MaaS multi-model publishing management method and system, and storage medium

Abstract

The application relates to the technical field of model service, and particularly discloses a MaaS multi-model publishing management method and system and a storage medium, the method comprises the following steps: matching a processing model in a preset processing model library according to demand processing information, generating feedback information, and sending the feedback information to a demand party; counting demand processing information of all demand parties in a preset time period, determining a global information range, counting historical demand processing information of each demand party, determining an individual information range of each demand party, and then constructing a model group of each processing model; the demand processing information is processed based on the processing model library, in the processing process, the global demand processing information in the preset time period and the individual demand processing information of each demand party are determined according to the demand processing information, and then a lightweight model suitable for different demand parties is determined, and when facing a demand, the demand processing information is processed through the lightweight model, so that the response speed is faster.

Description

Technical Field

[0001] This invention relates to the field of model service technology, specifically a MaaS multi-model release management method, system, and storage medium. Background Technology

[0002] As artificial intelligence models continue to develop, enterprises' daily work may involve multiple different types of models, such as NLP models, CV models, and recommendation models. For some small enterprises, they do not have the ability to build large models. Even if they build data processing models, the processing capacity of the data processing models will not be very high because the number of samples is small. Therefore, many model service providers have emerged, namely, Mass multi-model management providers. Mass stands for Model as a Service.

[0003] Existing model service providers build large, high-performance models to process data uploaded by clients. After the clients upload the data, the model service providers also expand the sample based on this data to further optimize the model performance. However, during the service process, a large model is used for each client. When there are many clients, the response speed will decrease, affecting the user experience. Therefore, how to improve the response speed of the model service process is the technical problem that the present invention aims to solve. Summary of the Invention

[0004] The purpose of this invention is to provide a MaaS multi-model release management method, system, and storage medium to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A MaaS multi-model release management method, the method comprising:

[0007] Receive the demand processing information uploaded by the demand party, match the processing model in the preset processing model library according to the demand processing information, generate feedback information based on the matched processing model, and send it to the demand party.

[0008] Collect demand processing information from all demanders within a preset time period, determine the global information scope, verify the model uploaded by any uploader based on the global information scope, and update the processing model library according to the verification results.

[0009] The historical demand processing information of each demander is statistically analyzed to determine the amount of demand information and the range of individual information for each demander. When the amount of demand information reaches a preset threshold, the processing model is distilled according to the range of individual information to obtain the individual model of that demander.

[0010] Individual models from different demanders are statistically analyzed, and the individual models are merged based on the amount of demand information for each demander to obtain a model group for each processing model.

[0011] Furthermore, the steps of receiving the demand processing information uploaded by the demand party, matching a processing model in a preset processing model library based on the demand processing information, generating feedback information based on the matched processing model, and sending it to the demand party include:

[0012] Receive request processing information uploaded by the requester;

[0013] The requirement processing information is formatted and then divided into sub-tasks based on the format recognition results.

[0014] Based on the type of subtask, a processing model is matched in the preset processing model library, and the predicted processing time of the matched processing model is obtained synchronously.

[0015] Select the processing model with the shortest prediction processing time, identify the sub-tasks, obtain the processing results, statistically analyze all processing results, generate feedback information, and send it to the requesting party.

[0016] Furthermore, the steps of statistically analyzing the demand processing information of all demanders within a preset time period, determining the global information scope, verifying the model uploaded by any uploader based on the global information scope, and updating the processing model library according to the verification results include:

[0017] Statistics on the processing information of all demanders within a preset time period;

[0018] The text data in the requirement processing information is segmented into words to obtain a word list; the word list contains word entries and word count entries.

[0019] For any word, query the word vector of the word, determine the vector radius of the word vector according to the number of words, obtain the vector range of each word, calculate the union of the vector ranges of each word, and obtain the global information range;

[0020] For image data in the demand processing information, obtain the image size of the image data, and create a global information range based on the image size;

[0021] The model uploaded by any uploader is verified based on the global information scope, and the processing model library is updated according to the verification results.

[0022] Furthermore, the step of verifying the model uploaded by any uploader based on the global information scope and updating the processing model library according to the verification results includes:

[0023] The model is received from the uploader via the model receiving port;

[0024] Randomly generate text and image data of a preset amount, and remove data that is included in the global information scope;

[0025] Input the data into the uploaded model to determine the processing speed;

[0026] The processing speed of all data is statistically analyzed. When the processing speed meets the preset speed conditions, the model is inserted into the processing model library.

