WO2020143610A1

WO2020143610A1 - Data processing method and apparatus, computer device, and storage medium

Info

Publication number: WO2020143610A1
Application number: PCT/CN2020/070651
Authority: WO
Inventors: 何德裕
Original assignee: 鲁班嫡系机器人（深圳）有限公司
Priority date: 2019-01-07
Filing date: 2020-01-07
Publication date: 2020-07-16
Also published as: CN109886415A

Abstract

A data processing method and apparatus, a computer device, and a storage medium. The method comprises: obtaining data to be processed (202); inputting the data into a data processing model (204); obtaining a preprocessing result output by each preprocessing sub-model in the data processing model (206); counting a pre-determination probability corresponding to each pre-processing result (208); and generating a processing result corresponding to the data according to the pre-determination probability corresponding to each preprocessing result (210). According to the pre-determination probability corresponding to each preprocessing result, the consistency of a plurality of preprocessing results can be verified, and the processing result corresponding to the data to be processed is generated according to each pre-determination probability, so that the model data processing accuracy is improved.

Description

Data processing method, device, computer equipment and storage medium

Technical field

This application relates to the field of computer technology, and in particular, to a data processing method, device, computer equipment, and storage medium.

Background technique

With the development of computer technology, machine learning technology has emerged. In machine learning, you first need to build a model, provide training data to the model for training, and use the trained model to predict unknown data. Machine learning is the core of artificial intelligence and has been widely used in recognition and classification.

However, in traditional machine learning techniques, in order to improve the accuracy of the model in processing input data, a large amount of training data is often input to the model during training so that the training data can cover various situations. Even so, there is still the possibility of errors in the data processing of a single model, and the processing results of the model can only be checked manually by humans with low accuracy.

Summary of the invention

Based on this, it is necessary to provide a data processing method, device, computer device, and storage medium that can improve the accuracy of data processing by the model in view of the above technical problems.

A data processing method, the method includes:

Obtain pending data;

Input the data to be processed into a data processing model;

Obtaining preprocessing results respectively output by each preprocessing sub-model in the data processing model;

Count the pre-judgement probability corresponding to each pre-processing result;

The processing results corresponding to the data to be processed are generated according to the pre-judgment probabilities corresponding to the respective preprocessing results.

A data processing device, the device includes:

Data acquisition module for acquiring data to be processed;

A data input module for inputting the data to be processed into a data processing model;

A result obtaining module, configured to obtain the preprocessing results respectively output by each preprocessing sub-model in the data processing model;

Probability statistics module, used to count the pre-judgment probability corresponding to each pre-processing result;

The result generation module is configured to generate a processing result corresponding to the data to be processed according to the pre-judgement probabilities corresponding to the respective pre-processing results.

A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor implements the computer program to implement the following steps:

Obtain pending data;

Input the data to be processed into a data processing model;

A computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the following steps are realized:

Obtain pending data;

Input the data to be processed into a data processing model;

Count the pre-judgment probability corresponding to each pre-processing result;

The above data processing method, device, computer equipment and storage medium obtain data to be processed, input the data to be processed into multiple pre-processing sub-models in the data processing model, and the multiple pre-processing sub-models simultaneously process the data to be processed; The pre-processing results output by each pre-processing sub-model separately, and the pre-judgement probability corresponding to each pre-processing result is counted; according to the pre-judgement probability corresponding to each pre-processing result, the consistency of multiple pre-processing results can be verified and Each prediction probability generates a processing result corresponding to the data to be processed, which improves the accuracy of model data processing.

BRIEF DESCRIPTION

FIG. 1 is an application environment diagram of a data processing method in an embodiment;

2 is a schematic flowchart of a data processing method in an embodiment;

3 is a schematic flowchart of the steps of training an initial sub-model in an embodiment;

4 is a schematic flowchart of steps of training an initial sub-model in another embodiment;

5 is a schematic flowchart of steps of constructing a data processing model in an embodiment;

6 is a schematic structural diagram of a structural source model in an embodiment;

7 is a schematic flowchart of steps of generating a processing result in an embodiment;

8 is a schematic flowchart of steps for generating an exception notification in an embodiment;

9 is a schematic diagram of training data in an embodiment;

10 is a schematic diagram of data processing in an embodiment;

11 is a structural block diagram of a data processing device in an embodiment;

FIG. 12 is an internal structure diagram of a computer device in an embodiment.

detailed description

In order to make the purpose, technical solutions and advantages of the present application more clear, the following describes the present application in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

The data processing method provided by the present application may be applied to the application environment shown in FIG. 1, and the application environment may include a terminal 102 and a server 104, and the terminal 102 communicates with the server 104 through a network. This method can be applied to both the terminal 102 and the server 104. Among them, the terminal 102 may be, but not limited to, various industrial computers, personal computers, notebook computers, smart phones, and tablet computers. The server 104 may be implemented by an independent server or a server cluster composed of multiple servers.

In one embodiment, as shown in FIG. 2, a data processing method is provided. The method is applied to the terminal in FIG. 1 as an example for illustration, and includes the following steps:

Step 202: Obtain data to be processed.

Among them, the data to be processed is the data input to the model for processing when the model is used.

Specifically, the terminal obtains the data processing instruction triggered by the user, analyzes the data processing instruction, and obtains the storage address of the data to be processed in the data processing instruction. The terminal accesses the storage space corresponding to the storage address, and extracts the stored data to be processed from the accessed storage space.

