Model generation method and apparatus, electronic device, and storage medium

Abstract

The application provides a model generation method and device, electronic equipment and storage medium, and relates to the technical field of machine learning. Through the AI algorithm platform, a business type and a training data set can be input, and annotation and exploration of the training data set are respectively implemented according to the business type and the data type in the training data set, and model configuration information is generated, so that the model is executed to perform a business corresponding to the business type. Through the method, a user only needs to input relevant information required for generating a model, and a model corresponding to the business type can be automatically generated, the human input during manual model training is reduced, and the user can obtain the required model without learning the details of the underlying algorithm, so that the threshold is low. In addition, a lightweight model meeting the precision loss threshold can be automatically generated based on the precision loss threshold input by the user, so that the model is deployed on an edge platform, and the generated model is more suitable for the business scenario of edge AI.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and more specifically, to a model generation method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the continuous development of machine learning, more and more enterprises are deploying high-performance neural network models in their business platforms to replace manual solutions to business problems.

[0003] Currently, most enterprises deploy neural network models on edge platforms, such as mobile / embedded devices. These platforms are characterized by limited memory resources, low processor performance, and power consumption. This makes it impossible for most high-precision models to be deployed and run in real time on these platforms. Lightweight models, on the other hand, offer faster computation speeds and lower hardware requirements, making them suitable for edge platforms. Therefore, how to generate lightweight neural network models has become an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this application is to address the shortcomings of the prior art by providing a model generation method, apparatus, electronic device, and storage medium to facilitate lightweight model processing.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows:

[0006] In a first aspect, embodiments of this application provide a model generation method, including:

[0007] Obtain the user-inputted business type and training dataset;

[0008] Based on the business type, determine the annotation method of the training dataset, and obtain the annotated training dataset according to the annotation method;

[0009] Based on the data type of the training dataset and the business type, perform the corresponding data exploration task on the labeled training sample set to obtain the data exploration results;

[0010] Based on the data exploration results and the business type, the model generation parameters are determined, and the model generation parameters include at least: artificial intelligence algorithms and parameters;

[0011] Based on the data exploration results, the model generation parameters, and the accuracy loss threshold input by the user, model configuration information is generated, and a target model is generated based on the model configuration information.

[0012] Optionally, determining the annotation method of the training dataset according to the business type, and obtaining the annotated training dataset according to the annotation method, includes:

[0013] Based on the business type, determine the target annotation method from the annotation methods that match the data type of the training dataset;

[0014] The training sample data in the training dataset is labeled using the target labeling method to obtain the labeled training dataset.

[0015] Optionally, after obtaining the labeled training dataset according to the labeling method, the method further includes:

[0016] Based on the data type of the training dataset, at least one data augmentation algorithm matching the data type is used to augment each training sample data in the training dataset to obtain augmented sample data corresponding to each training sample data. The data type includes: text data, image data, and structured data.

[0017] Based on the data augmentation algorithm used to generate the amplified sample data, the annotation information of the amplified sample data is generated;

[0018] The target training dataset is obtained based on the training sample data and the corresponding augmented sample data.

[0019] Optionally, determining the model generation parameters based on the data exploration results and the business type includes:

[0020] Based on the data exploration results and the business type, at least one artificial intelligence algorithm is determined from the artificial intelligence algorithm library;

[0021] The parameter information of each artificial intelligence algorithm is determined separately. The parameter information includes the operators used by the algorithm and / or the parameter values ​​of the algorithm.

[0022] The model generation parameters are obtained based on the various artificial intelligence algorithms and parameter information.

[0023] Optionally, generating the target model based on the model configuration information includes:

[0024] Execute the model configuration information to generate the initial model;

[0025] The importance of each convolutional kernel in each convolutional layer of the initial model is determined using the importance calculation formula.

[0026] Based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold, the convolutional layers of the initial model are pruned sequentially to obtain the target model.

[0027] Optionally, the step of pruning each convolutional layer of the initial model sequentially based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold to obtain the target model includes:

[0028] Based on the importance of each convolutional kernel in the target convolutional layer, the convolutional kernel with the lowest importance is pruned in sequence, and after each convolutional kernel is pruned, it is determined whether the accuracy loss value of the current model is greater than the accuracy loss threshold.

[0029] If yes, then determine that the number of convolutional kernels currently pruned is the number of pruning branches corresponding to the target convolutional layer; if no, then continue to prune the convolutional kernel with the lowest current importance from each convolutional kernel of the target convolutional layer until the accuracy loss value of the current model is greater than the accuracy loss threshold, then stop execution.

[0030] The initial model is pruned sequentially according to the number of pruning operations corresponding to each convolutional layer to obtain the target model.

[0031] Optionally, the step of pruning each convolutional layer of the initial model according to the pruning number corresponding to each convolutional layer to obtain the target model includes:

[0032] Based on the number of pruning branches corresponding to the current convolutional layer, the convolutional kernel with the lowest importance of the pruning branch number is pruned from the current convolutional layer to obtain the current pruned model;

[0033] Determine whether the accuracy loss value of the currently pruned model is greater than the accuracy loss threshold;

[0034] If so, the network parameters of the current pruned model are tuned until the accuracy loss value of the current pruned model is less than the accuracy loss threshold, and then a new model is obtained.

[0035] The next convolutional layer after the current convolutional layer is determined as the new convolutional layer. Based on the number of pruning branches corresponding to the new convolutional layer, the new convolutional layers of the current new model are pruned until all convolutional layers are pruned. The model obtained after pruning each convolutional layer is taken as the target model.

