WO2021139448A1 - 基于多个源模型修正新模型的方法、装置以及计算机设备 - Google Patents

基于多个源模型修正新模型的方法、装置以及计算机设备 Download PDF

Info

Publication number
WO2021139448A1
WO2021139448A1 PCT/CN2020/132596 CN2020132596W WO2021139448A1 WO 2021139448 A1 WO2021139448 A1 WO 2021139448A1 CN 2020132596 W CN2020132596 W CN 2020132596W WO 2021139448 A1 WO2021139448 A1 WO 2021139448A1
Authority
WO
WIPO (PCT)
Prior art keywords
vector
value
training data
model
new model
Prior art date
Application number
PCT/CN2020/132596
Other languages
English (en)
French (fr)
Inventor
徐玲玲
雷晨雨
张国辉
宋晨
Original Assignee
平安科技(深圳)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2021139448A1 publication Critical patent/WO2021139448A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/22Matching criteria, e.g. proximity measures
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Definitions

  • This application relates to the field of artificial intelligence, and in particular to a method, device and computer equipment for correcting a new model based on multiple source models.
  • the trained model works better in a certain scene, but the effect is poor in other scenes, and the data obtained from the model trained in the field cannot be used directly, but the model can only be reflowed, but the new model will continue to be optimized in the future.
  • Designing a new model requires cleaning the data and adding corresponding tags to different data. If the tags are not accurately added, the output results will not meet the accuracy requirements.
  • the inventor realized that the fusion model is directly used, especially if it is used more. When a model is used, the fused model is too large and the speed is too slow. Therefore, there is an urgent need for a method to modify a new model based on multiple source models.
  • the main purpose of this application is to provide a method, device and computer equipment for correcting a new model based on multiple source models, aiming to solve the technology that causes the fused model to be too large when the existing technology directly merges multiple models directly problem.
  • a method of revising a new model based on multiple source models including:
  • the first training data and the second training data obtained after reversing the first training data are respectively input into a plurality of preset source models for calculation to obtain feature vectors corresponding to each of the source models.
  • the feature vector of the source model includes multiple;
  • the parameters in the new model are corrected according to the gradient value.
  • This application also provides a new model training device based on multiple source models, including:
  • the training data calculation module is used to input the first training data into the new model for calculation to obtain the first current vector;
  • the first training data and the second training data obtained after reversing the first training data are respectively input into a plurality of preset source models for calculation to obtain feature vectors corresponding to each of the source models.
  • the feature vector of the source model includes multiple;
  • An average value calculation module configured to calculate the average value of the feature vector corresponding to each of the source models
  • An index vector calculation module configured to fuse and calculate the average value corresponding to each of the source models to obtain an index vector
  • a first similarity value calculation module configured to calculate a first similarity value between the first current vector and the index vector
  • the first similarity value judgment module is used to judge whether the first similarity value is less than a preset similarity value
  • a gradient value calculation module configured to calculate the gradient value between the current vector and the index vector if the first similarity value is less than a preset similarity value
  • the parameter update module is used to correct the parameters in the new model according to the gradient value.
  • the present application also provides a computer device, including a memory and a processor, the memory stores a computer program, and the processor implements the steps of a method for correcting a new model based on multiple source models when the computer program is executed: Input the first training data into the new model for calculation to obtain the first current vector; and,
  • the first training data and the second training data obtained after reversing the first training data are respectively input into a plurality of preset source models for calculation to obtain feature vectors corresponding to each of the source models.
  • the feature vector of the source model includes multiple;
  • the parameters in the new model are corrected according to the gradient value.
  • the present application also provides a computer-readable storage medium on which a computer program is stored.
  • the steps of implementing a method of correcting a new model based on multiple source models input first training data Perform calculations in the new model to obtain the first current vector; and,
  • the first training data and the second training data obtained after reversing the first training data are respectively input into a plurality of preset source models for calculation to obtain feature vectors corresponding to each of the source models.
  • the feature vector of the source model includes multiple;
  • the parameters in the new model are corrected according to the gradient value.
  • the beneficial effects of this application by inputting training data into multiple existing source models, multiple corresponding feature vectors are obtained, then the index vector is obtained by fusion calculation, and then the gradient between the index vector and the current vector obtained by the new model is calculated Value, the parameter in the new model is corrected by the gradient value.
  • the new model is trained based on multiple source models, and there is no need to merge multiple source models. While improving the calculation accuracy of the new model, the new model combines the advantages of multiple source models and avoids direct use. The problem of fusion model becoming larger and slower.
  • FIG. 1 is a schematic flowchart of a method for modifying a new model based on multiple source models according to an embodiment of the present application
  • FIG. 2 is a schematic block diagram of the structure of an apparatus for correcting a new model based on multiple source models according to an embodiment of the application;
  • FIG. 3 is a schematic block diagram of the structure of a computer device according to an embodiment of the application.
  • this application proposes a method for modifying a new model based on multiple source models, including:
  • S1 Input the first training data into the new model for calculation to obtain a first current vector; and input the first training data and the second training data obtained after flipping the first training data into multiple presets respectively It is assumed that the source model is calculated to obtain a feature vector corresponding to each of the source models, wherein the feature vector corresponding to each of the source models includes multiple;
  • multiple preset source models can be selected according to the actual data required. For example, model a is better in financial scenarios, model b is better in monitoring scenarios, and model c is better in attendance scenarios.
  • the acquired data needs to synthesize the above three scenarios, then the corresponding model a, model b, and model c can be selected.
  • a training model it may only be a trained model without its original training data. Input the training data and its flipped data into the preset source model, and you can get multiple feature vectors corresponding to each model.
  • This article takes only the left-right flip as an example, assuming it is the first training data input before flipping Get the feature vector aV1 from model a, and input the second training data after left and right flips into model a to get feature vector aV2.
  • model b feature vectors bV1 and bV2 are obtained accordingly, if model c is also available , Model d, etc., according to the corresponding method to obtain the two feature vectors of the corresponding model.
  • the flipped data is also similar to the training data. This application flips the training data. In order to obtain training data, obtain more values in each model, so as to improve the accuracy of the training model in the subsequent calculation process.
  • the first training data is also input into the new model to obtain the first current vector.
  • step S2 different source models obtain multiple feature vectors through the first training data and the flipped second training data, and then calculate the average value of the feature vectors corresponding to the same source model.
  • the calculated average value is an excellent value obtained by each model based on the training data. Generally speaking, the result of the average value will be better than the value obtained by directly inputting the first training data.
  • the average values corresponding to the source models are then merged to form a complete output, that is, the average value of the average value corresponding to each source model is calculated, or the average value is subjected to dimensionality reduction operation, for example,
  • the mean value of the value (indicator vector), then the mean value V (aV+bV)/2, if the fusion is to adopt a dimensionality reduction method, then the dimensionality reduction method can be missing value ratio, low variance filtering, high correlation filtering, random forest / Combination tree, principal component analysis and inverse feature elimination, finally get the index vector, which is used to compare the training of the new model.
  • the first current vector is compared with the index vector, that is, the first similarity value is calculated.
  • the calculation formula can be When the value of acc is closer to 1, it means that the current vector is more similar to the index vector, and when the value of acc is closer to 0, it means that the current vector is more dissimilar to the index vector.
  • the formula represents a first similarity value acc
  • zv i represents the i corresponding to the current feature vector dimension vector value
  • v i represents a vector of values corresponding to the index i-th dimension feature vector
  • n is the number of feature vector dimension.
  • the calculated first similarity value is compared with the similarity threshold. If the current similarity is greater than the similarity threshold, it can be explained that the current vector can be equivalent to the target vector; if the current similarity is Less than the similarity threshold, it indicates that the current vector is not the target vector, and the parameters in the model need to be adjusted to obtain the current vector again.
  • the gradient value is obtained according to the loss, and the formula for obtaining the original value of the gradient is Wherein Loss (zV, V) to the original value gradient, zv i represents the current vector corresponding to the i-th feature vector dimension, v i represents the index vector corresponding to the i-th feature vector dimension, n is the number of feature vector dimension. Then derivate Loss(zV,V) to obtain the gradient value, and then correct the parameters in the new model according to the acquired gradient value to achieve the technical effect of training the new model.
  • the parameters also include:
