Data training method and device

By combining iterative training business data and heterologous data samples, using edge detection technology and label equalization, the problem of insufficient sample number in deep learning model training is solved, and efficient and low-cost model training and prediction effects are achieved.

CN118941884APending Publication Date: 2024-11-12HUAWEI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310853647.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-11
Filing Date
2023-07-11
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In deep learning model training, due to the high cost of collecting business data and labeling, the number of training samples is insufficient. It is difficult for the existing technology to effectively use heterologous data for auxiliary training, improving the prediction accuracy and efficiency of the model.

Method used

By obtaining training samples including business data samples and heterologous data samples, iteratively train the target initial model, use edge detection thresholds and algorithms to detect edges in heterologous data, and combine the target attribute tags in business data samples and heterologous data samples in heterologous data samples. The calculation labels are trained for model training, and the positive and negative edge ratios are equalized to improve the prediction accuracy and efficiency of the model.

Benefits of technology

High prediction accuracy and high prediction efficiency of deep learning models assisted by heterologous data are achieved, reducing training costs and simplifying the training process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941884A_ABST
    Figure CN118941884A_ABST
Patent Text Reader

Abstract

The invention discloses a data training method and device, and relates to the technical field of data processing. The method comprises the following steps: acquiring a training sample comprising a business data sample and a heterogenous data sample; wherein the business data sample is marked with a target attribute tag of a target object for the user business, and the heterogeneous data sample is a data sample different from the business data sample in source. And determining a target initial model according to a target task, wherein the target task is a task for detecting a target object in the image. And iteratively training the target initial model according to the business data sample and the heterogeneous data sample to obtain a target model for executing the target task. Through the method, the deep learning model can be trained simply and efficiently through different-source data assistance, and the deep learning model obtained through the method has high prediction precision and high prediction efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] This application claims the priority of the Chinese patent application filed with the State Intellectual Property Office on May 11, 2023, with application number 202310531531.2 and application name “A Data Training Method”, all contents of which are incorporated by reference in this application. Technical Field

[0002] The present application relates to the field of data processing technology, and in particular to a data training method and device. Background Art

[0003] The training of deep learning models requires a large amount of labeled business data as training samples. However, due to the high cost of collecting business data and labeling, the actual scenario often faces the problem of insufficient number of training samples. Business data refers to the data obtained by users in their specific business and labeled.

[0004] To address this problem, heterogeneous data can be introduced to assist in model training during deep learning model training based on business data. This can reduce the customer's demand for business data when training deep learning models, thereby reducing the cost of training deep learning models for customers and improving the efficiency of developing deep learning models for customers. Heterogeneous data refers to data obtained by users from sources other than their own business, such as data obtained from the Internet. Compared with business data, heterogeneous data is not labeled, but the cost of collecting heterogeneous data is low.

[0005] Therefore, how to simply and efficiently use heterogeneous data to assist in training deep learning models is an urgent problem to be solved. Summary of the invention

[0006] The present application provides a data training method and device, which can simply and efficiently train a deep learning model through the assistance of heterogeneous data, and the obtained deep learning model has the advantages of high prediction accuracy and high prediction efficiency.

[0007] To achieve the above objectives, this application provides the following technical solutions:

[0008] In a first aspect, the present application provides a data training method, the method comprising: obtaining training samples including business data samples and heterogeneous data samples. The business data samples are annotated with target attribute labels of target objects for user businesses, and the heterogeneous data samples are data samples from sources different from those of the business data samples. A target initial model is determined according to a target task, and the target task is a task of detecting a target object in an image. The target initial model is iteratively trained according to the business data samples and heterogeneous data samples in the training samples to obtain a target model for performing the target task.

[0009] The target model for performing target detection tasks obtained by the method provided by the present application has greatly improved prediction efficiency and prediction accuracy, and the training process is simple. In addition, since a large amount of heterogeneous data with low collection cost is used in the method of the present application, the cost of training the target model by the method provided by the embodiment of the present application is low.

[0010] In a possible design, the method further includes: receiving at least one edge detection threshold input by a user, the edge detection threshold being used to detect edges in heterogeneous data samples. The iterative training of the target initial model based on the business data samples and heterogeneous data samples in the training samples includes: iteratively training the target initial model based on the business data samples, the heterogeneous data samples and at least one edge detection threshold.

[0011] Through this possible design, the computing device can obtain the edges in the heterogeneous data samples at different edge detection threshold granularities, which can ensure that the computing device fully detects all the edges in the heterogeneous data samples. And the aforementioned granularity is customized by the user, so the user experience is good.

[0012] In another possible design, the training sample includes a first training sample, the first training sample includes a first business data sample and a first heterogeneous data sample, the target attribute of the target object includes the position of the target object and the category of the target object, wherein the position of the target object is the position of the target object in the image. The iterative training of the target initial model according to the business data sample, the heterogeneous data sample and at least one edge detection threshold includes: determining the calculation label of the target object in the first heterogeneous data sample according to the target initial model to be trained and at least one edge detection threshold. The target initial model to be trained is trained according to the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculation label of the first heterogeneous data sample to obtain the target model. The calculation label is a label calculated for the target attribute of the target object in the first heterogeneous data sample, and the calculation label includes a calculation position box and a calculation category of the target object in the first heterogeneous data sample, the calculation position box is used to characterize the position of the target object calculated in the first heterogeneous data sample, and the calculation category is a category corresponding to the calculation position box.

[0013] Through this possible design method, the model is trained simultaneously with the target attribute labels of the target objects in the business data samples and the calculated labels of the target objects calculated in the heterogeneous data samples. Thus, by using the business data samples and a large number of heterogeneous data samples to jointly train the target initial model, the knowledge in a large number of heterogeneous data samples can be transferred to the converged target initial model that meets the user's business needs, thereby obtaining a target model with high prediction accuracy and high prediction efficiency.

[0014] In another possible design, the above-mentioned determination of the calculation label of the target object in the first heterogeneous data sample according to the target initial model to be trained and at least one edge detection threshold includes: determining the predicted position frame and the predicted category of the target object in the first heterogeneous data sample according to the target initial model to be trained. The predicted position frame is used to characterize the position of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained. The edge in the first heterogeneous data sample is determined according to the above-mentioned at least one edge detection threshold and edge detection algorithm. The calculation position frame is determined according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, and the calculation category is determined according to the predicted category.

[0015] In another possible design, the determining of the edge in the first heterogeneous data sample according to at least one edge detection threshold and an edge detection algorithm includes: detecting and obtaining an initial edge in the first heterogeneous data sample under each edge detection threshold according to the edge detection algorithm and at least one edge detection threshold. Fusing the initial edges according to a non-maximum suppression algorithm to obtain an edge in the first heterogeneous data sample.

[0016] In another possible design, the above-mentioned determining the calculated position frame according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame includes: determining the edge in the first heterogeneous data sample that is located in the predicted position frame as the positive edge in the first heterogeneous data sample according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, wherein the positive edge in the first heterogeneous data sample is the edge in the first heterogeneous data sample that is related to the target object. The positive edge in the first heterogeneous data sample is fused using a minimum bounding box fusion algorithm to obtain the calculated position frame.

[0017] Through the above-mentioned possible design methods, the computational labels of the target objects in the heterogeneous data samples can be determined.

[0018] In another possible design, the training of the target initial model to be trained based on the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculated label of the first heterogeneous data sample includes: inputting the first business data sample and the first heterogeneous data sample into the target initial model to be trained to obtain a prediction result. Determine a position loss function and a category loss function based on the prediction result, the target attribute label of the target object in the first business data sample, and the calculated label of the first heterogeneous data sample. Update the target initial model to be trained based on the position loss function and the category loss function.

[0019] Through this possible implementation method, it is possible to train a model based on both business data samples and heterogeneous data samples at the same time.

[0020] In another possible design, the method further includes: receiving a positive-negative edge ratio input by a user. The positive-negative edge ratio is used to balance the number of positive edges and negative edges in the heterogeneous data sample, the positive edge in the heterogeneous data sample is used to determine the calculated label of the target object in the heterogeneous data sample, and the negative edge in the heterogeneous data sample is used to determine the background label of the heterogeneous data sample, and the background label of the heterogeneous data sample includes the background box and the background category in the heterogeneous data sample. The above-mentioned iterative training of the target initial model based on the business data samples and the heterogeneous data samples in the training samples includes: iteratively training the target initial model based on the business data samples, the heterogeneous data samples, the calculated labels of the heterogeneous data samples, and the background labels of the heterogeneous data samples.

[0021] In another possible design, the method further includes: according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, the edge in the first heterogeneous data sample located outside the predicted position frame is used to determine the negative edge in the first heterogeneous data sample. Wherein, the negative edge in the first heterogeneous data sample is an edge in the first heterogeneous data sample that is not related to the target object. The negative edges in the first heterogeneous data sample are fused using the minimum bounding box fusion algorithm to obtain a background frame in the first heterogeneous data sample, and the category corresponding to the background frame in the first heterogeneous data sample is the background category. The iterative training of the target initial model according to the business data sample, the heterogeneous data sample, the calculated label of the heterogeneous data sample, and the background label of the heterogeneous data sample includes: training the target initial model to be trained according to the first business data sample, the first heterogeneous data sample, the calculated label of the first heterogeneous data sample, and the background frame and background category in the first heterogeneous data sample.

[0022] In another possible design, the method further includes: sampling in the negative edges in the first heterogeneous data sample so that the number of positive edges in the first heterogeneous data sample and the number of negative edges in the first heterogeneous data sample satisfy the positive-negative edge ratio. The above-mentioned fusing the negative edges in the first heterogeneous data sample with the minimum bounding box fusion algorithm to obtain the background frame in the first heterogeneous data sample includes: fusing the sampled negative edges with the minimum bounding box fusion algorithm to obtain the background frame of the first heterogeneous data sample.

[0023] Through the above-mentioned possible implementation methods, the present application can improve the accuracy of the model after convergence by adding a training object: background in the model training process. In addition, when the computing device determines the calculation label of the target object in the first heterogeneous data sample, it is necessary to use the predicted position frame obtained after the target initial model predicts the first heterogeneous data sample, and in the initial stage of training the target initial model, the predicted position frame of the target object in the first heterogeneous data sample predicted by the computing device based on the target initial model is not very accurate. Therefore, based on the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, the positive edges determined for obtaining the calculation label are relatively small, and the negative edges determined for obtaining the background label are relatively large. Therefore, the present application scheme can obtain a suitable number of positive edges and negative edges by balancing the ratio of the number of positive edges to the number of negative edges, and calculate the corresponding calculation labels and background labels for model training. In addition, the positive and negative edge ratio can be specified by the user, so the user experience is good.

[0024] In another possible design, before determining the target initial model according to the target task, the method further includes: receiving a first selection operation input by a user, and determining the target task from a plurality of preset tasks in response to an instruction of the first selection operation.

[0025] In another possible design, the above-mentioned determining the target initial model according to the target task includes: receiving a second selection operation input by the user. In response to an instruction of the second selection operation, determining the target initial model from a plurality of initial models corresponding to the target task.

[0026] Through the above two possible design methods, the computing device can determine the target initial model that meets the user's needs, so that the target model obtained after training the target initial model can also meet the user's usage needs. It can be seen that the present application solution is more in line with user needs, so the user experience is good.

[0027] In a second aspect, the present application provides a data training device.

[0028] In one possible design, the data training device is used to execute any one of the methods provided in the first aspect above. The present application may divide the data training device into functional modules according to any one of the methods provided in the first aspect above. For example, each functional module may be divided according to each function, or two or more functions may be integrated into one processing module. Exemplarily, the present application may divide the data training device into an acquisition unit, a processing unit, and a training unit, etc. according to the function. The description of the possible technical solutions and beneficial effects executed by the above-mentioned divided functional modules can refer to the technical solutions provided by the above-mentioned first aspect or its corresponding possible design, and will not be repeated here.