[0036] Secondly, embodiments of this application also provide a model generation apparatus, including: an acquisition module, an annotation module, a data exploration module, a determination module, and a generation module;

[0037] The acquisition module is used to acquire the business type and training dataset input by the user;

[0038] The annotation module is used to determine the annotation method of the training dataset according to the business type, and to obtain the annotated training dataset according to the annotation method;

[0039] The data exploration module is used to perform corresponding data exploration tasks on the labeled training sample set according to the data type of the training dataset and the business type, and obtain data exploration results;

[0040] The determining module is used to determine model generation parameters based on the data exploration results and the business type. The model generation parameters include at least: artificial intelligence algorithms and parameters.

[0041] The generation module is used to generate model configuration information based on the data exploration results, the model generation parameters, and the accuracy loss threshold input by the user, and to generate a target model based on the model configuration information.

[0042] Optionally, the annotation module is specifically used to determine the target annotation method from the annotation methods that match the data type of the training dataset according to the business type;

[0043] The training sample data in the training dataset is labeled using the target labeling method to obtain the labeled training dataset.

[0044] Optionally, it also includes: a data augmentation module;

[0045] The data augmentation module is used to augment each training sample data in the training dataset according to the data type of the training dataset, using at least one data augmentation algorithm that matches the data type, to obtain augmented sample data corresponding to each training sample data. The data types include: text data, image data, and structured data.

[0046] Based on the data augmentation algorithm used to generate the amplified sample data, the annotation information of the amplified sample data is generated;

[0047] The target training dataset is obtained based on the training sample data and the corresponding augmented sample data.

[0048] Optionally, the determining module is specifically used to determine at least one artificial intelligence algorithm from an artificial intelligence algorithm library based on the data exploration results and the business type;

[0049] The parameter information of each artificial intelligence algorithm is determined separately. The parameter information includes the operators used by the algorithm and / or the parameter values ​​of the algorithm.

[0050] The model generation parameters are obtained based on the various artificial intelligence algorithms and parameter information.

[0051] Optionally, the generation module is specifically used to execute the model configuration information and generate an initial model;

[0052] The importance of each convolutional kernel in each convolutional layer of the initial model is determined using the importance calculation formula.

[0053] Based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold, the convolutional layers of the initial model are pruned sequentially to obtain the target model.

[0054] Optionally, the generation module is specifically used to cut off the convolutional kernel with the lowest importance in sequence according to the importance of each convolutional kernel in the target convolutional layer, and after each convolutional kernel is cut off, to determine whether the accuracy loss value of the current model is greater than the accuracy loss threshold.

[0055] If yes, then determine that the number of convolutional kernels currently pruned is the number of pruning branches corresponding to the target convolutional layer; if no, then continue to prune the convolutional kernel with the lowest current importance from each convolutional kernel of the target convolutional layer until the accuracy loss value of the current model is greater than the accuracy loss threshold, then stop execution.

[0056] The initial model is pruned sequentially according to the number of pruning operations corresponding to each convolutional layer to obtain the target model.

[0057] Optionally, the generation module is specifically used to prune the convolutional kernel with the lowest importance from the current convolutional layer according to the number of prunings corresponding to the current convolutional layer, so as to obtain the current pruned model;

[0058] Determine whether the accuracy loss value of the currently pruned model is greater than the accuracy loss threshold;

[0059] If so, the network parameters of the current pruned model are tuned until the accuracy loss value of the current pruned model is less than the accuracy loss threshold, and then a new model is obtained.

[0060] The next convolutional layer after the current convolutional layer is determined as the new convolutional layer. Based on the number of pruning branches corresponding to the new convolutional layer, the new convolutional layers of the current new model are pruned until all convolutional layers are pruned. The model obtained after pruning each convolutional layer is taken as the target model.

[0061] Thirdly, embodiments of this application provide an electronic device, including: a processor, a storage medium, and a bus. The storage medium stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the model generation method provided in the first aspect.

[0062] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the model generation method provided in the first aspect.

[0063] The beneficial effects of this application are:

[0064] This application provides a model generation method, apparatus, electronic device, and storage medium. Through the provided AI algorithm platform, users can input a business type and training dataset. Based on the business type and the data types in the training dataset, the method annotates and explores the training dataset, generating model configuration information. Then, based on the execution model configuration information, a model for executing the corresponding business type can be generated. This method automatically generates a model corresponding to the business type by only requiring users to input relevant information, reducing the human effort required for manual model training. Users do not need to learn the details of the underlying algorithms to obtain the required model, making the barrier to entry low. Furthermore, based on the user-inputted accuracy loss threshold, a lightweight model that meets the accuracy loss threshold can be automatically generated, thus enabling deployment on edge platforms and making the generated model more suitable for edge AI business scenarios. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a schematic diagram of the architecture of a model generation system provided in an embodiment of this application;

[0067] Figure 2 A schematic flowchart illustrating a model generation method provided in an embodiment of this application;

[0068] Figure 3 A flowchart illustrating another model generation method provided in this application embodiment;

[0069] Figure 4A flowchart illustrating yet another model generation method provided in this application embodiment;

[0070] Figure 5 A flowchart illustrating another model generation method provided in this application embodiment;

[0071] Figure 6 A flowchart illustrating another model generation method provided in this application embodiment;

[0072] Figure 7 A flowchart illustrating yet another model generation method provided in this application embodiment;

[0073] Figure 8 A flowchart illustrating another model generation method provided in this application embodiment;

[0074] Figure 9 A schematic diagram of a model generation apparatus provided in an embodiment of this application;

[0075] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0077] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0078] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0079] Figure 1 This is a schematic diagram of the architecture of a model generation system provided in an embodiment of this application. The model generation system can refer to an AI (Artificial Intelligence) algorithm platform. The AI ​​algorithm platform provided by this method may include: a data management unit, a data exploration unit, a model training unit, and a model repository unit.