  • the corrected parameters can be tested again, that is, the current vector is recalculated according to the corrected parameters, and the second similarity value with the index vector is calculated. If the second similarity value is greater than the predicted value If the similarity value is set, the current vector after recalculation can be recorded as the target vector. If the second similarity value is less than or equal to the preset similarity value, you can continue to obtain the new gradient value through the above loss, change the corresponding parameter according to the new gradient value, and repeat for many times until the current vector and the index vector are the second If the similarity value is greater than the preset similarity value, it is determined that the training of the new model based on the first training data is completed. In addition, you can also input new training data, obtain new index vectors and current vectors, and perform repeated training on the new model. After reaching a certain number of times, it is deemed that the training of the new model has been completed.
  • the aforementioned step S3 of fusing the average values corresponding to each of the source models to obtain an index vector includes:
  • the relevant factors can be determined based on the role of the new model.
  • the new model is for the management of the company’s personnel, and the relevant factors are relatively large.
  • the application scenarios are monitoring scenarios, attendance scenarios, etc., and financial scenarios have relatively small related factors.
  • the value of the related factors can be determined according to the role of the new model.
  • the related factor of the monitoring scenario can be set to 0.5
  • the related factors of attendance can be set
  • the factor is set to 0.7
  • the relevant factor of the financial scene is set to 0.2
  • the index vector is calculated according to the relevant factors of each scene, so that the similarity value is calculated based on the index vector later, and the updated parameters can be closer to the application of the new model, so that the new model obtained by training is better.
  • the step S4 of calculating the first similarity value between the first current vector and the index vector as described above includes:
  • S401 Obtain parameter values of the index vector and the first current vector, where the parameter value includes at least the number of feature vector dimensions of the index vector and the current vector, and the index vector value and current value in each dimension.
  • Vector value includes at least the number of feature vector dimensions of the index vector and the current vector, and the index vector value and current value in each dimension.
  • the first similarity value will be calculated according to the formula according to the number of feature vector dimensions of the index vector and the current vector, and the index vector value and the current vector value in each dimension. The closer the value is to 1, the more similar the current vector and the index vector, and the closer the acc value is to 0, the less similar the current vector and the index vector.
  • acc represents the first similarity value
  • zvi represents the current vector value corresponding to the i-th feature vector dimension
  • vi represents the corresponding index vector value in the i-th feature vector dimension
  • n is the number of feature vector dimensions.
  • the first similarity value calculated according to the above formula can determine the current vector of the output of the new model and the similarity of the index vector after fusion with other models, and then determine whether the new model needs to adjust parameters according to the similarity. Among them, when the first similarity value approaches 1, it is considered that the current vector is more related to the index vector, and when the first similarity value approaches 0, it is considered that the current vector and the index vector are less related.
  • step S1 of inputting the first training data into the new model for calculation to obtain the first current vector it includes:
  • S002 Set pixels smaller than the preset pixel point threshold to 0, and set pixels greater than the preset pixel threshold to 1;
  • S003 Divide the picture into multiple regions on average, count the number of pixels in each region as 1, and form a matrix as the training data.
  • the picture is digitally processed to obtain each pixel in the picture, and then each pixel is compared with a preset pixel threshold.
  • the pixel threshold is a preset value.
  • the designer can set different values for the preset pixel threshold as needed, and then set the pixels smaller than the preset pixel threshold to 0, and set the pixels larger than the preset pixel threshold to 1.
  • Get an initial array, and then divide the image into multiple regions on average for example, divide the 32 ⁇ 32 initial array into multiple regions on average to obtain an 8 ⁇ 8 matrix, and then use the resulting matrix as training data for training .
  • this application can also perform other processing on other training data, such as text, customer information, etc., which will not be repeated here.
  • the method before the step S1 of inputting the first training data into the new model for calculation to obtain the first current vector, the method further includes:
  • S012 Compare the structure quantitative value with data in a preset list; wherein, the preset list includes the corresponding relationship between the numerical value of the structure quantitative value and the model;
  • the corresponding structural quantization value is calculated according to the application scenario of the new model.
  • the structural quantization value is used to express the structural complexity of the model. The larger the structural quantification value, the more complex the structure and the greater the structural quantification value. Smaller means the model is simpler. For example, when a new model is applied to a terminal, especially a smart terminal, when selecting a new model, the structure of the new model should not be too complicated, otherwise it will easily occupy the running memory of the terminal’s CPU and slow down the terminal. If it is on the server side, the server generally has a high running memory. In order to make the data obtained more accurate, it is possible to use a complex new model structure.
  • the step S1 of selecting a plurality of the source models for training the new model includes:
  • the feature set of each model in the source model database and the feature set of the new model are passed through the formula Perform calculations.
  • the features of the model and the new model can be algorithms, application scenarios, input values, and output values.
  • the correlation can be calculated according to their respective characteristics. When the calculated correlation approaches 1, it means that the model and the The more relevant the new model is, and the closer the calculated correlation degree is to 0, it means that the model is less relevant to the new model. Therefore, a preset relevance threshold can be set.
  • the preset relevance threshold is a value set in advance according to the actual situation.
  • the preset relevance threshold can be set correspondingly larger If the number of models in the source model database is small, the preset correlation threshold can be set smaller accordingly. Then, a model with a correlation greater than a preset correlation threshold is selected as the source model, so that the training effect of the new model based on the selected source model is better.
  • the beneficial effect of this application by inputting training data into multiple existing source models, multiple corresponding feature vectors are obtained, then the index vector is obtained by fusion calculation, and then the gradient between the index vector and the current vector obtained by the new model is calculated Value, the parameter in the new model is corrected by the gradient value.
  • the new model is trained based on multiple source models, and there is no need to merge multiple source models. While improving the calculation accuracy of the new model, the new model integrates the advantages of multiple source models and avoids direct use. The problem of fusion model becoming larger and slower.
  • this application proposes a new model training device based on multiple source models, including:
  • the training data calculation module 10 is configured to input the first training data into the new model for calculation to obtain the first current vector;
  • the first training data and the second training data obtained after reversing the first training data are respectively input into a plurality of preset source models for calculation to obtain feature vectors corresponding to each of the source models.
  • the feature vector of the source model includes multiple;
  • the average value calculation module 20 is configured to calculate the average value of the feature vector corresponding to each of the source models
  • the index vector calculation module 30 is configured to fuse and calculate the average value corresponding to each of the source models to obtain an index vector
  • a first similarity value calculation module 40 configured to calculate a first similarity value between the first current vector and the index vector
  • the first similarity value judgment module 50 is configured to judge whether the first similarity value is less than a preset similarity value
  • the gradient value calculation module 60 is configured to calculate the gradient value between the current vector and the index vector if the first similarity value is less than a preset similarity value;
  • the parameter update module 70 is configured to correct the parameters in the new model according to the gradient value.
  • model a is better in financial scenarios
  • model b is better in monitoring scenarios
  • model c is better in attendance scenarios.
  • the data that needs to be obtained needs to be integrated into the top three
  • model a, model b, and model c can be selected.
  • it may only be a trained model without its original training data. Input the training data and its flipped data into the preset source model, and you can get multiple feature vectors corresponding to each model.
  • This article takes only the left-right flip as an example, assuming it is the first training data input before flipping Get the feature vector aV1 from model a, and input the second training data after left and right flips into model a to get feature vector aV2.
  • model b feature vectors bV1 and bV2 are obtained accordingly, if model c is also available , Model d, etc., according to the corresponding method to obtain the two feature vectors of the corresponding model.
  • the flipped data is also similar to the training data. This application flips the training data. In order to obtain training data, obtain more values in each model, so as to improve the accuracy of the training model in the subsequent calculation process. Then input the first training data into the new model to obtain the first current vector.
  • Different source models obtain multiple feature vectors through the first training data and the flipped second training data input, and then calculate the average value of the feature vectors corresponding to the same source model.
  • the calculated average value is an excellent value obtained by each model based on the training data. Generally speaking, the result of the average value will be better than the value obtained by directly inputting the first training data.
  • the average value corresponding to each source model is merged to form a complete output, that is, the average value corresponding to each source model is calculated, or the average value is reduced in dimensionality.
  • the first current vector is compared with the index vector, that is, the first similarity value is calculated.
  • the calculation formula can be When the value of acc is closer to 1, it means that the current vector is more similar to the index vector, and when the value of acc is closer to 0, it means that the current vector is more dissimilar to the index vector.
