Data processing method and apparatus
By generating pseudo-labels and target weight sets to train the model, the problem of high labeling cost for training samples in segmentation models is solved, and labeling cost is reduced without affecting accuracy.
Patent Information
- Application Number
- CN202311219852.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-20
- Publication Date
- 2026-01-13
- Estimated Expiration
- 2043-09-20
AI Technical Summary
Existing segmentation models require a large number of manually labeled training samples during training, resulting in high labeling costs and potential decreases in model accuracy.
By generating pseudo-labels and target weight sets, the target model can be trained using pseudo-labels, reducing annotation costs without affecting model accuracy.
Without affecting model accuracy, the cost of labeling training samples was significantly reduced.
Smart Images
Figure CN117274753B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing, and more specifically to a data processing method and apparatus. Background Technology
[0002] Currently, with the continuous development of deep learning technology, more and more segmentation models, such as semantic segmentation models, object detection models, and instance segmentation models, are being applied to related fields such as intelligent driving, robot perception, or medical image analysis.
[0003] However, existing segmentation models typically require a large number of training samples during training, and all of these training samples need to be manually labeled by annotators, which makes the labeling cost of training samples very high. Summary of the Invention
[0004] In view of this, embodiments of the present invention provide a data processing method and apparatus to reduce the labeling cost of training samples without affecting the accuracy of the trained model.
[0005] In a first aspect, embodiments of the present invention provide a data processing method, the method comprising:
[0006] Obtain multiple training samples to be labeled, wherein the training samples include at least one target object without a real label;
[0007] Generate pseudo-labels for each target object to obtain target samples;
[0008] Determine the target weight set for the target object of each pseudo-label category;
[0009] The target model is trained based on the target samples carrying the pseudo-labels and the target weight set, wherein the target weight set includes the loss value weights of the target objects of the corresponding pseudo-label categories when the target model predicts that they belong to different object categories.
[0010] Furthermore, generating pseudo-tags for each of the target objects includes:
[0011] Pseudo-labels for each target object are generated based on high-precision map data.
[0012] Furthermore, generating pseudo-tags for each of the target objects includes:
[0013] The training samples are input into a preset annotation model to generate pseudo-labels for each target object.
[0014] Furthermore, determining the target weight set for the target object of each pseudo-label category includes:
[0015] Display the target object carrying the corresponding pseudo-label;
[0016] Receive the first weight set of the target object for each pseudo-label category;
[0017] The target weight set of the target object for each pseudo-label category is determined at least based on the first weight set.
[0018] Furthermore, determining the target weight set for the target object of each pseudo-label category also includes:
[0019] The training samples are input into the target model to generate predicted labels for each target object through the target model;
[0020] Based on the predicted labels and the pseudo-labels, a second set of weights for the target objects of each pseudo-label category is determined; and / or
[0021] The third weight set of the target objects for each pseudo-label category is determined based on the fully labeled data, wherein the fully labeled data only includes labeled objects with real labels.
[0022] Furthermore, determining the target weight set for each pseudo-label category of the target object based at least on the first weight set includes:
[0023] The target weight set of the target object for each pseudo-label category is determined based on the first weight set, the second weight set, and / or the third weight set.
[0024] Further, determining the second weight set of the target object for each pseudo-label category based on the predicted label and the pseudo-label includes:
[0025] A confusion matrix is generated based on the predicted labels and the pseudo-labels;
[0026] The second set of weights for the target object of each pseudo-label category is determined based on the confusion matrix.
[0027] Furthermore, the third weight set for the target object in each pseudo-label category, determined based on the fully labeled data, includes:
[0028] Generate pseudo-labels for each of the labeled objects;
[0029] Based on the pseudo-labels and real labels of the labeled objects, determine the set of labeled object weights for each pseudo-label category of the labeled objects;
[0030] For each pseudo-label category of the target object, the set of labeled object weights of labeled objects that are in the same pseudo-label category as the target object is determined as the third weight set of the target object.
[0031] Furthermore, the training samples include raw sample data that has not been labeled and / or historical sample data, wherein the historical sample data includes some labeled objects with real labels.
[0032] Furthermore, the training samples are image data, and the target object is a pixel.
[0033] Furthermore, the training samples are point cloud data, and the target objects are point cloud points.