[0080] The data management unit is used to collect the training dataset input by the user, and to complete the annotation and data augmentation processing of the training dataset to obtain the target training dataset; the data exploration unit is used to perform data exploration tasks on the target training dataset to obtain exploration results; the model training unit is used to generate model configuration information based on the data preprocessing method, feature engineering method and artificial intelligence algorithm selected by the user, and to execute the model configuration information on the computing cluster to generate the target model.

[0081] The model repository unit supports one-click publishing of model services, which can be used by users to publish model services to the generated target models. At the same time, it also supports user permission management for model services, configuring each user's access permissions for each model service to ensure the security of model service usage.

[0082] The model training unit supports automatic lightweighting, which can automatically generate a lightweight model that meets the accuracy loss threshold based on the user-input accuracy loss threshold. The generated lightweight model can be exported and deployed on edge platforms to meet the deployment requirements of models on edge platforms.

[0083] The method of this application will be described below through specific embodiments.

[0084] Figure 2 This is a flowchart illustrating a model generation method provided in an embodiment of this application; the execution subject of this method can be a computer or a terminal device, such as... Figure 2 As shown, the method may include:

[0085] S101. Obtain the business type and training dataset input by the user.

[0086] Optionally, users can input the business type and training dataset in the corresponding information input field through the user interface provided by the AI ​​algorithm platform. Here, "user" refers to an algorithm engineer.

[0087] The business type can include the specific execution content of the business, such as facial recognition, vehicle recognition, and data statistics. The training dataset is the set of training samples used to generate the model corresponding to the business type.

[0088] In some embodiments, the training dataset may be automatically uploaded by the user, while in other embodiments, the training dataset may be automatically collected from a big data platform according to the business type.

[0089] The acquired training datasets can be saved to the underlying database according to user ID, dataset number, and dataset version, respectively, to ensure that training datasets uploaded by different users are stored accordingly.

[0090] S102. Determine the annotation method for the training dataset according to the business type, and obtain the annotated training dataset according to the annotation method.

[0091] Optionally, based on the business type, users can select the annotation method for the training dataset from the visual annotation database provided by the algorithm platform. The selection of the annotation method is determined according to the business type.

[0092] For example, if the business type is to identify rectangular buildings in an image, then the annotation method selected here can be rectangular annotation, so that the target rectangular buildings in the training dataset can be annotated with rectangles.

[0093] For the recognition or detection of objects of different shapes or types, the shape of the corresponding bounding box can be different, such as circular annotation, polygonal annotation, etc.

[0094] Based on the selected annotation method, each sample in the training dataset can be annotated in the corresponding way to obtain the annotated training dataset.

[0095] S103. Based on the data type and business type of the training dataset, perform the corresponding data exploration task on the labeled training sample set to obtain the data exploration results.

[0096] In some embodiments, a data exploration task can also be performed on the labeled training dataset, and the results of the data exploration can determine the subsequent selection of data preprocessing methods.

[0097] This embodiment provides the ability to analyze and explore structured datasets, image datasets, and text datasets. The selection of data exploration tasks can be determined comprehensively based on the data type of the training dataset and the business type. The data type of the training dataset refers to the type of training sample data contained in the training dataset, which can be the aforementioned structured datasets, image datasets, and text datasets.

[0098] Different types of training datasets may be matched with different data exploration tasks, meaning that different data exploration tasks may support the processing of different data types. Furthermore, depending on the type of business, the data exploration tasks required may also differ, thus allowing for the selection of appropriate data exploration tasks for different business types and training datasets.

[0099] S104. Based on the data exploration results and business type, determine the model generation parameters. The model generation parameters shall include at least the artificial intelligence algorithm and its parameters.

[0100] In some embodiments, based on the data exploration results and business type described above, model generation parameters can be determined. Here, model generation parameters can refer to model-related parameter information used to generate and execute the business corresponding to the business type.

[0101] Optionally, the model generation parameters may include at least: artificial intelligence algorithms and parameters; however, in some cases, the model generation parameters may also include data preprocessing algorithms and parameters and / or feature engineering algorithms and parameters.

[0102] The data preprocessing algorithms and parameters, feature engineering algorithms and parameters, and artificial intelligence algorithms and parameters used to generate the model can be determined based on the data exploration results and business type.

[0103] S105. Based on the data exploration results, model generation parameters, and the accuracy loss threshold input by the user, generate model configuration information, and generate the target model based on the model configuration information.

[0104] In some embodiments, model configuration information can be generated based on the model generation parameters selected above and the obtained data exploration results, combined with the accuracy loss threshold input by the user. The model configuration information can be submitted to the computing cluster for execution, thereby obtaining the execution result, which is the target model.

[0105] Among them, based on the accuracy loss threshold input by the user, a lightweight model that meets the accuracy loss threshold can be automatically generated, thereby meeting the deployment requirements of the model on edge platforms.