  • the formula represents a first similarity value acc
  • zv i represents the i corresponding to the current feature vector dimension vector value
  • v i represents a vector of values corresponding to the index i-th dimension feature vector
  • n is the number of feature vector dimension.
  • the gradient value is obtained according to loss, and the formula for obtaining the original value of the gradient is Wherein Loss (zV, V) to the original value gradient, zv i represents the current vector corresponding to the i-th feature vector dimension, v i represents the index vector corresponding to the i-th feature vector dimension, n is the number of feature vector dimension. Then derivate Loss(zV,V) to obtain the gradient value, and then correct the parameters in the new model according to the acquired gradient value to achieve the technical effect of training the new model.
  • the device for training a new model based on multiple source models further includes:
  • the recalculation module is used to input the first training data into the new model after the correction parameters for calculation to obtain a second current vector, and calculate the second similarity value between the second current vector and the index vector ;
  • the second similarity value judgment module is configured to judge whether the second similarity value is greater than the preset similarity value
  • the training recognition module is configured to determine that the training of the new model based on the first training data is completed if the second similarity value is greater than the preset similarity value.
  • the corrected parameters can be tested again, that is, the current vector is recalculated according to the corrected parameters, and the second similarity value with the index vector is calculated. If the second similarity value is greater than the preset similarity value, the recalculation can be performed. The calculated current vector is recorded as the target vector. If the second similarity value is less than or equal to the preset similarity value, you can continue to obtain the new gradient value through the above loss, change the corresponding parameter according to the new gradient value, and repeat for many times until the current vector is the second of the index vector If the similarity value is greater than the preset similarity value, it is determined that the training of the new model based on the first training data is completed. In addition, you can also input new training data, obtain new index vectors and current vectors, and perform repeated training on the new model. After reaching a certain number of times, it is considered that the training of the new model is completed.
  • the index vector calculation module 30 includes:
  • the correlation factor acquisition sub-module is used to acquire correlation factors of the new model in different application scenarios
  • the index vector calculation sub-module is used to select the application scenario of the source model according to the formula Fusion vectors calculated metrics, wherein, V is the target vector, W i is the i th scenario correlation factor, f (w i) is the average value of the source model application scenario i-th scene.
  • the relevant factors can be determined based on the role of the new model.
  • the new model is for the company's personnel management, and the application scenarios with larger relevant factors are monitoring scenarios and attendance scenarios
  • the relevant factor of the financial scene can be determined according to the function of the new model.
  • the relevant factor of the monitoring scene can be set to 0.5
  • the relevant factor of attendance can be set to 0.7
  • the relevant factor of the financial scene can be set to 0.5.
  • the correlation factor is set to 0.2, and then according to the formula
  • the index vector is calculated according to the relevant factors of each scene, so that the similarity value is calculated based on the index vector later, and the updated parameters can be closer to the application of the new model, so that the new model obtained by training is better.
  • the first similarity value calculation module 40 includes:
  • the parameter value acquisition sub-module is used to acquire the parameter values of the index vector and the first current vector, where the parameter value includes at least the number of feature vector dimensions of the index vector and the current vector, and the number of dimensions in each dimension Medium index vector value and current vector value;
  • the first calculation sub-module is used according to the formula Calculating a first similarity value, which represents a first similarity value acc, ZV current vector i represents a first value corresponding to the i-th dimension feature vector, V i represents the index value of the vector corresponding to the i-th feature vector dimension, n is the number of eigenvector dimensions.
  • the first similarity value will be calculated according to the formula according to the number of eigenvector dimensions of the index vector and the current vector, as well as the index vector value and the current vector value in each dimension.
  • acc represents the first similarity value
  • zvi represents the current vector value corresponding to the i-th feature vector dimension
  • vi represents the corresponding index vector value in the i-th feature vector dimension
  • n is the number of feature vector dimensions.
  • the first similarity value calculated according to the above formula can determine the current vector of the output of the new model and the similarity of the index vector after fusion with other models, and then determine whether the new model needs to adjust parameters according to the similarity. Among them, when the first similarity value approaches 1, it is considered that the current vector is more related to the index vector, and when the first similarity value approaches 0, it is considered that the current vector and the index vector are less related.
  • a new model training device based on multiple source models includes:
  • a pixel point acquisition module configured to acquire each pixel point in the picture when the first training data is a picture, and compare each pixel point with a preset pixel point threshold;
  • a pixel point setting module configured to set pixels smaller than the preset pixel point threshold to 0, and set pixels larger than the preset pixel point threshold to 1;
  • the area division module is used to divide the picture into multiple areas on average, count the number of pixels in each area as 1, and form a matrix as the training data.
  • the picture is digitally processed to obtain each pixel in the picture, and then each pixel is compared with a preset pixel threshold.
  • the pixel threshold is a pre-set value. The designer can adjust the preset pixel as needed. Dot threshold is set to different values, and then pixels smaller than the preset pixel threshold are set to 0, and pixels larger than the preset pixel threshold are set to 1, to obtain an initial array, and then set
  • the picture is divided into multiple regions on average, for example, the 32 ⁇ 32 initial array is divided into multiple regions on average to obtain an 8 ⁇ 8 matrix, and then the obtained matrix is used as training data for training.
  • this application can also perform other processing on other training data, such as text, customer information, etc., which will not be repeated here.
  • the above-mentioned new model training device based on multiple source models further includes:
  • the structural quantitative value calculation module is used to calculate the corresponding structural quantitative value according to the application scenario of the new model
  • the structural quantification value comparison module is configured to compare the structure quantification value with data in a preset list; wherein the preset list includes the corresponding relationship between the numerical value of the structure quantification value and the model;
  • the new model screening module is used for screening the new model from the model database according to the comparison result.
  • the structural quantification value is used to express the structural complexity of the model.
  • the larger the structural quantification value the more complex the structure.
  • the smaller the structural quantification value the simpler the model.
  • the structure of the new model should not be too complicated, otherwise it will easily occupy the running memory of the terminal CPU and slow down the running speed of the terminal.
  • the server generally has a high running memory, in order to make the data more accurate, it can use a complex new model structure. It should be understood that the more complex the model, the higher the running memory it occupies, and the more accurate the calculation results of the data. Based on the above considerations, a new model should be selected for the application scenario, so a preset list or preset can be set in advance Function, you can select a suitable new model according to the application scenario, making the selected new model more practical.
  • the training data calculation module 10 includes:
  • Correlation calculation sub-module used to pass formula Calculate the correlation between each model in the source model database and the new model; where X represents the feature set of the model, and Y represents the feature set of the new model;
  • the correlation degree comparison module is used to compare the correlation degree between each model and the new model with a preset correlation degree threshold
  • the source model selection module is configured to select a model with the correlation degree greater than the preset correlation degree threshold as the source model.
  • the features of the model and the new model can be algorithms, application scenarios, input values, and output values.
  • the correlation can be calculated according to their respective characteristics. When the calculated correlation approaches 1, it means that the model and the The more relevant the new model is, and the closer the calculated correlation degree is to 0, it means that the model is less relevant to the new model. Therefore, a preset relevance threshold can be set.
  • the preset relevance threshold is a value set in advance according to the actual situation. For example, if there are enough models in the source model database, the preset relevance threshold can be set larger accordingly If the number of models in the source model database is small, the preset correlation threshold can be set smaller accordingly. Then, a model with a correlation greater than a preset correlation threshold is selected as the source model, so that the training effect of the new model based on the selected source model is better.
  • an embodiment of the present application also provides a computer device.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 3.
  • the computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor designed by the computer is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, a computer program, and a database.
  • the memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium.
  • the database of the computer equipment is used to store various training data and so on.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • FIG. 3 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied.
  • the embodiments of the present application also provide a computer-readable storage medium.
  • the above-mentioned storage medium may be a non-volatile storage medium or a volatile storage medium.
  • a computer program is stored thereon, and when the computer program is executed by the processor, the method for correcting a new model based on multiple source models described in any of the above embodiments can be implemented.