[0029] In another possible design, the data training device includes: one or more processors and a memory, and the one or more processors are configured to call program instructions stored in the memory so that the data training device executes any method provided by the first aspect and any possible design thereof.

[0030] In a third aspect, the present application provides a computer-readable storage medium, which includes program instructions. When the program instructions are executed on a computer or a processor, the computer or the processor executes any method provided by any possible implementation in the first aspect.

[0031] In a fourth aspect, the present application provides a computer program product, which, when executed on a data training device, enables any method provided by any possible implementation of the first aspect to be executed.

[0032] It can be understood that any of the above-mentioned devices, computer storage media or computer program products can be applied to the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding methods and will not be repeated here.

[0033] In this application, the name of the above-mentioned data training device does not limit the device or functional module itself. In actual implementation, these devices or functional modules may appear with other names. As long as the functions of each device or functional module are similar to those of this application, they belong to the scope of the claims of this application and their equivalent technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of the hardware structure of a computing device provided in an embodiment of the present application;

[0035] Figure 2 A schematic diagram of the architecture of a data training system provided in an embodiment of the present application;

[0036] Figure 3 A flowchart of a data training method provided in an embodiment of the present application;

[0037] Figure 4 A schematic diagram of receiving a first selection operation provided in an embodiment of the present application;

[0038] Figure 5 A schematic diagram of receiving a second selection operation provided in an embodiment of the present application;

[0039] Figure 6 A schematic diagram of receiving at least one edge detection threshold provided in an embodiment of the present application;

[0040] Figure 7 A schematic diagram of a process of training a target initial model to be trained according to a first training sample provided in an embodiment of the present application;

[0041] Figure 8 A schematic diagram of positive edges and negative edges in a first heterogeneous data sample provided in an embodiment of the present application;

[0042] Fig. 9 A schematic diagram of receiving a positive and negative edge ratio provided in an embodiment of the present application;

[0043] Fig.10 A schematic diagram of another process of training a target initial model to be trained according to a first training sample provided in an embodiment of the present application;

[0044] Fig.11 Another flowchart of the data training method provided in the embodiment of the present application;

[0045] Fig.12 A schematic diagram of another flow chart of the data training method provided in the embodiment of the present application;

[0046] Fig.13 A schematic diagram of another flow chart of the data training method provided in the embodiment of the present application;

[0047] Fig.14 A structural schematic diagram of a data training device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to more clearly understand the embodiments of the present application, some terms or technologies involved in the embodiments of the present application are explained below:

[0049] 1) Business data and heterogeneous data

[0050] When training a network model (or neural network model or deep learning model, etc.), a large number of labeled samples are needed to train the initial network model to obtain a target model that meets the requirements. The labeled samples are usually obtained by users labeling data collected from their own business scenarios according to their own business needs. In the embodiments of the present application, these sample data are referred to as business data.

[0051] As an example, the user's business requirement is to identify the human action in the image, and the business scenario is a scenario of guiding fitness action. In this case, the user can take multiple images including human fitness, and label the human action in the image. For example, the human action in image 1 is raising both hands to the side, so the label annotated for image 1 is "raising both hands to the side". For another example, the human action in image 2 is raising the left leg high, so the label annotated for image 2 is "raising the left leg high", and so on, without limitation.

[0052] Since labeling data is very costly and inefficient, it is not easy to obtain a large amount of business data for training network models. Too few samples used to train network models will result in poor model accuracy. Therefore, currently a large amount of data that is the same or related to user business needs is collected from various sources as heterogeneous data, and the heterogeneous data is also used as training samples to train network models, thereby increasing the number of training samples used to train network models.

[0053] In combination with the above example, the user can retrieve a large number of images including human actions from the Internet, and use these images as heterogeneous data to train the network model. Among them, the human actions in the images retrieved by the user are not limited to fitness actions, but can be any actions.

[0054] It should be pointed out that heterogeneous data does not need to be labeled, and heterogeneous data is not limited to data from user business scenarios, that is, the sources of heterogeneous data are diverse, so the collection cost of heterogeneous data is low and the quantity is large.

[0055] It can be understood that when training a network model through business data and heterogeneous data, the business data can also be called business data samples, and the heterogeneous data can also be called heterogeneous data samples.

[0056] 2) Other terms

[0057] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0058] In the embodiments of the present application, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features.

[0059] The term "at least one" in this application means one or more, and the term "plurality" in this application means two or more.

[0060] When the network model is trained by business data and heterogeneous data, in one possible implementation, a general model can be obtained by training with a large number of data in different application scenarios. Among them, the data in a large number of different application scenarios used to train the general model can be obtained by labeling the heterogeneous data after obtaining heterogeneous data from various sources (such as the Internet, enterprises, etc.). Then, on the basis of the general model, the general model is further trained using business data in a specific scenario to obtain a target model used by the user for a specific scenario. However, the target model trained in this way has low robustness in the user's specific business scenario. And because this method requires using a large number of data in different application scenarios to train the general model each time, the training cost is high.

[0061] In another possible implementation, a large amount of knowledge points of heterogeneous data are first learned to obtain a teacher model. Then, the network model designed for the user's business scenario is trained using business data, and the teacher model is used to guide the training process to obtain the target model required by the user. However, the process of training the target model in this way is complicated.

[0062] In addition, the methods described in the above two possible implementation methods can only be applied to image classification tasks. For the task of detecting targets / objects in images (i.e., object detection), the prediction speed and prediction accuracy of the target models trained based on the above two methods are both low. Among them, image classification refers to the classification of the image as a whole. For example, images are divided into human images and landscape images. Object detection refers to finding all targets / objects of interest in an image (including but not limited to people, animals, objects, etc.) and determining the category and location of the target / object. Object detection is one of the core issues in the field of computer vision.

[0063] Based on this, an embodiment of the present application provides a data training method, which uses business data samples and heterogeneous data samples to train a network model at the same time. Among them, the business data samples are annotated with target attribute labels of target objects for user business. Specifically, before starting to train the network model, the method first determines the target initial model corresponding to the target task. Among them, the target task is the task of detecting the target object in the image (i.e., the object detection task), and the target initial model is a network model designed for the target task. Then, the target initial model is iteratively trained by the business data samples and the heterogeneous data samples to obtain a target model that meets the user business. The target model obtained by the method provided by the embodiment of the present application has greatly improved prediction efficiency and prediction accuracy, and the training process is simple. In addition, since a large amount of heterogeneous data with low collection cost is used in the method of the embodiment of the present application, the cost of training the target model by the method provided by the embodiment of the present application is low.

[0064] The embodiment of the present application also provides a data training device, which is used to execute the data training method provided in the embodiment of the present application. Optionally, the device can be any computing device with data computing and processing capabilities. Alternatively, the device is integrated as a functional module / component in a computing device or computing system with data computing and processing capabilities.

[0065] In some examples, the data training device may be a computing device such as a general-purpose computer, a laptop, or a tablet. Alternatively, the device may be a network device such as a server. The present invention is not limited thereto.

[0066] In other examples, the data training device can be integrated into an intelligent hardware platform. In this case, the data training device is a functional module or component in the intelligent hardware platform, but is not limited thereto.

[0067] Take the data training device as an example, which is a computing device. Figure 1 , Figure 1 FIG. 1 is a schematic diagram showing a hardware structure of a computing device provided in an embodiment of the present application. Figure 1 As shown, Figure 1 As shown, the computing device 10 includes a processor 101, a memory 102, a network interface 103, and a bus 104. The processor 101, the memory 102, and the network interface 103 are connected via the bus 104. Optionally, the computing device 10 also includes an input / output interface 105, which communicates with the processor 101, the memory 102, and the network interface 103 via the bus 104.

[0068] The processor 101 is the control center of the computing device 10, and may be a general-purpose central processing unit (CPU). The processor 101 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, graphics processing units (GPU), neural processing units (NPU), tensor processing units (TPU) or artificial intelligence (artificial intelligent) chips, data processors (DPU), etc. As an example, the processor 101 may include one or more CPUs, such as Figure 1 In addition, the present application does not limit the number of processor cores in each processor.

[0069] The memory 102 is used to store program instructions or data to be accessed by the application process. The processor 101 can implement the data training method provided in the embodiment of the present application by executing the program instructions in the memory 102.

[0070] The memory 102 includes a volatile memory or a nonvolatile memory, or may include both volatile and nonvolatile memories. Among them, the volatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM) and direct rambus RAM (DR RAM). Non-volatile memory can be storage class memory (SCM), solid state drive (SSD), hard disk drive (HDD), etc. Among them, storage class memory can be, for example, non-volatile memory (NVM), phase-change memory (PCM), persistent memory, etc.

[0071] In one possible implementation, the memory 102 exists independently of the processor 101. The memory 102 is connected to the processor 101 via a bus 104 and is used to store data, instructions or program codes. When the processor 101 calls and executes the instructions or program codes stored in the memory 102, the data training method provided in the embodiment of the present application can be implemented.

[0072] In another possible implementation, the memory 102 and the processor 101 are integrated together.

[0073] The network interface 103 is used for the computing device 10 to connect to other nodes (such as the user terminal described below) through a communication network, and the communication network may be Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc. The network interface 103 includes a receiving unit for receiving data / messages and a sending unit for sending data / messages.

[0074] The bus 104 may be an industry standard architecture (ISA) bus, a peripheral component interconnect (PCI) bus, a high-speed serial computer expansion bus (PCIe), a compute express link (CXL) or an extended industry standard architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 1 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0075] The input / output interface 105 is used to implement human-computer interaction between the user and the computing device 10 , such as text interaction or voice interaction between the user and the computing device 10 .

[0076] The input and output interface 105 includes an input interface for enabling a user to input information to the computing device 10, and an output interface for enabling the computing device 10 to output information to the user. As an example, the input interface includes but is not limited to a touch screen, a keyboard, a mouse or a microphone, etc., and the output interface includes but is not limited to a display screen, a speaker, etc. The touch screen, keyboard or mouse is used to input text / image information, the microphone is used to input voice information, the display screen is used to output text / image information, and the speaker is used to output voice information.

[0077] It should be pointed out that Figure 1 The structure shown in the figure does not constitute a limitation on the computing device 10, except Figure 1 In addition to the components shown, the computing device 10 includes Figure 1 More or fewer components may be shown, or certain components may be combined, or the components may be arranged differently.

[0078] The embodiment of the present application further provides a data training system, which includes the above-mentioned data training device and terminal equipment. Optionally, the data training device and the terminal equipment can communicate with each other in a wired or wireless manner.

[0079] Among them, the description of the data training device in the system can be referred to the above description and will not be repeated here.

[0080] The terminal device includes an input and output interface, so that the terminal device can receive user input operations. Furthermore, the terminal device can instruct the data training device to perform parameter configuration before network model training according to the user input operation. For detailed description, please refer to the description of the method below, which will not be repeated here. Among them, the detailed description of the input and output interface of the terminal device can be referred to Figure 1 The description of the input and output interface 105 of the computing device 10 is not repeated here.

[0081] It should be understood that when the data training device receives the user's input operation through the terminal device to perform parameter configuration before training the network model, the terminal device is pre-installed with a client application for training the network model. In this way, the terminal device can receive the user's input operation through the operation interface of the client application, and instruct the data training device to perform parameter configuration before training the network model according to the input operation. For detailed description, please refer to the method described below, which will not be repeated here.

[0082] As an example, refer to Figure 2 , Figure 2 FIG. 1 is a schematic diagram showing the architecture of a data training system provided by an embodiment of the present application. Figure 2 As shown, the data training system 20 includes a data training device 21 and a terminal device 22. The data training device 21 may be a server. The terminal device 22 may be any terminal device capable of receiving user input operations, such as a mobile phone, a laptop computer, a general-purpose computer, a tablet, a car computer, etc., but is not limited thereto.

[0083] The data training method provided in the embodiment of the present application is described in detail below with reference to the accompanying drawings.