[0106] In summary, the model generation method provided in this embodiment allows users to input a business type and training dataset through the provided AI algorithm platform. Based on the business type and the data types in the training dataset, the method annotates and explores the training dataset, generating model configuration information. After executing the model configuration information, a model for performing the corresponding business type can be generated. This method automatically generates a model corresponding to the business type by simply inputting the relevant information involved in model generation, reducing the human effort required for manual model training. Users do not need to learn the details of the underlying algorithms to obtain the required model, making it relatively easy to learn. Furthermore, based on the user-inputted accuracy loss threshold, a lightweight model that meets the threshold can be automatically generated, thus enabling deployment on edge platforms and making the generated model more suitable for edge AI business scenarios.

[0107] Figure 3 The flowchart illustrates another model generation method provided in this application embodiment; optionally, in step S102, determining the annotation method of the training dataset according to the business type, and obtaining the annotated training dataset according to the annotation method, may include:

[0108] S301. Based on the business type, determine the target annotation method from the annotation methods that match the data type of the training dataset.

[0109] Data labeling is the process of adding metadata to a training dataset. This metadata typically takes the form of labels and can be added to any type of data, including text, images, and videos. Adding high-quality and accurate labels is a crucial process for developing training datasets for machine learning, as supervised machine learning models can learn to recognize recurring patterns in the labeled data.

[0110] The data management unit of the AI ​​algorithm platform supports visual annotation of image and text datasets. It uses front-end controls such as Canvas and jsPlumb to convert user-drawn graphics into JSON-structured annotation files. For image datasets, supported annotation methods include, but are not limited to: classification annotation, rectangle annotation, polygon annotation, circle annotation, keypoint annotation, and behavior / pose annotation. For text datasets, supported annotation methods include, but are not limited to: text classification and entity annotation.

[0111] The annotation method matching the training dataset can be determined based on the data type of the training dataset. Furthermore, based on the business type, the target annotation method can be determined from the annotation methods matching the data type of the training dataset.

[0112] S302. Use the target annotation method to annotate each training sample data in the training dataset to obtain the annotated training dataset.

[0113] Based on the determined target annotation method, each training sample data included in the training dataset can be annotated to obtain the annotated training dataset.

[0114] Figure 4 A flowchart illustrating another model generation method provided in this application embodiment; optionally, in step S302, after obtaining the labeled training dataset according to the annotation method, the method may further include:

[0115] S401. Based on the data type of the training dataset, at least one data augmentation algorithm matching the data type is used to augment the training sample data in the training dataset to obtain augmented sample data corresponding to each training sample data. The data types include: text data, image data, and structured data.

[0116] Data augmentation is a technique that artificially expands a training dataset by generating more equivalent data from a limited dataset. It is an effective means of overcoming insufficient training data and is currently widely used in various fields of deep learning.

[0117] The data management unit also supports automatic data augmentation for image and text datasets. For image datasets, data augmentation methods include, but are not limited to: flipping, rotating, scaling, cropping, blurring, adding noise, and color dithering. For text datasets, data augmentation methods include, but are not limited to: synonym replacement, random noise, reverse translation, and instance cross-referencing. Data augmentation of labeled image data will generate new annotation files. If blurring, adding noise, or color dithering are used, the content of the annotation file does not need to be changed; if flipping, rotating, scaling, or cropping are used, the content of the annotation file needs to be modified.

[0118] Optionally, based on the data type of the training dataset, at least one data augmentation algorithm can be determined from a variety of data augmentation algorithms that match the data type. Then, based on the determined data augmentation algorithm, each training sample in the training dataset is augmented to obtain augmented sample data corresponding to each training sample.

[0119] S402. Generate annotation information for the amplified sample data based on the data augmentation algorithm used to generate the amplified sample data.

[0120] For training datasets with text data type, manual annotation can be used to annotate the generated augmented sample data corresponding to each training sample data in the training dataset.

[0121] For training datasets with image data, when choosing blurring, noise, and color jitter as the three methods for data augmentation, since this data augmentation method does not interfere with the original size and content of the image, that is, the essence of the data does not change, the annotation information of the generated augmented sample data is the same as the original annotation information of the training sample data.

[0122] When using the four methods of flipping, rotating, scaling, and cropping for data augmentation, the size and shape of the generated amplified sample data will change. Therefore, new annotation information needs to be generated for each amplified sample data. In this embodiment, new annotation information can be automatically generated based on the original label information according to the amplification rules corresponding to the data amplification method.

[0123] S403. Based on each training sample data and the corresponding augmented sample data, the target training dataset is obtained.

[0124] Optionally, the target training dataset can be obtained based on the training sample data and corresponding annotation information in the training dataset, as well as the augmented sample data and corresponding annotation information corresponding to each training sample data.

[0125] Figure 5 This is a flowchart illustrating another model generation method provided in an embodiment of this application; optionally, in step S104, determining the model generation parameters based on the data exploration results and the business type may include:

[0126] S501. Based on the data exploration results and the business type, determine at least one artificial intelligence algorithm from the artificial intelligence algorithm library.

[0127] Optionally, the data exploration unit can provide analytical exploration capabilities for structured datasets, image datasets, and text datasets, including: category feature analysis, chi-square independence test, normality test, outlier analysis, missing value analysis, correlation analysis, correlation matrix analysis, statistical calculation analysis, channel mean analysis, and balance analysis.

[0128] The objectives of each data exploration method are as follows:

[0129] Category feature analysis: Explore the distribution of another continuous feature under different values ​​of a discrete feature.

[0130] Chi-square independence test: to explore whether two features are independent.

[0131] Normality test: to explore whether a feature conforms to a normal distribution.

[0132] Outlier detection: Outliers and their range of values ​​are identified using box plots or normal distribution methods.

