Large model training data generation method and device, equipment, storage medium and product

Abstract

The invention discloses a large model training data generation method and device, equipment, a storage medium and a product, and relates to the technical field of artificial intelligence, and the method comprises the steps: extracting semantic tag information and image feature information of data in an image data set for model training; generating a joint vector according to the semantic tag information and the image feature information; determining an aggregation category corresponding to the preset image data according to the joint vector and a preset classification model; and generating large model training data based on the aggregation category corresponding to the preset image data. The joint vector is generated according to the semantic label information and the image feature information of the preset image data, and then the category is determined based on the classification model, so that image classification can be performed based on the semantic information rich in the image data, and compared with the existing method based on single image information, the accuracy of an image classification result can be improved, and the image classification efficiency is improved. And the training data is automatically generated based on the category, so that the data generation efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment, storage medium and product for generating large model training data. Background Art

[0002] With the rapid development of artificial intelligence technology, deep learning large models with a large number of parameters, high computational complexity, and requiring a large amount of training data have been widely used. The training effect and performance of these models depend to a large extent on the richness and diversity of the training data. The amount of data required for large model training is very large. For example, the amount of training data for the GPT-3 model is as high as 45TB. These data often contain multiple modalities (text, images, videos, etc.) and have multiple sources. It is very difficult to effectively analyze these training data. In particular, for the image, video and other modal data required for multimodal large model training, it is necessary not only to clarify the main content contained in the data, but also to analyze the semantic information of the data. In the case of insufficient training data and low data quality, the cost of data expansion through manual collection and annotation is high and the efficiency is low. The current popular method is to use large models for data synthesis to further expand the training data set.

[0003] At present, a common method for analyzing image and video data is to use manual labeling, where business experts label each image and video with corresponding labels, determine the category to which the image data belongs, and then statistically analyze the quantity, scale and other information of image data in each category. This method is costly and inefficient. Another common method is to establish an image classification model, which only uses the single modal information of the image data and ignores the semantic connotation represented by the image data. This often leads to a one-sided model analysis result, making it difficult to cope with image data analysis in complex scenarios, resulting in low efficiency and accuracy in generating training image data. Summary of the invention

[0004] The main purpose of this application is to provide a method, device, equipment, storage medium and product for generating large model training data, aiming to solve the technical problems of low accuracy and efficiency when analyzing and generating image training sets of large models.

[0005] To achieve the above objectives, the present application proposes a method for generating large model training data, the method comprising:

[0006] Extracting semantic label information and image feature information of preset image data, wherein the preset image data is data in an image data set used for model training;

[0007] generating a joint vector according to the semantic label information and the image feature information;

[0008] Determining an aggregation category corresponding to the preset image data according to the joint vector and a preset classification model;

[0009] Generate large model training data based on the aggregation category corresponding to the preset image data.

[0010] In one embodiment, the step of extracting semantic tag information and image feature information of preset image data includes:

[0011] Build semantic understanding templates based on the business scenarios of the large model to be trained;

[0012] Inputting the preset image data and the semantic understanding template into a preset multimodal large language model to extract information and obtain semantic description text information;

[0013] Extracting information from the semantic description text information to obtain semantic label information;

[0014] Image feature information is generated according to the semantic label information, the preset instruction template and the preset multimodal large model.

[0015] In one embodiment, the step of generating a joint vector according to the semantic tag information and the image feature information includes:

[0016] Inputting the semantic tag information into a text encoder to generate a semantic text vector corresponding to the preset image data;

[0017] Inputting the image feature information into a visual encoder to generate an image feature vector corresponding to the preset image data;

[0018] The semantic text vector and the image feature vector are input into a preset vector fusion model for fusion to obtain a fused joint vector.

[0019] In one embodiment, the step of generating large model training data based on the aggregation category corresponding to the preset image data includes:

[0020] Performing statistical analysis on the preset image data based on the aggregation categories to obtain category statistical information corresponding to each aggregation category;

[0021] Determine an image generation template corresponding to each aggregation category according to the category statistical information;

[0022] Generate large model training data based on the image generation template.