[0027] Furthermore, the steps of statistically analyzing the historical demand processing information of each demander, determining the amount of demand information and the range of individual information for each demander, and performing model distillation on the processing model based on the range of individual information to obtain the individual model for that demander include:

[0028] For each demander, statistical analysis is performed on the demand processing information of each demander at different times.

[0029] The amount of data for calculating the demand processing information is taken as the demand information volume.

[0030] When the information to be processed is text data, query the vector range, calculate the union of the vector ranges at all times, and obtain the individual information range; when the information to be processed is image data, read the image size of each image data, and create the individual information range based on the image size.

[0031] The amount of demand information is compared with a preset information threshold. When the amount of demand information reaches the preset threshold, all demand processing information and corresponding feedback information within the individual information range are read, and a neural network model is trained as the individual model of the demander.

[0032] Furthermore, the step of merging individual models from different demanders based on the demand information of each demander to obtain a model group for each processing model includes:

[0033] The scope of individual information queries for different demanders;

[0034] Calculate the intersection-union ratio (IUU) of the individual information range, and determine the similarity of the demand side based on the IUU.

[0035] Based on the similarity, the demand sides are clustered, and the individual models of each demand side are merged to obtain model groups;

[0036] Query the processing model corresponding to each model in the model group, select the mode processing model, and connect the model group with the mode processing model.

[0037] The present invention also provides a MaaS multi-model release management system, the system comprising:

[0038] The demand processing module is used to receive demand processing information uploaded by the demander, match a processing model in a preset processing model library according to the demand processing information, generate feedback information based on the matched processing model, and send it to the demander.

[0039] The model library update module is used to collect the demand processing information of all demanders within a preset time period, determine the global information scope, verify the model uploaded by any uploader based on the global information scope, and update the processing model library according to the verification results.

[0040] The model distillation module is used to collect historical demand processing information for each demander, determine the amount of demand information and the range of individual information for each demander, and when the amount of demand information reaches a preset threshold, the processing model is distilled according to the range of individual information to obtain the individual model of that demander.

[0041] The model merging module is used to count the individual models of different demanders, and merge the individual models based on the amount of demand information of each demander to obtain a model group for each processing model.

[0042] Furthermore, the demand processing module includes:

[0043] The information receiving unit is used to receive demand processing information uploaded by the demander.

[0044] The task segmentation unit is used to perform format recognition on the requirement processing information and segment the requirement processing information into sub-tasks based on the format recognition result.

[0045] The time consumption prediction unit is used to match the processing model in the preset processing model library according to the type of the subtask, and synchronously obtain the predicted processing time of the matched processing model.

[0046] The results statistics unit is used to select the processing model with the shortest prediction processing time, identify sub-tasks, obtain processing results, statistically analyze all processing results, generate feedback information, and send it to the requesting party.

[0047] Furthermore, the model library update module includes:

[0048] The global information statistics unit is used to collect demand processing information from all demanders within a preset time period.

[0049] The text segmentation unit is used to segment the text data in the information to be processed into words, resulting in a word list; the word list contains word entries and word count entries.

[0050] The global scope determination unit is used to query the word vector of any word, determine the vector radius of the word vector based on the number of words, obtain the vector scope of each word, calculate the union of the vector scopes of each word, and obtain the global information scope.

[0051] The image analysis unit is used to obtain the image size of the image data in the information to be processed, and to create a global information range based on the image size;

[0052] The model verification unit is used to verify the model uploaded by any uploader based on the global information scope, and update the processing model library according to the verification results.

[0053] The present invention also provides a storage medium storing at least one line of program code, which, when loaded and executed by a processor, implements the MaaS multi-model release management method.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] This invention processes demand information based on a processing model library. During the processing, it determines the global demand processing information and the individual demand processing information of each demander within a preset time period based on the demand processing information. Then, it determines a lightweight model suitable for different demanders, constructs a model group, and inserts the model group into the processing model library. When faced with a demand, the demand processing information can be processed through the lightweight model, resulting in a faster response speed. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention.

[0057] Figure 1 The overall flowchart of the MaaS multi-model release management method is shown.