[0084] refer to Figure 3 , Figure 3 The flowchart of a data training method provided in an embodiment of the present application is shown. Optionally, the method can be applied to Figure 1 The computing device shown or Figure 2 In the data training system shown in FIG. Figure 1 Taking the execution of the computing device of the hardware structure shown as an example, the method includes the following steps.

[0085] S101: Obtain training samples.

[0086] The training samples include business data samples and heterogeneous data samples. Both business data samples and heterogeneous data samples are actually image samples. Usually, the number of business data samples is less than the number of heterogeneous data samples.

[0087] In the embodiment of the present application, the business data sample is annotated with a target attribute label of the target object for the user business, and the heterogeneous data sample is a data sample from a different source than the business data sample. Specifically, the detailed description of the business data sample and the heterogeneous data sample can refer to the above description and will not be repeated here.

[0088] Among them, the target object in the business data sample is the object that the user business is concerned about, or it is understood that the target object in the business data sample is the object that the user business needs to predict. As an example, the target object can be a human body, a face, an animal, a vehicle, smoke, etc. in the image, but is not limited to this. The target attribute of the target object includes one or more attributes of the target object. For example, the target attribute of the target object includes the position of the target object, the category of the target object, etc., but is not limited to this. Among them, the position of the target object is the position of the target object in the image. In practice, the position of the target object in the image is represented by a position box including the target object in the image. The category of the target object is the actual category to which the target object belongs, which can include multiple types. For example, the category of the target object can be a human action category, a human category, a human face category, an animal category, a smoke category, etc.

[0089] The target attribute label of the target object refers to: a label annotated with the target attribute of the target object concerned by the user's business in the business data sample. In some examples, the target attribute of the target object includes the location of the target object and the category of the target object, then the target attribute label of the target object in the business data sample includes a location label and a category label. Among them, the location label is a location box used to characterize the location of the target object in the business data sample, and the category label is a category value used to characterize the specific category of the target object in the business data sample, and the category value is, for example, a number, character, name, etc. corresponding to the actual category, but is not limited to this.

[0090] Optionally, the training samples are specifically image samples.

[0091] Optionally, the computing device has pre-stored training samples.

[0092] Optionally, the computing device obtains the training samples from other devices in advance, wherein the other devices may be database devices, storage devices (such as USB flash drives, hard disks, etc.), or any other devices storing the above training samples, which are not limited to this.

[0093] Optionally, when the computing device is a server (or an intelligent hardware platform, etc.), the server can receive training samples uploaded by the user through the terminal device.

[0094] For example, the user's terminal device is installed with the client application described above, so that the user's terminal device can receive the storage address of the training sample input / selected by the user through the input and output interface of the terminal device on the operation interface of the client application, and operate (such as clicking) the "upload" button, so as to upload the training sample pre-stored locally in the user's terminal device to the server. In response, the server receives the training sample uploaded by the user through the terminal device.

[0095] S102: Determine the target initial model according to the target task.

[0096] Among them, the target task is a task aimed at user business.

[0097] Specifically, the target task is to detect the target object in the image (i.e., the object detection task described above). In this case, the target attribute of the target object includes the position of the target object and / or the category of the target object. For the description of the target object and the target attribute of the target object, please refer to the description in S101, which will not be repeated here.

[0098] It should be understood that based on the different categories of target objects, the downstream tasks of the target detection task may also include multiple types. Here, the downstream task of the target detection task is the task for user services. For example, when the category of the target object is a human action category, the downstream task of the target detection is a motion recognition task, or it is called a task for motion recognition services. For another example, when the category of the target object is a smoke category, the downstream task of the target detection is a smoke detection task, or it is called a task for smoke detection services. For another example, when the category of the target object is a face category, the downstream task of the target detection is a face recognition task, or it is called a task for face recognition services. Not limited to this. Among them, motion recognition services, smoke detection services, and face recognition services are all user services.

[0099] The target initial model is a network model (or deep learning model) designed for the target task. For example, when the target task is an action recognition task, the target initial model is a network model designed for the action recognition task.

[0100] Optionally, the target initial model is a network model that has not been pre-trained.

[0101] Optionally, the target initial model is a pre-trained network model. That is, the target initial model is a network model obtained by pre-training a network model designed based on the target task. Optionally, the sample data used for pre-training can be a business data sample in the above training sample, or a part or all of the heterogeneous data samples with labels in the above training sample, without limitation.

[0102] Specifically, one or more tasks are preset in the computing device, and one or more initial models are preset for each task. One task may correspond to one initial model, or one task may correspond to multiple initial models, which is not limited. In this case, the computing device may determine a target task from the multiple preset tasks, and determine a target initial model from the initial models corresponding to the target task.

[0103] Optionally, the computing device may receive a first selection operation input by the user, and in response to an instruction of the first selection operation, determine a target task from a plurality of tasks preset by the computing device. In other words, the target task is a task that meets the business needs of the user.

[0104] In a possible implementation, the computing device directly uses its own input and output interfaces (such as Figure 1 The computing device may receive the first selection operation input by the user on the parameter configuration interface through its own input and output interface. The parameter configuration interface is a visual interface provided by the computing device when training the network model. Then, the computing device determines the target task from the plurality of tasks preset by itself in response to the instruction of the first selection operation.

[0105] In another possible implementation, the computing device is connected to the terminal device (such as Figure 2 The terminal device 22 shown in the figure receives the first selection operation input by the user through its input / output interface. Specifically, the terminal device receives the first selection operation input by the user on the operation interface of the client application through its own input / output interface, and sends the indication of the first selection operation to the computing device. Then, the computing device determines the target task from the plurality of tasks preset by itself in response to the indication of the received first selection operation.

[0106] As an example, refer to Figure 4 , Figure 4 A schematic diagram of receiving a first selection operation provided in an embodiment of the present application is shown.

[0107] like Figure 4 As shown, the task selection interface 40 can be a parameter configuration interface in a visual operation interface provided by the computing device when training a network model, or an operation interface of a client application installed on a terminal device, and this is not limited. Specifically, the computing device (or terminal device) first receives the user's operation on the control 41 on the task selection interface 40 (such as a mouse click, a touch click, etc.), so that the computing device (or terminal device) displays multiple task options (such as the control 42) on the task selection interface 40. Figure 4The computing device (or terminal device) then receives the user's operation on the task selection interface 40 for the task option to be selected (such as a mouse click, a touch click, etc.), so that the computing device (or terminal device) displays the task option selected by the user in the option box 43 of the task selection interface 40. Then, the computing device (or terminal device) receives the user's operation on the "OK" control on the task selection interface 40 (such as a mouse click, a touch click, etc.). In this way, the computing device (or terminal device) receives the first selection operation input by the user.

[0108] Furthermore, in one possible scenario, when the target task corresponds to only one initial model, after the computing device determines the target task in response to the first selection operation, it determines the initial model as the target initial model.

[0109] In another possible situation, when the target task corresponds to multiple initial models, after the computing device determines the target task in response to the first selection operation, it receives a second selection operation input by the user, and in response to the instruction of the second selection operation, determines the target initial model from the multiple initial models corresponding to the target task preset by itself. Optionally, the multiple initial models corresponding to the target task may include an initial model designed by a designer but not pre-trained, an initial model designed by a designer and pre-trained, etc., but is not limited thereto. It should be understood that when the initial model is a pre-trained model, when the initial model is iteratively trained based on the method described in the embodiment of the present application, the training efficiency of the model is higher, that is, the model converges faster.

[0110] Optionally, the computing device directly uses its own input and output interfaces (such as Figure 1 The computing device may receive the second selection operation input by the user on the parameter configuration interface through its own input and output interface. Then, the computing device determines the target initial model from a plurality of initial models corresponding to the target task preset by itself in response to the instruction of the second selection operation.

[0111] Optionally, the computing device is connected to a terminal device (such as Figure 2 The terminal device 22 shown in the figure receives the second selection operation input by the user through its own input and output interface. Specifically, the terminal device receives the second selection operation input by the user on the operation interface of the client application through its own input and output interface, and sends the indication of the second selection operation to the computing device. Then, the computing device determines the target initial model from a plurality of initial models corresponding to the target task preset by itself in response to the indication of the received second selection operation.

[0112] As an example, refer to Figure 5 , Figure 5 A schematic diagram of receiving a second selection operation provided in an embodiment of the present application is shown.

[0113] like Figure 5 As shown, the model selection interface 50 can be a parameter configuration interface of a visual operation interface provided by the computing device when training a network model, or an operation interface of a client application installed on a terminal device, and is not limited to this. Specifically, the computing device (or terminal device) first receives the user's operation on the control 51 on the model selection interface 50 (such as a mouse click, a touch click, etc.), so that the computing device (or terminal device) displays multiple model options (such as the control 52) on the model selection interface 50. Figure 5 The computing device (or terminal device) then receives the user's operation on the model selection interface 50 for the initial model option to be selected (such as a mouse click, a touch click, etc.), so that the computing device (or terminal device) displays the initial model option selected by the user in the option box 53 of the model selection interface 50. Then, the computing device (or terminal device) receives the user's operation on the "OK" control on the model selection interface 50 (such as a mouse click, a touch click, etc.). In this way, the computing device (or terminal device) receives the second selection operation input by the user.

[0114] Optionally, the embodiment of the present application does not limit the execution order of S101 and S102. For example, S101 may be executed first, and then S102. Alternatively, S101 and S102 may be executed simultaneously. Alternatively, S102 may be executed first, and then S101, and so on.

[0115] S103. Iteratively train the target initial model according to the business data samples and heterogeneous data samples in the training samples to obtain the target model.

[0116] Among them, the target model can be used to execute the above-mentioned target tasks to meet the user's business needs.

[0117] For the sake of simple description, the embodiment of the present application takes the target attribute of the target object in the training sample or the sample to be tested including the position of the target object and the category of the target object as an example to introduce the specific training process of the data training method provided by the embodiment of the present application. As can be seen from S101, when the target attribute of the target object includes the position of the target object and the category of the target object, the target attribute label of the target object in the business data sample includes the position label of the target object and the category label of the target object. Among them, the position label is a position box representation used to characterize the position of the target object in the business data sample, and the category label is a category value used to characterize the specific category of the target object in the business data sample.

[0118] Specifically, the computing device may first receive at least one edge detection threshold input by the user before “iteratively training the target initial model according to the business data samples and heterogeneous data samples in the training samples”. Thus, the computing device may iteratively train the target initial model according to the business data samples, heterogeneous data samples and the at least one edge detection threshold in the training samples to obtain the target model. The edge detection threshold is used to detect edges in heterogeneous data samples. Usually, the edge detection threshold is a gradient value of a pixel.

[0119] Of course, the at least one edge detection threshold mentioned above may also be preset in the computing device, which is not limited to this.

[0120] In a possible implementation, the computing device directly uses its own input and output interfaces (such as Figure 1 The input / output interface 105 shown in the figure receives at least one edge detection threshold value input by the user. Specifically, the computing device can receive at least one edge detection threshold value input by the user on the parameter configuration interface through its own input / output interface.

[0121] In another possible implementation, the computing device is connected to the terminal device (such as Figure 2 The terminal device 22 shown in the figure receives at least one edge detection threshold value input by the user through its input / output interface. Specifically, the terminal device receives at least one edge detection threshold value input by the user on the operation interface of the client application through its own input / output interface, and sends the at least one edge detection threshold value to the computing device. In response, the computing device receives at least one edge detection threshold value input by the user.

[0122] As an example, refer to Figure 6 , Figure 6 FIG. 2 shows a schematic diagram of receiving at least one edge detection threshold provided by an embodiment of the present application. Figure 6 As shown, the edge detection threshold configuration interface 60 can be a parameter configuration interface of a visual operation interface provided by a computing device when training a network model, or an operation interface of a client application installed on a terminal device, without limitation.