[0133] Missing feature analysis: Explore the missing feature level for each feature, expressed as a percentage.

[0134] Correlation analysis: Explore the correlation between two features to obtain the correlation coefficient and p-value.

[0135] Correlation matrix analysis: Explore the correlation coefficient between every two features among all features to obtain a correlation matrix.

[0136] Statistical calculation and analysis: Explore the mean, variance, minimum, maximum, quantiles, and histograms of each feature.

[0137] Channel mean analysis: Explore the mean of each color channel in an image type dataset.

[0138] Balance analysis: In a dataset of images or text, the total number of samples in each class is counted to determine whether the sample distribution is balanced.

[0139] The results of the data exploration are presented in chart form. Based on the results, the subsequent data preprocessing methods can be determined.

[0140] Optionally, based on the data exploration results and in combination with the business type, at least one artificial intelligence algorithm that constitutes the model generation parameters can be determined.

[0141] Here, the selection of artificial intelligence algorithms is also based on the algorithm library in the user interface of the provided AI algorithm platform.

[0142] S502. Determine the parameter information for each artificial intelligence algorithm. The parameter information includes: the operators used by the algorithm and / or the parameter values ​​of the algorithm.

[0143] In some embodiments, based on the selected algorithms, corresponding parameter information can also be set for each algorithm, wherein the parameter information may include: the specific parameter values ​​used by the algorithm and / or the specific operators used by the algorithm.

[0144] S503. Based on the information of each artificial intelligence algorithm and parameter, obtain the model generation parameters.

[0145] Therefore, the model generation parameters can be obtained from the parameter information of each determined artificial intelligence algorithm, which is to generate a model training task.

[0146] In another embodiment, in addition to determining the artificial intelligence algorithm as described above, any number of data preprocessing algorithms and any number of feature engineering algorithms can be determined based on the data exploration results and the business type.

[0147] The model training unit supports data preprocessing algorithms including but not limited to: sampling, data splitting, missing value imputation, data type conversion, data filtering, and sliding window. Supported feature engineering algorithms include but are not limited to: one-hot encoding, continuous feature discretization, normalization, standardization, scale conversion, shape modification, and mean reduction. It also includes over one hundred built-in AI algorithms, including but not limited to: logistic regression, decision trees, k-means clustering algorithm, MLP (Multilayer Perceptron), RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), MobileNet (Lightweight Convolutional Neural Network), ResNet (Residual Network), SSD (Single Shot MultiBox Detector), Faster R-CNN (Faster Region-CNN), EfficientDet (Scalable and Efficient Object Detection), deeplab (Semantic Segmentation), and Mask R-CNN (Mask Region-CNN).

[0148] For example, if a user creates a new data exploration task for outlier analysis, and the data exploration results show that outliers are those with a value greater than 100, then in the subsequent data preprocessing method selection, the data filtering method can be selected to filter out the sample data with a value greater than 100.

[0149] In this scenario, the model generation parameters can be obtained from the determined target data preprocessing algorithms and their parameter information, the feature engineering algorithms and their parameter information, and the artificial intelligence algorithms and their parameter information, thus generating a model training task.

[0150] In some embodiments, the model generation parameters obtained above can be submitted to the computing cluster of the AI ​​algorithm platform in JSON format for execution. The execution results can be saved in object storage according to user ID and model ID, and the execution result can be the generated target model.

[0151] The underlying model training unit is based on two machine learning computing engines, Spark ML and Tensorflow, and corresponding Spark clusters and GPU clusters are built.

[0152] Figure 6 This is a flowchart illustrating another model generation method provided in an embodiment of this application; optionally, in step S105, generating the target model according to the model configuration information may include:

[0153] S601. Execute the model configuration information to generate the initial model.

[0154] Optionally, after executing the model configuration information, an initial model can be obtained. In this implementation, the initial model is further lightweighted to obtain a model that can be deployed on an edge platform.

[0155] S602. Using the importance calculation formula, determine the importance of each convolution kernel in each convolutional layer of the initial model.

[0156] Optionally, the importance of each convolutional kernel in each convolutional layer of the initial model can be calculated using an importance calculation formula.

[0157] For each convolution kernel F i, Calculate its importance S i,j S i,j The calculation formula is as follows:

[0158]

[0159] in, i represents the i-th convolutional layer, j represents the j-th convolutional kernel of that convolutional layer, and n i Represents the convolution kernel F i, The number of channels, k×k represents the width and height of the convolution kernel, ∑| l | represents the sum of the absolute values ​​of all elements in the l-th channel.

[0160] S603. Based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold, prune each convolutional layer of the initial model in sequence to obtain the target model.

[0161] Based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold of the user input, the model can be lightweighted by pruning each convolutional layer of the initial model to obtain the target model.

[0162] Figure 7 The following is a flowchart illustrating another model generation method provided in this application embodiment; optionally, in step S603, pruning each convolutional layer of the initial model sequentially based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold to obtain the target model may include:

[0163] S701. Based on the importance of each convolutional kernel in the target convolutional layer, the convolutional kernel with the lowest importance is pruned in sequence. After each convolutional kernel is pruned, it is determined whether the accuracy loss value of the current model is greater than the accuracy loss threshold.

[0164] Each convolutional layer is pruned independently, and the pruned model is evaluated on the validation set to determine the number of prunings N_i for each convolutional layer.

[0165] When pruning a convolutional layer (taking the target convolutional layer as an example), the convolutional kernels in the convolutional layer are sorted from low to high importance, and the convolutional kernel with the lowest importance is pruned in turn. After each convolutional kernel is removed, the pruned model is evaluated on the validation set.