[0058] Figure 2 The diagram shows the structure of the MaaS multi-model release management system. Detailed Implementation

[0059] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0060] Figure 1The overall flowchart of the MaaS multi-model release management method is shown in this embodiment of the invention. The method includes:

[0061] Step S100: Receive the demand processing information uploaded by the demand party, match the processing model in the preset processing model library according to the demand processing information, generate feedback information based on the matched processing model, and send it to the demand party.

[0062] The requester is the party that needs to process data. The data that needs to be processed, uploaded by the requester, is called the request processing information. Under the current technological background, request processing information can be divided into text data and image data. For audio data, it can be converted into text data, and for video data, it can be converted into audio data and image data. Based on the request processing information, a processing model is matched in a preset processing model library. That is, a data processing model that can process the corresponding data is matched. The matched data processing model processes the specific data, generates feedback information, and sends it to the requester. This completes one interaction process.

[0063] MaaS stands for Model as a Service, which is used to uniformly manage models and then provide model services.

[0064] Step S200: Collect demand processing information from all demanders within a preset time period, determine the global information scope, verify the model uploaded by any uploader based on the global information scope, and update the processing model library according to the verification results.

[0065] Staff pre-set a time period, such as a week or a month, to collect all demand processing information from all demanders within that period. This creates a data scope that includes all demand processing information, known as the global information scope. The execution entity of this method also interfaces with the model uploader, receiving the uploaded model and verifying it based on the global information scope. When the verification is successful, the model is inserted into the processing model library, thus affecting the processing model matching process in step S100.

[0066] The verification process in this invention is a special verification. For the uploaded model, it checks whether the data it processes is within the global information range. If it is, the verification fails because the global information range is determined based on the existing processed data and is generated based on the existing processed model. If the range of data processed by the uploaded model (new model) is included in the global information range, it can be considered that the existing model can complete the functions that the new model can achieve, and there is no need to update the model library. Of course, if the processing speed of the new model is higher than that of the existing model, it can also replace a part of the existing model.

[0067] In practical applications, the execution entity of this method provides some basic models. These basic models have a wide data processing range and their processing speed for different types of data is average. For new models, if their data processing range differs significantly from the existing data processing range, they are inserted into the processing model library to expand the data processing range of the execution entity of this method. If the data processing range of the new model differs only slightly from the existing data processing range, but its data processing speed is relatively fast, it is also inserted into the processing model library. When data within this data processing range is received, the new model is applied for processing, and the original model is no longer used.

[0068] Step S300: Statistically analyze the historical demand processing information of each demander, determine the amount of demand information and the range of individual information for each demander, and when the amount of demand information reaches a preset threshold, perform model distillation on the processing model according to the range of individual information to obtain the individual model of the demander.

[0069] Step S200 compiles the demand processing information for all demanders, while step S300 compiles the demand processing information for each demander. The same scheme is used to determine the information range, which represents the demand situation of the demander at different times, and is called the individual information range. At the same time, based on the historical demand processing information, the amount of demand information for each demander can also be calculated, indicating how much information the demander needs to process. This indicates the amount of demand from the demander. When the amount of demand information reaches a preset threshold, it means that the demander needs to process a lot of data. At this time, the processing model is distilled according to the individual information range to obtain the individual model of the demander, that is, to generate a dedicated model for serving the demander.

[0070] Step S400: Statistically analyze the individual models of different demanders, and merge the individual models based on the amount of demand information of each demander to obtain a model group for each processing model.

[0071] After the individual model for each demander is generated, the individual models of different demanders are counted. Based on the amount of demand information for each demander, the individual models are merged. The merging process reduces the number of individual models. After merging, the individual models are inserted into the processing model library to obtain model combinations for each type of processing model.

[0072] Regarding step S100, the steps of receiving the demand processing information uploaded by the demand party, matching a processing model in a preset processing model library based on the demand processing information, generating feedback information based on the matched processing model, and sending it to the demand party include:

[0073] Receive request processing information uploaded by the requester;

[0074] The requirement processing information is formatted and then divided into sub-tasks based on the format recognition results.

[0075] Based on the type of subtask, a processing model is matched in the preset processing model library, and the predicted processing time of the matched processing model is obtained synchronously.

[0076] Select the processing model with the shortest prediction processing time, identify the sub-tasks, obtain the processing results, statistically analyze all processing results, generate feedback information, and send it to the requesting party.