[0123] like Figure 6As shown in (a) in FIG. 1 , the computing device (or terminal device) can directly receive at least one edge detection threshold value input by the user in the option box 61 of the edge detection threshold configuration interface 60 through the input and output interface. When there are multiple edge detection threshold values, the multiple edge detection threshold values ​​can be separated by punctuation marks (such as commas, semicolons, etc.) or spaces. Then, the computing device (or terminal device) receives the user's operation on the "OK" control on the edge detection threshold configuration interface 60 (such as a mouse click, a touch click, etc.). In this way, the computing device (or terminal device) receives at least one edge detection threshold value input by the user.

[0124] like Figure 6 As shown in (b) of FIG. 1 , the edge detection threshold configuration interface 60 of the computing device (or terminal device) includes a control showing multiple edge detection thresholds, such as Figure 6 As shown in (b) in the figure, there is a control showing an edge detection threshold value of "50", a control showing an edge detection threshold value of "60", a control showing an edge detection threshold value of "70", and a control showing an edge detection threshold value of "80". In this way, the computing device (or terminal device) receives the user's selection operation on at least one control showing an edge detection threshold value based on the user's own needs on the edge detection threshold configuration interface 60 (such as a selection operation by clicking the control with a mouse or touching the control). Then, the computing device (or terminal device) receives the user's operation on the "OK" control on the edge detection threshold configuration interface 60 (such as a mouse click, touch click, etc.). In this way, the computing device (or terminal device) receives at least one edge detection threshold value input by the user.

[0125] It should be understood that when the computing device detects multiple edges in the heterogeneous data sample according to the above-mentioned at least one edge detection threshold, the computing device can determine the calculation label of the target object in the heterogeneous data sample according to the detected edges (refer to the description below for the detailed process). In this way, the above-mentioned "iteratively training the target initial model according to the business data samples in the training samples, the heterogeneous data samples and the at least one edge detection threshold to obtain the target model" is specifically implemented as follows: the computing device iteratively trains the target initial model according to the business data samples in the training samples, the target attribute labels of the target objects in the business data samples, the heterogeneous data samples, and the calculation labels determined based on the above-mentioned at least one edge detection threshold to obtain the target model.

[0126] Among them, the calculated label is a label calculated for the target attribute of the target object in the heterogeneous data sample. The calculated label includes a calculated position box and a calculated category of the target object in the heterogeneous data sample. Among them, the calculated position box is used to characterize the position of the target object calculated in the heterogeneous data sample, and the calculated category is the category corresponding to the calculated position box. Specifically, the calculated category is obtained by predicting the heterogeneous data sample based on the target initial model to be trained. Here, the specific process of the computing device determining the calculated label can be referred to the description below and will not be repeated.

[0127] As an example, for a first training sample including a first business data sample and a first heterogeneous data sample in the training sample, the computing device first determines the calculation label of the target object in the first heterogeneous data sample according to the target initial model to be trained and the at least one edge detection threshold. Then, the computing device trains the target initial model to be trained according to the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculation label of the target object in the first heterogeneous data sample to obtain a target model.

[0128] The first training sample may also be referred to as a group of training samples including a first business data sample and a first heterogeneous data sample.

[0129] The target initial model to be trained can also be referred to as the target initial model to be trained this time. The target initial model to be trained this time refers to the target initial model that receives the training sample input this time. For example, the target initial model to be trained for the first time refers to the target initial model that receives the training sample input for the first time. After the first training process is completed, the target initial model trained once can be obtained, and the target initial model trained once is used as the target initial model to be trained for the second time, and the sample input is received again (i.e., the second time), and after the second training process is completed, the target initial model trained twice is obtained.

[0130] It should be understood that the target model obtained after the computing device iteratively trains the target initial model according to the training samples is the converged target initial model.

[0131] Optionally, the computing device may use the number of iterative training times greater than or equal to a first threshold as a convergence condition to determine whether the target initial model has converged. For example, when the number of iterative training times is greater than the first threshold, the computing device determines that the target initial model has converged.

[0132] Optionally, the computing device may use the loss function value less than or equal to the second threshold as a convergence condition to determine whether the target initial model has converged. For example, assuming that the loss function value obtained after the current training of the target initial model is less than the second threshold, the computing device determines that the target initial model has converged. At this time, the computing device no longer needs to update the target initial model based on the loss function obtained from the current training. Among them, the embodiments of the present application do not specifically limit the specific values ​​of the first threshold and the second threshold.

[0133] The following uses a first training sample including a first business data sample and a first heterogeneous data sample as an example to illustrate a process in which a computing device trains a target initial model to be trained according to the training sample. Figure 7 , Figure 7 A schematic diagram of a process of training a target initial model to be trained according to a first training sample provided by an embodiment of the present application is shown. Figure 7 As shown, the process includes the following steps.

[0134] S201. Determine an edge in a first heterogeneous data sample according to the at least one edge detection threshold and edge detection algorithm.

[0135] Specifically, the computing device may first detect the initial edge in the first heterogeneous data sample under each edge detection threshold in the at least one edge detection threshold according to the at least one edge detection threshold and the edge detection algorithm. The edge detection algorithm is used to detect edge features in an image. As an example, the edge detection algorithm is a Canny algorithm, but is not limited thereto.

[0136] Optionally, the edge detection algorithm may be any edge detection algorithm preset in the computing device.

[0137] Optionally, the edge detection algorithm may be any edge detection algorithm acquired by the computing device from other devices, wherein the other devices may be algorithm servers, storage devices, etc., but are not limited thereto.

[0138] It should be understood that in practice, when using an edge detection algorithm to detect edges in an image, it is necessary to determine the edge features in the image based on an edge detection threshold. Among them, the edge detection threshold is usually the gradient value of the pixel. When the gradient value of a pixel in the image is greater than the edge detection threshold, the edge detection algorithm determines the pixel as an edge pixel point. Then, in the direction perpendicular to the pixel gradient direction, edge pixels whose distance difference between adjacent edge pixels is less than the threshold can constitute an edge. It should be understood that when the edge detection threshold is larger, it means that the pixel difference between the pixel and the adjacent pixels in the gradient direction of the pixel is larger. Therefore, the larger the edge detection threshold, the fewer the number of edges detected in the image based on the edge detection algorithm. Conversely, the smaller the edge detection threshold, the more the number of edges detected in the image based on the edge detection algorithm.

[0139] Therefore, the computing device detects and obtains the initial edge in the first heterogeneous data sample under each edge detection threshold in the multiple edge detection thresholds according to the edge detection algorithm and multiple edge detection thresholds, which can also be understood as the computing device obtaining the initial edge in the first heterogeneous data sample under different edge detection threshold granularities. In this way, it can be ensured that the computing device comprehensively detects all initial edges in the first heterogeneous data sample.

[0140] In the embodiment of the present application, the process of detecting and obtaining the initial edge in the first heterogeneous data sample under the edge detection threshold according to the edge detection algorithm and the edge detection threshold by the computing device is not described in detail.

[0141] Next, the computing device optimizes the initial edge detected in the first heterogeneous data sample to obtain the edge in the first heterogeneous data sample.

[0142] Optionally, the computing device can optimize (or fuse) the initial edge detected in the first heterogeneous data sample according to the non-maximum suppression algorithm, thereby retaining the edge pixel points in the initial edge whose gradient value along the gradient direction is the maximum value, and removing the edge pixel points in the initial edge whose gradient value along the gradient direction is not the maximum value, thereby achieving the purpose of eliminating redundant edge pixels in the initial edge, and further optimizing the initial edge. The detailed description of the non-maximum suppression algorithm is not described in detail in the embodiment of the present application.

[0143] It is understandable that the embodiment of the present application does not limit the execution order of S201 and S202. For example, S201 may be executed first, and then S202. Alternatively, S201 and S202 may be executed at the same time. Alternatively, S202 may be executed first, and then S201.

[0144] S202: Determine a predicted location box of a target object and a predicted category of the target object in a first heterogeneous data sample according to the target initial model to be trained.

[0145] The predicted position frame of the target object in the first heterogeneous data sample is used to represent: the position of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained. The predicted category of the target object in the first heterogeneous data sample is: the category of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained. For ease of description, the predicted position frame of the target object in the heterogeneous data sample and the predicted category of the target object are referred to as the predicted position frame and the predicted category having a corresponding relationship.

[0146] Specifically, the computing device may input the first heterogeneous data sample into the target initial model to be trained, and the target initial model may output the predicted location box of the target object and the predicted category of the target object in the first heterogeneous data sample.

[0147] S203: Determine a calculation position frame of the target object in the first heterogeneous data sample according to the positional relationship between the edge in the first heterogeneous data sample and the above-mentioned predicted position frame, and determine a calculation category of the target object in the first heterogeneous data sample according to the above-mentioned predicted category.

[0148] Among them, the edge in the first heterogeneous data sample is the edge determined in the first heterogeneous data sample after the computing device executes S201. The predicted position box and the predicted category are the predicted position box and the predicted category obtained after the computing device executes S202. The calculated position box and the calculated category of the target object in the first heterogeneous data sample are the calculated labels of the target object in the first heterogeneous data sample. The detailed description of the calculated labels of the target object in the first heterogeneous data sample can be referred to the above description, which will not be repeated here.

[0149] Specifically, the process of determining the calculation label of the first heterogeneous data sample according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position box by the computing device can be implemented through S2031 to S2302.

[0150] S2031. Determine a positive edge in the first heterogeneous data according to a positional relationship between an edge in the first heterogeneous data sample and the predicted position box.

[0151] The positive edge in the first heterogeneous data sample includes an edge in the first heterogeneous data sample that is located within the prediction position frame. In this case, the computing device can determine the edge in the first heterogeneous data sample that is located within the prediction position frame as the positive edge of the first heterogeneous data sample based on the positional relationship between the prediction position frame and the edge in the first heterogeneous data sample. The positive edge of the first heterogeneous data sample is also called an edge in the first heterogeneous data sample that is related to the target object, so this process can also be called a process of analyzing edge correlation.

[0152] As an example, refer to Figure 8 , Figure 8 FIG. 1 is a schematic diagram showing a positive edge in a first heterogeneous data sample provided by an embodiment of the present application. Figure 8 As shown, the predicted position frame of the target object in the heterogeneous data sample 80 determined by the computing device in step S201 is a rectangular frame 81. The computing device determines four edges in the heterogeneous data sample 80 in step S202, namely edge 82, edge 83, edge 84 and edge 85. Then, by comparing the positional relationship between the rectangular frame 81 and the four edges, the computing device determines the edge 82 located in the rectangular frame 81 and the portion of the edge 83 located inside the rectangular frame 81 as positive edges in the heterogeneous data sample 80.

[0153] S2032: Merge the positive edges in the first heterogeneous data sample to obtain a calculated position frame of the target object in the first heterogeneous data sample, and determine the calculated category of the target object in the first heterogeneous data sample based on the predicted category.

[0154] Optionally, the computing device may fuse the positive edges in the first heterogeneous data sample using a minimum bounding box fusion algorithm to obtain a calculated position frame of the target object in the first heterogeneous data sample, where the calculated position frame is the minimum bounding box of the positive edges in the first heterogeneous data sample.

[0155] Among them, the minimum bounding box fusion algorithm is specifically: a rectangular frame enclosed by the maximum horizontal coordinate, minimum horizontal coordinate, maximum vertical coordinate and minimum vertical coordinate in a given two-dimensional shape (the positive edge in the embodiment of the present application) is determined as the minimum bounding box of the two-dimensional shape.

[0156] In addition, the computing device also determines the category of the positive edge in the first heterogeneous data sample located in the prediction position frame as the prediction category corresponding to the prediction position frame (i.e., the prediction category determined in S202). Furthermore, the computing device determines the calculation category corresponding to the calculation position frame obtained by positive edge fusion in the first heterogeneous data sample as the prediction category.

[0157] In this way, the computing device determines the computing position box and computing category of the target object in the first heterogeneous data sample through S201 to S203, that is, determines the computing label of the target object in the first heterogeneous data sample.