[0034] In a second aspect, embodiments of the present invention provide a data processing apparatus, the apparatus comprising:
[0035] An acquisition unit is used to acquire multiple training samples to be labeled, wherein the training samples include at least one target object without a real label;
[0036] A generation unit is used to generate pseudo-labels for each of the target objects to obtain target samples;
[0037] A determining unit is used to determine the target weight set of the target object for each pseudo-label category;
[0038] The training unit is used to train a target model based on target samples carrying the pseudo-labels and the target weight set, wherein the target weight set includes the loss value weights of target objects of the corresponding pseudo-label categories when the target model predicts that they belong to different object categories.
[0039] Thirdly, embodiments of the present invention provide a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, implement the method as described in any one of the first aspects.
[0040] Fourthly, embodiments of the present invention provide an electronic device, the device comprising:
[0041] Memory is used to store one or more computer program instructions;
[0042] A processor, wherein the one or more computer program instructions are executed by the processor to implement the method as described in any one of the first aspects.
[0043] Fifthly, embodiments of the present invention provide a computer program product that, when run on a computer, causes the computer to perform the method as described in any one of the first aspects.
[0044] The data processing method of this invention, after acquiring multiple training samples to be labeled, generates pseudo-labels for each target object in the training samples, determines the target weight set for each pseudo-label category of the target object, and then trains a target model based on the target samples carrying pseudo-labels and the target weight set. The target weight set includes the loss value weights of the target objects of the corresponding pseudo-label categories when predicted by the target model to belong to different object categories. Therefore, by using target samples carrying pseudo-labels and the target weight set to train the target model, this embodiment can reduce the labeling cost of training samples without affecting the accuracy of the trained model. Attached Figure Description
[0045] The above and other objects, features and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which:
[0046] Figure 1 This is a flowchart of a data processing method according to an embodiment of the present invention;
[0047] Figure 2 This is a flowchart of the target weight set determination method according to an embodiment of the present invention;
[0048] Figure 3 This is a schematic diagram illustrating the process of determining the target weight set according to an embodiment of the present invention;
[0049] Figure 4 This is a schematic diagram of a data processing apparatus according to an embodiment of the present invention;
[0050] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation
[0051] The present application is described below based on embodiments, but it is not limited to these embodiments. In the detailed description of the present application below, certain specific details are described in detail. Those skilled in the art can fully understand the present application without these details. To avoid obscuring the substance of the present application, well-known methods, processes, flows, elements, and circuits are not described in detail.
[0052] Furthermore, those skilled in the art should understand that the accompanying drawings provided herein are for illustrative purposes only and are not necessarily drawn to scale.
[0053] Unless the context explicitly requires it, words such as "including" or "contains" throughout the application should be interpreted as including rather than exclusive or exhaustive; that is, meaning "including but not limited to".
[0054] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means two or more.
[0055] The solutions described in this specification and embodiments, if involving the processing of personal information, will be processed only under the premise of having a legal basis (such as obtaining the consent of the personal information subject, or being necessary for the performance of a contract), and will only be processed within the scope stipulated or agreed upon. A user's refusal to process personal information beyond what is necessary for basic functions will not affect the user's use of basic functions.
[0056] In the following description, the field of autonomous driving will be used as an example. However, it should be understood that the solutions of the embodiments of the present invention are also applicable to other related fields, such as the field of robot perception or the field of medical image analysis, and this application does not limit them.
[0057] During autonomous driving, vehicles typically use multiple deep learning models working together. Each model has its own annotation rules for the training samples used during training. As a result, a large amount of sample data with similar tasks but different annotation types is retained. For example, some sample data may only annotate individual objects such as vehicles, pedestrians, bicycles, and traffic cones, while other sample data may annotate not only individual objects but also the ground, flower beds, and fences.
[0058] To address the high cost of labeling training samples, related technologies utilize previously labeled historical sample data to train the current model. Specifically, these technologies mix historical sample data from different labeling rules together. However, because they do not consider the label differences of historical sample data under different labeling rules, while these technologies can reduce the cost of labeling training samples, they can decrease the accuracy of the trained model.
[0059] To address this, embodiments of the present invention provide a data processing method and apparatus to reduce the labeling cost of training samples without affecting the accuracy of the trained model.
[0060] Figure 1 This is a flowchart of a data processing method according to an embodiment of the present invention. Figure 1 As shown, the data processing method may specifically include the following steps:
[0061] It should be understood that the execution subject of the data processing method can be any general-purpose data processing device, such as a tablet computer or a computer. Optionally, in some embodiments, depending on the application field, the execution subject of the data processing method can also be a data processing terminal mounted on a machine device such as a vehicle, robot, or aircraft, and this application does not impose any limitations on this.