In one embodiment, the terminal acquires the entered data identifier, generates a data acquisition request according to the data identifier, and sends the data acquisition request to the server through the network. The server receives the data acquisition request, extracts the data to be processed from the database according to the data identifier in the data acquisition request, and sends the data to be processed to the terminal through the network.

In one embodiment, the terminal is equipped with an image acquisition device. After the terminal acquires the data processing instruction, the image acquisition device is started, and the terminal uses the image data collected by the image acquisition device as data to be processed.

Step 204: Input the data to be processed into the data processing model.

Among them, the data processing model is a model composed of multiple pre-processing sub-models, which is used to process the input data to be processed.

Specifically, after acquiring the data to be processed, the terminal triggers a data input instruction, and inputs the acquired data to be processed into the data processing model according to the data input instruction.

Step 206: Obtain the preprocessing results respectively output by each preprocessing sub-model in the data processing model.

Among them, the preprocessing result is the processing result of the data to be processed in the preprocessing sub-model in the data processing model.

Specifically, the data processing model is composed of multiple preprocessing sub-models. After the terminal inputs the data to be processed into the data processing model, each pre-processing sub-model in the data processing model will process the data to be processed and output its respective pre-processing results. In the data processing model obtained by the terminal, each pre-processing sub-model preprocesses the data to be processed.

In step 208, the pre-judgment probability corresponding to each pre-processing result is counted.

The pre-judgement probability is the probability that each pre-processing result appears in the pre-processing results output by all pre-processing sub-models, and may be the ratio of the number of times each pre-processing result appears to the total number of pre-processing results.

Specifically, the terminal reads the preprocessing results output by each preprocessing sub-model, and counts the number of occurrences of different preprocessing results. The terminal calculates the ratio of the number of occurrences of each pre-processing result to the total number of pre-processing results, and uses the calculated ratio as the prediction probability of each pre-processing result.

Step 210: Generate a processing result corresponding to the data to be processed according to the pre-judgement probability corresponding to each pre-processing result.

The processing result is the output result of the data processing model to the input data to be processed.

Specifically, the terminal statistically obtains the pre-judgement probabilities corresponding to each pre-processing result, obtains the preset probability conditions, compares the pre-judgement probabilities corresponding to the pre-processing results with the pre-set probability conditions one by one, and selects The pre-processing result corresponding to the prediction probability.

When the terminal filters out the preprocessing result corresponding to the pre-judgement probability that meets the preset probability condition, the terminal completes the processing of the data to be processed, and uses the filtered preprocessing result as the processing result corresponding to the input data to be processed.

Wherein, the preset probability condition is a condition set in advance for screening specific pre-processing results from each pre-processing result. The preset probability condition may be that the predicted probability is greater than or equal to a preset probability threshold.

In one embodiment, the terminal sorts the pre-judgement probabilities of the pre-processing results through a sorting algorithm. After the sorting is completed, the pre-processing result corresponding to the highest pre-judgement probability is selected. The sorting algorithm may be at least one of bubble sorting, selection sorting, and merge sorting.

In one embodiment, the terminal generates a processing result notification according to each preprocessing result and the pre-judgment probability and processing result corresponding to each preprocessing result, and displays the processing result notification through the display screen.

In one embodiment, the data processing model may be any kind of model, and the data to be processed may be any kind of data corresponding to the data processing model. For example, when the data processing model is a recognition model or a classification model, the data to be processed may be image data, and the processing result of the data processing model is the identified product defect, item type, and quantity. When the data processing model is a trajectory planning model, the data to be processed may be image data or posture data of the object, and the processing result is the moving path trajectory of the object.

In this embodiment, by acquiring the data to be processed, the data to be processed is input into multiple pre-processing sub-models in the data processing model, and the multiple pre-processing sub-models simultaneously process the data to be processed; each pre-processing sub-model is obtained and output separately The pre-processing results of each pre-processing result are counted, and the pre-judging probabilities corresponding to each pre-processing result are counted; according to the pre-judge probabilities corresponding to each pre-processing result, the consistency of multiple pre-processing results can be verified, and the to-be-processed is generated according to each pre-judgement probability The processing results corresponding to the data improve the accuracy of model data processing.

As shown in FIG. 3, in one embodiment, before step 202, a step of training an initial sub-model is further included. This step specifically includes the following steps:

Step 302: Acquire multiple different initial sub-models and training data.

Among them, the initial sub-model is the initial model without parameter adjustment. Training data are data samples used to train the initial sub-model.

Specifically, before using the model, a machine learning model for data processing needs to be trained first. The terminal obtains the model training instruction triggered by the user, analyzes the model training instruction, and obtains the initial sub-model storage address and the training data storage address. The terminal reads multiple initial sub-models and training data from the storage space according to the initial sub-model storage address and the training data storage address, and loads the read initial sub-models and training data into the memory. The multiple initial sub-models may be different types of models, or models of the same type but with different initial model parameters, or different types of models and models of the same type but with different initial model parameters.

Step 304: Train each initial sub-model with training data to obtain multiple pre-processing sub-models.

Among them, the pre-processing sub-model is the model obtained after the initial sub-model training is completed.

Specifically, the terminal inputs the training data into each initial sub-model. The initial sub-model is trained according to the input training data, and the model parameters are adjusted until the training stop condition is met, and the training is stopped, and multiple pre-processing sub-models are obtained. The training data may include label data. The initial sub-model outputs initial results based on the training data. The initial results and label data determine the prediction error. If the prediction error is greater than or equal to the preset error threshold, the model parameters are adjusted in the direction that minimizes the prediction error, and Iteratively execute the above training process until the prediction error is less than the preset error threshold, then stop training.