[0158] S204: Train the target initial model to be trained according to the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculated label of the target object in the first heterogeneous data sample.

[0159] For the sake of simple description, the "target initial model to be trained" is referred to as the first model below. Specifically, the computing device can train the first model according to the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculation label of the target object in the first heterogeneous data sample through steps S2041 to S2043.

[0160] S2041. Input the first business data sample and the first heterogeneous data sample into the first model to obtain a prediction result.

[0161] Specifically, after the computing device inputs the first business data sample and the first heterogeneous data sample into the first model, the first model outputs the prediction results of the first business data sample and the first heterogeneous data sample. The prediction results include a first prediction value of the first business data sample and a second prediction value of the first heterogeneous data sample.

[0162] The first prediction value includes a first prediction position box and a first prediction category of the target object in the first business data sample. The first prediction position box is used to characterize the position of the target object in the first business data sample predicted by the first model, and the first prediction category is the category of the target object in the first business data sample predicted by the first model.

[0163] The second prediction value includes a second predicted position frame and a second predicted category of the target object in the first heterogeneous data sample. The second predicted position frame is used to characterize the position of the target object in the first heterogeneous data sample predicted by the first model, and the second predicted category is the category of the target object in the first heterogeneous data sample predicted by the first model.

[0164] S2042. Determine a position loss function and a category loss function according to the above prediction results, the target attribute label of the target object in the first business data sample, and the calculated label of the target object in the first heterogeneous data sample.

[0165] The position loss function includes the position loss function of the first service data sample and the position loss function of the first heterogeneous data sample. The category loss function includes the category loss function of the first service data sample and the category loss function of the first heterogeneous data sample.

[0166] Specifically, the computing device determines the position loss function of the first business data sample based on the offset between the first predicted position box in the above prediction result and the position box in the target attribute label as the position label. And, the computing device determines the category loss function of the first business data sample based on the first predicted category in the above prediction result and the category value in the target attribute label as the category label. In some examples, the computing device uses a cross entropy loss function to calculate the first predicted category and the category value in the target attribute label as the category label, thereby obtaining the category loss function of the first business data sample. For ease of distinction, the position loss function of the first business data sample is referred to as the first loss function, and the category loss function of the first business data sample is referred to as the second loss function.

[0167] The computing device also determines the position loss function of the first heterogeneous data sample based on the offset between the second predicted position box in the above prediction result and the calculated position box in the calculated label. And, the computing device determines the category loss function of the first heterogeneous data sample based on the second predicted category in the above prediction result and the calculated category in the calculated label. In some examples, the computing device uses a cross entropy loss function to calculate the second predicted category and the calculated category in the calculated label to obtain the category loss function of the first heterogeneous data sample. For ease of distinction, the position loss function of the first heterogeneous data sample is called the third loss function, and the category loss function of the first heterogeneous data sample is called the fourth loss function.

[0168] S2043. Update the first model according to the position loss function and the category loss function.

[0169] The computing device sums the position loss function and the category loss function determined above to obtain an overall loss function after the first model predicts the first business data sample and the first business data sample.

[0170] Then, the computing device adjusts the network parameters of the first model according to the determined overall loss function to update the first model, thereby obtaining the second model. It should be understood that the second model is the model obtained after the first model is trained once with the first business data sample and the first heterogeneous data sample.

[0171] Exemplarily, the computing device may update the first model according to the overall loss function and the back propagation algorithm to obtain the second model.

[0172] At this point, through the method described in S201 to S204, it is possible to implement a training of the first model through the first business data sample and the first heterogeneous data sample, thereby obtaining a trained second model. Furthermore, the computing device uses the second model as the target initial model to be trained, and executes the above methods of S201 to S204 again for another group of training samples (such as the second training sample) to implement training of the second model according to the second training sample, thereby obtaining a third model. Among them, the second training sample is a group of samples including a heterogeneous data sample and a business data sample in the training sample, and the second training sample and the first training sample can be the same or different, and there is no limitation on this. Similarly, the target initial model is iteratively trained by the training sample obtained in S101. When the target initial model converges, the target model that meets the user's business needs can be obtained.

[0173] Based on the data training method described above, the embodiment of the present application uses business data samples and a large number of heterogeneous data samples to jointly train the target initial model, thereby migrating the knowledge in the large number of heterogeneous data samples to the converged target initial model that meets the user's business needs, thereby obtaining a target model with high prediction accuracy and high prediction efficiency. And because the embodiment of the present application uses a large number of heterogeneous data samples with very low collection costs when training the model, the training cost is low when training the model using the method of the present application.

[0174] In other embodiments, in order to improve the prediction accuracy of the target model obtained by the above training, training objects in the heterogeneous data samples can be added in the process of training the target initial model. That is, in addition to training the target initial model according to the calculated labels of the target objects in the heterogeneous data samples, the target initial model is also trained according to the background labels of the background in the heterogeneous data samples.

[0175] In this embodiment, the computing device may first receive the positive and negative edge ratio input by the user before training the target initial model. The positive and negative edge ratio is used to balance the number of positive edges and negative edges in the heterogeneous data sample. The positive edge in the heterogeneous data sample is the positive edge described above, and the positive edge is used to determine the calculation label of the target object in the heterogeneous data sample. The negative edge in the heterogeneous data sample is used to determine the background label of the heterogeneous data sample, and the background label includes the background frame and background category in the heterogeneous data sample. Among them, the detailed description of the negative edge, background frame and background category can be referred to the description below, which will not be repeated here.

[0176] In this way, the "iterative training of the target initial model according to the business data samples and heterogeneous data samples in the training samples" in S103 above is specifically implemented as: iteratively training the target initial model according to the business data samples, heterogeneous data samples, the calculated labels of the heterogeneous data samples and the background labels of the heterogeneous data samples to obtain the target model.

[0177] In a possible implementation, the computing device directly uses its own input and output interfaces (such as Figure 1 The input and output interface 105 shown in the figure receives the positive and negative edge ratio input by the user. Specifically, the computing device can receive the positive and negative edge ratio input by the user on the positive and negative edge ratio configuration interface through its own input and output interface.

[0178] In another possible implementation, the computing device is connected to the terminal device (such as Figure 2 The input and output interface of the terminal device 22 shown in the figure receives the positive and negative edge ratio input by the user. Specifically, the terminal device receives the positive and negative edge ratio input by the user on the operation interface of the client application through its own input and output interface, and sends the positive and negative edge ratio to the computing device. In response, the computing device receives the positive and negative edge ratio input by the user.

[0179] As an example, refer to Fig. 9 , Fig. 9 FIG. 1 shows a schematic diagram of receiving a positive and negative edge ratio provided by an embodiment of the present application. Fig. 9 As shown, the positive and negative edge ratio configuration interface 90 can be the positive and negative edge ratio configuration interface in the visual operation interface provided by the computing device when training the network model, or the operation interface of the client application installed on the terminal device, which is not limited to this.

[0180] like Fig. 9 As shown in (a) in FIG. 1 , the computing device (or terminal device) can directly receive the positive and negative edge ratio input by the user in the option box 91 of the positive and negative edge ratio configuration interface 90 through the input and output interface. Then, the computing device (or terminal device) receives the user's operation (such as a mouse click, a touch click, etc.) on the "OK" control on the positive and negative edge ratio configuration interface 90. In this way, the computing device (or terminal device) receives the positive and negative edge ratio input by the user.

[0181] like Fig. 9 As shown in (b) of FIG. 1 , the positive and negative edge ratio configuration interface 90 of the computing device (or terminal device) includes a control showing multiple positive and negative edge ratios, such as Fig. 9As shown in (b) in the figure, there is a control showing a ratio of "1", a control showing a ratio of "1.5", a control showing a ratio of "2", and a control showing a ratio of "2.5". In this way, the computing device (or terminal device) receives the user's selection operation on the positive and negative edge ratio configuration interface 90 based on the user's own needs for the control showing the expected value of the positive and negative edge ratio (such as a selection operation by clicking the control with a mouse or touching the control). Then, the computing device (or terminal device) receives the user's operation on the "OK" control on the positive and negative edge ratio configuration interface 90 (such as a mouse click, touch click, etc.). In this way, the computing device (or terminal device) receives the positive and negative edge ratios input by the user.

[0182] Specifically, refer to Fig.10 , the data training method provided in the embodiment of the present application executes step S201 to S202 described above, in addition to executing step S203, also executes step S301. Among them, the embodiment of the present application does not limit the execution sequence of S203 and S301. For example, the embodiment of the present application may execute S203 first, and then execute S301. Alternatively, the embodiment of the present application may execute S301 first, and then execute S203. Alternatively, the embodiment of the present application may execute S203 and S301 at the same time.

[0183] S301. Determine a background frame of a first heterogeneous data sample according to a positional relationship between an edge in a first heterogeneous data sample and a predicted position frame, wherein a category corresponding to the background frame is a background category.

[0184] The edge in the first heterogeneous data sample is an edge determined in the first heterogeneous data sample after the computing device executes S201. The predicted position frame is a predicted position frame obtained after the computing device executes S202.

[0185] The background frame and background category of the first heterogeneous data sample are the background label of the first heterogeneous data sample.

[0186] Specifically, the process of the computing device determining the background frame of the first heterogeneous data sample according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame can be implemented through S3011 to S3012.

[0187] S3011. Determine a negative edge in the first heterogeneous data according to a positional relationship between an edge in the first heterogeneous data sample and a predicted position box.

[0188] Among them, the negative edge in the first heterogeneous data sample includes the edge in the first heterogeneous data sample that is located outside the prediction position box. In this case, the computing device can determine the edge in the first heterogeneous data sample that is located outside the prediction position box as the negative edge of the first heterogeneous data sample based on the positional relationship between the prediction position box and the edge in the first heterogeneous data sample. The negative edge of the first heterogeneous data sample is also called the edge in the first heterogeneous data sample that is not related to the target object, so this process is also called the root edge correlation process.

[0189] As an example, Figure 8 As shown, when the predicted position frame of the target object in the heterogeneous data sample 80 determined by the computing device in step S201 is a rectangular frame 81, and the computing device determines four edges in the heterogeneous data sample 80 in step S202, namely edge 82, edge 83, edge 84, and edge 85. The computing device determines edge 84 and edge 85 located outside the rectangular frame 81, and the portion of edge 83 located outside the rectangular frame 81 as negative edges in the heterogeneous data sample 80 by comparing the positional relationship between the rectangular frame 81 and the four edges.

[0190] In practice, when the computing device determines the computational label of the target object in the first heterogeneous data sample, it needs to use the predicted position frame obtained after the target initial model predicts the first heterogeneous data sample, and in the initial stage of training the target initial model, the predicted position frame of the target object in the first heterogeneous data sample predicted by the computing device based on the target initial model is not very accurate, so based on the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, the positive edges determined are relatively few, while the negative edges are relatively many. For this reason, the computing device can balance the ratio of the number of positive edges to the number of negative edges.

[0191] Specifically, the computing device may sample the negative edge of the first heterogeneous data sample based on the positive-negative edge ratio so that the number of positive edges and the number of negative edges in the first heterogeneous data sample satisfy the positive-negative edge ratio. The positive-negative edge ratio is a preset value, and the positive-negative edge ratio is an expected value of the ratio of the number of positive edges to the number of negative edges in the first heterogeneous data sample, and the expected value may be determined based on actual experience, for example, and is not limited to this.

[0192] As an example, when the positive-to-negative edge ratio is 1, and the number of positive edges in heterogeneous data sample 1 is 5 and the number of negative edges is 9, the computing device may sample 5 negative edges among the negative edges in heterogeneous data sample 1 so that the ratio of the number of positive edges to the number of negative edges in heterogeneous data sample 1 is 1, that is, the ratio satisfies the positive-to-negative edge ratio of 1.

[0193] Optionally, when the computing device samples the negative edge of the first heterogeneous data sample based on the positive-negative edge ratio, the sampling may be random or based on a preset strategy, and there is no limitation on this.