[0062] S100: Obtain multiple training samples to be labeled.
[0063] Specifically, the processing device can acquire multiple training samples to be labeled. These training samples include at least one target object that has not been labeled with a real tag.
[0064] Optionally, the training samples can be image data or point cloud data acquired through corresponding acquisition devices (e.g., cameras or radar sensors). Further, based on whether they are labeled with real-world tags, the pixels or point cloud points in the training samples can be divided into target objects and labeled objects. Target objects are pixels or point cloud points without real-world tags, while labeled objects are pixels or point cloud points with real-world tags. The real-world tags can be object category labels assigned to pixels or point cloud points by annotators; typically, these real-world tags can be used to characterize the real-world object category to which the labeled object belongs.
[0065] Optionally, the training samples may include unlabeled raw sample data and / or historical sample data. The raw sample data includes only the target object, i.e., pixels or point cloud points without actual labels. The historical sample data includes some labeled objects with actual labels. Therefore, the processing device in this embodiment can obtain training samples that meet the current model's labeling requirements not only by supplementing historical sample data with additional labels, but also by directly labeling the raw sample data.
[0066] It should be noted that when the training samples are image data, the relevant model involved in this embodiment can refer to the corresponding image segmentation model. When the training samples are point cloud data, the relevant model involved in this embodiment can refer to the corresponding point cloud segmentation model.
[0067] S200: Generate pseudo-labels for each target object to obtain target samples.
[0068] Specifically, after acquiring training samples, the processing device can generate pseudo-labels for each target object in the training samples to obtain target samples. These pseudo-labels are object category labels assigned to each target object by the processing device. Compared to real labels, these pseudo-labels only represent the real object category to which the target object belongs to to a certain extent; that is, the object category represented by the pseudo-label may or may not be the real object category of the target object. The target samples can be training samples that, after being supplemented and labeled by the processing device, meet the labeling requirements of the current model and can be used to train the current model.
[0069] Optionally, in step S200, the target object for generating the pseudo-label can be any of the target objects in the training samples, or it can be any target object in the training samples belonging to a specific category. Specifically, if the current model is a semantic segmentation model, since the semantic segmentation model needs to label the object category to which each pixel or point cloud point in the training samples belongs, the target object for generating the pseudo-label can be any of the target objects in the training samples. If the current model is an object detection model or an instance segmentation model, since the object detection model or instance segmentation model only needs to label the pixels or point cloud points belonging to the target category, the target object for generating the pseudo-label can be any target object in the training samples belonging to a specific category.
[0070] Furthermore, the specific category of objects can be determined by the processing device based on the current model annotation requirements and the training samples themselves. For example, if the training samples are historical sample data that have been annotated for individual objects such as vehicles, pedestrians, bicycles, and traffic cones, and the current model needs to annotate individual objects such as vehicles, pedestrians, bicycles, and traffic cones, as well as the ground, flower beds, and fences, then the specific category of objects can be the ground, flower beds, and fences. As another example, if the training samples are raw sample data, and the current model needs to annotate individual objects such as vehicles, pedestrians, bicycles, and traffic cones, then the specific category of objects can be individual objects such as vehicles, pedestrians, bicycles, and traffic cones.
[0071] Optionally, in step S200, the processing device can input the training samples into a preset annotation model to generate pseudo-labels for each target object in the training samples. It should be understood that the preset annotation model can be a multi-class object annotation model (i.e., an annotation model for annotating multiple classes of objects) or multiple single-class object annotation models (i.e., annotation models for annotating single-class objects). Thus, through the multi-class object annotation model or multiple single-class object annotation models, the processing device can annotate each target object in the training samples to generate pseudo-labels for each target object.
[0072] In some embodiments, the processing device can also generate pseudo-labels for each target object based on high-precision map data. Specifically, the training samples may have corresponding acquisition location information. When labeling target objects in the training samples, the processing device can first acquire high-precision map data corresponding to the acquisition location of the training samples. After acquiring the high-precision map data, the processing device can match the high-precision map data with the training samples and determine the corresponding pseudo-label for each target object based on the approximate location of each target object in the matched high-precision map data. Taking point cloud data as an example, the processing device can label point cloud points that are higher or lower than the corresponding ground distance threshold in the high-precision map data as "ground" pseudo-labels, label point cloud points located in the fence area of the high-precision map data as "fence" pseudo-labels, and label point cloud points located in the flower bed area of the high-precision map data as "flower bed" pseudo-labels.
