CN111753914A

CN111753914A - Model optimization method and device, electronic equipment and storage medium

Info

Publication number: CN111753914A
Application number: CN202010605252.2A
Authority: CN
Inventors: 赵雪鹏; 聂磊; 黄锋
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2020-06-29
Filing date: 2020-06-29
Publication date: 2020-10-09
Anticipated expiration: 2040-06-29
Also published as: CN111753914B

Abstract

The application discloses a model optimization method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence, deep learning and image processing. The specific implementation scheme is as follows: acquiring a training data set and a misjudgment data set of a model to be optimized; carrying out sample mining on a training data set based on a current model to be optimized to obtain first effective sample data; taking the first effective sample data and the misjudgment data as training samples, and performing iterative training on the model to be optimized; testing the first performance of the model to be optimized after iterative training based on the training sample by adopting a test data set; repeatedly executing model optimization operation until the first performance of the model to be optimized after iterative training meets a first preset condition, wherein the model optimization operation comprises the following steps: and carrying out sample mining on the training data set based on the current model to be optimized to obtain second effective sample data, taking the second effective sample data and the misjudgment data as new training samples, and carrying out iterative training on the current model to be optimized to update the model to be optimized.

Description

Model optimization method and device, electronic equipment and storage medium

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to artificial intelligence, deep learning, and image processing technologies, and in particular, to a model optimization method and apparatus, an electronic device, and a storage medium.

Background

With the development of artificial intelligence technology and the continuous expansion of the application field of artificial intelligence, higher requirements are put forward on a neural network model for executing a deep learning task. The fast iteration of the model is a method for optimizing the model by using the existing data to improve the performance of the model, and the iterative optimization of the model depends on a large amount of data, needs to consume a long time and cannot meet the requirement of fast optimization on-line. The existing data has some redundant data, which cannot optimize the model but needs to occupy more resources for training.

Disclosure of Invention

The disclosure provides a model optimization method, a model optimization device, an electronic device and a storage medium.

According to an aspect of the present disclosure, there is provided a model optimization method, including: acquiring a training data set and a misjudgment data set of a model to be optimized; carrying out sample mining on a training data set based on a current model to be optimized to obtain first effective sample data; taking the first effective sample data and the misjudgment data in the misjudgment data set as training samples, and performing iterative training on the model to be optimized to update the model to be optimized; testing the first performance of the model to be optimized after iterative training based on the training sample by adopting a preset test data set; repeatedly executing model optimization operation until the first performance of the model to be optimized after iterative training meets a first preset condition, wherein the model optimization operation comprises the following steps: and carrying out sample mining on the training data set based on the current model to be optimized to obtain second effective sample data, taking the second effective sample data and the misjudgment data in the misjudgment data set as new training samples, and carrying out iterative training on the current model to be optimized to update the model to be optimized.

According to a second aspect of the present disclosure, there is provided a model optimization apparatus comprising: the acquisition unit is configured to acquire a training data set and a misjudgment data set of a model to be optimized; the sample mining unit is configured to perform sample mining on the training data set based on the current model to be optimized to obtain first effective sample data; the first training unit is configured to take the first effective sample data and the misjudgment data in the misjudgment data set as training samples, and perform iterative training on the model to be optimized to update the model to be optimized; the first testing unit is configured to test the first performance of the model to be optimized after iterative training based on the training sample by adopting a preset testing data set; a second training unit configured to repeatedly perform model optimization operations until a first performance of the iteratively trained model to be optimized satisfies a first preset condition, wherein the model optimization operations include: and carrying out sample mining on the training data set based on the current model to be optimized to obtain second effective sample data, taking the second effective sample data and the misjudgment data in the misjudgment data set as new training samples, and carrying out iterative training on the current model to be optimized to update the model to be optimized.

According to a third aspect of the present disclosure, there is provided an electronic device comprising: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the model optimization method provided in the first aspect.

According to a fourth aspect of the present disclosure, there is provided a non-transitory computer readable storage medium having stored thereon computer instructions for causing a computer to perform the model optimization method provided by the first aspect.

According to the technology of the application, effective sample mining is realized, the training efficiency of the model is improved, and the calculation resources occupied by model training are reduced.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.

Drawings

The drawings are included to provide a better understanding of the present solution and are not intended to limit the present application. Wherein:

FIG. 1 is a schematic flow chart diagram of one embodiment of a model optimization method according to the present application;

FIG. 2 is a schematic flow chart diagram of another embodiment of a model optimization method according to the present application;

FIG. 3 is another schematic flow diagram of a model optimization method according to the present application;

FIG. 4 is a schematic flow chart diagram illustrating an alternative implementation of sample mining in an embodiment of the present application;

FIG. 5 is a distribution diagram of the prediction results of the confidence of the model to be optimized for the sample data;

FIG. 6 is a schematic block diagram of one embodiment of a model optimization apparatus according to the present application;

FIG. 7 is a block diagram of an electronic device for implementing a model optimization method according to an embodiment of the present application.