[0077] In one example of the technical solution of this invention, the processing procedure for demand processing information is described. The demand processing information uploaded by the demander is received, and the format of the demand processing information is identified. Based on the format identification result, the demand processing information is divided into sub-tasks. In this technical solution, there are only two types of sub-tasks: text tasks and image tasks. A processing model is matched in a preset processing model library according to the type of sub-task. The predicted processing time of the matched processing model is obtained synchronously. The prediction of processing time is a basic function of the model. The processing model with the shortest predicted processing time is selected, the sub-task is identified, the processing result is obtained, all processing results are statistically analyzed, feedback information is generated, and sent to the demander.

[0078] The above process is the matching and application of the model, and it is not complicated.

[0079] Regarding step S200, the steps of statistically analyzing the demand processing information of all demanders within a preset time period, determining the global information scope, verifying the model uploaded by any uploader based on the global information scope, and updating the processing model library according to the verification results include:

[0080] Statistics on the processing information of all demanders within a preset time period;

[0081] The text data in the requirement processing information is segmented into words to obtain a word list; the word list contains word entries and word count entries.

[0082] For any word, query the word vector of the word, determine the vector radius of the word vector according to the number of words, obtain the vector range of each word, calculate the union of the vector ranges of each word, and obtain the global information range;

[0083] For image data in the demand processing information, obtain the image size of the image data, and create a global information range based on the image size;

[0084] The model uploaded by any uploader is verified based on the global information scope, and the processing model library is updated according to the verification results.

[0085] In one example of the technical solution of this invention, the update process of the processing model library is described. The process involves statistically analyzing the demand processing information of all demanders within a preset time period. This demand processing information includes two types: text data and image data. For the text data in the demand processing information, word segmentation is performed to obtain a word table. The word table contains word entries and word count entries, indicating how many times each word appears. For any word, its word vector is queried. The vector radius of the word vector is determined based on the word count, which is proportional to the word count. This yields the vector range of each word. The union of the vector ranges of each word is calculated to obtain the global information range corresponding to the text data.

[0086] Correspondingly, for the image data in the demand processing information, the image size of the image data is obtained, and a global information range of the image data is created based on the image size. Since the color values ​​of the images themselves are similar, such as RGB images, the complexity of each pixel is the same. The only difference is the number of pixels in the image itself, which is determined by the size. The minimum width and minimum length serve as one boundary, and the maximum width and maximum length serve as another boundary, thereby determining an image size range as the global information range of the image data.

[0087] After determining the scope of global information, the model uploaded by any uploader is verified based on the scope of global information, and the processing model library is updated according to the verification results.

[0088] As a preferred embodiment of the technical solution of the present invention, the step of verifying the model uploaded by any uploader based on the global information range and updating the processing model library according to the verification result includes:

[0089] The model is received from the uploader via the model receiving port;

[0090] Randomly generate text and image data of a preset amount, and remove data that is included in the global information scope;

[0091] Input the data into the uploaded model to determine the processing speed;

[0092] The processing speed of all data is statistically analyzed. When the processing speed meets the preset speed conditions, the model is inserted into the processing model library.

[0093] In one example of the technical solution of this invention, the model uploading process of the uploader is defined. Based on the model receiving port, the model uploaded by the uploader is received, and a preset amount of text data and image data are randomly generated. Data contained in the global information range is removed, and the data is input into the uploaded model. The processing speed is determined. If the processing speed is zero, it means that it cannot be processed. If the processing speed is high, it means that the model has a good processing representation for data that is not frequently processed. Based on this, when the processing speed is high enough, the model is regarded as a valid model and inserted into the processing model library.

[0094] Regarding step S300, the steps of statistically analyzing the historical demand processing information of each demander, determining the amount of demand information and the range of individual information for each demander, and performing model distillation on the processing model based on the range of individual information to obtain the individual model for that demander include:

[0095] For each demander, statistical analysis is performed on the demand processing information of each demander at different times.

[0096] The amount of data for calculating the demand processing information is taken as the demand information volume.

[0097] When the information to be processed is text data, query the vector range, calculate the union of the vector ranges at all times, and obtain the individual information range; when the information to be processed is image data, read the image size of each image data, and create the individual information range based on the image size.