[0073] It should be understood that the pseudo-tag generation method given above is only for illustration. In actual application, the processing device may also use other methods to generate pseudo-tags for each target object, and this application does not limit this.
[0074] S300. Determine the target weight set of the target object for each pseudo-label category.
[0075] Specifically, since pseudo-labels can only represent the object category to which the target object belongs to a certain extent, directly using target samples carrying the pseudo-labels to train the current model will reduce the accuracy of the trained model. Therefore, in order not to affect the accuracy of the trained model, the processing device will first determine the target weight set of the target object of each pseudo-label category before using the target samples to train the current model.
[0076] The target weight set includes the loss value weights of target objects of corresponding pseudo-label categories when predicted by the current model to belong to different object categories. Specifically, the loss value weight refers to the coefficient that the loss value of each target object needs to be multiplied before summing when the current model weights and sums the loss values of each pixel or point cloud point in the target sample. For example, for a target object with the pseudo-label category of "road surface," the loss value weight of this pseudo-label category target object when predicted by the current model to be a flower bed can be 0.5. That is, when the current model weights and sums the loss values of each pixel or point cloud point in the target sample, the loss value of the target object of this pseudo-label category will be multiplied by 0.5 before participating in the summation.
[0077] Figure 2 This is a flowchart illustrating the method for determining the target weight set according to an embodiment of the present invention. Figure 2 As shown, the method for determining the target weight set may specifically include the following steps:
[0078] It should be understood that the method for determining the target weight set can be specifically used to implement the above step S300.
[0079] S310. Display the target object carrying the corresponding pseudo-label.
[0080] Specifically, the processing device can display target objects carrying the corresponding pseudo-labels so that the labeler can determine the first weight set of target objects for each pseudo-label category based on the displayed content.
[0081] Each pseudo-label can have a corresponding display color. Therefore, the processing equipment allows labelers to quickly identify pseudo-labels with labeling errors.
[0082] It should be understood that the specific numerical value of the loss value weight can be inversely proportional to the probability or number of target objects of the corresponding real object category being incorrectly labeled as the corresponding pseudo-label category by the processing device. For example, when the processing device labels target objects, the greater the probability or number of target objects of the real object category of flower bed being incorrectly labeled as the pseudo-label category of road surface, the smaller the loss value weight of the target object with the pseudo-label category of road surface when predicted by the current model to belong to flower bed can be when calculating the loss value. This can reduce the impact of incorrectly labeled pseudo-labels on the accuracy of the trained model.
[0083] S320, Receive the first weight set of the target object for each pseudo-label category.
[0084] Specifically, the processing device can determine the first weight set of the target objects for each pseudo-label category by receiving input information from the labeler.
[0085] S330. Determine the target weight set of the target object for each pseudo-label category based at least on the first weight set.
[0086] Specifically, for each target object, the processing device can directly determine the first weight set as the target weight set for each target object. Alternatively, to improve the accuracy of the target weight set, the processing device can also determine different weight sets separately through other methods, and then jointly determine the target weight set based on the determined different weight sets.
[0087] Optionally, in some embodiments, the processing device can input training samples into a target model to generate predicted labels for each target object based on the output of the target model. The target model can be a current model pre-trained using partially fully labeled data, and the predicted labels can be object category labels assigned by the target model for each target object. After generating the predicted labels for each target object, since the predicted labels have higher labeling accuracy than pseudo-labels, the processing device can use the object categories labeled by the predicted labels as the true object categories for each target object to determine a second weight set for each pseudo-labeled target object category based on the predicted labels and pseudo-labels.