[0023] In one embodiment, the category statistics include the number of images and data similarity;

[0024] The step of performing statistical analysis on the preset image data based on the aggregation categories to obtain category statistical information corresponding to each aggregation category includes:

[0025] Counting the number of preset image data in each aggregation category to obtain the number of images;

[0026] The average similarity between the joint vectors corresponding to the preset image data in each aggregation category is determined to obtain the data similarity corresponding to each aggregation category.

[0027] In one embodiment, the category statistical information also includes data dispersion;

[0028] After the step of determining the average similarity between the joint vectors corresponding to the preset image data in each aggregation category and obtaining the data similarity corresponding to each aggregation category, the method further includes:

[0029] One of the aggregation categories is taken as the current category in turn;

[0030] Determine the target image data with the highest average similarity to the joint vector of the preset image data in the current category, and obtain the category center vector;

[0031] The average distance between the joint vector of each preset image data in the current category and the category center vector is determined to obtain the data dispersion corresponding to the current category.

[0032] In addition, to achieve the above purpose, the present application also proposes a large model training data generation device, the large model training data generation device comprising:

[0033] A semantic information extraction module, used to extract semantic label information and image feature information of preset image data, wherein the preset image data is data in an image data set used for model training;

[0034] A semantic information fusion module, used to generate a joint vector according to the semantic label information and the image feature information;

[0035] An aggregation category determination module, used to determine the aggregation category corresponding to the preset image data according to the joint vector and a preset classification model;

[0036] The training data generation module is used to generate large model training data based on the aggregation category corresponding to the preset image data.

[0037] In addition, to achieve the above-mentioned purpose, the present application also proposes a large model training data generation device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the large model training data generation method as described above.

[0038] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by the processor, the steps of the large model training data generation method described above are implemented.

[0039] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the large model training data generation method as described above.

[0040] The present application provides a method for generating large model training data, by extracting semantic label information and image feature information of preset image data, wherein the preset image data is data in an image data set used for model training; generating a joint vector according to the semantic label information and the image feature information; determining the aggregation category corresponding to the preset image data according to the joint vector and a preset classification model; and generating large model training data based on the aggregation category corresponding to the preset image data. The present application generates a joint vector according to the semantic label information and image feature information of the preset image data, and then determines the category based on the classification model, so that image classification can be performed based on the rich semantic information of the image data. Compared with the existing method based on single image information, the accuracy of image classification results can be improved, and training data is automatically generated based on categories, thereby improving the efficiency of data generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0042] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0043] Figure 1 A flowchart of the first embodiment of the method for generating large model training data in this application is provided;

[0044] Figure 2A flowchart diagram of the second embodiment of the method for generating large model training data in this application;

[0045] Figure 3 This is a flowchart of the image-semantic text information fusion process and data classification process of this application;

[0046] Figure 4 A schematic diagram of the overall process of determining aggregation categories in the method for generating training data for a large model in this application;

[0047] Figure 5 A flowchart diagram of the third embodiment of the method for generating large model training data in this application;

[0048] Figure 6 This is a schematic diagram of the overall process of generating image data in the large model training data generation method of this application;

[0049] Figure 7 This is a schematic diagram of the module structure of the large model training data generating device according to an embodiment of the present application;

[0050] Figure 8 Schematic diagram of the device structure of the hardware operating environment involved in the large model training data generation method in the embodiment of the present application.

[0051] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0052] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0053] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0054] The main solution of the embodiment of the present application is: extracting semantic label information and image feature information of preset image data, wherein the preset image data is data in the image data set used for model training; generating a joint vector based on the semantic label information and the image feature information; determining the aggregation category corresponding to the preset image data based on the joint vector and a preset classification model; and generating large model training data based on the aggregation category corresponding to the preset image data.

[0055] In the prior art, a common method for analyzing image and video data is to use manual labeling, where business experts label each image and video with corresponding labels, determine the category to which the image data belongs, and then statistically analyze the quantity, scale and other information of the image data in each category. This method is costly and inefficient. Another common method is to establish an image classification model. This method only uses the single modal information of the image data and ignores the semantic connotation represented by the image data. This often leads to a one-sided model analysis result, making it difficult to cope with image data analysis in complex scenarios, resulting in low efficiency and accuracy in generating training image data.

