIGC model optimization method with adaptive learning capability according to claim 1, characterized in that: When selecting the architecture information of the AIGC model, the optimization module extracts the known optimization information and the corresponding result information, analyzes the usage direction of the known optimization information, and converts the usage direction into a retrieval feature. The retrieval features are J1, J2, J3, ..., J T , let the direction features on the folder label be F1, F2, F3, ..., F Z , let the matching feature be X 符合 ; <h2 style=";text-align:left;direction:ltr">X<h2 style=";text-align:left;direction:ltr"> 符合 <h2 style=";text-align:left;direction:ltr"> (J1, J2, J3,……, J)<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> )∩(F1、F2、F3、……、F<h2 style=";text-align:left;direction:ltr"> Z According to the above formula, the matching characteristics can be calculated and the number of matching characteristics can be analyzed.

5. The AIGC model optimization method with adaptive learning capability according to claim 4, characterized in that: After calculating the matching features, the optimization module will extract the number of matching features, analyze the number of search features, and set the number of matching features as F 数量 , let the number of retrieval features be J 数量 , let the compliance index be F 指数 ; According to the above formula, the compliance index can be calculated, and the highest compliance index and corresponding architecture information can be extracted.

6. The AIGC model optimization method with adaptive learning capability according to claim 5, characterized in that: When optimizing the test model, the optimization module establishes a standard optimization unit, and the standard optimization unit analyzes the standard optimization times according to the time period for extracting the optimization information and the known optimization information, and optimizes the test model according to the number of standard optimization times.

7. The AIGC model optimization method with adaptive learning capability according to claim 2, characterized in that: The formula for calculating the difference between the verification result information and the known result information in the verification module is as follows: Z , let the known result information be X Z , let the number of verification results be M, let the verification average difference be C 差异 ; According to the above formula, the verification average difference can be calculated. Let the optimization increase threshold be Y 阈值 ; When C 差异 ≥Y 阈值 , the standard optimization times will be increased by one, the optimization times of the test model will be increased, and the current verification model information will be deleted; When C 差异 <Y 阈值 , the current validation model is transferred to the data analysis module.

8. The AIGC model optimization method with adaptive learning capability according to claim 7, characterized in that: When the architecture change module reuses the architecture selection module to select the architecture information of the AIGC model, it will simultaneously transmit an architecture reselection signal to the data analysis module. When the number of optimizations is lower than the architecture threshold, it will transmit a confirmation analysis signal to the data analysis module.

9. The AIGC model optimization method with adaptive learning capability according to claim 8, characterized in that: The data analysis module can extract the verification model information. When receiving the architecture reselection signal, the extracted verification model information will be deleted. When receiving the confirmation analysis signal, the information to be analyzed will be input into the verification model. At the same time, when checking the gap between the result information and the model prediction result information, the actual difference will be obtained, and a prediction difference threshold will be established. The prediction difference threshold will extract the optimization addition threshold and establish a time index. The time index is controlled by the staff. Then, the prediction difference threshold is analyzed according to the time index and the optimization addition threshold. Let the time index be S 指数 , let the prediction error threshold be Y 差阈值 ; AND 差阈值 =And 阈值 (1+S 指数 ) According to the above formula, the prediction difference threshold can be calculated. When the actual difference is lower than the prediction difference threshold, the verification model will be transferred to the impact analysis module. When the actual difference is higher than the prediction difference threshold, the standard optimization times will be increased by one, and the optimization times of the test model will be increased, so that it has a certain adaptive learning ability.

Images