[0088] Furthermore, when determining the second weight set, the processing device can determine the corresponding confusion matrix based on the predicted labels and pseudo-labels, display the confusion matrix to the annotators, and then determine the second weight set of target objects for each pseudo-label category by receiving input information from the annotators. The confusion matrix is a table used to measure the performance of the classification model, showing the difference between the predicted results and the actual results on the test set. Alternatively, for target objects of each real object category, the processing device can also determine the second weight set of target objects for each pseudo-label category based on the ratio of the number of correctly labeled target objects of that real object category to the number of incorrectly labeled target objects of the corresponding pseudo-label category. For example, for a target object whose real object category is flower bed, if the ratio of the number of correctly labeled target objects of that real object category to the number of incorrectly labeled target objects of the pseudo-label category "road surface" is 1:2, it indicates that most target objects of that real object category have been incorrectly labeled as "road surface" by the processing device. Therefore, the loss value weight of a target object with the pseudo-label "road surface" when predicted by the current model to belong to flower bed can be set relatively small, for example, to 0.5. It should be understood that the method for determining the specific value of the loss value weight can be set and adjusted by relevant personnel according to actual needs, and this application does not impose any restrictions on this.
[0089] In other embodiments, the processing device may also utilize a small amount of fully labeled data to determine a third weight set for target objects in each pseudo-label category. Specifically, the processing device may generate pseudo-labels for each labeled object in the fully labeled data in the same manner as generating pseudo-labels for the target objects. After generating the pseudo-labels for each labeled object, the processing device can determine the third weight set based on the pseudo-labels and the actual labels of each labeled object.
[0090] It should be understood that since the same method is used to generate pseudo-labels, the labeling errors that occur when the processing device labels the target object will also occur when labeling the labeled object. Therefore, the processing device can determine the set of labeled object weights for each pseudo-label category based on the pseudo-labels of the labeled object and the real label. Then, for each pseudo-label category of the target object, the set of labeled object weights for the labeled objects that are in the same pseudo-label category as the target object is determined as the third set of weights for the target object.
[0091] Furthermore, after determining the first weight set, the second weight set, and the third weight set, the processing device can determine the target weight set based on any one or a combination of the first weight set, the second weight set, or the third weight set. Optionally, when determining the target weight set, the processing device can determine the target weight set by weighted summation of the loss values in the corresponding weight sets.
[0092] S400. Train the target model based on the target samples carrying the pseudo-labels and the target weight set.
[0093] Specifically, after determining the target samples and the target weight set, the processing device can train the target model based on the target samples carrying the pseudo-labels and the target weight set.
[0094] It should be understood that the loss value calculated by the loss function is often used as a metric to compare the difference between the predicted output of a machine learning model and the true value of the sample (also known as the supervised value). That is, it is used to measure the difference between the predicted output of the machine learning model and the true value of the sample. The loss value can be used to determine the adjustment range of model parameters during model training. In this embodiment, the currently trained model is a segmentation model. During training, the segmentation model predicts the object category to which each pixel or point cloud point in the target sample belongs. When calculating the loss value using the loss function, the segmentation model calculates the loss value corresponding to each pixel or point cloud point in the target sample based on the prediction results. Then, it weights and sums the loss values of each pixel or point cloud point to calculate the loss value of the target sample. The model parameters are then adjusted using the loss value of the target sample. By using the target weight set to determine the weight of the loss value of each target object when the current model performs a weighted sum of the loss values of each pixel or point cloud point in the target sample, this embodiment can reduce the impact of incorrectly labeled pseudo-labels on the model training process, thereby reducing the labeling cost of training samples without affecting the accuracy of the trained model.
[0095] Figure 3 This is a schematic diagram illustrating the process of determining the target weight set according to an embodiment of the present invention. Figure 3As shown, after acquiring the training samples 31 to be labeled, the processing device can generate pseudo-labels 32 for each target object in the training samples 31 using high-precision map data or a preset labeling model. After generating the pseudo-labels 32, the processing device can perform visualization processing on the pseudo-labels 32 to determine the first weight set 33 of the target objects for each pseudo-label category. The processing device can also input the pseudo-labels 32 into the target model 34 to generate and display the corresponding confusion matrix, thereby determining the second weight set 35 of the target objects for each pseudo-label category. Simultaneously, the processing device can also determine the second weight set 37 of the target objects for each pseudo-label category based on the fully labeled data. After determining the first weight set 33, the second weight set 35, and the second weight set 37, the processing device can determine the target weight set 38 based on the first weight set 33, the second weight set 35, and the second weight set 37.
[0096] Furthermore, the target model 34 can be a current model that has been pre-trained using partially fully labeled data. After determining the target weight set 38, the processing device can retrain or continue training the target model 34 based on the training samples 31, pseudo-labels 32, and target weight set 38, combined with partially fully labeled data. This application does not impose any restrictions on this.