[0056] The present application provides a solution, by extracting semantic label information and image feature information of preset image data, the preset image data is data in the image data set used for model training; generating a joint vector according to the semantic label information and the image feature information; determining the aggregation category corresponding to the preset image data according to the joint vector and a preset classification model; generating large model training data based on the aggregation category corresponding to the preset image data. The present application generates a joint vector according to the semantic label information and image feature information of the preset image data, and then determines the category based on the classification model, so that the image can be classified based on the rich semantic information of the image data. Compared with the existing method based on single image information, the accuracy of the image classification result can be improved, and the training data is automatically generated based on the category, thereby improving the efficiency of data generation.

[0057] It should be noted that the execution subject of the method of this embodiment can be a computing service device with large model training data generation, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc.; it can also be a large model training data generation device with the same or similar functions. This embodiment and the following embodiments will be described by taking the large model training data generation device as an example.

[0058] Based on this, the present application embodiment provides a method for generating large model training data, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the method for generating large model training data in the present application.

[0059] In this embodiment, the large model training data generation method includes steps S10 to S40:

[0060] Step S10, extracting semantic label information and image feature information of preset image data, wherein the preset image data is data in an image data set used for model training.

[0061] It is understandable that, considering that most existing image data classification methods are based on single graphic information without considering the semantic information of image data, the image classification results are not accurate enough. Therefore, this embodiment proposes a large model training data generation method, which is mainly divided into two stages: (1) image data classification based on multimodal information fusion, and (2) image data generation based on the aggregation category corresponding to the preset image data. In this embodiment, semantic description text information is extracted from the preset image data in the image data set used for model training, semantic label information is extracted from the semantic description text information, and then image feature information is extracted based on the semantic label information.

[0062] Step S20: generating a joint vector according to the semantic tag information and the image feature information.

[0063] It should be understood that the semantic label information and the image feature information may be fused to obtain a joint vector including the semantic label and the image feature.

[0064] Step S30: determining the aggregation category corresponding to the preset image data according to the joint vector and a preset classification model.

[0065] It should be noted that the joint vector obtained in the above steps is input into the classification model based on the Multilayer Perceptron (MLP) to output the aggregated category of the image data. The classification model here is multi-label classification, and the loss function of the classification model is defined as:

[0066]

[0067] Among them, Loss represents the loss function, y represents the true label of the sample, y' is the predicted label of the model, N is the number of samples, and w i is the weight of sample i, y i represents the true label of sample i, y i ` represents the predicted label of the model for sample i, which can be set based on the proportion of samples in each category.

[0068] Step S40, generating large model training data based on the aggregation category corresponding to the preset image data.

[0069] It is understandable that, considering that the existing methods adopt an independent approach for the analysis and generation of image data, the efficiency is low. Therefore, in the method of this embodiment, the analysis and generation of training data are performed simultaneously, which can effectively improve the rationality, effectiveness, accuracy and efficiency of data analysis and generation tasks. In the above-mentioned first stage process, the original image characteristics and semantic information of the image data are comprehensively considered, and the image data is divided into the aggregation categories to which it belongs. In the second stage, the data sets of each category are statistically analyzed, and the constituent characteristic indicators of the data sets of each category are calculated from multiple dimensions such as quantity and quality. According to the calculation results of the indicators, the prompt instruction (Prompt Instruction) for image data synthesis is dynamically selected according to the data differences of each category, and the image data required for large model training is generated based on the few-shot learning (FSL) method based on the multimodal large language model.

[0070] The present embodiment provides a method for generating large model training data, by extracting semantic label information and image feature information of preset image data, wherein the preset image data is data in an image data set used for model training; generating a joint vector according to the semantic label information and the image feature information; determining the aggregation category corresponding to the preset image data according to the joint vector and a preset classification model; and generating large model training data based on the aggregation category corresponding to the preset image data. The present application generates a joint vector according to the semantic label information and image feature information of the preset image data, and then determines the category based on the classification model, thereby being able to perform image classification based on the rich semantic information of the image data, which can improve the accuracy of image classification results compared to the existing method based on single image information, and automatically generates training data based on categories, thereby improving the efficiency of data generation.