Detailed Description

The following description of the exemplary embodiments of the present application, taken in conjunction with the accompanying drawings, includes various details of the embodiments of the application for the understanding of the same, which are to be considered exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present application. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

The model optimization method provided by the application can be applied to machine learning models such as neural network models, probability-based models and decision trees, and can be applied to a server side.

In an actual scenario, iterative optimization may be performed on a model running on a line or a model not running on the line, and the purpose of the iterative optimization is to improve the performance of the model, for example, improve the classification accuracy of the model, improve the running efficiency of the model, and the like. The model may be a preliminarily trained model. The server side can obtain initial information of the model, including structure information and parameter information of the model, and can also obtain sample data for training the model. And then constructing a supervision function, and iteratively training the model by using the sample data to realize the optimization of the model.

Here, the server may be hardware or software, may be implemented as a single hardware module or a single software module, and may also be implemented as a plurality of hardware or software modules providing distributed services.

The method can also be applied to terminal equipment with data processing capability, and the terminal equipment can also optimize the parameters of the model to be optimized through iterative training by utilizing a processor (such as a graphic processor GPU, a central processing unit CPU and the like).

Referring to fig. 1, a schematic flow chart diagram of an embodiment of a model optimization method of the present application is shown. As shown in fig. 1, a flow 100 of the model optimization method of the present embodiment includes the following steps:

step 101, acquiring a training data set and a misjudgment data set of a model to be optimized.

In this embodiment, the executing agent (e.g., a server) of the model optimization method may first obtain a training data set. The training data set is a set of sample data for training the model to be optimized, and may include media data such as images, text, voice, and the like. The sample data in the training data set may include labeling information of a class to which the sample data belongs, such as a class identifier of a target object in the image, quality class information of the image, language class information of the text sentence, a speaker identity identifier corresponding to the voice signal, and the like. The model to be optimized may be a model for performing a classification task, in practice, an image classification model running on-line, a speech recognition model, etc.

The execution main body can also obtain a misjudgment data set of the model to be optimized. The misjudgment data set is a set of sample data or actual data to be processed with wrong judgment results or prediction results of the model to be optimized. The misjudged data set may include misjudged sample data, and may further include a judgment result of the misjudged sample data and a corresponding true value.

In practice, misjudgment data can be collected to construct a misjudgment data set in the operation process of the model to be optimized. For example, when the classification prediction result of the to-be-optimized model on the to-be-processed data is inconsistent with the real category of the to-be-processed data, the to-be-processed data is added to the misjudgment data set as misjudgment data.

The misjudgment data set may be stored in the execution subject or the database in advance, and associated with a corresponding model. When the execution main body executes the model optimization, the misjudgment data set of the model to be optimized is read from the memory or the database.

102, carrying out sample mining on the training data set based on the current model to be optimized to obtain first effective sample data.

The training data set contains a large amount of training data, and the model is iteratively trained by using the large amount of training data, so that a large amount of training time and computing resources are consumed, and the requirement for quick online cannot be met. In addition, redundant data may exist in the training data, and the redundant data cannot positively contribute to the optimization of the model and may even cause the performance of the model to be poor. In this embodiment, sample mining may be performed on the training data set obtained in step 201, effective samples are mined from the training data set, and redundant samples are removed, so as to accelerate the iterative optimization speed of the model.

Specifically, the obtained online model to be optimized or the model to be replaced is used as the current model to be optimized corresponding to the first iterative training, each sample data in the training data set can be predicted based on the current model to be optimized, and each sample data or the value of the loss function corresponding to each sample data sampled from the training data set is calculated based on the preset loss function representing the error of the model to be optimized. And respectively selecting partial sample data as first effective sample data in each interval of the distribution of the loss function values.

Alternatively, hard samples in the training data may be mined based on the current model to be optimized. The method comprises the steps of predicting each sample data in a training data set or each sample data sampled from the training data set by using a current model to be optimized, determining a sample with a wrong prediction result as a difficult sample of the current model to be optimized, sampling partial sample data from other sample data except the difficult sample in the training data set according to a certain sampling rate, and taking the difficult sample and the sampled partial sample data as first effective sample data.

And 103, taking the first effective sample data and the misjudgment data in the misjudgment data set as training samples, and performing iterative training on the model to be optimized to update the model to be optimized.