[0097] The data processing method of this invention, after acquiring multiple training samples to be labeled, generates pseudo-labels for each target object in the training samples, determines the target weight set for each pseudo-label category of the target object, and then trains a target model based on the target samples carrying pseudo-labels and the target weight set. The target weight set includes the loss value weights of the target objects of the corresponding pseudo-label categories when predicted by the target model to belong to different object categories. Therefore, by using target samples carrying pseudo-labels and the target weight set to train the target model, this embodiment can reduce the labeling cost of training samples without affecting the accuracy of the trained model.
[0098] Figure 4 This is a schematic diagram of a data processing apparatus according to an embodiment of the present invention. Figure 4 As shown, the data processing device of this embodiment includes an acquisition unit 41, a generation unit 42, a determination unit 43, and a training unit 44.
[0099] Specifically, the acquisition unit 41 is used to acquire multiple training samples to be labeled, wherein the training samples include at least one target object that is not labeled with a real label;
[0100] The generation unit 42 is used to generate pseudo-labels for each of the target objects to obtain target samples;
[0101] The determining unit 43 is used to determine the target weight set of the target object for each pseudo-label category;
[0102] The training unit 44 is used to train a target model based on target samples carrying the pseudo-labels and the target weight set, wherein the target weight set includes the loss value weights of target objects of the corresponding pseudo-label categories when the target model predicts that they belong to different object categories.
[0103] The data processing apparatus of this invention, after acquiring multiple training samples to be labeled, generates pseudo-labels for each target object in the training samples, determines the target weight set for each pseudo-label category of the target object, and then trains a target model based on the target samples carrying pseudo-labels and the target weight set. The target weight set includes the loss value weights of the target objects of the corresponding pseudo-label categories when predicted by the target model to belong to different object categories. Therefore, by using target samples carrying pseudo-labels and the target weight set to train the target model, this embodiment can reduce the labeling cost of training samples without affecting the accuracy of the trained model.
[0104] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of the present invention. (For example...) Figure 5 As shown, Figure 5 The illustrated electronic device can be a data processing device as described in the above embodiments, comprising a general computer hardware architecture, including at least a processor 51 and a memory 52. The processor 51 and memory 52 are connected via a bus 53. The memory 52 is adapted to store instructions or programs executable by the processor 51. The processor 51 can be a standalone microprocessor or a collection of one or more microprocessors. Thus, the processor 51 executes the instructions stored in the memory 52 to perform the method flow of the embodiments of the present invention as described above, thereby realizing data processing and control of other devices. The bus 53 connects the aforementioned components together, and also connects the aforementioned components to a display controller 54, a display device, and an input / output (I / O) device 55. The input / output (I / O) device 55 can be a mouse, keyboard, modem, network interface, touch input device, motion-sensing input device, printer, and other devices known in the art. Typically, the input / output device 55 is connected to the system via an input / output (I / O) controller 56.
[0105] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus (devices), or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0106] This application is described with reference to flowchart illustrations of methods, apparatus (devices), and computer program products according to embodiments of this application. It should be understood that each step in the flowchart can be implemented by computer program instructions.
[0107] These computer program instructions may be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction means, the implementation process of which is described in the instruction means. Figure 1 The function specified in one or more processes.
[0108] These computer program instructions may also be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, produce instructions for implementing processes. Figure 1 A device for a function specified in one or more processes.
[0109] Another embodiment of the present invention relates to a non-volatile storage medium for storing a computer-readable program for use by a computer to execute some or all of the above-described method embodiments.
[0110] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program specifying the relevant hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0111] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A data processing method, characterized by, The method comprises: obtaining a plurality of training samples to be labeled, the training samples comprising at least one target object without a labeled true label, the training samples being image data, the target objects being pixel points, or the training samples being point cloud data, and the target objects being point cloud points; generating pseudo labels for each target object to obtain target samples; determining a target weight set of the target objects of each pseudo label category; training a target model according to the target samples carrying the pseudo labels and the target weight set, the target weight set comprising a loss value weight of the target objects of a corresponding pseudo label category when being predicted by the target model to belong to different object categories; wherein determining the target weight set of the target objects of each pseudo label category comprises: determining the target weight set of the target objects of each pseudo label category according to at least a first weight set; determining the target weight set of the target objects of each pseudo label category according to at least the first weight set comprises: determining the target weight set of the target objects of each pseudo label category according to the first weight set and a second weight set; the second weight set is determined by: inputting the training samples into the pre-trained target model to generate predicted labels for each target object according to the output of the target model, the target model being pre-trained based on full-labeled data, the full-labeled data comprising only labeled objects labeled with true labels; determining a second weight set of the target objects of each pseudo label category according to the predicted labels and the pseudo labels; determining the second weight set of the target objects of each pseudo label category according to the predicted labels and the pseudo labels comprises: determining the second weight set of the target objects of each pseudo label category according to a ratio of the number of the target objects of a true object category that are correctly labeled to the number of the target objects of each pseudo label that are incorrectly labeled, the true object category being determined according to the predicted labels.