[0071] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can refer to the above introduction, and will not be repeated later. Figure 2 , step S10, the large model training data generation method further includes steps S101 to S104:

[0072] Step S101: construct a semantic understanding template according to the business scenario of the large model to be trained.

[0073] It is worth noting that, considering that the existing data generation prompt template is a predefined static template, it is impossible to filter the irrelevant pixel data in the image according to the actual distribution of the training data set, and the existing image data classification method does not filter the irrelevant pixel data in the image according to the requirements of different business scenarios, which affects the accuracy of the classification results. Therefore, in this embodiment, the dynamic change of the prompt template is realized, so as to perform a more reasonable and efficient data generation task. When extracting the semantic label information and image feature information of the preset image data, the prompt template for the semantic understanding of the image data can be formulated based on the business scenario faced by the large model training. For example, in the field of autonomous driving, the richness and effectiveness of the training data set are mainly concentrated in the time period, road environment, surrounding buildings, traffic signals, etc. Based on the characteristics of the business scenario, a semantic understanding template can be constructed as follows: "Please refer to the following aspects to describe the current {image}: whether the image shows a daytime or night scene, whether the image shows that the vehicle is driving on a closed road section or an open road section, what types of buildings are included in the image, the pedestrian movements on both sides of the road in the image, and what traffic signs are included in the image. . . ".

[0074] Step S102: input the preset image data and the semantic understanding template into a preset multimodal large language model to extract information and obtain semantic description text information.

[0075] Step S103: extracting information from the semantic description text information to obtain semantic tag information.

[0076] It can be understood that the image data in the training data set can be obtained, combined with the semantic understanding template constructed above, and input into the preset multimodal large language model to obtain the semantic description text information of the image data that meets the business scenario, and extract the corresponding semantic labels from the semantic description text information. For example, the extracted semantic labels can be [daytime, open road section, office buildings on both sides, pedestrians walking normally, green light for passage, ...].

[0077] Step S104: Generate image feature information according to the semantic label information, the preset instruction template and the preset multimodal large model.

[0078] It is understandable that the semantic label information of the acquired image data can be inserted as a word slot into the predefined instruction template to obtain the instruction for image feature extraction, for example: "Extract the feature objects in the following pictures, respectively extract [main road], [office building], [pedestrians on the roadside], [traffic light]..., and output each feature object separately." The instruction is input into the preset multimodal large model to extract the key image features corresponding to each semantic label, thereby obtaining the image feature information. For different business scenarios, the key image features in the image data are extracted to reduce the interference of other irrelevant pixel data on the classification results and improve the classification accuracy.

[0079] In a feasible implementation, step S20 may include steps S201 to S203:

[0080] Step S201: input the semantic tag information into a text encoder to generate a semantic text vector corresponding to the preset image data.

[0081] It should be noted that you can refer to Figure 3 The fusion process and data classification process are explained. Figure 3 The figure is a flowchart of the image-semantic text information fusion process and data classification process. The bidirectional encoder representation model (Bidirectional Encoder Representations from Transformers, BERT) model can be selected as the text encoder (Text Encoder), and the semantic label information is input into the BERT model to obtain the output text embedding (i.e., semantic text vector), which is expressed as: Y = (Y 1 , …, Y N ).

[0082] Step S202: input the image feature information into a visual encoder to generate an image feature vector corresponding to the preset image data.

[0083] It is understandable that the extracted image feature information can be input into the vision encoder (VisionEncoder) to obtain the embedding (i.e., image feature vector) corresponding to each image feature. Here, the CNN model is selected as the Vision Encoder, and the image feature embedding output by CNN is expressed as: X = (X 1 , …, X N ), N represents the number of image features corresponding to the semantic label.

[0084] Step S203: input the semantic text vector and the image feature vector into a preset vector fusion model for fusion to obtain a fused joint vector.