The model to be optimized can be iteratively trained by utilizing the first effective sample and the misjudgment data set, and parameters of the model to be optimized are optimized. Specifically, the misjudgment data in the first valid sample and the misjudgment data set may include labeling information of a real category, a supervision function may be constructed based on a classification error of the model to be optimized, and parameter updating of the model to be optimized may be supervised based on the supervision function. And when the value of the supervision function is converged, the updating of the model to be optimized can be stopped, and the model to be optimized which completes iterative training based on the first effective sample data and the misjudgment data set is obtained.

And 104, testing the first performance of the model to be optimized after iterative training based on the training sample by adopting a preset test data set.

The preset test data set is a data set for testing the performance of the model to be optimized, and may be extracted from the training data set, or may be constructed and stored separately from the training data set. The test data set contains sample data for testing. The sample data for testing may include labeling information of the corresponding real category. Sample data for testing in the test data set may be input to the model to be optimized that is iteratively trained in step 203, and the first performance of the model to be optimized is determined according to the prediction result of the model or the consumption of computational resources in which the model operates. For example, a performance indicator characterizing the first performance of the model to be optimized may be calculated based on at least one of a false positive rate of the model to the test data set, a hardware latency of the model, and a memory occupancy rate.

And 105, repeatedly executing the model optimization operation until the first performance of the iteratively trained model to be optimized meets a first preset condition.

If the first performance of the model to be optimized, which is obtained through the test in step 104, does not satisfy the first preset condition, the model optimization operation may be repeatedly performed until the first performance of the model to be optimized after the iterative training satisfies the first preset condition. Here, the first preset condition may include: the performance index of the model does not reach a preset index threshold, for example, the precision of the model is lower than the precision threshold, the hardware delay of the model exceeds the time threshold, or a comprehensive index calculated according to the precision, the hardware delay and other indexes of the model does not reach the preset index threshold.

The model optimization operation comprises the following steps: step 1051, sample mining is carried out on the training data set based on the current model to be optimized to obtain second effective sample data, the second effective sample data and the misjudgment data in the misjudgment data set are used as new training samples, and iterative training is carried out on the current model to be optimized to update the model to be optimized.

Specifically, based on the updated current model to be optimized, the second valid sample data may be re-mined in the training data set by the sample mining method as in step 102. After step 103, the parameters of the current model to be optimized are updated. In step 105, the parameters of the current model to be optimized continue to be updated each time step 1051 is performed. In this way, in the first model optimization operation, the model to be optimized updated based on step 103 is used as a new current model to be optimized to perform sample mining, and in each subsequent model optimization operation, the model to be optimized updated after the last model optimization operation is completed is used as a new current model to be optimized to perform sample mining again.

In each model optimization operation, a new training sample can be generated based on the currently mined second effective sample data and the misjudgment data set, and iterative training is continuously performed on the model to be optimized based on the new training sample so as to update the parameters of the model to be optimized.

Because the parameters of the model to be optimized in each model optimization operation are updated, sample data excavated based on the new model to be optimized in each model optimization operation may have differences, so that the model to be optimized can be updated iteratively to update the excavated second effective sample data through multiple iterations, the effectiveness of the excavated second effective sample data on model iterative training is gradually improved, and the iterative optimization efficiency of the model is further improved.

After each model optimization operation, testing the first performance of the updated model to be optimized based on the test data set, and if the first performance does not satisfy the first preset condition, returning to the step 1051 based on the current model to be optimized to execute the next model optimization operation; if the first performance meets the first preset condition, the model optimization operation can be stopped, and an optimized model is obtained.

In the model optimization method of this embodiment, the training data set is subjected to sample mining, the misjudged data set and the sample mining result are combined to serve as the training sample to perform iterative optimization on the model to be optimized, and when the performance of the model after iterative optimization does not meet the first preset condition, the operations of re-mining the sample based on the current model to be optimized, combining the mined effective sample and the misjudged data set into a new training sample to train the model to be optimized are repeatedly performed until the performance of the model to be optimized meets the preset condition, so that the optimization of the model is realized. In the method, sample data which has a positive effect on model optimization can be extracted from a large amount of sample data by sample mining, and the model is iteratively optimized by combining misjudgment of the data set, so that the waste of computing resources caused by redundant sample data on model iteration can be avoided, hardware resources occupied by model optimization are reduced, and the optimization effect of the model can be improved.

The above-described method may be applied to optimize a model for performing image processing tasks, such as image-based object classification tasks. Because the image processing model needs to process a large amount of complex matrix operations, the training of the model needs to consume more time, in order to quickly complete the optimization of the online model, the training of the image processing model is executed by mining effective samples and combining misjudgment data sets, the total operation amount of the matrix operations in the training process can be effectively reduced, and the optimization efficiency and the optimization effect of the image processing model are improved.

Referring to fig. 2, a schematic flow chart diagram of another embodiment of a model optimization method according to the present disclosure is shown. As shown in fig. 2, a flow 200 of the model optimization method of the present embodiment includes the following steps:

step 201, a training data set and a misjudgment data set of a model to be optimized are obtained.