[0166] It's important to distinguish between the concepts of model accuracy and model accuracy loss. Assuming the model's current accuracy is 99% and the accuracy loss is 5%, then after a 5% accuracy loss, the model's accuracy becomes 94%. This embodiment uses the model's accuracy loss threshold for calculation, not the model's accuracy threshold.

[0167] S702. If yes, then determine that the number of convolutional kernels currently pruned is the number of pruning branches corresponding to the target convolutional layer; if no, then continue to prune the convolutional kernels with the lowest current importance from each convolutional kernel of the target convolutional layer until the accuracy loss value of the current model is greater than the accuracy loss threshold, then stop execution.

[0168] If the accuracy loss of the model is less than the accuracy loss threshold set by the user, the next convolutional kernel with the lowest importance will be pruned. If the accuracy loss of the model is greater than or equal to the accuracy loss threshold set by the user, the pruning of this layer ends. At this time, the total number of convolutional kernels pruned in this layer is the number of pruned kernels N_i.

[0169] S703. Prune each convolutional layer of the initial model sequentially according to the number of pruning operations corresponding to each convolutional layer to obtain the target model.

[0170] The number of pruning branches for each convolutional layer can be determined through the above steps. Then, each convolutional layer can be pruned according to the number of pruning branches for each convolutional layer to obtain a lightweight target model.

[0171] Figure 8 This is a flowchart illustrating another model generation method provided in an embodiment of this application; in step S703, the initial model's convolutional layers are pruned sequentially according to the pruning number corresponding to each convolutional layer to obtain the target model, which may include:

[0172] S801. Based on the number of pruning branches corresponding to the current convolutional layer, prune the convolutional kernel with the lowest importance of the number of pruning branches from the current convolutional layer to obtain the current pruned model.

[0173] The current convolutional layer can refer to the convolutional layer that is currently being pruned. Pruning can start from the first convolutional layer of the initial model. After pruning the first layer, the second convolutional layer will be pruned, and so on, to complete the pruning of each convolutional layer.

[0174] The pruned model can be obtained by subtracting the N_i least important convolutional kernels from the current convolutional layer. N_i is the number of pruned kernels corresponding to the current convolutional layer determined above.

[0175] S802. Determine whether the accuracy loss value of the currently pruned model is greater than the accuracy loss threshold.

[0176] Optionally, the current pruned model can be evaluated on a validation set.

[0177] S803. If so, then the network parameters of the current pruned model are tuned until the accuracy loss value of the current pruned model is less than the accuracy loss threshold, and then the current new model is obtained.

[0178] If the accuracy loss of the current pruned model is greater than or equal to the accuracy loss threshold set by the user, the model will continue to be tuned on the training set. Tuning the model involves adjusting the network parameters of the model while keeping the network structure unchanged. When the tuning result makes the accuracy loss of the model less than the accuracy loss threshold, the tuning stops and a new model is obtained.

[0179] S804. Determine the next convolutional layer of the current convolutional layer as the new convolutional layer, and prune the new convolutional layer of the current model according to the number of pruning branches corresponding to the new convolutional layer, until the pruning of all convolutional layers is completed, and take the model obtained after the pruning of each convolutional layer as the target model.

[0180] At this point, the pruning of the current convolutional layer is complete. The next layer of the current convolutional layer can be taken as a new convolutional layer, and the new convolutional layer can be pruned in the same way as described above. This process continues until all convolutional layers have been pruned, and then the lightweight target model is obtained.

[0181] In some embodiments, if the accuracy loss value of the currently pruned model is less than the accuracy loss threshold set by the user, then no model tuning is performed, and instead, the next convolutional layer of the current convolutional layer is pruned directly.

[0182] Alternatively, in one approach, users can download the generated target model file from the AI ​​algorithm platform and deploy the model file on the platform to be deployed for executing corresponding business logic. Users can download not only their own target models but also target models belonging to other users with access permissions, thus achieving model sharing.

[0183] In another approach, the model repository unit of the AI ​​algorithm platform can support one-click deployment of model services. Using Docker container technology, hot updates of models can be achieved.

[0184] The model repository unit supports user permission management for model services, configuring each user's access permissions for each model service. For authorized users, when a user successfully logs in for the first time with their username and password, the platform generates a token and a token expiration time and returns this to the client. Subsequently, the client only needs to bring this token with it when requesting data within the valid time, without having to bring the username and password again, thus ensuring the security of using model services.

[0185] When users call the model service, they can log in through the client on the terminal. After successful login, they can call the model service with the appropriate permissions to process business online. This method allows for real-time online model calls without deploying and loading the target model on the terminal.

[0186] In summary, the model generation method provided in this application allows users to input a business type and training dataset through the provided AI algorithm platform. Based on the business type and the data types in the training dataset, the method annotates and explores the training dataset, generating model configuration information. After executing the model configuration information, a model for performing the corresponding business type can be generated. This method automatically generates a model corresponding to the business type by simply inputting the relevant information involved in model generation, reducing the human effort required for manual model training. Users do not need to learn the details of the underlying algorithms to obtain the required model, making the barrier to entry low. Furthermore, based on the user-inputted accuracy loss threshold, a lightweight model that meets the accuracy loss threshold can be automatically generated, thus satisfying the deployment requirements of the model on edge platforms and making the generated model more suitable for edge AI business scenarios.

[0187] The following describes the apparatus, device, and storage medium used to execute the model generation method provided in this application. The specific implementation process and technical effects are described above and will not be repeated below.