Abstract

本申请及人工智能领域,提供了一种基于多个源模型修正新模型的方法、装置以及计算机设备,其中方法包括:将第一训练数据输入至新模型中得到第一当前向量;以及,将第一训练数据分别输入至多个预设的源模型中进行计算,得到对应各源模型的特征向量;并融合计算得到指标向量;计算第一当前向量与指标向量之间的梯度值;根据梯度值校正新模型中的参数。本申请的有益效果:通过将训练数据输入现有的多个源模型中,得到对应的多个特征向量,然后融合计算得到指标向量,然后计算指标向量与新模型得到的当前向量之间的梯度值,通过梯度值校正新模型中的参数。使新模型综合了多个源模型融合后的优点,避免了直接使用融合模型,体积变大,速度变慢的问题。

Description

基于多个源模型修正新模型的方法、装置以及计算机设备
本申请要求于2020年07月31日提交中国专利局、申请号为2020107609139,发明名称为“基于多个源模型修正新模型的方法、装置以及计算机设备”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能领域,特别涉及一种基于多个源模型修正新模型的方法、装置以及计算机设备。
背景技术
目前已经训练好的模型在某个场景下效果比较好,在其他场景下效果比较差,并且在外场训练的模型,其得到的数据不能直接利用,只能回流模型,但是后续继续优化新模型的时候缺少数据。设计新的模型需要清洗数据,给不同的数据添加相应的标签,若标签添加的不准确,则会导致输出的结果达不到精度要求,而发明人意识到直接使用融合模型,特别是使用多个模型时,融合后的模型过大,速度过慢。因此,亟需一种基于多个源模型修正新模型的方法。
技术问题
本申请的主要目的为提供一种基于多个源模型修正新模型的方法、装置以及计算机设备,旨在解决现有技术直接对多个模型直接进行融合时,导致融合后的模型过大的技术问题。
技术解决方案
一种基于多个源模型修正新模型的方法,包括:
将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
分别计算每个所述源模型对应的所述特征向量的平均值;
将各所述源模型对应的平均值融合计算得到指标向量;
计算所述第一当前向量与所述指标向量的第一相似度值;
判断所述第一相似度值是否小于预设相似度值;
若所述第一相似度值小于预设相似度值,则计算所述第一当前向量与所述指标向量之间的梯度值;
根据所述梯度值校正所述新模型中的参数。
本申请还提供了一种基于多个源模型的新模型训练装置,包括:
训练数据计算模块,用于将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
平均值计算模块,用于分别计算每个所述源模型对应的所述特征向量的平均值;
指标向量计算模块,用于将各所述源模型对应的平均值融合计算得到指标向量;
第一相似度值计算模块,用于计算所述第一当前向量与所述指标向量的第一相似度值;
第一相似度值判断模块,用于判断所述第一相似度值是否小于预设相似度值;
梯度值计算模块,用于若所述第一相似度值小于预设相似度值,则计算所述当前向量与所述指标向量之间的梯度值;
参数更新模块,用于根据所述梯度值校正所述新模型中的参数。
本申请还提供了一种计算机设备,包括存储器和处理器,所述存储器存储有计算机程序,所述处理器执行所述计算机程序时实现一种基于多个源模型修正新模型的方法的步骤:将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
分别计算每个所述源模型对应的所述特征向量的平均值;
将各所述源模型对应的平均值融合计算得到指标向量;
计算所述第一当前向量与所述指标向量的第一相似度值;
判断所述第一相似度值是否小于预设相似度值;
若所述第一相似度值小于预设相似度值,则计算所述第一当前向量与所述指标向量之间的梯度值;
根据所述梯度值校正所述新模型中的参数。
本申请还提供了一种计算机可读存储介质,其上存储有计算机程序,所述计算机程序被处理器执行时实现种基于多个源模型修正新模型的方法的步骤:将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
分别计算每个所述源模型对应的所述特征向量的平均值;
将各所述源模型对应的平均值融合计算得到指标向量;
计算所述第一当前向量与所述指标向量的第一相似度值;
判断所述第一相似度值是否小于预设相似度值;
若所述第一相似度值小于预设相似度值,则计算所述第一当前向量与所述指标向量之间的梯度值;
根据所述梯度值校正所述新模型中的参数。
有益效果
本申请的有益效果:通过将训练数据输入现有的多个源模型中,得到对应的多个特征向量,然后融合计算得到指标向量,然后计算指标向量与新模型得到的当前向量之间的梯度值,通过梯度值校正新模型中的参数。使新模型基于多个源模型训练而成,且无需将多个源模型进行融合,在提高新模型的计算精度的同时,使新模型综合了多个源模型融合后的优点,避免了直接使用融合模型,体积变大,速度变慢的问题。
附图说明
图1是本申请一实施例的一种基于多个源模型修正新模型的方法的流程示意图;
图2为本申请一实施例的基于多个源模型修正新模型的装置的结构示意框图;
图3为本申请一实施例的计算机设备的结构示意框图。
本发明的最佳实施方式
参照图1,本申请提出一种基于多个源模型修正新模型的方法,包括:
S1:将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
S2:分别计算每个所述源模型对应的所述特征向量的平均值;
S3:将各所述源模型对应的平均值融合计算得到指标向量;
S4:计算所述第一当前向量与所述指标向量的第一相似度值;
S5:判断所述第一相似度值是否小于预设相似度值;
S6:若所述第一相似度值小于预设相似度值,则计算所述当前向量与所述指标向量之间的梯度值;
S7:根据所述梯度值校正所述新模型中的参数。
如上述步骤S1所述,可以根据实际需要的数据选取多个预设的源模型,例如模型a在金融场景比较好,模型b在监控场景表现比较好,模型c在考勤场景比较比较好,需要获取的数据需要综合上边三个场景,那么就可以选取对应的模型a、模型b和模型c。需要说明的是,对于选取训练模型,其可以只是训练完毕的模型,而不需要其原始训练数据。将训练数据和其翻转后的数据输入至预设的源模型中,可以得到每个模型对应的多个特征向量,本文以只进行了左右翻转为例,假设为翻转前的第一训练数据输入至模型a中得到特征向量aV1,将左右翻转后的第二训练数据输入至模型a中得到特征向量aV2,当然,若输入至模型 b中,相应得到特征向量bV1和bV2,若还具有模型c,模型d等,按照相应的方法得到对应模型的两个特征向量,需要说明的是,由于训练数据相同,故而其翻转后的数据与该训练数据也是相似的,本申请将训练数据翻转,是为了获取训练数据在各模型中得到更多的值,以便于后续计算过程中提升训练模型的精度。与此同时,还将第一训练数据输入至新模型中,得到第一当前向量。
如上述步骤S2所述,不同的源模型通过第一训练数据和其翻转后的第二训练数据输计算得到多个特征向量,然后计算得到对应同一个源模型的特征向量的平均值。求出的平均值是各模型基于训练数据得到的一个优值,一般而言,该平均值的结果会优于将第一训练数据直接输入后得到的值。
如上述步骤S3所述,然后再将各源模型对应的平均值进行融合,形成一个完整的输出,即求各源模型对应的平均值的均值,或者将各平均值进行降维操作,举例而言,假设具有两个模型那么各模型的平均值分别为aV=(aV1+aV2)/2,bV=(bV1+bV2)/2,然后将其融合,假设融合是求各源模型对应的平均值的均值(指标向量),那么均值V=(aV+bV)/2,若融合是采取降维的方式,那么降维的方法可以是缺失值比率、低方差滤波、高相关滤波、随机森林/组合树、主成分分析和反特征消除中的一种,最终得到指标向量,用于对新模型训练的对比。
如上述步骤S4所述,将第一当前向量与指标向量进行比较,即计算第一相似度值,计算的公式可以是
Figure PCTCN2020132596-appb-000001
当acc的值越接近于1,表明当前向量与指标向量越相似,当acc的值越接近于0时,表明当前向量与指标向量越不相似。其中公式中acc表示第一相似度值,zv i表示在第i特征向量维度中对应的当前向量数值,v i表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。
如上述步骤S5-S7所述,将计算的第一相似度值与相似度阈值进行比较,若当前的相似度大于相似度阈值,则可以说明当前向量可以相当于目标向量;若当前的相似度小于相似度阈值,则表明当前向量不是目标向量,还需要调整模型中的参数,重新获取当前向量,具体地,根据loss获取梯度值,获取梯度原始值的公式为
Figure PCTCN2020132596-appb-000002
其中Loss(zV,V)为梯度的原始值,zv i表示在第i特征向量维度中对应的当前向量,v i表示在第i特征向量维度中对应的指标向量,n为特征向量维度数量。然后再对Loss(zV,V)进行求导,得到梯度值,然后再根据获取到的梯度值校正新模型中的参数,以达到训练新模型的技术效果。
本实施例中,上述若所述第一相似度值小于预设相似度值,则计算所述当前向量与所述指标向量之间的梯度值,并根据所述梯度值校正所述新模型中的参数的步骤S7之后,还包括:
S8:将所述第一训练数据输入到校正参数后的新模型中进行计算,得到第二当前向量,并计算所述第二当前向量与所述指标向量的第二相似度值;
S9:判断所述第二相似度值是否大于所述预设相似度值;
S10,若第二相似度值大于所述预设相似度值,则认定所述新模型基于所述第一训练数据的训练完成。