2. The method of claim 1, wherein, generating pseudo labels for each target object comprises: generating pseudo labels for each target object according to high-definition map data.
3. The method of claim 1, wherein, generating pseudo labels for each target object comprises: inputting the training samples into a preset labeling model to generate pseudo labels for each target object through the preset labeling model.
4. The method of claim 1, wherein, the first weight set is determined by: displaying target objects carrying corresponding pseudo labels; receiving a first weight set of the target objects of each pseudo label category.
5. The method of claim 1, wherein, determining the target weight set of the target objects of each pseudo label category according to at least the first weight set further comprises: determining the target weight set of the target objects of each pseudo label category according to the first weight set, the second weight set, and a third weight set.
6. The method of claim 5, wherein, the third weight set is determined by: determining a third weight set of the target objects of each pseudo label category according to full-labeled data.
7. The method of claim 1, wherein, determining the second weight set of the target objects of each pseudo label category according to the predicted labels and the pseudo labels further comprises: generating a confusion matrix according to the predicted labels and the pseudo labels; The second weight set of the target object of each pseudo-label category is determined according to the confusion matrix.
8. The method of claim 6, wherein, The third weight set of the target object of each pseudo-label category is determined according to full-labeled data, including: generating a labeled object pseudo-label of each labeled object; determining a labeled object weight set of each labeled object of each pseudo-label category according to the labeled object pseudo-label and the real label; for the target object of each pseudo-label category, determining the labeled object weight set of the labeled object of the same pseudo-label category as the target object as the third weight set of the target object.
9. The method of claim 1, wherein, The training sample includes original sample data and / or historical sample data that have not been subjected to a labeling operation, and the historical sample data include part of the labeled objects labeled with real labels.
10. A data processing apparatus, characterized by, The device includes: an acquisition unit configured to acquire a plurality of training samples to be labeled, the training samples including at least one target object without a labeled real label, the training sample being image data, the target object being a pixel point, or the training sample being point cloud data, and the target object being a point cloud point; a generation unit configured to generate a pseudo-label of each target object to obtain a target sample; a determination unit configured to determine a target weight set of the target object of each pseudo-label category; a training unit configured to train a target model according to the target sample carrying the pseudo-label and the target weight set, the target weight set including a loss value weight of the target object of a corresponding pseudo-label category when being predicted by the target model to belong to different object categories; wherein the determination unit is further configured to: determine the target weight set of the target object of each pseudo-label category according to at least a first weight set; determine the target weight set of the target object of each pseudo-label category according to at least the first weight set, including: determine the target weight set of the target object of each pseudo-label category according to the first weight set and a second weight set; the second weight set is determined by: inputting the training sample into the pre-trained target model to generate a predicted label of each target object according to an output result of the target model, the target model being pre-trained based on full-labeled data, the full-labeled data including only labeled objects labeled with real labels; determine a second weight set of the target object of each pseudo-label category according to the predicted label and the pseudo-label; determine the second weight set of the target object of each pseudo-label category according to the predicted label and the pseudo-label, including: determine the second weight set of the target object of each pseudo-label category according to a ratio of the number of the target object of a real object category correctly labeled to the number of the target object of each pseudo-label incorrectly labeled, the real object category being determined according to the predicted label.
11. A computer readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by a processor, implement the method of any one of claims 1-9.
12. An electronic device, comprising: The device includes: a memory configured to store one or more computer program instructions; a processor, the one or more computer program instructions being executed by the processor to implement the method of any one of claims 1-9.
13. A computer program product, characterised in that, when the computer program product is run on a computer, causing the computer to perform the method of any one of claims 1-9.
Citation Information
Patent Citations
Training sample processing method, and voice quality inspection method and device
CN116153299A