[0194] Optionally, the above-mentioned positive and negative edge ratios may be preset in the computing device.

[0195] Optionally, the computing device may receive the positive and negative edge ratios input by the user in advance (as described above).

[0196] S3012: Fuse the negative edges in the first heterogeneous data sample to obtain a background frame of the first heterogeneous data sample, and determine a category corresponding to the background frame.

[0197] Optionally, the computing device may fuse the negative edges in the first heterogeneous data sample using a minimum bounding box fusion algorithm to obtain a background frame of the first heterogeneous data sample, where the background frame is the minimum bounding box of the negative edges in the first heterogeneous data sample. The specific description of the minimum bounding box fusion algorithm can be referred to the above description and will not be repeated here.

[0198] It is understandable that when the computing device performs a sampling action on the negative edge in the first heterogeneous data sample based on the positive-negative edge ratio in S3011, the negative edge sampled by the computing device can be called a target negative edge. In this case, the computing device fuses the target negative edge in the first heterogeneous data sample using a minimum bounding box fusion algorithm to obtain a background frame of the first heterogeneous data sample.

[0199] In addition, the computing device further determines the category of the negative edge outside the prediction position frame in the first heterogeneous data sample as the background category. Furthermore, the computing device determines the category corresponding to the background frame obtained by negative edge fusion as the background category.

[0200] Thus, the background box and the background category corresponding to the background box determined by the computing device in the first heterogeneous data sample can participate in the training of the target initial model as another calculation label of the first heterogeneous data sample, which can improve the prediction accuracy of the target model.

[0201] Furthermore, after executing S203 and S301, the computing device executes S302 to replace S204 described above.

[0202] S302. Train the target initial model to be trained according to the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, the calculated label of the target object in the first heterogeneous data sample, and the background frame and background category in the first heterogeneous data sample.

[0203] The “target initial model to be trained” is referred to as the first model. The background frame and background category in the first heterogeneous data sample can also be regarded as another calculation label of the first heterogeneous data sample.

[0204] Specifically, S302 can be implemented through S3021 to S3023.

[0205] S3021. Input the first business data sample and the first heterogeneous data sample into the first model to obtain a prediction result.

[0206] Specifically, the computing device inputs the first business data sample and the first heterogeneous data sample into the first model, and the first model outputs the prediction results of the first business data sample and the first heterogeneous data sample. The prediction results include the first prediction value of the first business data sample and the second prediction value and the third prediction value of the first heterogeneous data sample.

[0207] Among them, the detailed description of the first prediction value and the second prediction value can refer to the description in S2041, which will not be repeated here. The third prediction value includes a predicted background frame and a predicted background category of the first heterogeneous data sample. Here, the predicted background frame is used to characterize the position of the background of the target object in the first heterogeneous data sample predicted by the first model, and the predicted background category is the category of the background of the target object in the first heterogeneous data sample predicted by the first model.

[0208] S3022. Determine a position loss function and a category loss function based on the above prediction results, the target attribute label of the target object in the first business data sample, the calculated label of the target object in the first heterogeneous data sample, and the background frame and background category of the first heterogeneous data sample.

[0209] Among them, the detailed description of the position loss function can be referred to the description of S2042 and will not be repeated here.

[0210] Specifically, the position loss function is obtained by summing the first function, the third function and the fifth function, and the category loss function is obtained by summing the second function, the fourth function and the sixth function. The detailed description of the first function, the second function, the third function and the fourth function can refer to the relevant description in S2042, which will not be repeated here.

[0211] Among them, the fifth function is: the computing device calculates the background position loss function of the first heterogeneous data sample according to the offset between the predicted background box in the above prediction result and the background box of the first heterogeneous data sample (i.e., the background box determined in S3012). The sixth function is: the computing device calculates the background category loss function of the first heterogeneous data sample according to the predicted background category in the above prediction result and the background category of the first heterogeneous data sample (i.e., the background category determined in S3012). In some examples, the computing device uses a cross entropy loss function to calculate the predicted background category and the background category in the first heterogeneous data sample, thereby obtaining the background category loss function of the first heterogeneous data sample.

[0212] S3023. Update the first model according to the position loss function and the category loss function.

[0213] Among them, the detailed description of S3023 can refer to the description of S2043 and will not be repeated here.

[0214] In this way, by adding a prediction object (or training object): background to the heterogeneous data sample during the training of the target model, the prediction accuracy of the target model can be improved.

[0215] In some other embodiments, reference Fig.11 , Fig.11 Another flow chart of the data training method provided in the embodiment of the present application is shown. Optionally, the method can be applied to Figure 1 The computing device shown or Figure 2 The system is trained with the data shown. Figure 1 The computing device execution of the hardware result shown is as an example, Fig.11 As shown, the method includes the following steps.

[0216] S401: Obtain training samples.

[0217] The description of the computing device acquiring the training samples may refer to the description of S101 and will not be repeated here.

[0218] As an example, the training sample includes 5,000 business data labeled with target attribute tags and 118,000 heterogeneous data. In one example, the heterogeneous data is a public data set Coco. When the data training device in the data training system is deployed on an intelligent hardware platform, the training sample is pre-stored in the intelligent hardware platform.

[0219] S402: Obtain parameters.

[0220] The parameters include the target task, target initial model, edge detection threshold, positive and negative edge ratio, etc. mentioned above.

[0221] The specific process of obtaining parameters of the computing device can refer to the detailed description of obtaining the target task, the target initial model, the edge detection threshold, and the positive and negative edge ratio in the above text, which will not be repeated here.

[0222] S403: Perform edge detection on a heterogeneous data sample in a group of training samples to obtain edges in the heterogeneous data sample.

[0223] It should be understood that in the process of the computing device iteratively training the target initial model according to the training samples, the computing device can schedule a heterogeneous data sample from all heterogeneous data samples and schedule a business data sample from all business data samples, so that the computing device uses the heterogeneous data sample and the business data sample as a group of training samples (such as the first training sample, the second training sample, etc.) to train the target initial model once. When each heterogeneous data sample in the training sample participates in the training of the target initial model once, the computing device is said to have completed a round of iterative training based on the training sample.

[0224] Since the number of business data samples in the training samples is usually much smaller than the number of heterogeneous data samples, during a round of iterative training, for any two groups of training samples (such as the first training sample and the second training sample), the heterogeneous data samples in the first training sample and the heterogeneous data samples in the second training sample are different, while the business data samples in the first training sample and the business data samples in the second training sample may be the same or different.

[0225] The computing device performs edge detection on the heterogeneous data sample to obtain a detailed description of the edge in the heterogeneous data sample. Please refer to the detailed description of S201 above, which will not be repeated here.

[0226] S404: According to the target initial model to be trained, determine the positive edges and negative edges among the edges detected by the heterogeneous data samples, and balance the ratio of the number of positive edges and the number of negative edges.

[0227] Among them, for detailed instructions on how the computing device determines positive edges among edges detected by heterogeneous data samples based on the target initial model to be trained, please refer to the descriptions of S202 and S2031. For detailed instructions on how the computing device determines negative edges among edges detected by heterogeneous data samples based on the target initial model to be trained, and balancing the ratio of the number of positive edges to negative edges, please refer to the descriptions of S202 and S3011, which will not be repeated here.

[0228] S405 : Determine a first calculation label based on the positive edge and the category of the positive edge, and determine a second calculation label based on the negative edge and the category of the negative edge.

[0229] The first calculation label is the calculation label mentioned above, including the calculation position frame and calculation category of the target object in the heterogeneous data sample. The second calculation label includes the background label mentioned above, including the background frame and background category of the background in the heterogeneous data sample.

[0230] For a detailed description of how the computing device determines the first computing label based on the positive edge and the category of the positive edge, please refer to S203 for the description of determining the computing position box and the computing category of the target object in the first heterogeneous data sample. For a detailed description of how the computing device determines the second computing label based on the negative edge and the category of the negative edge, please refer to S301 for the description of determining the background box and the background category of the first heterogeneous data sample, which will not be repeated here.

[0231] S406. Train the target initial model to be trained according to the heterogeneous data samples, the first calculation labels and the second calculation labels of the heterogeneous data samples, the business data samples, and the target attribute labels of the business data samples to update the target initial model to be trained.

[0232] Among them, the computing device trains the target initial model to be trained according to the heterogeneous data samples, the first computing labels and the second computing labels of the heterogeneous data samples, the business data samples, and the target attribute labels of the business data samples to update the description of the target initial model to be trained. Please refer to the description of S302 and will not repeat it here.

[0233] S407: Determine whether the current training meets the convergence condition.

[0234] For a detailed description of the convergence condition, please refer to the relevant description in S103 and will not be repeated here.

[0235] When the computing device determines that the current training meets the convergence condition, the target model is output.

[0236] When the computing device determines that the current training does not meet the convergence condition, in one possible case, if the next training still belongs to the first round of iterative training of the target initial model, the computing device executes S403 to S407 for the next set of training samples. In another possible case, if the next training belongs to the second and subsequent rounds of iterative training of the target initial model, the computing device executes S404 to S407 for the next set of training samples.

[0237] This is because: during the first round of iterative training, the computing device has executed S403 for each heterogeneous data sample in the training sample and determined the edge in each heterogeneous data sample. Therefore, in the second and subsequent rounds of iterative training, the computing device does not need to execute S403 again. The computing device only needs to call the edge determined for each heterogeneous data sample during the first round of iterative training, and perform correlation analysis on the edge in the heterogeneous data sample according to the latest updated target initial model (the specific process can be referred to S203) to determine the positive edge and the negative edge.

[0238] In this way, based on the method described in S401 to S407, it is possible to train the target initial model based on business data and a large amount of heterogeneous data at the same time to obtain a target model with high prediction accuracy and high prediction efficiency, and the process is simple.

[0239] In order to deepen the understanding of the method provided in the embodiments of the present application, the method provided in the embodiments of the present application is further explained with examples below.

[0240] refer to Fig.12 , Fig.12 Another flow chart of the data training method provided in the embodiment of the present application is shown. In this example, a downstream task is configured with an initial model as an example for description.

[0241] like Fig.12 As shown in the figure, the intelligent hardware platform is pre-installed with business data labeled with target object attribute labels and unlabeled heterogeneous data. The intelligent hardware platform is also pre-installed with initial models pre-trained for three downstream tasks, namely: the initial model for the action recognition task, the initial model for the smoke detection task, and the initial model for the license plate detection task.

[0242] like Fig.12 As shown, for a user terminal communicating with the intelligent hardware platform, the terminal is installed with a client application for training the network model. The operation interface of the client application can also be called Fig.12 The remote control interface for network model training is shown.

[0243] In a possible implementation, on the remote control interface displayed by the user terminal, the user terminal receives an operation of selecting a parameter template input by the user to receive a target parameter template input by the user. It should be understood that the target parameter template includes a set of pre-configured parameters, which include the edge detection threshold, positive and negative edge ratio, target task, etc. mentioned above, but are not limited thereto. Since one downstream task corresponds to one initial model in this example, the terminal only needs to receive the target task input by the user.

[0244] In another possible implementation, on the remote control interface displayed by the user terminal, the user terminal receives multiple parameter input operations input by the user to receive multiple parameters input by the user. It should be understood that each of the multiple parameter input operations is used to input a parameter. For example, one of the edge detection threshold, the positive and negative edge ratio, and the target task is input through one input operation.

[0245] Furthermore, the user terminal sends corresponding parameter configurations to the data training device in the intelligent hardware platform in response to the user's operation on the remote control interface, such as edge detection thresholds, positive and negative edge ratios, and identification of target tasks, etc.

[0246] Next, the data training device in the intelligent hardware platform first determines the target initial model corresponding to the target task according to the identifier of the target task received from the user terminal. Then the data training device receives the edge detection threshold and the positive and negative edge ratio from the user terminal, the determined target initial model, and the preset training samples including business data and heterogeneous data and inputs them into the training module, so that the training module trains the target initial model according to the edge detection threshold, the positive and negative edge ratio, and the training samples including business data and heterogeneous data (corresponding to S103 described above), and finally obtains a converged target model.