如上述步骤S8-S10所述,可以对校正后的参数再次进行检测,即将根据校正后的参数重新计算当前向量,并计算与指标向量的第二相似度值,若第二相似度值大于预设相似度值,则可以将重新计算后的当前向量记为目标向量。若第二相似度值小于或等于预设相似度值,则可以通过上述loss继续获取新的梯度值,根据新的梯度值更改对应的参数,重复多次,直至当前向量与指标向量的第二相似度值大于预设相似度值,认定所述新模型基于所述第一训练数据的训练完成。另外,还可以输入新的训练数据,获取新的指标向量和当前向量,对新模型进行多次反复的训练,达到一定次数后,视为完成了新模型的训练。
本实施例中,上述将各所述源模型对应的平均值融合计算得到指标向量的步骤S3,包括:
S301:获取所述新模型在不同应用场景的相关因子;
S302:根据选取的所述源模型的应用场景,依照公式
Figure PCTCN2020132596-appb-000003
融合计算得到指标向量,其中,V为目标向量,w i为第i个应用场景的相关因子,f(w i)为应用场景为第i个场景的源模型的平均值。
如上述步骤S301-S302所述,获取新模型在不同应用场景下的相关因子,该相关因子可以是基于新模型的作用进行确定的,例如新模型是为了公司的人员管理,则相关因子较大的应用场景为监控场景、考勤场景等,而金融场景的相关因子较小,其中相关因子的数值可以根据新模型的作用进行确定,例如可以将监控场景的相关因子设置为0.5,将考勤的相关因子设置为0.7,将金融场景的相关因子设置为0.2,然后再根据公式
Figure PCTCN2020132596-appb-000004
依据各个场景的相关因子计算得到指标向量,使后续基于指标向量计算相似度值,以及更新的参数可以更加贴近新模型的应用,使训练得到的新模型更好。
本实施例中,如上述计算所述第一当前向量与所述指标向量的第一相似度值的步骤S4,包括:
S401:获取所述指标向量和所述第一当前向量的参数值,其中所述参数值至少包括所述指标向量和所述当前向量的特征向量维度数量,以及在各维度中指标向量数值和当前向量数值;
S402:根据公式
Figure PCTCN2020132596-appb-000005
计算第一相似度值,其中acc表示第一相似度值,zv i表示在第i特征向量维度中对应的第一当前向量数值,v i表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。
如上述步骤S401-S402所述,将根据所述指标向量和所述当前向量的特征向量维度数量,以及在各维度中指标向量数值和当前向量数值依照公式计算第一相 似度值,当acc的值越接近于1,表明当前向量与指标向量越相似,当acc的值越接近于0时,表明当前向量与指标向量越不相似。其中公式中acc表示第一相似度值,zvi表示在第i特征向量维度中对应的当前向量数值,vi表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。根据上述公式计算的第一相似度值可以判断新模型的输出的当前向量,与其他模型融合后的指标向量的相似度,再根据相似度判断新模型是否需要调整参数。其中,第一相似度值越趋近于1时,则认为当前向量与指标向量越相关,第一相似度值越趋近于0时,则认为当前向量与指标向量越不相关。
本实施例中,上述将第一训练数据输入至所述新模型中进行计算,得到第一当前向量的步骤S1之前,包括:
S001:当所述第一训练数据为图片时,获取图片中的每个像素点,将每个所述像素点和预设像素点阈值进行比较;
S002:将小于所述预设像素点阈值的像素点设置为0,将大于所述预设像素点阈值的像素点设置为1;
S003:将所述图片平均划分为多个区域,并统计每个区域中的像素点为1的个数,并构成矩阵作为所述训练数据。
如上述步骤S001-S003所述,将图片进行数字化处理,得到图片中的每个像素点,然后将每个像素点和预设像素点阈值进行比较,该像素点阈值为事先设定的值,设计人员可以根据需要对预设像素点阈值进行设定不同的值,然后将小于所述预设像素点阈值的像素点设置为0,将大于所述预设像素点阈值的像素点设置为1,得到一个初始阵列,然后再将图片平均划分为多个区域,例如将32×32的初始阵列平均划分为多个区域,以得到8×8的矩阵,然后将得到的矩阵作为训练数据进行训练。当然,上述指的是对图片进行训练数据的处理,本申请还可以对其他训练数据进行其他的处理,例如文本,客户信息等,此处不再赘述。
本实施例中,上述将第一训练数据输入至所述新模型中进行计算,得到第一当前向量的步骤S1之前,还包括:
S011:根据所述新模型的应用场景计算对应的结构量化值;
S012:根据所述结构量化值与预设列表中的数据进行比对;其中,所述预设列表包括了所述结构量化值的数值与模型的对应关系;
S013:根据比对结果从模型数据库中筛选出所述新模型。
如上述步骤S011-S013所述,根据新模型的应用场景计算对应的结构量化值,其中结构量化值用于表示模型的结构复杂度,结构量化值越大,表示结构越复杂,结构量化值越小,表示模型越简单,举例而言,当新模型应用在终端,尤其是智能终端时,选择新模型时,新模型的结构不宜太复杂,否则容易占用终端CPU 的运行内存,减慢了终端的运行速度,若在服务器端,服务器一般有很高的运行内存,为了使得到的数据更加精确,故而可以使用复杂的新模型结构。应当理解的是,模型越复杂,占用的运行内存越高,对于数据的计算结果也更为准确,基于上述考虑,应该针对应用场景选择新模型,故而可以事先设置一个预设列表,或者预设函数,可以根据应用场景选择合适的新模型,使选择的新模型更加具有实用性。
本实施例中,所述选取多个所述源模型用于训练所述新模型的步骤S1,包括:
S111:通过公式
Figure PCTCN2020132596-appb-000006
计算源模型数据库中各模型与所述新模型的相关度;其中X表示模型的特征集合,Y表示新模型的特征集合;
S112:将各个模型与所述新模型的相关度与预设相关度阈值进行比较;
S113:选取所述相关度大于所述预设相关度阈值模型作为所述源模型。
如上述步骤S111-S113所述,将源模型数据库中各模型的特征集合和新模型的特征集合通过公式
Figure PCTCN2020132596-appb-000007
进行计算,其中模型的特征和新模型的特征可以是算法、应用场景、输入值以及输出值等,可以根据各自的特征计算相关度,当计算的相关度越趋近于1时,表示模型与新模型越相关,当计算的相关度越趋近于0时,表示模型与新模型越不相关。因此,可以设置一个预设相关度阈值,该预设相关度阈值为事先根据实际情况设置的值,例如若源模型数据库中的模型数量足够多,那么预设相关度阈值可以相应的设置大一些,若源模型数据库中的模型数量较少,那么预设相关度阈值可以相应的设置小一些。然后选取相关度大于预设相关度阈值的模型作为源模型,使基于选取的源模型对新模型训练的效果更好。
本申请的有益效果:通过将训练数据输入现有的多个源模型中,得到对应的多个特征向量,然后融合计算得到指标向量,然后计算指标向量与新模型得到的当前向量之间的梯度值,通过梯度值校正新模型中的参数。使新模型基于多个源模型训练而成,且无需将多个源模型进行融合,在提高新模型的计算精度的同时,使新模型综合了多个源模型融合后的优点,避免了直接使用融合模型,体积变大,速度变慢的问题。
参照图2,本申请提出一种基于多个源模型的新模型训练装置,包括:
训练数据计算模块10,用于将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
平均值计算模块20,用于分别计算每个所述源模型对应的所述特征向量的 平均值;
指标向量计算模块30,用于将各所述源模型对应的平均值融合计算得到指标向量;
第一相似度值计算模块40,用于计算所述第一当前向量与所述指标向量的第一相似度值;
第一相似度值判断模块50,用于判断所述第一相似度值是否小于预设相似度值;
梯度值计算模块60,用于若所述第一相似度值小于预设相似度值,则计算所述当前向量与所述指标向量之间的梯度值;
参数更新模块70,用于根据所述梯度值校正所述新模型中的参数。
可以根据实际需要的数据选取多个预设的源模型,例如模型a在金融场景比较好,模型b在监控场景表现比较好,模型c在考勤场景比较比较好,需要获取的数据需要综合上边三个场景,那么就可以选取对应的模型a、模型b和模型c。需要说明的是,对于选取训练模型,其可以只是训练完毕的模型,而不需要其原始训练数据。将训练数据和其翻转后的数据输入至预设的源模型中,可以得到每个模型对应的多个特征向量,本文以只进行了左右翻转为例,假设为翻转前的第一训练数据输入至模型a中得到特征向量aV1,将左右翻转后的第二训练数据输入至模型a中得到特征向量aV2,当然,若输入至模型b中,相应得到特征向量bV1和bV2,若还具有模型c,模型d等,按照相应的方法得到对应模型的两个特征向量,需要说明的是,由于训练数据相同,故而其翻转后的数据与该训练数据也是相似的,本申请将训练数据翻转,是为了获取训练数据在各模型中得到更多的值,以便于后续计算过程中提升训练模型的精度。然后将第一训练数据输入至新模型中,得到第一当前向量。
不同的源模型通过第一训练数据和其翻转后的第二训练数据输计算得到多个特征向量,然后计算得到对应同一个源模型的特征向量的平均值。求出的平均值是各模型基于训练数据得到的一个优值,一般而言,该平均值的结果会优于将第一训练数据直接输入后得到的值。
然后再将各源模型对应的平均值进行融合,形成一个完整的输出,即求各源模型对应的平均值的均值,或者将各平均值进行降维操作,举例而言,假设具有两个模型那么各模型的平均值分别为aV=(aV1+aV2)/2,bV=(bV1+bV2)/2,然后将其融合,假设融合是求各源模型对应的平均值的均值(指标向量),那么均值V=(aV+bV)/2,若融合是采取降维的方式,那么降维的方法可以是缺失值比率、低方差滤波、高相关滤波、随机森林/组合树、主成分分析和反特征消除中的一种,最终得到指标向量,用于对新模型训练的对比。