In one embodiment, the training stop condition may be the number of iterations of the initial sub-model in training. When the number of iterations of the initial sub-model during training is greater than or equal to the preset threshold of the number of iterations, the training is stopped and the pre-processed sub-model is obtained.

Step 306: Construct a data processing model according to multiple pre-processing sub-models.

Specifically, after the terminal trains each initial sub-model, multiple pre-processing sub-models are obtained. Based on each pre-processing sub-model, the terminal forms a pre-processing sub-model cluster, and uses the obtained pre-processing sub-model cluster as a data processing model.

In this embodiment, multiple different initial sub-models and training data are obtained, and the same training data is input into multiple different initial sub-models for training to obtain multiple pre-processing sub-models, which improves the reliability of model training , Based on multiple pre-processing sub-models to build data processing models, improving the efficiency of obtaining data processing models.

As shown in FIG. 4, in another embodiment, before step 202, a step of training an initial sub-model is further included. This step specifically includes the following steps:

Step 402: Acquire multiple identical initial submodels and training data.

Specifically, the terminal obtains the model training instruction, and extracts the initial sub-model storage address, the training data storage address, and the training parameters from the model training instruction. The terminal reads the initial sub-model and training data from the storage space according to the initial sub-model storage address and the training data storage address. The terminal copies the initial sub-model according to the number of preset sub-models in the training parameters to obtain multiple identical initial sub-models that match the number of preset sub-models.

Step 404: Extract multiple training sample sets corresponding to multiple initial sub-models from the training data.

Specifically, for each initial sub-model, the terminal extracts part of the data from the training data in a preset manner, and obtains the training sample set according to the extracted part of the data. The terminal may randomly extract partial data from the training data, and construct a training sample set according to the randomly extracted partial data.

In one embodiment, the terminal divides the training data to obtain multiple training data subsets that match the number of initial sub-models, and uses the multiple training data subsets as multiple trainings corresponding to the multiple initial sub-models. Sample set. For example, there are 5 initial sub-models, the training data contains 10000 pictures, the terminal according to the first 1-2000, 2001-4000, 4001-6000, 6001-8000, 8001-10000 5 partitions of the training data are obtained, and the five training data subsets are used as the training sample set corresponding to the five initial submodels.

Step 406: Train the corresponding initial sub-model according to each training sample set to obtain multiple pre-processing sub-models.

Specifically, for each initial sub-model, the terminal obtains a training sample set corresponding to the initial sub-model, and trains the initial sub-model according to the training sample set. After the terminal trains the multiple initial sub-models, multiple pre-processing sub-models are obtained.

In one embodiment, when the terminal trains the initial sub-model, the training data to be input into the initial sub-model is divided into multiple groups. The terminal first enters a set of training data into the initial sub-model to obtain the initial result of the initial sub-model output, compares the initial result with the label data and calculates the prediction error, adjusts the model parameters according to the prediction error; then inputs another set of training data to adjust After the initial sub-model, repeat the above process until the prediction error converges to obtain the pre-processing sub-model.

Step 408: Construct a data processing model according to multiple pre-processing sub-models.

Specifically, after the terminal trains multiple identical initial sub-models, multiple pre-processing sub-models are obtained. The terminal constructs a sub-model cluster according to multiple pre-processing sub-models, and uses the constructed sub-model cluster as a data processing model.

In this embodiment, multiple identical initial submodels and training data are acquired, and multiple training sample sets corresponding to the multiple initial submodels are extracted from the training data. For multiple identical initial submodels, control variables are used. The method of inputting different training sample sets for training respectively obtains multiple pre-processing sub-models, which improves the reliability of model training. Building a data processing model based on multiple pre-processing sub-models improves the efficiency of obtaining data processing models.

As shown in FIG. 5, in one embodiment, step 204 specifically includes the step of building a data processing model. The step specifically includes the following steps:

Step 502: Extract multiple different model substructures from the trained structure source model.

Among them, the model substructure is a substructure randomly extracted from the source model of the structure, and can normally perform the function of the model. The structure of the structure source model contains multiple layers of links, and the data to be processed can be processed.

Specifically, the terminal pre-stores a trained structure source model. The terminal parses the obtained data processing instruction to obtain the storage address of the structure source model, and reads the structure source model from the storage space according to the storage address of the structure source model. After reading the structure source model, the terminal performs multiple extractions from the structure source model, and each time a model substructure is extracted. During each extraction, the terminal randomly extracts a link from the structure source model to obtain multiple different model substructures.

In one embodiment, the structure of the structure source model is shown in FIG. 6. The structure source model may be a neural network model, including an input layer, a hidden layer, and an output layer. The terminal can extract a certain link between any two layers to obtain multiple different model substructures. For example, the terminal may remove the link from x ₁ to h ₁ in FIG. 6 or the link from h ₂ to o ₂ .

In one embodiment, before acquiring the data to be processed, the terminal first obtains the structure source model through training. The terminal obtains the initial sub-model and training data according to the model training instructions, and trains the initial sub-model according to the training data. After the training, the structure source model is obtained.

Step 504: Use each model substructure as a preprocessing submodel in the data processing model to construct a data processing model.

Specifically, after the terminal obtains a plurality of different model substructures through extraction, each model substructure is used as a preprocessing submodel. The terminal constructs a sub-model cluster according to multiple pre-processing sub-models to obtain a data processing model.