[0247] refer to Fig.13 , Fig.13 Another flow chart of the data training method provided in the embodiment of the present application is shown.

[0248] like Fig.13 As shown, for any group of training samples including heterogeneous data samples and business data samples, in step 1, the data training device running in the intelligent hardware platform determines the edges of the heterogeneous data samples under different edge detection thresholds (i.e., at different granularities) through an edge detection algorithm and multiple edge detection thresholds (refer to S201 described above for the specific process).

[0249] In step 2, the data training device performs correlation analysis on the edges determined in step 1 based on the target initial model to determine positive edges and negative edges. The data training device also balances the number of positive edges and negative edges (for specific processes, refer to S202, S2031 and S3011).

[0250] In step 3, the data training device determines the calculation position frame according to the positive edge of the heterogeneous data sample, and determines the background frame according to the negative edge of the heterogeneous data sample, and determines the calculation position frame and the calculation category corresponding to the calculation position frame as the calculation label 1 of the heterogeneous data sample, and determines the background frame and the background category corresponding to the background frame as the calculation label 2 of the heterogeneous data sample (the specific process can be referred to S203 and S301).

[0251] In step 4, the data training device trains the target initial model to be trained through the heterogeneous data samples, the calculation labels 1 and 2 of the heterogeneous data samples, the business data samples and the target attribute labels of the target objects in the business data samples to update the parameters of the target initial model to be trained (the detailed process can be referred to the description of S302).

[0252] If the current training meets the convergence conditions, the updated target initial model, that is, the target model, is output.

[0253] If the current training does not meet the convergence condition and the next training still belongs to the first round of iterative training, the data training device executes the above steps 1 to 4 for the next group of samples to achieve one training of the target initial model to be trained by this group of samples.

[0254] If the current training does not meet the convergence conditions, and the next training belongs to the second round or subsequent rounds of iterative training, the data training device performs the above steps 2 to 4 for the next group of samples to achieve one training of the target initial model to be trained by this group of samples.

[0255] In summary, the data training method provided in the embodiment of the present application can complete the data training process through a simple process. In addition, since a large amount of heterogeneous data is introduced to participate in model training, the trained target model has high prediction accuracy and high prediction speed, and the training cost is low.

[0256] The above mainly introduces the solution provided in the embodiment of the present application from the perspective of method.

[0257] In order to achieve the above functions, Fig.14 As shown, Fig.14 FIG. 1 is a schematic diagram showing a structure of a data training device 140 provided in an embodiment of the present application. The data training device 140 can be used to execute the above-mentioned data training method, for example, to execute Figure 3 , Figure 7 , Fig.10 , Fig.11 , Fig.12 or Fig.13 The method shown in the figure. The data training device 140 may include an acquisition unit 141, a processing unit 142 and a training unit 143.

[0258] The acquisition unit 141 is used to acquire training samples including business data samples and heterogeneous data samples, wherein the business data samples are annotated with target attribute labels of target objects for user businesses, and the heterogeneous data samples are data samples from sources different from the business data samples. The processing unit 142 is used to determine the target initial model according to the target task, and the target task is to detect the target object in the image. The training unit 143 is used to iteratively train the target initial model according to the business data samples and heterogeneous data samples in the training samples to obtain a target model for executing the target task.

[0259] As an example, combining Figure 3 , the acquisition unit 141 can be used to execute S101, the processing unit 142 can be used to execute S102, and the training unit 143 can be used to execute S103.

[0260] Optionally, the data training device 140 further includes: a receiving unit 144, configured to receive at least one edge detection threshold input by a user. The edge detection threshold is used to detect edges in heterogeneous data samples. A training unit 143 is specifically configured to iteratively train a target initial model based on the business data sample, the heterogeneous data sample, and at least one edge detection threshold.

[0261] Optionally, the training sample includes a first training sample, and the first training sample includes a first business data sample and a first heterogeneous data sample. The target attribute of the target object includes the position of the target object and the category of the target object, wherein the position of the target object is the position of the target object in the image. The processing unit 142 is also used to determine the calculation label of the target object in the first heterogeneous data sample according to the target initial model to be trained and at least one edge detection threshold. The training unit 143 is specifically used to train the target initial model to be trained according to the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculation label of the first heterogeneous data sample to obtain the target model. Among them, the calculation label is a label calculated for the target attribute of the target object in the first heterogeneous data sample, and the calculation label includes a calculation position box and a calculation category of the target object in the first heterogeneous data sample, the calculation position box is used to characterize the position of the target object calculated in the first heterogeneous data sample, and the calculation category is a category corresponding to the calculation position box.

[0262] As an example, combining Figure 7 The processing unit 142 can be used to execute S201 to S203, and the training unit 143 can be used to execute S204.

[0263] Optionally, the processing unit 142 is further used to: determine a predicted position frame and a predicted category of a target object in the first heterogeneous data sample according to the target initial model to be trained, wherein the predicted position frame is used to characterize the position of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained. Determine an edge in the first heterogeneous data sample according to at least one edge detection threshold and an edge detection algorithm. Determine a calculated position frame according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame. And, determine a calculated category according to the predicted category.

[0264] As an example, combining Figure 7 , the processing unit 142 can be used to execute S201 to S203.

[0265] Optionally, the processing unit 142 is further specifically configured to: detect an initial edge in the first heterogeneous data sample under each edge detection threshold according to an edge detection algorithm and at least one edge detection threshold, and fuse the initial edges according to a non-maximum suppression algorithm to obtain an edge in the first heterogeneous data sample.

[0266] As an example, combining Figure 7 , the processing unit 142 can be used to execute S201.

[0267] Optionally, the processing unit 142 is further specifically configured to: determine the edge in the first heterogeneous data sample that is located within the predicted position frame as a positive edge in the first heterogeneous data sample according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, where the positive edge in the first heterogeneous data sample is an edge in the first heterogeneous data sample that is related to the target object. And, fuse the positive edge in the first heterogeneous data sample using a minimum bounding box fusion algorithm to obtain a calculated position frame.

[0268] As an example, combining Figure 7 , the processing unit 142 can be used to execute S2031~S2032.

[0269] Optionally, the training unit 143 is specifically used to: input the first business data sample and the first heterogeneous data sample into the target initial model to be trained to obtain a prediction result. Determine a position loss function and a category loss function based on the prediction result, the target attribute label of the target object in the first business data sample, and the calculation label of the first heterogeneous data sample. Update the target initial model to be trained based on the position loss function and the category loss function.

[0270] As an example, combining Figure 7 , the training unit 143 can be used to execute S2041~S2043.

[0271] Optionally, the data training device 140 also includes: a receiving unit 144, which is used to receive the positive and negative edge ratio input by the user, and the positive and negative edge ratio is used to balance the number of positive edges and negative edges in the heterogeneous data sample. Among them, the positive edge in the heterogeneous data sample is used to determine the calculation label of the target object in the heterogeneous data sample, and the negative edge in the heterogeneous data sample is used to determine the background label of the heterogeneous data sample, and the background label of the heterogeneous data sample includes the background box and background category in the heterogeneous data sample. The training unit 143 is specifically used to iteratively train the target initial model according to the business data sample, the heterogeneous data sample, the calculation label of the heterogeneous data sample, and the background label of the heterogeneous data sample.

[0272] Optionally, the processing unit 142 is also used to: determine the negative edge in the first heterogeneous data sample by merging the edge in the first heterogeneous data sample that is outside the predicted position box according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position box. And, fuse the negative edge in the first heterogeneous data sample with the minimum bounding box fusion algorithm to obtain the background box in the first heterogeneous data sample. Among them, the negative edge in the first heterogeneous data sample is an edge in the first heterogeneous data sample that is not related to the target object, the predicted position box is used to characterize the position of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained, and the category corresponding to the background box in the first heterogeneous data sample is the background category. The training unit 143 is also specifically used to train the target initial model to be trained according to the first business data sample, the first heterogeneous data sample, the calculated label of the first heterogeneous data sample, and the background box and background category in the first heterogeneous data sample.

[0273] As an example, combining Fig.10 The processing unit 142 can be used to execute S3011 to S3012, and the training unit 143 can be used to execute S302.

[0274] Optionally, the processing unit 142 is further specifically configured to: sample the negative edges in the first heterogeneous data sample so that the number of positive edges in the first heterogeneous data sample and the number of negative edges in the first heterogeneous data sample satisfy the positive-negative edge ratio. And, fuse the sampled negative edges using a minimum bounding box fusion algorithm to obtain a background frame of the first heterogeneous data sample.

[0275] As an example, combining Fig.10 , the processing unit 142 can be used to execute S3011~S3012.

[0276] Optionally, the data training device 140 further includes: a receiving unit 144, configured to receive a first selection operation input by a user before determining the target initial model according to the target task. A processing unit 142, configured to determine the target task from a plurality of preset tasks in response to an instruction of the first selection operation.

[0277] Optionally, the data training device 140 further includes: a receiving unit 144, configured to receive a second selection operation input by a user. A processing unit 142, specifically configured to determine a target initial model from a plurality of initial models corresponding to the target task in response to an instruction of the second selection operation.

[0278] For the specific description of the above optional methods, please refer to the above method embodiments, which will not be repeated here. In addition, the explanation of any data training device 140 provided above and the description of the beneficial effects can refer to the above corresponding method embodiments, which will not be repeated here.

[0279] As an example, combining Figure 1 The functions implemented by the processing unit 142 and the training unit 143 in the data training device 140 can be realized by Figure 1 Processor 101 in the Figure 1 The functions implemented by the acquisition unit 141 can be implemented by the program code in the memory 102 in FIG. Figure 1 The processor 101 is shown reading Figure 1 The sample data stored in the memory 102 is implemented, or by Figure 1 The network interface 103 shown in FIG. 1 is implemented. The functions implemented by the receiving unit 144 can be Figure 1 The input and output interface 105 is implemented in.

[0280] Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0281] It should be noted that Fig.14 The division of modules in the example is schematic and is only a logical function division. There may be other division methods in actual implementation. For example, two or more functions may be integrated into one processing module. The above integrated modules may be implemented in the form of hardware or software function modules.

[0282] The present application also provides a computer program product and a computer-readable storage medium for storing the computer program product. The computer program product may include one or more program instructions, which, when executed by one or more processors, can provide the above Figure 3 , Figure 7 , Fig.10 , Fig.11 , Fig.12 or Fig.13 Thus, for example, reference to Figure 3 One or more features of S101 to S103 may be undertaken by one or more instructions in the computer program product.

[0283] In some examples, for execution Figure 3 , Figure 7 , Fig.10 , Fig.11 , Fig.12 or Fig.13 The data training apparatus of the illustrated method may be configured to provide various operations, functions, or actions in response to one or more program instructions stored in a computer-readable storage medium.

[0284] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using a software program, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When executing instructions on a computer and executing a computer, a process or function according to an embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. Computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server, data center, etc. that contains one or more servers that can be integrated with a medium. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a DVD), or a semiconductor medium (eg, a solid state disk (SSD)).

[0285] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A data training method, characterized in that: include: Acquire training samples, wherein the training samples include business data samples and heterogeneous data samples, wherein the business data samples are annotated with target attribute labels of target objects for user businesses, and the heterogeneous data samples are data samples from sources different from those of the business data samples; Determine a target initial model according to a target task, wherein the target task is a task of detecting the target object in an image; The target initial model is iteratively trained according to the business data samples and heterogeneous data samples in the training samples to obtain a target model, and the target model is used to perform the target task.

2. The method according to claim 1, characterized in that The method further comprises: receiving at least one edge detection threshold input by a user, wherein the edge detection threshold is used to detect edges in the heterogeneous data sample; The iterative training of the target initial model according to the business data samples and the heterogeneous data samples in the training samples comprises: The target initial model is iteratively trained according to the business data samples, the heterogeneous data samples and the at least one edge detection threshold.