[0098] The amount of demand information is compared with a preset information threshold. When the amount of demand information reaches the preset threshold, all demand processing information and corresponding feedback information within the individual information range are read, and a neural network model is trained as the individual model of the demand party.

[0099] In one example of the technical solution of this invention, the analysis process for each demander is described. For each demander, the demand processing information of each demander at different times is statistically analyzed, that is, the historical demand processing information, and the data volume of the demand processing information is calculated as the demand information volume. Then, when the demand processing information is text data, the vector range is queried, and the union of the vector ranges at all times is calculated to obtain the individual information range. When the demand processing information is image data, the image size of each image data is read, and the individual information range is created according to the image size. This process is similar to the generation process of the global information range, and will not be described again here.

[0100] The amount of demand information is compared with a preset information threshold. When the amount of demand information reaches the preset threshold, all demand processing information and corresponding feedback information within the individual information range are read to train a neural network model as the individual model for that demander. This process is called model distillation. A small model is trained from the input and output data of a large processing model. When the model size is small, its response speed is faster, but it can only process data within a smaller range. This is a lightweight model unique to that demander.

[0101] Regarding step S400, the step of statistically analyzing individual models from different demanders and merging these individual models based on the demand information of each demander to obtain a model group for each processing model includes:

[0102] The scope of individual information queries for different demanders;

[0103] Calculate the intersection-union ratio (IUU) of the individual information range, and determine the similarity of the demand side based on the IUU.

[0104] Based on the similarity, the demand sides are clustered, and the individual models of each demand side are merged to obtain model groups;

[0105] Query the processing model corresponding to each model in the model group, select the mode processing model, and connect the model group with the mode processing model.

[0106] In one example of the technical solution of this invention, the process of merging individual models is described. The individual information ranges of different demanders are queried, and the intersection-union ratio (IUR) of the individual information ranges is calculated. The similarity of the demanders is determined based on the IUR; a larger IUR indicates a higher degree of overlap between the two individual information ranges. Therefore, the IUR is used as the similarity of the demanders. The demanders are then clustered based on the similarity. A conventional clustering scheme, such as k-means clustering, can be used. For each group of demanders, their individual models are merged, and the merged individual models are statistically analyzed to obtain a model group. The merging process involves combining two sufficiently similar models into one. Within a class of models, there may be some models with extremely high similarity; in this case, only one needs to be retained. The merging process can either retain only one model or merge its sample sets to obtain a new model. The several models obtained after merging all sufficiently similar individual models within the same class are called a model group.

[0107] The model is queried sequentially in the model group, and the corresponding processing model is queried in the processing model library (matching is performed based on the format of the input information or the architecture of the model). After the query is completed, each processing model corresponds to a batch of individual models. The processing model with the largest number of corresponding models is called the mode processing model. The model group is connected with the mode processing model, that is, the model group is connected with the corresponding processing model in the processing model library. In this way, the processing model library is expanded.

[0108] Figure 2 A structural diagram of a MaaS multi-model release management system is shown. In a preferred embodiment of the technical solution of the present invention, a MaaS multi-model release management system is also provided, the system 10 comprising:

[0109] The demand processing module 11 is used to receive demand processing information uploaded by the demander, match a processing model in a preset processing model library according to the demand processing information, generate feedback information based on the matched processing model, and send it to the demander.

[0110] The model library update module 12 is used to collect the demand processing information of all demanders within a preset time period, determine the global information range, verify the model uploaded by any uploader based on the global information range, and update the processing model library according to the verification results.

[0111] The model distillation module 13 is used to collect historical demand processing information for each demander, determine the amount of demand information and the range of individual information for each demander, and when the amount of demand information reaches a preset threshold, the processing model is distilled according to the range of individual information to obtain the individual model of the demander.

[0112] The model merging module 14 is used to count the individual models of different demanders, and merge the individual models based on the amount of demand information of each demander to obtain a model group for each processing model.

[0113] Furthermore, the demand processing module 11 includes:

[0114] The information receiving unit is used to receive demand processing information uploaded by the demander.

[0115] The task segmentation unit is used to perform format recognition on the requirement processing information and segment the requirement processing information into sub-tasks based on the format recognition result.

[0116] The time consumption prediction unit is used to match the processing model in the preset processing model library according to the type of the subtask, and synchronously obtain the predicted processing time of the matched processing model.