Step 506: Input the data to be processed into each pre-processing sub-model in the data processing model.

Specifically, after the data processing model is constructed by the terminal, the data to be processed is copied according to the number of pre-processing sub-models to obtain multiple pieces of the same data to be processed matching the number of pre-processing sub-models. The terminal inputs each piece of data to be processed into each preprocessing sub-model in the data processing model.

In one embodiment, the pre-stored models in the terminal are identified as the first model, the second model, and/or the third model, respectively. Among them, the first model and the second model are data processing models, each pre-processing sub-model of the first model is trained by multiple different initial sub-models, and each pre-processing sub-model of the second model is obtained by multiple identical initial sub-models The model is trained. The third model is the structure source model. After the terminal obtains the data to be processed, it obtains the model identifier of the pre-stored model. When the obtained model identifier is the first model identifier or the second model identifier, the terminal inputs the data to be processed into each of the first model or the second model. Processing sub-models; when the acquired model identifier is the third model identifier, the terminal extracts multiple different model sub-structures from the third model, and uses each model sub-structure as a pre-processing sub-model to construct a data processing model. The data to be processed is input to each pre-processing sub-model in the data processing model. The source of each pre-processing sub-model of the data processing model may be one of the first model, the second model, and/or the third model.

In this embodiment, multiple different model substructures are obtained by extracting from the trained structure source model, each model substructure is used as a preprocessing submodel, a data processing model is constructed, and data to be processed are input into the data processing model respectively Each pre-processing sub-model in. By randomly extracting the model substructure from the source model of the structure and using the model substructure as the preprocessing submodel, the reliability of the selected preprocessing submodel is guaranteed.

As shown in FIG. 7, in one embodiment, step 210 specifically includes the step of generating a processing result. The step specifically includes the following steps:

In step 702, from each preprocessing result, a preprocessing result corresponding to a pre-judgement probability that meets a preset probability condition is selected.

Specifically, each pre-processing sub-model can obtain multiple candidate pre-processing results and candidate probabilities corresponding to each candidate pre-processing result. The terminal reads the pre-stored preset conversion probabilities. For each pre-processing sub-model, the terminal may use the candidate pre-processing result with a candidate probability greater than or equal to the preset conversion probability as the pre-processing result of the pre-processing sub-model. The terminal counts the pre-judgement probability corresponding to each pre-processing result of each pre-processing sub-model, and obtains the preset probability condition, compares the pre-judgement probability corresponding to the pre-processing result with the preset probability condition one by one, and selects the preset probability The preprocessing result corresponding to the conditional prediction probability.

Step 704: Calculate the uncertainty of the screened preprocessing result.

Among them, the uncertainty is a quantitative evaluation value of the uncertainty of the pre-processing results screened by the terminal. The lower the uncertainty, the higher the reliability of the preprocessing results.

Specifically, the terminal may add uncertainty to the filtered preprocessing results based on the Bayesians theory. The terminal obtains a preset uncertainty calculation method, and calculates the uncertainty of the screened preprocessing result according to the uncertainty calculation method and the obtained preprocessing results or candidate preprocessing results.

Step 706: Generate a processing result corresponding to the data to be processed according to the uncertainty and the filtered preprocessing result.

Specifically, the terminal uses the filtered preprocessing result as the processing result corresponding to the input to-be-processed data. When the terminal displays the processing result, it simultaneously displays the uncertainty corresponding to the processing result. The terminal can also obtain the pre-judgement probability of the pre-processed result screened, and simultaneously display the processing result, the pre-judgement probability corresponding to the processing result and the uncertainty.

In one embodiment, the terminal may be based on frequency theory (Frequentists) theory, do not calculate the uncertainty, output the determined processing result, and output the pre-filtered result and the pre-judged probability.

In one embodiment, when the calculated uncertainty is less than the preset uncertainty threshold, the terminal displays the filtered preprocessing result and the calculated uncertainty.

The terminal calculates the uncertainty based on the Bayesian theory, and provides a quantitative evaluation value for the credibility of the preprocessing results. When the uncertainty is greater than or equal to the preset uncertainty threshold, the terminal displays a notification of the uncertainty of abnormality. The uncertainty anomaly notification can remind the user whether the input data to be processed has been used to train the initial sub-model or structure source model, and remind the user to supplement the training data; the uncertainty anomaly notification can also remind the user, the preset uncertainty threshold Whether it is reasonable, and remind the user to reset the uncertainty threshold.

Table 1:

For example, the data processing model includes four pre-processing sub-models. The data to be processed is image data. Each pre-processing sub-model obtains two candidate pre-processing results. The candidate pre-processing results are the identified animal species. The candidate pre-processing results of each pre-processing sub-model can be shown in Table 1, where A, B and C respectively represent three types of candidate pre-processing results, the numbers in the table are the candidate probabilities of the candidate pre-processing results, and the type C pre-processing results It is obtained by converting the pre-processing result of the A-type candidate or the pre-processing result of the B-type candidate according to the preset conversion probability. The terminal takes the result of the candidate preprocessing result of class C as the preprocessing result of the preprocessing sub-model. For example, in the type A candidate preprocessing result of the preprocessing submodel 1, the candidate probability of the candidate preprocessing result is cat, the candidate probability of the candidate preprocessing result is dog is 0.1, and the preset conversion probability may be 0.5; 0.9 is greater than 0.5, so after converting the A-type candidate pre-processing result to the C-type candidate pre-processing result, the candidate probability for the candidate pre-processing result is cat, and the candidate probability for the dog pre-processing result is 0. The terminal uses the cat as the pre-processing result of the pre-processing sub-model 1.