[0085] It is understandable that X and Y can be input into the preset vector fusion model to obtain the joint vector after the fusion of image and semantic information as Z=(Z 1 , …, Z N ),in:

[0086] Z i =(X i T U) T A(Y i T V),

[0087] Among them, U∈R n×k , V∈R m×k is a linear matrix, n is the dimension of image embedding, m is the dimension of text embedding, and k is the dimension of joint vector. A is the attention matrix, defined as:

[0088] A=P T ((U T X i )·(V T Y i )),

[0089] Where P is the pooling matrix.

[0090] Understandably, one can refer to Figure 4 The first phase of the process is described below. Figure 4 The overall process diagram for determining the aggregation category in the large model training data generation method is shown in FIG. A data set for model training can be obtained, and a prompt template for semantic understanding of image data can be formulated based on business scenarios, and input into the multimodal large language model to generate semantic description text information and semantic label information of image data. According to the extraction instructions corresponding to the semantic label information and the multimodal large model, the key image features corresponding to each semantic label are extracted, and these extracted image features are input into the visual encoder to obtain the image feature vector corresponding to each image feature. The semantic label text is input into the Bert model to obtain the output semantic text vector, and the semantic text vector and the image feature vector are input into the preset vector fusion model for fusion to obtain the fused joint vector. The semantic information and key image information are fused together, which can reduce the interference of irrelevant pixel data in the image data and realize more reasonable and accurate data classification for different business scenarios. Finally, the joint vector obtained in the above steps is input into the classification model based on MLP to obtain the aggregation category of the image data.

[0091] In this embodiment, a prompt template for semantic understanding of image data is formulated according to the business scenario faced by the current model training, and a semantic description text and label information of the image data are generated; an instruction for image extraction is generated based on the semantic information, and a multimodal large language model is used to extract key image features and generate image embeddings; a semantic text embedding of the image data is generated and fused with the image embedding; finally, a classification model is constructed based on the fused vector, so as to obtain the aggregated category of the image data, effectively improving the rationality, effectiveness, accuracy and efficiency of data analysis and generation tasks.

[0092] Based on the first embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the first embodiment can be referred to the above description, and will not be described in detail later. Figure 5 , step S40, the large model training data generation method further includes steps S401 to S403:

[0093] Step S401: performing statistical analysis on the preset image data based on the aggregation categories to obtain category statistical information corresponding to each aggregation category.

[0094] It is understandable that based on the image data classification results of the first stage, statistical analysis can be performed on the data sets of each category, and the category statistical information of the data sets in each aggregation category can be calculated from multiple dimensions such as quantity and quality.

[0095] Step S402: determining an image generation template corresponding to each aggregation category according to the category statistical information.

[0096] It can be understood that the category statistical information of different category data sets can be obtained based on the above steps, and different image data generation prompt templates can be set according to the category statistical information. For example, when the number of category data sets is less than the threshold, the prompt template selected for quantity expansion is: "Please refer to {image samples} and the corresponding {semantic description text} to generate {number} pieces of image data that are similar to the reference image data in both graphics and semantics". When the data similarity of the category data set is greater than the threshold, the prompt template selected for data difference expansion is: "Please refer to {image samples} and the corresponding {semantic description text} to generate image data that is different from the reference image data in {local features} but similar in {overall features}".

[0097] Step S403: Generate large model training data based on the image generation template.

[0098] It should be understood that the reference {image sample} and {semantic description text} in the prompt template use a few-shot method to extract picture-text example data from the category data set to complete the information. The completed image data generation instruction is input into the multimodal large language model to obtain the generated large model training data. In this embodiment, the generation prompt template corresponding to different category data is dynamically selected based on the category statistical information, and the image data is generated based on the prompt template. Compared with the data generation method based on the static prompt template, the characteristics of the training data set itself can be used to generate data structures missing in the current training data, so as to better ensure the rationality and effectiveness of the data generation task.

[0099] In a feasible implementation manner, the category statistical information includes the number of images and data similarity; step S401 may include steps S4011 to S4012:

[0100] Step S4011, counting the number of preset image data in each aggregation category to obtain the number of images.

[0101] Step S4012, determining the average similarity between the joint vectors corresponding to the preset image data in each aggregation category, and obtaining the data similarity corresponding to each aggregation category.