[0117] The results statistics unit is used to select the processing model with the shortest prediction processing time, identify sub-tasks, obtain processing results, statistically analyze all processing results, generate feedback information, and send it to the requesting party.

[0118] Specifically, the model library update module 12 includes:

[0119] The global information statistics unit is used to collect demand processing information from all demanders within a preset time period.

[0120] The text segmentation unit is used to segment the text data in the information to be processed into words, resulting in a word list; the word list contains word entries and word count entries.

[0121] The global scope determination unit is used to query the word vector of any word, determine the vector radius of the word vector based on the number of words, obtain the vector scope of each word, calculate the union of the vector scopes of each word, and obtain the global information scope.

[0122] The image analysis unit is used to obtain the image size of the image data in the information to be processed, and to create a global information range based on the image size;

[0123] The model verification unit is used to verify the model uploaded by any uploader based on the global information scope, and update the processing model library according to the verification results.

[0124] Furthermore, the model distillation module 13 includes:

[0125] The individual information statistics unit is used to collect the demand processing information of each demander at different times for each demander.

[0126] The information calculation unit is used to calculate the amount of data for demand processing information, which is then used as the demand information amount.

[0127] The individual range determination unit is used to query the vector range and calculate the union of the vector ranges at all times when the information to be processed is text data, so as to obtain the individual information range; when the information to be processed is image data, it reads the image size of each image data and creates the individual information range according to the image size.

[0128] The distillation execution unit is used to compare the amount of demand information with a preset information threshold. When the amount of demand information reaches the preset threshold, it reads all demand processing information and corresponding feedback information within the individual information range, trains a neural network model, and uses it as the individual model for that demand party.

[0129] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A MaaS multi-model release management method, characterized in that, The method includes: Receive the demand processing information uploaded by the demand party, match the processing model in the preset processing model library according to the demand processing information, generate feedback information based on the matched processing model, and send it to the demand party. Collect demand processing information from all demanders within a preset time period, determine the global information scope, verify the model uploaded by any uploader based on the global information scope, and update the processing model library according to the verification results. The historical demand processing information of each demander is statistically analyzed to determine the amount of demand information and the range of individual information for each demander. When the amount of demand information reaches a preset threshold, the processing model is distilled according to the range of individual information to obtain the individual model of that demander. Individual models of different demanders are statistically analyzed, and individual models are merged based on the amount of demand information of each demander to obtain model groups for each processing model. The steps of receiving the demand processing information uploaded by the demand party, matching the processing model in a preset processing model library according to the demand processing information, generating feedback information based on the matched processing model, and sending it to the demand party include: Receive request processing information uploaded by the requester; The requirement processing information is formatted and then divided into sub-tasks based on the format recognition results. Based on the type of subtask, a processing model is matched in the preset processing model library, and the predicted processing time of the matched processing model is obtained synchronously. Select the processing model with the shortest prediction processing time, identify the sub-tasks, obtain the processing results, statistically analyze all processing results, generate feedback information, and send it to the requesting party. The steps of statistically analyzing the demand processing information of all demanders within a preset time period, determining the global information scope, verifying the model uploaded by any uploader based on the global information scope, and updating the processing model library according to the verification results include: Statistics on the processing information of all demanders within a preset time period; The text data in the requirement processing information is segmented into words to obtain a word list; the word list contains word entries and word count entries. For any word, query the word vector of the word, determine the vector radius of the word vector according to the number of words, obtain the vector range of each word, calculate the union of the vector ranges of each word, and obtain the global information range; For image data in the demand processing information, obtain the image size of the image data, and create a global information range based on the image size; The model uploaded by any uploader is verified based on the global information scope, and the processing model library is updated according to the verification results. The step of verifying the model uploaded by any uploader based on the global information scope, and updating the processing model library according to the verification results includes: The model is received from the uploader via the model receiving port; Randomly generate text and image data of a preset amount, and remove data that is included in the global information scope; Input the data into the uploaded model to determine the processing speed; The processing speed of all data is statistically analyzed, and when the processing speed meets the preset speed conditions, the model is inserted into the processing model library. The steps of statistically analyzing the historical demand processing information of each demander, determining the amount of demand information and the range of individual information for each demander, and performing model distillation on the processing model based on the range of individual information to obtain the individual model for that demander include: For each demander, statistical analysis is performed on the demand processing information of each demander at different times. The amount of data for calculating the demand processing information is taken as the demand information volume. When the information to be processed is text data, query the vector range, calculate the union of the vector ranges at all times, and obtain the individual information range; when the information to be processed is image data, read the image size of each image data, and create the individual information range based on the image size. The amount of demand information is compared with a preset information threshold. When the amount of demand information reaches the preset threshold, all demand processing information and corresponding feedback information within the individual information range are read, and a neural network model is trained as the individual model of the demander.