Assuming that the candidate pre-processing results and candidate probabilities obtained by each pre-processing sub-model in two uses are the A-type candidate pre-processing results and the B-type candidate pre-processing results in Table 1, the selected pre-processing results are cats, And the prediction probability is 75%. Assume that the terminal calculates the uncertainty in the class A candidate preprocessing result is 0.2; in the class B candidate preprocessing result, the uncertainty is 0.1. For the class A candidate preprocessing result or the class B candidate preprocessing result, when based on the frequency school theory, the terminal outputs the determined processing result, and outputs "the preprocessed result selected is a cat and the prediction probability is 75%". Based on the Bayesian theory, for the A-type candidate pre-processing results, the terminal outputs "screened pre-processing results are cats, the prediction probability is 75% and the uncertainty is 0.2"; for the B-type candidate pre-processing results ，The terminal outputs “The pre-screened result is cat, the prediction probability is 75% and the uncertainty is 0.1”. When based on frequency school theory, the results of class A candidate preprocessing are the same as those of class B candidate; when based on Bayesian theory, the certainty of class B candidate preprocessing results is greater than the certainty of class A candidate preprocessing results .

In one embodiment, when the terminal calculates the uncertainty based on the Bayesian theory, it can first convert the A-type candidate pre-processing results or the B-type candidate pre-processing results into the C-type candidate pre-processing results, and calculate Certainty:

Uncertainty = 1-the number of occurrences of the pre-processing results with the most occurrences / total number of pre-processing results

For example, after the terminal converts the A-type candidate pre-processing results into C-type candidate pre-processing results, the pre-processing results of each pre-processing sub-model are "cat, cat, cat, dog" respectively, and the most frequent pre-processing results are cats , The number of occurrences is 3, and the total number of preprocessing results is 4. According to formula 1, the uncertainty is: 1-3/4=0.25. When calculating the uncertainty through formula 1, the minimum uncertainty is 0 and the maximum uncertainty is 0.5.

In one embodiment, the terminal calculates the uncertainty according to Equation 2:

Uncertainty = final uncertainty-the average value of the uncertainty of each pre-processing sub-model (2)

Among them, the terminal calculates the average value of the uncertainty of each pre-processing sub-model through Equation 3:

The average value of the uncertainty of each pre-processing sub-model = the sum of the uncertainty of each pre-processing sub-model / the total number of pre-processing sub-models (3)

Among them, the terminal calculates the uncertainty H of each pre-processing sub-model through Equation 4:

Where, p _i is the candidate probability of the candidate pre-processing result in the pre-processing sub-model, and i is a positive integer, which is the number of candidate pre-processing results of the pre-processing sub-model. For example, the candidate pre-processing results of a pre-processing sub-model include cats and dogs, where the cat candidate probability is 0.5 and the dog candidate probability is 0.5, then the uncertainty of the pre-processing sub-model h=-[0.5log( 0.5)+0.5log(0.5)=1. When the cat's candidate probability is 0 and the dog's candidate probability is 1, the uncertainty is h=-[0×log0+1×log1=0. The range of uncertainty is [0,1]. After obtaining the uncertainty of each pre-processing sub-model separately, the terminal adds the uncertainty of each pre-processing sub-model to obtain the addend sum, calculates the ratio of the add-on sum to the number of pre-processing sub-models, and uses the obtained ratio as each The average value of the uncertainty of the pre-processing submodel.

When calculating the final uncertainty, the terminal calculates the simple arithmetic average of the candidate probabilities of various candidate preprocessing results according to each preprocessing submodel

i is a positive integer, which is the number of candidate pre-processing results of the pre-processing sub-model, and then the final uncertainty is calculated according to Equation 5:

For example, in the candidate pre-processing results of Class A in Table 1, in each pre-processing sub-model, the candidate probabilities of the candidate pre-processing results are 0.9, 0.8, 0.7, and 0.1, respectively, and the simple arithmetic average

The candidate preprocessing result is that the dog's candidate probabilities are 0.1, 0.2, 0.3, and 0.9, respectively, and the simple arithmetic average

The final uncertainty H=-(0.625×log0.625+0.375×log0.375)=0.63.

The data processing model obtained by the third model from the data processing model constructed by the terminal according to the first model, the second model, and the third model, that is, a method of extracting multiple different model substructures from the trained structure source model , Can be better applied to the uncertainty based on Bayesian theory.

In this embodiment, from each preprocessing result, the preprocessing result corresponding to the pre-judgement probability that meets the preset probability condition is selected, and then the uncertainty of the pre-processing result screened is calculated, and the uncertainty reflects the screening to The credibility of the pre-processing results of the system; when generating the processing results corresponding to the data to be processed according to the pre-processing results, the uncertainty is added to improve the accuracy of the processing results output by the model.

As shown in FIG. 8, in one embodiment, after step 208, a step of generating an exception notification is included. This step specifically includes the following steps:

Step 802: Obtain data to be processed.

Step 804: Input the data to be processed into the data processing model.

Step 806: Obtain the preprocessing results respectively output by each preprocessing sub-model in the data processing model.

In step 808, the pre-judge probability corresponding to each pre-processing result is counted.