Step 202, sample mining is carried out on the training data set based on the current model to be optimized, and first effective sample data is obtained.

And 203, taking the first effective sample data and the misjudgment data in the misjudgment data set as training samples, and performing iterative training on the model to be optimized to update the model to be optimized.

And 204, testing the first performance of the model to be optimized after iterative training based on the training sample by adopting a preset test data set.

Step 205, repeatedly executing model optimization operation until the first performance of the iteratively trained model to be optimized meets a first preset condition, wherein the model optimization operation includes: and carrying out sample mining on the training data set based on the current model to be optimized to obtain second effective sample data, taking the second effective sample data and the misjudgment data in the misjudgment data set as new training samples, and carrying out iterative training on the current model to be optimized to update the model to be optimized.

The above steps 201 to 205 are the same as steps 101 to 105 in the foregoing embodiment, and specific implementation manners may refer to descriptions of steps 101 to 105 in the foregoing embodiment, which are not described herein again.

And step 206, in response to determining that the first performance of the model to be optimized after iterative training based on the current training sample meets a first preset condition, determining the second performance of the model to be optimized after iterative optimization based on the current training sample based on the misjudgment data set.

If the first performance of the model to be optimized satisfies the first predetermined condition after the model optimization operations are repeatedly performed for several times in step 205, the misjudgment data set may be continuously used to test the second performance of the current model to be optimized. Specifically, the misjudgment data in the misjudgment data set may be input to a model to be optimized after iterative optimization based on the current training sample, and the second performance of the model to be optimized may be determined based on one or more of the prediction accuracy, the model operation efficiency, the memory occupancy rate, and the hardware delay of the model to be optimized. The second performance may be a single performance index or a total performance index calculated according to one or at least two of prediction accuracy of the model to be optimized, model operation efficiency, memory occupancy rate, and hardware delay.

Step 207, in response to determining that the second performance of the model to be optimized after iterative optimization based on the current training sample does not satisfy the second preset condition, copying and adding the misjudgment data set to the current training sample, and performing iterative training on the model to be optimized based on the current training sample to which the misjudgment data set is added.

Then, it may be determined whether the second performance satisfies a second preset condition, and specifically, whether the second performance index reaches an expected index may be determined. For example, it may be determined whether the prediction accuracy of the model to be optimized reaches a preset accuracy threshold, whether the operation delay of the model to be optimized does not exceed a preset delay range, and the like.

If the second performance does not meet the second preset condition, a misjudgment data set can be copied and added into the current training sample, namely, a misjudgment data set is added into the training sample, and iterative training is continuously carried out on the model to be optimized based on the current training sample added with the misjudgment data set.

After the misjudgment data set is added into the current training sample, the model to be optimized can better learn the prediction of the misjudgment data, so that the prediction effect of the model to be optimized on the misjudgment data set is improved. After the iteration training of the model to be optimized based on the current training sample added with the misjudgment data set reaches the preset times, the iteration training can be stopped to obtain the optimized model.

Alternatively, the method may further include: and in response to the fact that the second performance of the model to be optimized after iterative optimization based on the current training sample meets a second preset condition, determining the model to be optimized after iterative optimization based on the current training sample as the optimized model after several rounds of iterative training. That is, in each iterative training of the model to be optimized based on the current training sample added with the misjudgment data set, it may be determined whether the second performance of the model to be optimized satisfies the second preset condition, and when the second performance satisfies the second preset condition, the iterative training may be stopped to obtain the optimized model.

Referring to FIG. 3, another flow diagram of a model optimization method according to the present disclosure is shown. As shown in fig. 3, initial data is first acquired, including a training data set, an initial model (initial model to be optimized), and a misjudgment data set. And then, carrying out sample mining on the training data set, copying a part of misjudgment data in the misjudgment data set, and merging the part of misjudgment data into iterative data. Model iteration is then performed on the initial model using the iteration data. And after the model iteration, judging whether the model reaches the standard on the test set, if so, continuously judging whether the model reaches the standard on the misjudged data set, and if so, ending the optimization process. And if the model does not reach the standard on the test set after iteration, carrying out sample mining again and forming new iteration data with the misjudged data set to iterate the model. If the model reaches the standard on the test set but does not reach the standard on the misjudged data set, the misjudged data set can be copied, a part of misjudged data is added into the iterative data, and the model iteration is continued.