[0102] It is understandable that the category statistics may include the number of images and data similarity. The number of images may be obtained by counting the number of preset image data in each aggregate category. The data similarity is obtained by calculating the average similarity between the joint vectors corresponding to the preset image data in each aggregate category, and the specific method is:

[0103]

[0104] Among them, ‖Z i ‖ represents the vector Z i Model.

[0105] In a feasible implementation manner, the category statistical information further includes data dispersion; after step S4012, steps S101 to S103 may also be included:

[0106] Step S4013: taking one of the aggregation categories in turn as the current category.

[0107] Step S4014, determining the target image data having the highest average similarity with the joint vector of the preset image data in the current category, and obtaining the category center vector.

[0108] Step S4015, determining the average distance between the joint vector of each preset image data in the current category and the category center vector, and obtaining the data dispersion corresponding to the current category.

[0109] It is worth noting that the category statistics also include data dispersion, which can be calculated based on data similarity. One of the aggregation categories can be taken as the current category in turn, and the data dispersion of one category can be calculated each time until the data dispersion of all the aggregation categories is calculated. Find the target image data with the highest average similarity between the joint vectors of the preset image data in the current category, and record it as the category center vector Z 0 , and then calculate the average distance between the joint vector of each preset image data in the current category and the category center vector to obtain the data dispersion:

[0110]

[0111] In the specific implementation, you can refer to Figure 6 The overall process of generating image data in the second stage is described. Statistical analysis is performed on the data sets of each category to calculate indicators such as the number of data sets of each category, data similarity, and data dispersion. According to the indicator calculation results, the prompt instruction for image data synthesis is dynamically selected according to the data differences of each category, and the required image data is generated based on the few-shot method and the multimodal large language model.

[0112] In this embodiment, based on the data quantity, data similarity, data discreteness and other category statistical information of each category data set, a generation prompt template corresponding to different category data is dynamically selected, and image data is generated based on the prompt template. Compared with the data generation method based on the static prompt template, the characteristics of the training data set itself can be used to generate data structures missing in the current training data, thereby better ensuring the rationality and effectiveness of the data generation task.

[0113] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the method for generating large model training data in the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0114] This application also provides a large model training data generation device, please refer to Figure 7 , the large model training data generating device comprises:

[0115] A semantic information extraction module 10, used to extract semantic label information and image feature information of preset image data, wherein the preset image data is data in an image data set used for model training;

[0116] A semantic information fusion module 20, configured to generate a joint vector according to the semantic label information and the image feature information;

[0117] An aggregation category determination module 30, configured to determine an aggregation category corresponding to the preset image data according to the joint vector and a preset classification model;

[0118] The training data generation module 40 is used to generate large model training data based on the aggregation category corresponding to the preset image data.

[0119] The large model training data generation device provided by the present application adopts the large model training data generation method in the above embodiment to solve the technical problem. Compared with the prior art, the beneficial effects of the large model training data generation device provided by the present application are the same as the beneficial effects of the large model training data generation method provided by the above embodiment, and the other technical features in the large model training data generation device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0120] The present application provides a large model training data generation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the large model training data generation method in the above-mentioned embodiment one.

[0121] Reference below Figure 8 , which shows a schematic diagram of the structure of a large model training data generation device suitable for implementing the embodiment of the present application. The large model training data generation device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 8 The large model training data generation device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0122] like Figure 8As shown, the large model training data generation device may include a processing device 1001 (such as a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the large model training data generation device are also stored. The processing device 1001, ROM1002 and RAM1004 are connected to each other via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the large model training data generation device to communicate wirelessly or wired with other devices to exchange data. Although the diagram shows a large model training data generation device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or have instead.

[0123] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0124] The large model training data generation device provided by the present application adopts the large model training data generation method in the above embodiment, which can solve the technical problem of large model training data generation. Compared with the prior art, the beneficial effects of the large model training data generation device provided by the present application are the same as the beneficial effects of the large model training data generation method provided by the above embodiment, and the other technical features in the large model training data generation device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0125] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0126] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0127] The present application provides a computer-readable storage medium having computer-readable program instructions (i.e., computer programs) stored thereon, and the computer-readable program instructions are used to execute the large model training data generation method in the above-mentioned embodiment.