2. The MaaS multi-model release management method according to claim 1, characterized in that, The steps of statistically analyzing individual models from different demanders and merging these individual models based on the demand information of each demander to obtain a model group for each processing model include: The scope of individual information queries for different demanders; Calculate the intersection-union ratio (IUU) of the individual information range, and determine the similarity of the demand side based on the IUU. Based on the similarity, the demand sides are clustered, and the individual models of each demand side are merged to obtain model groups; Query the processing model corresponding to each model in the model group, select the mode processing model, and connect the model group with the mode processing model.

3. A MaaS multi-model release management system, characterized in that, The system includes: The demand processing module is used to receive demand processing information uploaded by the demander, match a processing model in a preset processing model library according to the demand processing information, generate feedback information based on the matched processing model, and send it to the demander. The model library update module is used to collect the demand processing information of all demanders within a preset time period, determine the global information scope, verify the model uploaded by any uploader based on the global information scope, and update the processing model library according to the verification results. The model distillation module is used to collect historical demand processing information for each demander, determine the amount of demand information and the range of individual information for each demander, and when the amount of demand information reaches a preset threshold, the processing model is distilled according to the range of individual information to obtain the individual model of that demander. The model merging module is used to count individual models from different demanders and merge individual models based on the amount of demand information from each demander to obtain model groups for each processing model. The demand processing module includes: The information receiving unit is used to receive demand processing information uploaded by the demander. The task segmentation unit is used to perform format recognition on the requirement processing information and segment the requirement processing information into sub-tasks based on the format recognition result. The time consumption prediction unit is used to match the processing model in the preset processing model library according to the type of the subtask, and synchronously obtain the predicted processing time of the matched processing model. The results statistics unit is used to select the processing model with the shortest prediction processing time, identify sub-tasks, obtain processing results, statistically analyze all processing results, generate feedback information, and send it to the requesting party. The model library update module includes: The global information statistics unit is used to collect demand processing information from all demanders within a preset time period. The text segmentation unit is used to segment the text data in the information to be processed into words, resulting in a word list; the word list contains word entries and word count entries. The global scope determination unit is used to query the word vector of any word, determine the vector radius of the word vector based on the number of words, obtain the vector scope of each word, calculate the union of the vector scopes of each word, and obtain the global information scope. The image analysis unit is used to obtain the image size of the image data in the information to be processed, and to create a global information range based on the image size; The model verification unit is used to verify the model uploaded by any uploader based on the global information scope, and update the processing model library according to the verification results. The process of validating models uploaded by any uploader based on a global information scope, and updating the content of the model library according to the verification results, includes: The model is received from the uploader via the model receiving port; Randomly generate text and image data of a preset amount, and remove data that is included in the global information scope; Input the data into the uploaded model to determine the processing speed; The processing speed of all data is statistically analyzed, and when the processing speed meets the preset speed conditions, the model is inserted into the processing model library. The process involves statistically analyzing the historical demand processing information of each demander to determine the amount of demand information and the range of individual information for each demander. When the amount of demand information reaches a preset threshold, the processing model is distilled based on the range of individual information to obtain the content of the individual model for that demander, which includes: For each demander, statistical analysis is performed on the demand processing information of each demander at different times. The amount of data for calculating the demand processing information is taken as the demand information volume. When the information to be processed is text data, query the vector range, calculate the union of the vector ranges at all times, and obtain the individual information range; when the information to be processed is image data, read the image size of each image data, and create the individual information range based on the image size. The amount of demand information is compared with a preset information threshold. When the amount of demand information reaches the preset threshold, all demand processing information and corresponding feedback information within the individual information range are read, and a neural network model is trained as the individual model of the demander.

4. A storage medium, characterized in that, The storage medium stores at least one piece of program code, which, when loaded and executed by the processor, implements the MaaS multi-model release management method as described in claim 1 or 2.

Images