As can be seen from fig. 3, the model optimization method preferentially ensures the effect of the model on the test set after sample mining and training the model by using the mined effective samples and the misjudgment data. And if the test set effect does not reach the standard, the sample mining is carried out again, and the effectiveness of the sample is further improved. If the test set meets the standard. And if the misjudgment data set does not reach the standard, supplementing the misjudgment data into the iteration data to continue model iteration, thereby improving the discrimination capability of the model on the misjudgment data. Repeated iteration is carried out, the effect of the model on the training data set and the misjudgment data set is gradually improved, and the model is optimized. In addition, by adding the step of testing the second performance of the model based on the misjudgment data set, the optimized model can be ensured to have more accurate judgment on misjudgment data, and the accuracy of the model is improved.

In some optional implementations of the above embodiments, the training data set may be sample mined based on a confidence of the sample data. Specifically, please refer to fig. 4, which shows a schematic flow chart of an alternative implementation of sample mining in the embodiment of the present application. The process is applicable to the step of mining the first valid sample data and the second valid sample data in the foregoing embodiment.

As shown in FIG. 4, a process 400 for sample mining a training data set based on a current model to be optimized may include:

step 401, predicting confidence of each sample data in the training data set based on the current model to be optimized.

The confidence of the sample data represents the reliability of the classification result of the sample data. Sample data with higher confidence coefficient is easier to learn and is a simple sample; the sample data with lower confidence coefficient is difficult to learn and is a difficult sample. Here, the confidence of the sample data is related to a classification probability value of the current model to be optimized for the sample data, and the classification probability value is a probability value of dividing the sample data into a preset class. The higher the classification probability value, the lower the confidence of the sample data. For example, the confidence level of the sample data may be: 1-Classification probability value. It should be noted that, if the classification probability values of the sample data calculated by the model to be optimized for at least two different preset categories are different, the confidence degrees of the sample data for the at least two different preset categories are also different accordingly. That is, for the same sample data, the difficulty levels of different preset categories have differences according to the classification probability values of the sample data in the corresponding categories.

Step 402, extracting sample data with the confidence coefficient lower than a first threshold interval in the training data set as effective sample data.

After the confidence of the sample data is determined, statistics may be performed on the distribution of the confidence of each sample data in the training data set. And taking the sample data with the confidence coefficient lower than a first threshold interval, such as the sample data with the confidence coefficient lower than a certain number of bits before the confidence coefficient sorting, as the extracted valid sample data.

Fig. 5 shows the distribution of the prediction results of the confidence of the model to be optimized on the sample data in one iteration, wherein the abscissa represents the confidence score of the sample data, and the ordinate represents the probability distribution of the sample data. As can be seen from fig. 5, the model has a higher confidence for most of the sample data, meaning that the model can predict more accurately for most of the data. The model needs to learn a small amount of sample data with low confidence. In this embodiment, the difficulty level of the sample data is determined according to the confidence level of the sample data, and the top m% with the lowest confidence level is selected as the extracted difficult sample, where m is a positive number.

Step 403, sample data meeting the preset quantity condition is sampled from the sample data of which the confidence coefficient of the training data set is higher than the first threshold interval, and the sample data is used as effective sample data.

Then, a certain number of samples may be extracted from the sample data with confidence above the first threshold interval. For example, in the sample data of the confidence coefficient distribution shown in fig. 5, n% of sample data among 1-m% of sample data is randomly selected as the extracted simple sample, where n is a positive number.

The effective sample data extracted in

steps

402 and 403 may be used as the effective sample data mined based on the current model to be optimized. In this way, the mined effective sample data includes hard samples and simple samples. The method avoids the model malformation caused by only containing difficult samples in the training samples, and the number of the difficult samples is properly increased so that the model can fully learn the difficult samples.

Further, in step 403, the sample data whose confidence of the training data set is higher than the first threshold interval may be divided into at least two data intervals according to the confidence, and the sample data meeting the preset quantity condition is uniformly sampled in each data interval.

Specifically, according to the confidence distribution shown in fig. 5, for sample data distributed in 1-m% (i.e., other sample data except the difficult sample extracted in step 402), q intervals are divided from low to high according to the confidence, and p samples are respectively filtered in each interval, where p and q are positive integers. The random sampling mode is adopted in each interval. Thus, the difficulty degree of the samples extracted in the step 402 can be ensured to be uniformly distributed, and the accuracy of the model to be optimized is improved.

Further, before step 402, invalid training data with a confidence level lower than a first threshold interval in the training data set may be removed, where the invalid training data is ranked from low to high and located at a first preset ratio.

In an actual scene, aiming at massive sample data, it is difficult to ensure that the labeling information is correct. These sample data are low confidence, but in practice are not necessarily difficult samples. Therefore, the top c% data with the lowest confidence can be culled for the actual situation, where 0 < c < m. And then mining the sample data after c%. Therefore, the phenomenon that the labeled information is not in line with the expected sample data is added into a difficult sample and the iterative optimization of the model is interfered can be avoided.