[0128] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0129] The above-mentioned computer-readable storage medium may be included in the large model training data generation device; or it may exist independently without being assembled into the large model training data generation device.

[0130] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the large model training data generation device, the large model training data generation device: extracts semantic label information and image feature information of preset image data, wherein the preset image data is data in the image data set used for model training; generates a joint vector according to the semantic label information and the image feature information; determines the aggregation category corresponding to the preset image data according to the joint vector and a preset classification model; and generates large model training data based on the aggregation category corresponding to the preset image data.

[0131] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0132] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0133] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0134] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned large model training data generation method, and can solve technical problems. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the large model training data generation method provided in the above-mentioned embodiment, and will not be repeated here.

[0135] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the large model training data generation method as described above.

[0136] The computer program product provided in this application can solve the technical problem. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the large model training data generation method provided in the above embodiment, which will not be repeated here.

[0137] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for generating large model training data, characterized in that: The method includes: Extracting semantic label information and image feature information of preset image data, wherein the preset image data is data in an image data set used for model training; generating a joint vector according to the semantic label information and the image feature information; Determining an aggregation category corresponding to the preset image data according to the joint vector and a preset classification model; Generate large model training data based on the aggregation category corresponding to the preset image data.

2. The method according to claim 1, characterized in that The step of extracting semantic tag information and image feature information of preset image data includes: Build semantic understanding templates based on the business scenarios of the large model to be trained; Inputting the preset image data and the semantic understanding template into a preset multimodal large language model to extract information and obtain semantic description text information; Extracting information from the semantic description text information to obtain semantic label information; Image feature information is generated according to the semantic label information, the preset instruction template and the preset multimodal large model.

3. The method according to claim 1, characterized in that The step of generating a joint vector according to the semantic label information and the image feature information comprises: Inputting the semantic tag information into a text encoder to generate a semantic text vector corresponding to the preset image data; Inputting the image feature information into a visual encoder to generate an image feature vector corresponding to the preset image data; The semantic text vector and the image feature vector are input into a preset vector fusion model for fusion to obtain a fused joint vector.

4. The method according to claim 1, characterized in that The step of generating large model training data based on the aggregation category corresponding to the preset image data includes: Performing statistical analysis on the preset image data based on the aggregation categories to obtain category statistical information corresponding to each aggregation category; Determine an image generation template corresponding to each aggregation category according to the category statistical information; Generate large model training data based on the image generation template.

5. The method according to claim 4, characterized in that The category statistics include the number of images and data similarity; The step of performing statistical analysis on the preset image data based on the aggregation categories to obtain category statistical information corresponding to each aggregation category includes: Counting the number of preset image data in each aggregation category to obtain the number of images; The average similarity between the joint vectors corresponding to the preset image data in each aggregation category is determined to obtain the data similarity corresponding to each aggregation category.

6. The method according to claim 5, characterized in that The category statistical information also includes data dispersion; After the step of determining the average similarity between the joint vectors corresponding to the preset image data in each aggregation category and obtaining the data similarity corresponding to each aggregation category, the method further includes: One of the aggregation categories is taken as the current category in turn; Determine the target image data with the highest average similarity to the joint vector of the preset image data in the current category, and obtain the category center vector; The average distance between the joint vector of each preset image data in the current category and the category center vector is determined to obtain the data dispersion corresponding to the current category.

7. A large model training data generation device, characterized in that: The large model training data generating device comprises: A semantic information extraction module, used to extract semantic label information and image feature information of preset image data, wherein the preset image data is data in an image data set used for model training; A semantic information fusion module, used to generate a joint vector according to the semantic label information and the image feature information; An aggregation category determination module, used to determine the aggregation category corresponding to the preset image data according to the joint vector and a preset classification model; The training data generation module is used to generate large model training data based on the aggregation category corresponding to the preset image data.

8. A large model training data generation device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the large model training data generation method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the large model training data generation method as described in any one of claims 1 to 6 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, which, when executed by a processor, implements the steps of the large model training data generation method according to any one of claims 1 to 6.

Images