将第一当前向量与指标向量进行比较,即计算第一相似度值,计算的公式可 以是
Figure PCTCN2020132596-appb-000008
当acc的值越接近于1,表明当前向量与指标向量越相似,当acc的值越接近于0时,表明当前向量与指标向量越不相似。其中公式中acc表示第一相似度值,zv i表示在第i特征向量维度中对应的当前向量数值,v i表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。
将计算的第一相似度值与相似度阈值进行比较,若当前的相似度大于相似度阈值,则可以说明当前向量可以相当于目标向量;若当前的相似度小于相似度阈值,则表明当前向量不是目标向量,还需要调整模型中的参数,重新获取当前向量,具体地,根据loss获取梯度值,获取梯度原始值的公式为
Figure PCTCN2020132596-appb-000009
Figure PCTCN2020132596-appb-000010
其中Loss(zV,V)为梯度的原始值,zv i表示在第i特征向量维度中对应的当前向量,v i表示在第i特征向量维度中对应的指标向量,n为特征向量维度数量。然后再对Loss(zV,V)进行求导,得到梯度值,然后再根据获取到的梯度值校正新模型中的参数,以达到训练新模型的技术效果。
本实施例中,基于多个源模型的新模型训练装置,还包括:
重新计算模块,用于将所述第一训练数据输入到校正参数后的新模型中进行计算,得到第二当前向量,并计算所述第二当前向量与所述指标向量的第二相似度值;
第二相似度值判断模块,用于判断所述第二相似度值是否大于所述预设相似度值;
训练认定模块,用于若第二相似度值大于所述预设相似度值,则认定所述新模型基于所述第一训练数据的训练完成。
可以对校正后的参数再次进行检测,即将根据校正后的参数重新计算当前向量,并计算与指标向量的第二相似度值,若第二相似度值大于预设相似度值,则可以将重新计算后的当前向量记为目标向量。若第二相似度值小于或等于预设相似度值,则可以通过上述loss继续获取新的梯度值,根据新的梯度值更改对应的参数,重复多次,直至当前向量与指标向量的第二相似度值大于预设相似度值,认定所述新模型基于所述第一训练数据的训练完成。另外,还可以输入新的训练数据,获取新的指标向量和当前向量,对新模型进行多次反复的训练,达到一定次数后,视为完成了新模型的训练。
本实施例中,指标向量计算模块30,包括:
相关因子获取子模块,用于获取所述新模型在不同应用场景的相关因子;
指标向量计算子模块,用于根据选取的所述源模型的应用场景,依照公式
Figure PCTCN2020132596-appb-000011
融合计算得到指标向量,其中,V为目标向量,w i为第i个应用场景的相关因子,f(w i)为应用场景为第i个场景的源模型的平均值。
获取新模型在不同应用场景下的相关因子,该相关因子可以是基于新模型的作用进行确定的,例如新模型是为了公司的人员管理,则相关因子较大的应用场景为监控场景、考勤场景等,而金融场景的相关因子较小,其中相关因子的数值 可以根据新模型的作用进行确定,例如可以将监控场景的相关因子设置为0.5,将考勤的相关因子设置为0.7,将金融场景的相关因子设置为0.2,然后再根据公式
Figure PCTCN2020132596-appb-000012
依据各个场景的相关因子计算得到指标向量,使后续基于指标向量计算相似度值,以及更新的参数可以更加贴近新模型的应用,使训练得到的新模型更好。
本实施例中,第一相似度值计算模块40,包括:
参数值获取子模块,用于获取所述指标向量和所述第一当前向量的参数值,其中所述参数值至少包括所述指标向量和所述当前向量的特征向量维度数量,以及在各维度中指标向量数值和当前向量数值;
第一计算子模块,用于根据公式
Figure PCTCN2020132596-appb-000013
计算第一相似度值,其中acc表示第一相似度值,zv i表示在第i特征向量维度中对应的第一当前向量数值,v i表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。
将根据所述指标向量和所述当前向量的特征向量维度数量,以及在各维度中指标向量数值和当前向量数值依照公式计算第一相似度值,当acc的值越接近于1,表明当前向量与指标向量越相似,当acc的值越接近于0时,表明当前向量与指标向量越不相似。其中公式中acc表示第一相似度值,zvi表示在第i特征向量维度中对应的当前向量数值,vi表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。根据上述公式计算的第一相似度值可以判断新模型的输出的当前向量,与其他模型融合后的指标向量的相似度,再根据相似度判断新模型是否需要调整参数。其中,第一相似度值越趋近于1时,则认为当前向量与指标向量越相关,第一相似度值越趋近于0时,则认为当前向量与指标向量越不相关。
本实施例中,基于多个源模型的新模型训练装置,包括:
像素点获取模块,用于当所述第一训练数据为图片时,获取图片中的每个像素点,将每个所述像素点和预设像素点阈值进行比较;
像素点设置模块,用于将小于所述预设像素点阈值的像素点设置为0,将大于所述预设像素点阈值的像素点设置为1;
区域划分模块,用于将所述图片平均划分为多个区域,并统计每个区域中的像素点为1的个数,并构成矩阵作为所述训练数据。
将图片进行数字化处理,得到图片中的每个像素点,然后将每个像素点和预设像素点阈值进行比较,该像素点阈值为事先设定的值,设计人员可以根据需要对预设像素点阈值进行设定不同的值,然后将小于所述预设像素点阈值的像素点设置为0,将大于所述预设像素点阈值的像素点设置为1,得到一个初始阵列,然后再将图片平均划分为多个区域,例如将32×32的初始阵列平均划分为多个 区域,以得到8×8的矩阵,然后将得到的矩阵作为训练数据进行训练。当然,上述指的是对图片进行训练数据的处理,本申请还可以对其他训练数据进行其他的处理,例如文本,客户信息等,此处不再赘述。
本实施例中,上述基于多个源模型的新模型训练装置,还包括:
结构量化值计算模块,用于根据所述新模型的应用场景计算对应的结构量化值;
结构量化值比对模块,用于根据所述结构量化值与预设列表中的数据进行比对;其中,所述预设列表包括了所述结构量化值的数值与模型的对应关系;
新模型筛选模块,用于根据比对结果从模型数据库中筛选出所述新模型。
根据新模型的应用场景计算对应的结构量化值,其中结构量化值用于表示模型的结构复杂度,结构量化值越大,表示结构越复杂,结构量化值越小,表示模型越简单,举例而言,当新模型应用在终端,尤其是智能终端时,选择新模型时,新模型的结构不宜太复杂,否则容易占用终端CPU的运行内存,减慢了终端的运行速度,若在服务器端,服务器一般有很高的运行内存,为了使得到的数据更加精确,故而可以使用复杂的新模型结构。应当理解的是,模型越复杂,占用的运行内存越高,对于数据的计算结果也更为准确,基于上述考虑,应该针对应用场景选择新模型,故而可以事先设置一个预设列表,或者预设函数,可以根据应用场景选择合适的新模型,使选择的新模型更加具有实用性。
本实施例中,训练数据计算模块10,包括:
相关度计算子模块,用于通过公式
Figure PCTCN2020132596-appb-000014
计算源模型数据库中各模型与所述新模型的相关度;其中X表示模型的特征集合,Y表示新模型的特征集合;
相关度比较模块,用于将各个模型与所述新模型的相关度与预设相关度阈值进行比较;
源模型选取模块,用于选取所述相关度大于所述预设相关度阈值模型作为所述源模型。
将源模型数据库中各模型的特征集合和新模型的特征集合通过公式
Figure PCTCN2020132596-appb-000015
进行计算,其中模型的特征和新模型的特征可以是算法、应用场景、输入值以及输出值等,可以根据各自的特征计算相关度,当计算的相关度越趋近于1时,表示模型与新模型越相关,当计算的相关度越趋近于0时,表示模型与新模型越不相关。因此,可以设置一个预设相关度阈值,该预设相关度阈值为事先根据实际情况设置的值,例如若源模型数据库中的模型数量足够多,那么预设相关度阈值可以相应的设置大一些,若源模型数据库中的模型数量较少,那么预设相关度阈值可以相应的设置小一些。然后选取相关度大于预设相关度阈值的模型作为源模型,使基于选取的源模型对新模型训练的效果更好。
参照图3,本申请实施例中还提供一种计算机设备,该计算机设备可以是服务器,其内部结构可以如图3所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中,该计算机设计的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机程序和数据库。该内存器为非易失性存储介质中的操作系统和计算机程序的运行提供环境。该计算机设备的数据库用于存储各种训练数据等。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机程序被处理器执行时可以实现上述任一实施例所述的基于多个源模型的新模型训练方法。
本领域技术人员可以理解,图3中示出的结构,仅仅是与本申请方案相关的部分结构的框图,并不构成对本申请方案所应用于其上的计算机设备的限定。本申请实施例还提供一种计算机可读存储介质,上述存储介质可以是非易失性存储介质,也可以是易失性存储介质。其上存储有计算机程序,计算机程序被处理器执行时可以实现上述任一实施例所述的基于多个源模型修正新模型的方法。