Optionally, the training data set includes at least two training data of preset categories. The class of the training data may be determined according to its labeling information. In practice, a classification task, such as an object recognition task in an image, will typically specify at least two preset categories, and for each preset category, sample data containing labeling information for the preset category is constructed. When sample data with the confidence coefficient lower than the first threshold interval is extracted in the step 402, according to the confidence coefficient classified as the target preset class by the current model to be optimized, the first preset number of sample data with the confidence coefficient lower than the first threshold interval can be extracted as effective positive sample data of the target preset class; and for the sample data of the non-target preset type in the training data set, extracting second preset quantity of sample data with the highest confidence coefficient as effective negative sample data of the target preset type according to the confidence coefficient of the target preset type classified by the current model to be optimized.

Specifically, the classification probability values of the sample data corresponding to the preset categories may be respectively calculated by using the current model to be optimized, so as to determine the confidence levels of the sample data corresponding to the preset categories. Then, for a target preset category in the plurality of preset categories, a first preset number of samples with a corresponding confidence coefficient lower than a first preset interval may be extracted from the sample data of the target preset category as effective positive sample data of the target preset category. And extracting second preset quantity sample data with the highest confidence coefficient from the sample data of the non-target preset type to serve as effective negative sample data of the target preset type. The first preset number may be equal to the second preset number.

Assuming that k preset classes (k is a positive integer) are total, sorting target preset class samples according to the predicted confidence coefficients of the predicted target preset classes, and selecting the top 0.5/k m% with the lowest confidence coefficient as the hard samples of the positive samples of the classes; and selecting the top 0.5/k m% with the highest confidence as the hard sample of the negative sample of the category.

The method can ensure that the mined sample data contains the positive and negative difficult samples of each category, so that the model can be ensured to carry out sufficient training on the samples of each category, the trained model is not influenced by the quantity difference of the samples of different categories, the bias of some categories is avoided, and the effectiveness of the model on each category is improved.

In some optional implementations of the foregoing embodiments, in two adjacent model optimization operations, a data amount of second valid sample data mined in a next model optimization operation is greater than a data amount of second valid sample data mined in a previous model optimization operation. That is, if the first performance of the model in the current model optimization operation does not satisfy the first preset condition, the number of the mined effective samples is increased in the next model optimization operation. Optionally, the number of second valid samples mined in the first model optimization operation is greater than the number of first valid samples. Therefore, the number of effective samples is gradually increased when the optimization effect of the model is not expected, so that the convergence speed of the model can be increased, and the optimization efficiency of the model is improved.

Further, the data amount of the second valid sample data may increase by a preset scaling factor as the number of times of performing the model optimization operation. The preset scaling factor is, for example, 0.2, that is, the data size of the sample data obtained by each sample mining is increased by 0.2 times than the data size of the sample data obtained by the last sample mining.

Optionally, an iteration strategy may be designed, and when the first valid sample is mined, m% with the lowest confidence coefficient is extracted as a difficult sample, and n% of the simplest samples are extracted from the rest non-difficult samples. The mining proportions of the initial hard sample and the simple sample are m% and n% respectively, and the misjudgment data in the initial training sample is copied. In each model optimization operation, if the first performance obtained by testing the data set does not meet the first preset condition, the hard sample and the simple sample are respectively increased by gxm% and gxn% when sample mining is carried out again, wherein g is more than 0 and less than 0.5. Accordingly, if the first performance still does not satisfy the first preset condition in the e-th (e is a positive integer) model optimization operation, the excavation ratios of the hard sample and the simple sample are (e × g +1) × m% and (e × g +1) × n%, respectively.

Further, if the first performance of the model to be optimized after multiple times of model optimization operations satisfies a first preset condition, if the second performance of the model to be optimized determined based on the misjudged data does not satisfy a second preset condition, when the misjudged data is copied again, the misjudged data is increased by w (w is a positive integer), that is, w copies of the misjudged data in the whole misjudged data set are added to a new training sample, and the model to be optimized is continuously iteratively trained. And when the second performance of the model to be optimized, which is determined based on the misjudgment data at the e-th time, does not meet the second preset condition, the updated training samples contain (1+ e multiplied by w) parts of misjudgment data in total.

According to the implementation mode, the number of samples excavated in each sample excavation is increased according to the preset proportionality coefficient, the misjudgment data is increased according to the preset proportionality coefficient, and the number of training samples can be increased step by step in a targeted manner by a fixed step length, so that the model optimization efficiency is improved, and the waste of computing resources caused by introducing excessive invalid samples is avoided.

Referring to fig. 6, as an implementation of the model optimization method, the present disclosure provides an embodiment of a model optimization apparatus, where the embodiment of the apparatus corresponds to the above method embodiments, and the apparatus may be applied to various electronic devices.