Step 810: When a processing result corresponding to the data to be processed is not generated according to each pre-judgment probability, a processing exception notification is generated according to each pre-processing result.

Among them, the processing exception notification is notification information generated when the terminal does not filter the preprocessing result.

Specifically, when the terminal does not filter the preprocessing result corresponding to the pre-judgement probability that meets the preset probability condition, it cannot generate the processing result corresponding to the data to be processed. The terminal according to each preprocessing result and the corresponding preprocessing result Predict the probabilities and generate notifications for handling exceptions.

Step 812: Display the exception notification.

Specifically, after generating the processing exception notification, the terminal sends the processing exception notification to the display screen of the terminal, and displays the processing exception notification through the display screen.

In one embodiment, the processing exception notification may prompt the user to re-enter the data to be processed, so that after the user inputs new data to be processed, the new data to be processed is processed to obtain a processing result. Handling exception notifications can also prompt users to retrain the model.

In this embodiment, when the processing result corresponding to the data to be processed is not generated according to each pre-judgment probability, that is, the preprocessing result is not filtered, a processing exception notification is generated according to each preprocessing result and the processing exception notification is displayed so as to receive the input again Data to be processed improves the reliability of data processing.

The method provided by this application based on Bayesian theory and calculating the uncertainty of the preprocessing results selected from the preprocessing results of multiple preprocessing sub-models can be applied to various machine learning techniques, such as supervised learning , Semi-supervised learning, reinforcement learning and imitation learning, etc.; various machine learning techniques based on this application can solve various problems related to classification or regression in various fields, as described below:

1. Supervised learning

Take defect detection based on supervised learning as an example. Defect detection can be applied in various fields, such as defect detection of processed products (such as scratches, bubbles and integrity), AOI inspection (Automated Optical Inspection, automatic optical inspection), etc.

When performing defect detection, the pretreatment result is a defect in the product. When the uncertainty value of the screened pretreatment result is low, the accuracy of the screened pretreatment result is high; when the pretreatment screened As a result, it is judged that the product is not defective, but when the uncertainty is high, the product is likely to be defective. The reason may be that the defects in the product are marked for supervised learning, or rarely appear in the training data; this type of defect can be added to the supervised learning training data to make the model learn again. The reason may also be that the preset uncertainty threshold is set unreasonably, and the uncertainty threshold needs to be reset.

2. Semi-supervised learning

Semi-Supervised Learning (SSL) refers to the use of some unlabeled training data and the already labeled training data for model training.

During training, input some labeled training data to train the initial sub-model or the original structure source model. After the training is completed, input the data to be processed into each pre-processing sub-model and filter the pre-processing results; when the uncertainty of the filtered pre-processing results is high, the unlabeled training data needs to be labeled manually, and then Train the model.

Taking the regression problem as an example, the data processing model is applied to target object position or pose recognition to identify the pose of the target object. FIG. 9 is a schematic diagram of training data in an embodiment. Specifically, referring to FIG. 9, the uncertainty in identifying the position or pose of the target object in FIG. 9(a) is low; in FIG. 9(b), the position or position of the target object is recognized due to the occlusion relationship between the objects The pose uncertainty is high, you may need to re-label the training data.

3. Reinforcement learning and imitation learning

In reinforcement learning and imitation learning, an action (behavior) needs to be performed for the current state, and the current strategy will provide an action option for the current state (reinforcement learning and imitation learning require agents with exploration capabilities, such as robots, but the agent is doing When the trajectory is adopted, the action option may not only be executed), and the behavior value function will provide a desired return for the current state and the action option. When the behavior value function uses Bayesian inference of multiple models, it not only provides an expected return for the current state and the current action option, but also provides the uncertainty of the expected return; if the uncertainty of the expected return is low, that is After executing the current action option in the current state, the expected return of the trajectory is relatively certain, indicating that the combination of the current state and the current action option has been effectively explored in the agent's past learning process; the agent needs to be explored in reinforcement learning or imitation learning It is not recommended to select the current action option at the stage; when the uncertainty of the expected return is high, that is, the expected return of the trajectory after the current action option is executed in the current state is relatively unknown, indicating that the combination of the current state and the current action option has passed Is not fully explored during the learning process, it is recommended to choose the current action option in the stage where reinforcement learning or imitation learning requires agent exploration.

10 is a schematic diagram of data processing in an embodiment. Specifically, referring to FIG. 10, the data processing model may be an image recognition model, which is composed of four pre-processing sub-models. The data to be processed acquired by the terminal is image data, and the image data is input to 4 pre-processing sub-models in the data processing model, and the 4 pre-processing sub-models identify the animals in the image. The terminal obtains the preprocessing results of the three preprocessing sub-models as cats, with a pre-judgment probability of 75%, and the pre-processing results of one pre-processing sub-model as dogs, with a pre-judgment probability of 25%. If the preset probability condition is that the pre-judgment probability is greater than or equal to 75%, the terminal can filter out the preprocessing result that meets the preset probability condition, that is, a cat, and use the cat as a data processing model to process image data. If the preset probability condition is that the pre-judgment probability is equal to 100%, the terminal cannot select the pre-processing result.

It should be understood that although the steps in the flowcharts of FIGS. 2-5 and 7-8 are sequentially displayed according to the arrows, the steps are not necessarily executed in the order indicated by the arrows. Unless clearly stated in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least some of the steps in FIGS. 2-5 and 7-8 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. The execution order of the sub-steps or stages is not necessarily sequential, but may be executed in turn or alternately with at least a part of other steps or sub-steps or stages of other steps.