Claims (20)

  1. 一种基于多个源模型修正新模型的方法,其中,包括:
    将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
    将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
    分别计算每个所述源模型对应的所述特征向量的平均值;
    将各所述源模型对应的平均值融合计算得到指标向量;
    计算所述第一当前向量与所述指标向量的第一相似度值;
    判断所述第一相似度值是否小于预设相似度值;
    若所述第一相似度值小于预设相似度值,则计算所述第一当前向量与所述指标向量之间的梯度值;
    根据所述梯度值校正所述新模型中的参数。
  2. 如权利要求1所述的基于多个源模型修正新模型的方法,其中,所述根据所述梯度值校正所述新模型中的参数的步骤之后,还包括:
    将所述第一训练数据输入到校正参数后的新模型中进行计算,得到第二当前向量,并计算所述第二当前向量与所述指标向量的第二相似度值;
    判断所述第二相似度值是否大于所述预设相似度值;
    若第二相似度值大于所述预设相似度值,则认定所述新模型基于所述第一训练数据的训练完成。
  3. 如权利要求1所述的基于多个源模型修正新模型的方法,其中,所述将各所述源模型对应的平均值融合计算得到指标向量的步骤,包括:
    获取所述新模型在不同应用场景的相关因子;
    根据选取的所述源模型的应用场景,依照公式
    Figure PCTCN2020132596-appb-100001
    融合计算得到指标向量,其中,V为目标向量,w i为第i个应用场景的相关因子,f(w i)为应用场景为第i个场景的源模型的平均值。
  4. 如权利要求1所述的基于多个源模型修正新模型的方法,其中,所述计算所述第一当前向量与所述指标向量的第一相似度值的步骤,包括:
    获取所述指标向量和所述第一当前向量的参数值,其中所述参数值至少包括所述指标向量和所述当前向量的特征向量维度数量,以及在各维度中指标向量数值和当前向量数值;
    根据公式
    Figure PCTCN2020132596-appb-100002
    计算第一相似度值,其中acc表示第一相似度值,zv i表示在第i特征向量维度中对应的第一当前向量数值,v i表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。
  5. 如权利要求1所述的基于多个源模型修正新模型的方法,其中,所述将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量的步骤之前,包括:
    当所述第一训练数据为图片时,获取所述图片中的每个像素点,将每个所述像素点和预设像素点阈值进行比较;
    将小于所述预设像素点阈值的像素点设置为0,将大于所述预设像素点阈值的像素点设置为1;
    将所述图片平均划分为多个区域,并统计每个区域中的像素点为1的个数,并构成矩阵作为所述训练数据。
  6. 如权利要求1所述的基于多个源模型修正新模型的方法,其中,所述将第一训练数据输入至所述新模型中进行计算,得到第一当前向量的步骤之前,还包括:
    根据所述新模型的应用场景计算对应的结构量化值;
    根据所述结构量化值与预设列表中的数据进行比对;其中,所述预设列表包括了所述结构量化值的数值与模型的对应关系;
    根据比对结果从模型数据库中筛选出所述新模型。
  7. 如权利要求1所述的基于多个源模型修正新模型的方法,其中,所述将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量的步骤之前,包括:
    通过公式
    Figure PCTCN2020132596-appb-100003
    计算源模型数据库中各模型与所述新模型的相关度;其中X表示模型的特征集合,Y表示新模型的特征集合;
    将各个模型与所述新模型的相关度与预设相关度阈值进行比较;
    选取所述相关度大于所述预设相关度阈值的模型作为所述源模型。
  8. 一种基于多个源模型的新模型训练装置,其中,包括:
    训练数据计算模块,用于将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
    将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
    平均值计算模块,用于分别计算每个所述源模型对应的所述特征向量的平均值;
    指标向量计算模块,用于将各所述源模型对应的平均值融合计算得到指标向量;
    第一相似度值计算模块,用于计算所述第一当前向量与所述指标向量的第一 相似度值;
    第一相似度值判断模块,用于判断所述第一相似度值是否小于预设相似度值;
    梯度值计算模块,用于若所述第一相似度值小于预设相似度值,则计算所述当前向量与所述指标向量之间的梯度值;
    参数更新模块,用于根据所述梯度值校正所述新模型中的参数。
  9. 一种计算机设备,包括存储器和处理器,所述存储器存储有计算机程序,其中,所述处理器执行所述计算机程序时实现一种基于多个源模型的新模型训练方法的步骤:将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
    将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
    分别计算每个所述源模型对应的所述特征向量的平均值;
    将各所述源模型对应的平均值融合计算得到指标向量;
    计算所述第一当前向量与所述指标向量的第一相似度值;
    判断所述第一相似度值是否小于预设相似度值;
    若所述第一相似度值小于预设相似度值,则计算所述第一当前向量与所述指标向量之间的梯度值;
    根据所述梯度值校正所述新模型中的参数。
  10. 如权利要求9所述的计算机设备,其中,所述根据所述梯度值校正所述新模型中的参数的步骤之后,还包括:
    将所述第一训练数据输入到校正参数后的新模型中进行计算,得到第二当前向量,并计算所述第二当前向量与所述指标向量的第二相似度值;
    判断所述第二相似度值是否大于所述预设相似度值;
    若第二相似度值大于所述预设相似度值,则认定所述新模型基于所述第一训练数据的训练完成。
  11. 如权利要求9所述的计算机设备,其中,所述将各所述源模型对应的平均值融合计算得到指标向量的步骤,包括:
    获取所述新模型在不同应用场景的相关因子;
    根据选取的所述源模型的应用场景,依照公式
    Figure PCTCN2020132596-appb-100004
    融合计算得到指标向量,其中,V为目标向量,wi为第i个应用场景的相关因子,f(wi)为应用场景为第i个场景的源模型的平均值。
  12. 如权利要求9所述的计算机设备,其中,所述计算所述第一当前向量与所述指标向量的第一相似度值的步骤,包括:
    获取所述指标向量和所述第一当前向量的参数值,其中所述参数值至少包括所述指标向量和所述当前向量的特征向量维度数量,以及在各维度中指标向量数值和当前向量数值;
    根据公式
    Figure PCTCN2020132596-appb-100005
    计算第一相似度值,其中acc表示第一相似度值,zv i表示在第i特征向量维度中对应的第一当前向量数值,v i表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。
  13. 如权利要求9所述的计算机设备,其中,所述将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量的步骤之前,包括:
    当所述第一训练数据为图片时,获取所述图片中的每个像素点,将每个所述像素点和预设像素点阈值进行比较;
    将小于所述预设像素点阈值的像素点设置为0,将大于所述预设像素点阈值的像素点设置为1;
    将所述图片平均划分为多个区域,并统计每个区域中的像素点为1的个数,并构成矩阵作为所述训练数据。
  14. 如权利要求9所述的计算机设备,其中,所述将第一训练数据输入至所述新模型中进行计算,得到第一当前向量的步骤之前,还包括:
    根据所述新模型的应用场景计算对应的结构量化值;
    根据所述结构量化值与预设列表中的数据进行比对;其中,所述预设列表包括了所述结构量化值的数值与模型的对应关系;
    根据比对结果从模型数据库中筛选出所述新模型。
  15. 如权利要求9所述的计算机设备,其中,所述将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量的步骤之前,包括:
    通过公式
    Figure PCTCN2020132596-appb-100006
    计算源模型数据库中各模型与所述新模型的相关度;其中X表示模型的特征集合,Y表示新模型的特征集合;
    将各个模型与所述新模型的相关度与预设相关度阈值进行比较;
    选取所述相关度大于所述预设相关度阈值的模型作为所述源模型。
  16. 一种计算机可读存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时实现一种基于多个源模型修正新模型的方法的步骤。
    将第一训练数据输入至所述新模型中进行计算,得到第一当前向量;以及,
    将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量,其中,对应每一个所述源模型的特征向量包括多个;
    分别计算每个所述源模型对应的所述特征向量的平均值;
    将各所述源模型对应的平均值融合计算得到指标向量;
    计算所述第一当前向量与所述指标向量的第一相似度值;
    判断所述第一相似度值是否小于预设相似度值;
    若所述第一相似度值小于预设相似度值,则计算所述第一当前向量与所述指标向量之间的梯度值;
    根据所述梯度值校正所述新模型中的参数。
  17. 如权利要求16所述的计算机可读存储介质,其中,所述根据所述梯度值校正所述新模型中的参数的步骤之后,还包括:
    将所述第一训练数据输入到校正参数后的新模型中进行计算,得到第二当前向量,并计算所述第二当前向量与所述指标向量的第二相似度值;
    判断所述第二相似度值是否大于所述预设相似度值;
    若第二相似度值大于所述预设相似度值,则认定所述新模型基于所述第一训练数据的训练完成。
  18. 如权利要求16所述的计算机可读存储介质,其中,所述将各所述源模型对应的平均值融合计算得到指标向量的步骤,包括:
    获取所述新模型在不同应用场景的相关因子;
    根据选取的所述源模型的应用场景,依照公式
    Figure PCTCN2020132596-appb-100007
    融合计算得到指标向量,其中,V为目标向量,wi为第i个应用场景的相关因子,f(wi)为应用场景为第i个场景的源模型的平均值。
  19. 如权利要求16所述的计算机可读存储介质,其中,所述计算所述第一当前向量与所述指标向量的第一相似度值的步骤,包括:
    获取所述指标向量和所述第一当前向量的参数值,其中所述参数值至少包括所述指标向量和所述当前向量的特征向量维度数量,以及在各维度中指标向量数值和当前向量数值;
    根据公式
    Figure PCTCN2020132596-appb-100008
    计算第一相似度值,其中acc表示第一相似度值,zvi表示在第i特征向量维度中对应的第一当前向量数值,vi表示在第i特征向量维度中对应的指标向量数值,n为特征向量维度数量。
  20. 如权利要求16所述的计算机可读存储介质,其中,所述将第一训练数据和翻转所述第一训练数据后得到的第二训练数据分别输入至多个预设的源模型中进行计算,得到对应各所述源模型的特征向量的步骤之前,包括:
    当所述第一训练数据为图片时,获取所述图片中的每个像素点,将每个所述像素点和预设像素点阈值进行比较;
    将小于所述预设像素点阈值的像素点设置为0,将大于所述预设像素点阈值 的像素点设置为1;
    将所述图片平均划分为多个区域,并统计每个区域中的像素点为1的个数,并构成矩阵作为所述训练数据。
PCT/CN2020/132596 2020-07-31 2020-11-30 基于多个源模型修正新模型的方法、装置以及计算机设备 WO2021139448A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN2020107609139 2020-07-31
CN202010760913.9A CN111898682B (zh) 2020-07-31 2020-07-31 基于多个源模型修正新模型的方法、装置以及计算机设备