As shown in fig. 6, the model optimization apparatus 600 of the present embodiment includes an obtaining unit 601, a sample mining unit 602, a first training unit 603, a first testing unit 604, and a second training unit 605. Wherein, the obtaining unit 601 is configured to obtain a training data set and a misjudgment data set of the model to be optimized; the sample mining unit 602 is configured to perform sample mining on a training data set based on a current model to be optimized to obtain first effective sample data; the first training unit 603 is configured to perform iterative training on the model to be optimized to update the model to be optimized, with the first valid sample data and the misjudgment data in the misjudgment data set as training samples; the first testing unit 604 is configured to test a first performance of the model to be optimized after iterative training based on the training samples by using a preset testing data set; the second training unit 605 is configured to repeatedly perform a model optimization operation until the first performance of the iteratively trained model to be optimized satisfies a first preset condition, where the model optimization operation includes: and carrying out sample mining on the training data set based on the current model to be optimized to obtain second effective sample data, taking the second effective sample data and the misjudgment data in the misjudgment data set as new training samples, and carrying out iterative training on the current model to be optimized to update the model to be optimized.

In some embodiments, the apparatus 600 further comprises: a second testing unit configured to determine, based on the misjudgment data set, a second performance of the model to be optimized after iterative optimization based on the current training sample in response to determining that the first performance of the model to be optimized after iterative training based on the current training sample satisfies a first preset condition; and the third training unit is configured to copy and add the misjudged data set to the current training sample in response to the fact that the second performance of the model to be optimized after iterative optimization based on the current training sample does not meet a second preset condition, and conduct iterative training on the model to be optimized based on the current training sample after the misjudged data set is added.

In some embodiments, the apparatus 600 further comprises: the determining unit is configured to determine the model to be optimized after iterative optimization based on the current training sample as the optimized model in response to determining that the second performance of the model to be optimized after iterative optimization based on the current training sample meets a second preset condition.

In some embodiments, the sample mining unit 602 includes: the prediction module is configured to predict the confidence coefficient of each sample data in the training data set based on the current model to be optimized, and the confidence coefficient represents the reliability of the classification result of the sample data; the first extraction module is configured to extract sample data with the confidence coefficient lower than a first threshold interval in the training data set as effective sample data; and the second extraction module is configured to sample data meeting a preset quantity condition from the sample data of which the confidence coefficient of the training data set is higher than the first threshold interval, and the sample data is used as effective sample data.

In some embodiments, the second decimation module is further configured to sample data satisfying a preset number condition from sample data of which the confidence of the training data set is higher than the first threshold interval as follows: dividing the sample data of which the confidence coefficient is higher than the first threshold interval of the training data set into at least two data intervals according to the confidence coefficient, and uniformly sampling the sample data meeting the preset quantity condition in each data interval.

In some embodiments, the first extraction module is further configured to extract sample data in the training data set with confidence below a first threshold interval as follows: invalid training data with the confidence coefficient lower than a first threshold value range in the training data set are extracted from the training data set with the invalid training data removed.

In some embodiments, the training data set includes at least two predetermined categories of training data; the first extraction module includes: the first extraction submodule is configured to extract a first preset number of sample data with the confidence coefficient lower than a first threshold interval as effective positive sample data of a target preset category according to the confidence coefficient of the target preset category classified by the current model to be optimized aiming at the sample data of the target preset category in the training data set; and the second extraction sub-module is configured to extract a second preset number of sample data with the highest confidence as effective negative sample data of the target preset category according to the confidence of the sample data of the non-target preset category classified into the target preset category by the current model to be optimized.

In some embodiments, in the two adjacent model optimization operations, the data size of the second valid sample data mined in the next model optimization operation is larger than the data size of the second valid sample data mined in the previous model optimization operation.

In some embodiments, the data amount of the second valid sample data increases according to a preset scaling factor as the number of times of performing the model optimization operation.

The above-described apparatus 600 corresponds to the steps in the foregoing method embodiments. Thus, the operations, features and technical effects described above for the model optimization method are also applicable to the apparatus 600 and the units included therein, and are not described herein again.

According to an embodiment of the present application, an electronic device and a readable storage medium are also provided.

Fig. 7 is a block diagram of an electronic device according to an embodiment of the present application. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application that are described and/or claimed herein.

As shown in fig. 7, the electronic apparatus includes: one or more processors 701, a memory 702, and interfaces for connecting the various components, including a high-speed interface and a low-speed interface. The various components are interconnected using different buses and may be mounted on a common motherboard or in other manners as desired. The processor may process instructions for execution within the electronic device, including instructions stored in or on the memory to display graphical information of a GUI on an external input/output apparatus (such as a display device coupled to the interface). In other embodiments, multiple processors and/or multiple buses may be used, along with multiple memories and multiple memories, as desired. Also, multiple electronic devices may be connected, with each device providing portions of the necessary operations (e.g., as a server array, a group of blade servers, or a multi-processor system). In fig. 7, one processor 701 is taken as an example.