In one embodiment, as shown in FIG. 11, a data processing apparatus 1100 is provided, including: a data processing module 1102, a data input module 1104, a result acquisition module 1106, a probability statistics module 1108, and a result generation module 1110 wherein:

The data obtaining module 1102 is used to obtain data to be processed.

The data input module 1104 is used to input the data to be processed into the data processing model.

The result obtaining module 1106 is used to obtain the preprocessing results respectively output by each preprocessing sub-model in the data processing model.

The probability statistics module 1108 is used to count the pre-judgement probability corresponding to each pre-processing result.

The result generating module 1110 is configured to generate processing results corresponding to the data to be processed according to the pre-judgement probabilities corresponding to the respective pre-processing results.

In one embodiment, the data processing apparatus 1100 further includes a model training module, the model training module is used to obtain a plurality of different initial sub-models and training data; the training data is used to train each initial sub-model to obtain multiple pre-processors Model; build a data processing model based on multiple preprocessing submodels.

In another embodiment, the model training module is also used to obtain multiple identical initial sub-models and training data; extract multiple training sample sets corresponding to the multiple initial sub-models from the training data; The training sample set trains the corresponding initial sub-model to obtain multiple pre-processing sub-models; the data processing model is constructed according to the multiple pre-processing sub-models.

In one embodiment, the data input module 1104 specifically includes: a structure extraction module, a model construction module, and an input module, where:

The structure extraction module is used to extract multiple different model substructures from the trained processing model.

The model building module is used to construct the data processing model by using each model substructure as a preprocessing submodel in the data processing model.

The input module is used to input the data to be processed into each pre-processing sub-model in the data processing model.

In one embodiment, the result generation module 1110 is configured to filter out the pre-processing results corresponding to the pre-judgement probability that meets the preset probability condition from each pre-processing result; calculate the uncertainty of the screened pre-processing results; Uncertainty and the pre-processed results screened to generate processing results corresponding to the data to be processed.

In one embodiment, the data processing apparatus 1100 further includes: a notification generation module and a notification display module, where:

The notification generating module is configured to generate a processing exception notification according to each preprocessing result when the processing result corresponding to the data to be processed is not generated according to each pre-judgment probability.

Notification display module, used to display exception notifications.

For the specific limitation of the data processing device, reference may be made to the above limitation on the data processing method, and details are not described herein again. Each module in the above data processing device may be implemented in whole or in part by software, hardware, and a combination thereof. The above modules may be embedded in the hardware form or independent of the processor in the computer device, or may be stored in the memory in the computer device in the form of software so that the processor can call and execute the operations corresponding to the above modules.

In one embodiment, a computer device is provided. The computer device may be a terminal, and an internal structure diagram thereof may be as shown in FIG. The computer equipment includes a processor, a memory, a network interface, a display screen, an input device, and an image acquisition device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with external terminals through a network connection. The computer program is executed by the processor to implement a data processing method. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or may be a button, a trackball, or a touchpad provided on the computer device housing , Can also be an external keyboard, touchpad or mouse. The image collection device is used to collect image data, and the collected image data will be used as data to be processed.

Those skilled in the art can understand that the structure shown in FIG. 12 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer equipment to which the solution of the present application is applied. The specific computer equipment may Include more or less components than shown in the figure, or combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the following steps are implemented: acquiring data to be processed; The data to be processed is input into the data processing model; the preprocessing results respectively output by the preprocessing sub-models in the data processing model are obtained; the pre-judgement probabilities corresponding to each pre-processing result are counted; The processing result corresponding to the data to be processed.

In one embodiment, before acquiring the data to be processed, the processor also implements the following steps when executing the computer program: acquiring multiple different initial sub-models and training data; training each initial sub-model with training data to obtain multiple Processing sub-models; construct data processing models based on multiple pre-processing sub-models.

In another embodiment, before acquiring the data to be processed, the processor also implements the following steps when executing the computer program: acquiring multiple identical initial sub-models and training data; extracting from the training data one-to-one correspondence with the multiple initial sub-models Multiple training sample sets; train the corresponding initial sub-model according to each training sample set respectively to obtain multiple pre-processing sub-models; construct a data processing model based on multiple pre-processing sub-models.

In one embodiment, inputting the data to be processed into the data processing model includes: extracting a plurality of different model substructures from the trained structure source model; using each model substructure as a preprocessing submodel in the data processing model, Build a data processing model; input the data to be processed into each pre-processing sub-model in the data processing model.

In one embodiment, generating the processing result corresponding to the data to be processed according to the pre-judgement probability corresponding to each pre-processing result includes: filtering out the pre-judgement corresponding to the pre-judgement probability that meets the preset probability condition from each pre-processing result Processing results; Calculate the uncertainty of the screened preprocessing results; generate processing results corresponding to the data to be processed based on the uncertainty and the screened preprocessing results.

In one embodiment, after pre-judgement probabilities corresponding to each pre-treatment result are counted from each pre-treatment result, the processor also implements the following steps when the computer program is executed: when the pre-judgement probability is not generated according to each pre-judgement probability When the processing result is processed, a processing exception notification is generated according to each preprocessing result; the processing exception notification is displayed.