3. The method according to claim 2, characterized in that The training samples include a first training sample, the first training sample includes a first business data sample and a first heterogeneous data sample, the target attribute of the target object includes a position of the target object and a category of the target object, and the position is a position of the target object in an image; The iterative training of the target initial model according to the business data sample, the heterogeneous data sample and the at least one edge detection threshold comprises: Determine, according to the target initial model to be trained and the at least one edge detection threshold, a calculated label of the target object in the first heterogeneous data sample, the calculated label being a label calculated for a target attribute of the target object in the first heterogeneous data sample, the calculated label comprising a calculated position frame and a calculated category of the target object in the first heterogeneous data sample, the calculated position frame being used to characterize a position of the target object calculated in the first heterogeneous data sample, and the calculated category being a category corresponding to the calculated position frame; The target initial model to be trained is trained according to the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculated label of the first heterogeneous data sample to obtain the target model.

4. The method according to claim 3, characterized in that The step of determining the calculated label of the target object in the first heterogeneous data sample according to the target initial model to be trained and the at least one edge detection threshold comprises: Determine, according to the target initial model to be trained, a predicted position frame and a predicted category of a target object in the first heterogeneous data sample, wherein the predicted position frame is used to represent a position of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained; Determining an edge in the first heterogeneous data sample according to the at least one edge detection threshold and the edge detection algorithm; Determining the calculated position frame according to a positional relationship between an edge in the first heterogeneous data sample and the predicted position frame; The calculation category is determined according to the prediction category.

5. The method according to claim 4, characterized in that The determining the edge in the first heterogeneous data sample according to the at least one edge detection threshold and the edge detection algorithm comprises: According to the edge detection algorithm and the at least one edge detection threshold, detecting and obtaining an initial edge in the first heterogeneous data sample under each edge detection threshold; The initial edges are fused according to a non-maximum suppression algorithm to obtain edges in the first heterogeneous data sample.

6. The method according to claim 4 or 5, characterized in that: The determining the calculated position frame according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame comprises: According to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, an edge in the first heterogeneous data sample located within the predicted position frame is determined as a positive edge in the first heterogeneous data sample, where the positive edge in the first heterogeneous data sample is an edge in the first heterogeneous data sample that is related to the target object; The positive edges in the first heterogeneous data sample are fused using a minimum bounding box fusion algorithm to obtain the calculated position frame.

7. The method according to any one of claims 3 to 6, characterized in that The training of the target initial model to be trained according to the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculated label of the first heterogeneous data sample includes: Inputting the first business data sample and the first heterogeneous data sample into the target initial model to be trained to obtain a prediction result; Determine a position loss function and a category loss function according to the prediction result, the target attribute label of the target object in the first business data sample, and the calculated label of the first heterogeneous data sample; The target initial model to be trained is updated according to the position loss function and the category loss function.

8. The method according to any one of claims 1 to 7, characterized in that The method further comprises: Receive a positive-negative edge ratio input by a user, wherein the positive-negative edge ratio is used to balance the number of positive edges and negative edges in the heterogeneous data sample, wherein the positive edge in the heterogeneous data sample is used to determine a calculated label of a target object in the heterogeneous data sample, and the negative edge in the heterogeneous data sample is used to determine a background label of the heterogeneous data sample, wherein the background label of the heterogeneous data sample includes a background frame and a background category in the heterogeneous data sample; The iterative training of the target initial model according to the business data samples and the heterogeneous data samples in the training samples comprises: The target initial model is iteratively trained according to the business data samples, the heterogeneous data samples, the calculated labels of the heterogeneous data samples, and the background labels of the heterogeneous data samples.

9. The method according to claim 8, characterized in that The method further comprises: According to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, the edge in the first heterogeneous data sample that is outside the predicted position frame is used to determine the negative edge in the first heterogeneous data sample; the negative edge in the first heterogeneous data sample is an edge in the first heterogeneous data sample that is not related to the target object, and the predicted position frame is used to represent the position of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained; The negative edge in the first heterogeneous data sample is fused by the minimum bounding box fusion algorithm to obtain a background box in the first heterogeneous data sample; the category corresponding to the background box in the first heterogeneous data sample is the background category; The iterative training of the target initial model according to the business data sample, the heterogeneous data sample, the calculated label of the heterogeneous data sample and the background label of the heterogeneous data sample comprises: The target initial model to be trained is trained according to the first business data sample, the first heterogeneous data sample, the calculated label of the first heterogeneous data sample, and the background frame and background category in the first heterogeneous data sample.

10. The method according to claim 9, characterized in that The method further comprises: Sampling in the negative edges in the first heterogeneous data sample so that the number of positive edges in the first heterogeneous data sample and the number of negative edges in the first heterogeneous data sample satisfy the positive-to-negative edge ratio; The step of fusing the negative edge in the first heterogeneous data sample with the minimum bounding box fusion algorithm to obtain a background frame in the first heterogeneous data sample comprises: The sampled negative edges are fused using the minimum bounding box fusion algorithm to obtain a background frame of the first heterogeneous data sample.

11. The method according to any one of claims 1 to 10, characterized in that Before determining the target initial model according to the target task, the method further includes: Receive a first selection operation input by the user; In response to the instruction of the first selection operation, the target task is determined from a plurality of preset tasks.

12. The method according to any one of claims 1 to 11, characterized in that Determining the target initial model according to the target task includes: Receiving a second selection operation input by the user; In response to an instruction of the second selection operation, the target initial model is determined from a plurality of initial models corresponding to the target task.

13. A data training device, characterized in that: include: An acquisition unit, configured to acquire training samples, wherein the training samples include business data samples and heterogeneous data samples, wherein the business data samples are annotated with target attribute labels of target objects for user businesses, and the heterogeneous data samples are data samples from sources different from those of the business data samples; A processing unit, configured to determine a target initial model according to a target task, wherein the target task is a task of detecting the target object in an image; A training unit is used to iteratively train the target initial model according to the business data samples and heterogeneous data samples in the training samples to obtain a target model, and the target model is used to perform the target task.

14. The device according to claim 13, characterized in that The device also includes: A receiving unit, configured to receive at least one edge detection threshold input by a user, wherein the edge detection threshold is used to detect edges in the heterogeneous data sample; The training unit is specifically used to iteratively train the target initial model according to the business data samples, the heterogeneous data samples and the at least one edge detection threshold.

15. The device according to claim 14, characterized in that The training samples include a first training sample, the first training sample includes a first business data sample and a first heterogeneous data sample, the target attribute of the target object includes a position of the target object and a category of the target object, and the position is a position of the target object in an image; The processing unit is further used to determine a calculated label of a target object in the first heterogeneous data sample according to the target initial model to be trained and the at least one edge detection threshold, wherein the calculated label is a label calculated for a target attribute of the target object in the first heterogeneous data sample, and the calculated label includes a calculated position frame and a calculated category of the target object in the first heterogeneous data sample, wherein the calculated position frame is used to represent a position of the target object calculated in the first heterogeneous data sample, and the calculated category is a category corresponding to the calculated position frame; The training unit is specifically used to train the target initial model to be trained based on the first business data sample, the target attribute label of the target object in the first business data sample, the first heterogeneous data sample, and the calculation label of the first heterogeneous data sample to obtain the target model.

16. The device according to claim 15, characterized in that The processing unit is further configured to: Determine, according to the target initial model to be trained, a predicted position frame and a predicted category of a target object in the first heterogeneous data sample, wherein the predicted position frame is used to represent a position of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained; Determining an edge in the first heterogeneous data sample according to the at least one edge detection threshold and the edge detection algorithm; Determining the calculated position frame according to a positional relationship between an edge in the first heterogeneous data sample and the predicted position frame; The calculation category is determined according to the prediction category.

17. The device according to claim 16, characterized in that The processing unit is further specifically configured to: According to the edge detection algorithm and the at least one edge detection threshold, detecting and obtaining an initial edge in the first heterogeneous data sample under each edge detection threshold; The initial edges are fused according to a non-maximum suppression algorithm to obtain edges in the first heterogeneous data sample.

18. The device according to claim 16 or 17, characterized in that The processing unit is further specifically configured to: According to the positional relationship between the edge in the first heterogeneous data sample and the predicted position frame, an edge in the first heterogeneous data sample located within the predicted position frame is determined as a positive edge in the first heterogeneous data sample, where the positive edge in the first heterogeneous data sample is an edge in the first heterogeneous data sample that is related to the target object; The positive edges in the first heterogeneous data sample are fused using a minimum bounding box fusion algorithm to obtain the calculated position frame.

19. The device according to any one of claims 15 to 18, characterized in that The training unit is specifically used for: Inputting the first business data sample and the first heterogeneous data sample into the target initial model to be trained to obtain a prediction result; Determine a position loss function and a category loss function according to the prediction result, the target attribute label of the target object in the first business data sample, and the calculated label of the first heterogeneous data sample; The target initial model to be trained is updated according to the position loss function and the category loss function.

20. The device according to any one of claims 13 to 19, characterized in that The device also includes: A receiving unit, configured to receive a positive-negative edge ratio input by a user, wherein the positive-negative edge ratio is used to balance the number of positive edges and negative edges in the heterogeneous data sample, wherein the positive edge in the heterogeneous data sample is used to determine a calculated label of a target object in the heterogeneous data sample, and the negative edge in the heterogeneous data sample is used to determine a background label of the heterogeneous data sample, wherein the background label of the heterogeneous data sample includes a background frame and a background category in the heterogeneous data sample; The training unit is specifically used to iteratively train the target initial model based on the business data samples, the heterogeneous data samples, the calculated labels of the heterogeneous data samples, and the background labels of the heterogeneous data samples.

21. The device according to claim 20, characterized in that The processing unit is further used to determine the negative edge in the first heterogeneous data sample by taking the edge in the first heterogeneous data sample that is outside the predicted position box according to the positional relationship between the edge in the first heterogeneous data sample and the predicted position box; the negative edge in the first heterogeneous data sample is an edge in the first heterogeneous data sample that is irrelevant to the target object, and the predicted position box is used to characterize the position of the target object obtained after the first heterogeneous data sample is predicted by the target initial model to be trained; and, fuse the negative edge in the first heterogeneous data sample with the minimum bounding box fusion algorithm to obtain the background box in the first heterogeneous data sample; The category corresponding to the background frame in the first heterogeneous data sample is the background category; The training unit is further specifically used to train the target initial model to be trained based on the first business data sample, the first heterogeneous data sample, the calculated label of the first heterogeneous data sample, and the background frame and background category in the first heterogeneous data sample.

22. The device according to claim 21, characterized in that The processing unit is further specifically configured to: Sampling in the negative edges in the first heterogeneous data sample so that the number of positive edges in the first heterogeneous data sample and the number of negative edges in the first heterogeneous data sample satisfy the positive-to-negative edge ratio; The sampled negative edges are fused using the minimum bounding box fusion algorithm to obtain a background frame of the first heterogeneous data sample.

23. The device according to any one of claims 13 to 22, characterized in that The device also includes: A receiving unit, used for receiving a first selection operation input by a user before determining a target initial model according to a target task; The processing unit is used to determine the target task from a plurality of preset tasks in response to an instruction of the first selection operation.

24. The device according to any one of claims 13 to 23, characterized in that The device also includes: A receiving unit, used for receiving a second selection operation input by a user; The processing unit is specifically configured to determine the target initial model from a plurality of initial models corresponding to the target task in response to an instruction of the second selection operation.

25. A data training device, characterized in that: include: One or more processors and a memory, wherein the one or more processors are configured to read program instructions stored in the memory to execute the method according to any one of claims 1 to 12.

26. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes program instructions. When the program instructions are executed on a computer or a processor, the computer or the processor is caused to perform the method according to any one of claims 1 to 12.