Publications (1)

Publication Number Publication Date
WO2021139448A1 true WO2021139448A1 (zh) 2021-07-15

Family

ID=73183460

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/132596 WO2021139448A1 (zh) 2020-07-31 2020-11-30 基于多个源模型修正新模型的方法、装置以及计算机设备

Country Status (2)

Country Link
CN (1) CN111898682B (zh)
WO (1) WO2021139448A1 (zh)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111898682B (zh) * 2020-07-31 2023-08-25 平安科技(深圳)有限公司 基于多个源模型修正新模型的方法、装置以及计算机设备
CN112434323A (zh) * 2020-12-01 2021-03-02 Oppo广东移动通信有限公司 模型参数获取方法、装置、计算机设备及存储介质
CN113139609B (zh) * 2021-04-29 2023-12-29 国网甘肃省电力公司白银供电公司 基于闭环反馈的模型校正方法、装置和计算机设备
CN116663648B (zh) * 2023-04-23 2024-04-02 北京大学 模型训练方法、装置、设备及存储介质

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103336880A (zh) * 2013-04-03 2013-10-02 温州大学 一种用于求解面向操作优化的模型修正问题高效方法
CN103837159A (zh) * 2014-03-04 2014-06-04 中国科学院光电技术研究所 一种经纬仪指向修正模型正交化解耦修正方法
WO2017167543A1 (en) * 2016-03-30 2017-10-05 British Telecommunications Public Limited Company Multiform persistence abstraction
CN108304354A (zh) * 2018-01-25 2018-07-20 腾讯科技(深圳)有限公司 一种预测模型训练方法及装置、存储介质、电子设备
CN109815988A (zh) * 2018-12-27 2019-05-28 北京奇艺世纪科技有限公司 模型生成方法、分类方法、装置及计算机可读存储介质
CN110363239A (zh) * 2019-07-04 2019-10-22 中国人民解放军国防科技大学 一种面向多模态数据的小样本机器学习方法、系统和介质
CN111177446A (zh) * 2019-12-12 2020-05-19 苏州科技大学 一种用于足迹图像检索的方法
CN111898682A (zh) * 2020-07-31 2020-11-06 平安科技(深圳)有限公司 基于多个源模型修正新模型的方法、装置以及计算机设备

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10423850B2 (en) * 2017-10-05 2019-09-24 The Climate Corporation Disease recognition from images having a large field of view
CN110826581B (zh) * 2018-08-10 2023-11-07 京东科技控股股份有限公司 一种动物数量识别方法、装置、介质及电子设备
CN110738182A (zh) * 2019-10-21 2020-01-31 四川隧唐科技股份有限公司 一种用于高精度识别中标金额的lstm模型单元训练方法及装置
CN111353076B (zh) * 2020-02-21 2023-10-10 华为云计算技术有限公司 训练跨模态检索模型的方法、跨模态检索的方法和相关装置

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103336880A (zh) * 2013-04-03 2013-10-02 温州大学 一种用于求解面向操作优化的模型修正问题高效方法
CN103837159A (zh) * 2014-03-04 2014-06-04 中国科学院光电技术研究所 一种经纬仪指向修正模型正交化解耦修正方法
WO2017167543A1 (en) * 2016-03-30 2017-10-05 British Telecommunications Public Limited Company Multiform persistence abstraction
CN108304354A (zh) * 2018-01-25 2018-07-20 腾讯科技(深圳)有限公司 一种预测模型训练方法及装置、存储介质、电子设备
CN109815988A (zh) * 2018-12-27 2019-05-28 北京奇艺世纪科技有限公司 模型生成方法、分类方法、装置及计算机可读存储介质
CN110363239A (zh) * 2019-07-04 2019-10-22 中国人民解放军国防科技大学 一种面向多模态数据的小样本机器学习方法、系统和介质
CN111177446A (zh) * 2019-12-12 2020-05-19 苏州科技大学 一种用于足迹图像检索的方法
CN111898682A (zh) * 2020-07-31 2020-11-06 平安科技(深圳)有限公司 基于多个源模型修正新模型的方法、装置以及计算机设备

Also Published As

Publication number Publication date
CN111898682A (zh) 2020-11-06
CN111898682B (zh) 2023-08-25

Similar Documents

Publication Publication Date Title
WO2021139448A1 (zh) 基于多个源模型修正新模型的方法、装置以及计算机设备
JP6994588B2 (ja) 顔特徴抽出モデル訓練方法、顔特徴抽出方法、装置、機器および記憶媒体
US11587356B2 (en) Method and device for age estimation
US10019823B2 (en) Combined composition and change-based models for image cropping
JP2023523029A (ja) 画像認識モデル生成方法、装置、コンピュータ機器及び記憶媒体
CN110909663B (zh) 一种人体关键点识别方法、装置及电子设备
CN112232426B (zh) 目标检测模型的训练方法、装置、设备及可读存储介质
CN112861659B (zh) 一种图像模型训练方法、装置及电子设备、存储介质
CN113221645B (zh) 目标模型训练方法、人脸图像生成方法以及相关装置
US20230273914A1 (en) Data augmentation method and apparatus, computing device, and computer-readable storage medium
JP2018163444A (ja) 情報処理装置、情報処理方法及びプログラム
CN113658091A (zh) 一种图像评价方法、存储介质及终端设备
CN112417947A (zh) 关键点检测模型的优化及面部关键点的检测方法及装置
WO2020168807A1 (zh) 图像亮度的调节方法、装置、计算机设备和存储介质
CN111445487A (zh) 图像分割方法、装置、计算机设备和存储介质
WO2024041108A1 (zh) 图像矫正模型训练及图像矫正方法、装置和计算机设备
CN114332556A (zh) 训练样本筛选方法、装置、计算机设备和存储介质
CN114627397A (zh) 行为识别模型构建方法以及行为识别方法
WO2021035979A1 (zh) 边缘学习的图像填充方法、装置、终端及可读存储介质
CN112383824A (zh) 视频广告过滤方法、设备及存储介质
US11688175B2 (en) Methods and systems for the automated quality assurance of annotated images
CN116258923A (zh) 图像识别模型训练方法、装置、计算机设备和存储介质
WO2022127603A1 (zh) 一种模型处理方法及相关装置
CN115941871A (zh) 视频插帧方法、装置、计算机设备和存储介质
WO2021169604A1 (zh) 动作信息识别方法、装置、电子设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20912198

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20912198

Country of ref document: EP

Kind code of ref document: A1