In one embodiment, a computer-readable storage medium is provided on which a computer program is stored. When the computer program is executed by a processor, the following steps are achieved: acquiring data to be processed; inputting the data to be processed into a data processing model; acquiring data The pre-processing results output by each pre-processing sub-model in the processing model; the pre-judgment probability corresponding to each pre-processing result is counted; according to the pre-judge probability corresponding to each pre-processing result, the processing result corresponding to the data to be processed is generated. In one embodiment, before acquiring the data to be processed, the computer program is executed by the processor to implement the following steps: acquiring multiple different initial sub-models and training data; training each initial sub-model with training data to obtain multiple Pre-processing sub-model; build a data processing model based on multiple pre-processing sub-models.

In another embodiment, before acquiring the data to be processed, the computer program is executed by the processor to implement the following steps: acquiring multiple identical initial sub-models and training data; extracting from the training data one by one with the multiple initial sub-models Corresponding multiple training sample sets; train the corresponding initial sub-model according to each training sample set respectively to obtain multiple pre-processing sub-models; construct a data processing model according to multiple pre-processing sub-models.

In one embodiment, inputting the data to be processed into the data processing model includes: extracting multiple different model substructures from the trained structure source model; using each model substructure as a preprocessing submodel in the data processing model, Build a data processing model; input the data to be processed into each pre-processing sub-model in the data processing model. In one embodiment, generating the processing result corresponding to the data to be processed according to the pre-judgement probability corresponding to each pre-processing result includes: filtering out the pre-judgement corresponding to the pre-judgement probability that meets the preset probability condition from each pre-processing result Processing results; Calculate the uncertainty of the screened preprocessing results; generate processing results corresponding to the data to be processed based on the uncertainty and the screened preprocessing results.

In one embodiment, after pre-judgement probabilities corresponding to each pre-treatment result are counted from each pre-treatment result, when the computer program is executed by the processor, the following steps are also realized: when the data to be processed is not generated according to each pre-judgement probability For the corresponding processing result, generate a processing exception notification according to each preprocessing result; display the processing exception notification.

A person of ordinary skill in the art may understand that all or part of the processes in the method of the above embodiments may be completed by instructing relevant hardware through a computer program, and the computer program may be stored in a non-volatile computer readable storage In the medium, when the computer program is executed, the process of the foregoing method embodiments may be included. Wherein, any reference to the memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and/or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

The technical features of the above embodiments can be arbitrarily combined. In order to simplify the description, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be It is considered as the scope described in this specification.

The above-mentioned embodiments only express several implementation manners of the present application, and their descriptions are more specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that, for those of ordinary skill in the art, without departing from the concept of the present application, a number of modifications and improvements can also be made, which all fall within the protection scope of the present application. Therefore, the protection scope of the patent of this application shall be subject to the appended claims.

Claims

A data processing method, the method includes:

Obtain pending data;

Input the data to be processed into a data processing model;

Obtaining preprocessing results respectively output by each preprocessing sub-model in the data processing model;

Count the pre-judgment probability corresponding to each pre-processing result;

The processing results corresponding to the data to be processed are generated according to the pre-judgment probabilities corresponding to the respective preprocessing results.
The method according to claim 1, wherein before acquiring the data to be processed, the method further comprises:

Obtain multiple different initial sub-models and training data;

Training each initial sub-model with the training data to obtain multiple pre-processing sub-models;

Construct a data processing model according to the multiple preprocessing sub-models.
The method according to claim 1, wherein before acquiring the data to be processed, the method further comprises:

Obtain multiple identical initial submodels and training data;

Extract multiple training sample sets corresponding to multiple initial submodels from the training data;

Train the corresponding initial sub-model according to each training sample set respectively to obtain multiple pre-processing sub-models;

Construct a data processing model according to the multiple preprocessing sub-models.
The method according to claim 1, wherein the inputting the data to be processed into a data processing model comprises:

Extract multiple different model substructures from the trained structure source model;

Use each model substructure as a preprocessing submodel in the data processing model to construct the data processing model;

The data to be processed is input into each pre-processing sub-model in the data processing model.
The method according to claim 1, wherein the generating a processing result corresponding to the data to be processed according to the pre-judgement probabilities corresponding to the respective preprocessing results includes:

From the pre-processing results, select the pre-processing results corresponding to the pre-judgement probability that meets the preset probability conditions;

Calculate the uncertainty of the pre-processing results screened;

Generate a processing result corresponding to the data to be processed according to the uncertainty and the filtered preprocessing result.
The method according to claim 1, wherein after counting the pre-judgement probability corresponding to each pre-processing result, the method further comprises:

When a processing result corresponding to the data to be processed is not generated according to each pre-judgment probability, a processing exception notification is generated according to each pre-processing result;

Demonstrate the handling exception notification.
A data processing device, characterized in that the device includes:

Data acquisition module for acquiring data to be processed;

A data input module for inputting the data to be processed into a data processing model;

A result obtaining module, configured to obtain the preprocessing results respectively output by each preprocessing sub-model in the data processing model;

Probability statistics module, used to count the pre-judgment probability corresponding to each pre-processing result;

The result generation module is configured to generate a processing result corresponding to the data to be processed according to the pre-judgement probabilities corresponding to the respective pre-processing results.
The device according to claim 6, wherein the data input module comprises:

Structure extraction module, used to extract multiple different model substructures from the trained processing model;

The model building module is used to construct each data processing model by using each model sub-structure as a pre-processing sub-model in the data processing model;

The input module is used to input the data to be processed into each pre-processing sub-model in the data processing model.
A computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, when the processor executes the computer program, any one of claims 1 to 6 is realized The steps of the method.
A computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are realized.