[0188] Figure 9This is a schematic diagram of a model generation apparatus provided in an embodiment of this application. The function implemented by this model generation apparatus corresponds to the steps performed by the method described above. This apparatus can be understood as the aforementioned computer device, such as... Figure 9 As shown, the device may include: an acquisition module 910, a labeling module 920, a data exploration module 930, a determination module 940, and a generation module 950;

[0189] Module 910 is used to acquire the business type and training dataset input by the user.

[0190] The annotation module 920 is used to determine the annotation method of the training dataset according to the business type, and to obtain the annotated training dataset according to the annotation method;

[0191] The data exploration module 930 is used to perform corresponding data exploration tasks on the labeled training sample set according to the data type and business type of the training dataset, and obtain data exploration results.

[0192] The determination module 940 is used to determine the model generation parameters based on the data exploration results and the business type. The model generation parameters include at least the artificial intelligence algorithm and its parameters.

[0193] The generation module 950 is used to generate model configuration information based on the data exploration results, model generation parameters, and the accuracy loss threshold input by the user, and to generate the target model based on the model configuration information.

[0194] Optionally, the annotation module 920 is specifically used to determine the target annotation method from the annotation methods that match the data type of the training dataset, based on the business type;

[0195] The target annotation method is used to annotate each training sample in the training dataset to obtain the annotated training dataset.

[0196] Optionally, it also includes: a data augmentation module;

[0197] The data augmentation module is used to augment the training sample data in the training dataset according to the data type of the training dataset, using at least one data augmentation algorithm that matches the data type, to obtain augmented sample data corresponding to each training sample data. The data types include: text data, image data, and structured data.

[0198] Based on the data augmentation algorithm used to generate the amplified sample data, generate annotation information for the amplified sample data;

[0199] The target training dataset is obtained based on each training sample data and the corresponding augmented sample data.

[0200] Optionally, the determining module 940 is specifically used to determine at least one artificial intelligence algorithm from an artificial intelligence algorithm library based on the data exploration results and the business type;

[0201] Determine the parameter information for each artificial intelligence algorithm. The parameter information includes the operators used by the algorithm and / or the parameter values ​​of the algorithm.

[0202] Based on the information of various artificial intelligence algorithms and parameters, the model generation parameters are obtained.

[0203] Optionally, the generation module 950 is specifically used to execute the model configuration information and generate the initial model;

[0204] The importance of each convolutional kernel in each convolutional layer of the initial model is determined using the importance calculation formula.

[0205] Based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold, the initial model's convolutional layers are pruned sequentially to obtain the target model.

[0206] Optionally, the generation module 950 is specifically used to cut off the convolutional kernel with the lowest importance in turn according to the importance of each convolutional kernel in the target convolutional layer, and after cutting off a convolutional kernel each time, to determine whether the accuracy loss value of the current model is greater than the accuracy loss threshold.

[0207] If yes, then determine that the number of convolutional kernels currently pruned is the number of pruning branches corresponding to the target convolutional layer; if not, then continue to prune the convolutional kernels with the lowest current importance from each convolutional kernel of the target convolutional layer until the accuracy loss value of the current model is greater than the accuracy loss threshold, then stop execution.

[0208] The initial model is pruned sequentially according to the number of pruning operations corresponding to each convolutional layer to obtain the target model.

[0209] Optionally, the generation module 950 is specifically used to prune the convolutional kernel with the lowest pruning importance from the current convolutional layer according to the pruning number corresponding to the current convolutional layer, so as to obtain the current pruned model;

[0210] Determine whether the accuracy loss value of the currently pruned model is greater than the accuracy loss threshold;

[0211] If so, the network parameters of the current pruned model are tuned until the accuracy loss of the current pruned model is less than the accuracy loss threshold, and then a new model is obtained.

[0212] The next convolutional layer after the current convolutional layer is determined as the new convolutional layer. Based on the number of pruning branches corresponding to the new convolutional layer, the new convolutional layer of the current model is pruned until all convolutional layers are pruned. The model obtained after pruning each convolutional layer is taken as the target model.

[0213] The above-described device is used to execute the method provided in the foregoing embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

[0214] These modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more digital signal processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). Alternatively, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a system-on-a-chip (SOC).

[0215] The modules described above can be connected or communicate with each other via wired or wireless connections. Wired connections can include metal cables, optical fibers, hybrid cables, or any combination thereof. Wireless connections can include connections via LAN, WAN, Bluetooth, ZigBee, or NFC, or any combination thereof. Two or more modules can be combined into a single module, and any module can be divided into two or more units. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here.

[0216] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The device includes: a processor 801 and a storage medium 802.

[0217] Storage medium 802 is used to store programs, and processor 801 calls the programs stored in storage medium 802 to execute the above method embodiments. The specific implementation and technical effects are similar, and will not be described in detail here.

[0218] The storage medium 802 stores program code, which, when executed by the processor 801, causes the processor 801 to perform various steps in the model generation method according to various exemplary embodiments of this application as described in the "Exemplary Methods" section above.

[0219] The processor 801 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0220] Storage medium 802, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. The storage medium can include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type storage medium, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage medium, magnetic disk, optical disk, etc. The storage medium is any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In the embodiments of this application, storage medium 802 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0221] Optionally, this application also provides a program product, such as a computer-readable storage medium, including a program that, when executed by a processor, performs the above-described method embodiments.

[0222] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0223] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0224] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0225] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A model generation method, characterized in that, include: Obtain the user-inputted business type and training dataset; Based on the business type, determine the annotation method of the training dataset, and obtain the annotated training dataset according to the annotation method; Based on the data type of the training dataset and the business type, a corresponding data exploration task is performed on the labeled training sample set to obtain data exploration results; the data exploration task is used to provide the ability to analyze and explore structured datasets, image datasets, and text datasets; Based on the data exploration results and the business type, the model generation parameters are determined, and the model generation parameters include at least: artificial intelligence algorithms and parameters; Based on the data exploration results, the model generation parameters, and the accuracy loss threshold input by the user, model configuration information is generated, and a target model is generated based on the model configuration information. The step of determining the annotation method of the training dataset according to the business type, and obtaining the annotated training dataset according to the annotation method, includes: Based on the business type, determine the target annotation method from the annotation methods that match the data type of the training dataset; The training sample data in the training dataset is labeled using the target annotation method to obtain the labeled training dataset; The step of determining the model generation parameters based on the data exploration results and the business type includes: Based on the data exploration results and the business type, at least one artificial intelligence algorithm is determined from the artificial intelligence algorithm library; The parameter information of each artificial intelligence algorithm is determined, and the parameter information includes: the operators used by the algorithm and / or the parameter values ​​of the algorithm; The model generation parameters are obtained based on the various artificial intelligence algorithms and parameter information.

2. The method according to claim 1, characterized in that, After obtaining the labeled training dataset according to the labeled method, the process further includes: Based on the data type of the training dataset, at least one data augmentation algorithm matching the data type is used to augment each training sample data in the training dataset to obtain augmented sample data corresponding to each training sample data. The data type includes: text data, image data, and structured data. Based on the data augmentation algorithm used to generate the amplified sample data, the annotation information of the amplified sample data is generated; The target training dataset is obtained based on the training sample data and the corresponding augmented sample data.

3. The method according to claim 1, characterized in that, The step of generating the target model based on the model configuration information includes: Execute the model configuration information to generate the initial model; The importance of each convolutional kernel in each convolutional layer of the initial model is determined using the importance calculation formula. Based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold, the convolutional layers of the initial model are pruned sequentially to obtain the target model.

4. The method according to claim 3, characterized in that, The step of pruning each convolutional layer of the initial model sequentially based on the importance of each convolutional kernel in each convolutional layer and the accuracy loss threshold to obtain the target model includes: Based on the importance of each convolutional kernel in the target convolutional layer, the convolutional kernel with the lowest importance is pruned in sequence, and after each convolutional kernel is pruned, it is determined whether the accuracy loss value of the current model is greater than the accuracy loss threshold. If yes, then determine that the number of convolutional kernels currently pruned is the number of pruning branches corresponding to the target convolutional layer; if no, then continue to prune the convolutional kernel with the lowest current importance from each convolutional kernel of the target convolutional layer until the accuracy loss value of the current model is greater than the accuracy loss threshold, then stop execution. The initial model is pruned sequentially according to the number of pruning operations corresponding to each convolutional layer to obtain the target model.

5. The method according to claim 4, characterized in that, The step of pruning each convolutional layer of the initial model according to the pruning number corresponding to each convolutional layer to obtain the target model includes: Based on the number of pruning branches corresponding to the current convolutional layer, the convolutional kernel with the lowest importance of the pruning branch number is pruned from the current convolutional layer to obtain the current pruned model; Determine whether the accuracy loss value of the currently pruned model is greater than the accuracy loss threshold; If so, the network parameters of the current pruned model are tuned until the accuracy loss value of the current pruned model is less than the accuracy loss threshold, and then a new model is obtained. The next convolutional layer after the current convolutional layer is determined as the new convolutional layer. Based on the number of pruning branches corresponding to the new convolutional layer, the new convolutional layers of the current new model are pruned until all convolutional layers are pruned. The model obtained after pruning each convolutional layer is taken as the target model.

6. A model generation apparatus, characterized in that, include: Acquisition module, annotation module, data exploration module, determination module, generation module; The acquisition module is used to acquire the business type and training dataset input by the user; The annotation module is used to determine the annotation method of the training dataset according to the business type, and to obtain the annotated training dataset according to the annotation method; The data exploration module is used to perform corresponding data exploration tasks on the labeled training sample set according to the data type of the training dataset and the business type, and obtain data exploration results; the data exploration tasks are used to provide the ability to analyze and explore structured datasets, image datasets, and text datasets; The determining module is used to determine model generation parameters based on the data exploration results and the business type. The model generation parameters include at least: artificial intelligence algorithms and parameters. The generation module is used to generate model configuration information based on the data exploration results, the model generation parameters, and the accuracy loss threshold input by the user, and to generate a target model based on the model configuration information. The annotation module is specifically used to determine the target annotation method from the annotation methods that match the data type of the training dataset, based on the business type. The training sample data in the training dataset is labeled using the target annotation method to obtain the labeled training dataset; The determining module is specifically used to determine at least one artificial intelligence algorithm from an artificial intelligence algorithm library based on the data exploration results and the business type. The parameter information of each artificial intelligence algorithm is determined, and the parameter information includes: the operators used by the algorithm and / or the parameter values ​​of the algorithm; The model generation parameters are obtained based on the various artificial intelligence algorithms and parameter information.

7. An electronic device, characterized in that, include: The device includes a processor, a storage medium, and a bus, wherein the storage medium stores program instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the program instructions to perform the steps of the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, performs the steps of the method as described in any one of claims 1 to 5.

Images