The memory 702 is a non-transitory computer readable storage medium as provided herein. The memory stores instructions executable by the at least one processor to cause the at least one processor to perform the model optimization methods provided herein. The non-transitory computer readable storage medium of the present application stores computer instructions for causing a computer to perform the model optimization method provided herein.

The memory 702, which is a non-transitory computer readable storage medium, may be used to store non-transitory software programs, non-transitory computer executable programs, and modules, such as program instructions/modules corresponding to the model optimization method in the embodiments of the present application (for example, the obtaining unit 601, the sample mining unit 602, the first training unit 603, the first testing unit 604, and the second training unit 605 shown in fig. 6). The processor 701 executes various functional applications of the server and data processing by executing non-transitory software programs, instructions, and modules stored in the memory 702, that is, implements the model optimization method in the above method embodiment.

The memory 702 may include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function; the storage data area may store data created according to usage of the model-optimized electronic device, and the like. Further, the memory 702 may include high speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, memory 702 may optionally include memory located remotely from processor 701, which may be connected to a model optimized electronic device via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The electronic device of the model optimization method may further include: an input device 703 and an output device 704. The processor 701, the memory 702, the input device 703 and the output device 704 may be connected by a bus or other means, and are exemplified by a bus 705 in fig. 7.

The input device 703 may receive input numeric or character information and generate key signal inputs related to user settings and function control of the model-optimized electronic device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, a pointer, one or more mouse buttons, a track ball, a joystick, or other input device. The output devices 704 may include a display device, auxiliary lighting devices (e.g., LEDs), and tactile feedback devices (e.g., vibrating motors), among others. The display device may include, but is not limited to, a Liquid Crystal Display (LCD), a Light Emitting Diode (LED) display, and a plasma display. In some implementations, the display device can be a touch screen.

Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, application specific ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

These computer programs (also known as programs, software applications, or code) include machine instructions for a programmable processor, and may be implemented using high-level procedural and/or object-oriented programming languages, and/or assembly/machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), Wide Area Networks (WANs), and the Internet.

The computer system may include clients and servers. The client may be, but is not limited to, a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and the like. The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as cloud computing, cloud service, a cloud database, cloud storage and the like. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

According to the technical scheme of the embodiment of the application, the model optimization efficiency is improved through sample mining and misjudgment data copying.

It should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present application may be executed in parallel, sequentially, or in different orders, and the present invention is not limited thereto as long as the desired results of the technical solutions disclosed in the present application can be achieved.

The above-described embodiments should not be construed as limiting the scope of the present application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of model optimization, comprising:

acquiring a training data set and a misjudgment data set of a model to be optimized;

performing sample mining on the training data set based on a current model to be optimized to obtain first effective sample data;

taking the first effective sample data and the misjudgment data in the misjudgment data set as training samples, and performing iterative training on a model to be optimized to update the model to be optimized;

testing the first performance of the model to be optimized after iterative training based on the training sample by adopting a preset test data set;

repeatedly executing model optimization operation until the first performance of the iteratively trained model to be optimized meets a first preset condition, wherein the model optimization operation comprises the following steps: and carrying out sample mining on the training data set based on the current model to be optimized to obtain second effective sample data, taking the second effective sample data and misjudgment data in the misjudgment data set as new training samples, and carrying out iterative training on the current model to be optimized to update the model to be optimized.

2. The method of claim 1, wherein the method further comprises:

in response to determining that a first performance of the model to be optimized after iterative training based on the current training sample meets a first preset condition, determining a second performance of the model to be optimized after iterative optimization based on the current training sample based on the misjudgment data set;

and copying and adding the misjudgment data set into the current training sample in response to the fact that the second performance of the model to be optimized after iterative optimization based on the current training sample does not meet a second preset condition, and performing iterative training on the model to be optimized based on the current training sample after the misjudgment data set is added.

3. The method of claim 2, wherein the method further comprises:

and in response to determining that the second performance of the model to be optimized after iterative optimization based on the current training sample meets a second preset condition, determining the model to be optimized after iterative optimization based on the current training sample as an optimized model.

4. The method of claim 1, wherein the sample mining of the training data set based on the current model to be optimized comprises:

predicting the confidence coefficient of each sample data in the training data set based on the current model to be optimized, wherein the confidence coefficient represents the reliability of the classification result of the sample data;

extracting sample data with the confidence coefficient lower than a first threshold interval in the training data set as effective sample data;

and sampling sample data meeting a preset number condition from the sample data of which the confidence coefficient of the training data set is higher than a first threshold interval to serve as effective sample data.

5. The method of claim 4, wherein the sampling sample data satisfying a preset number condition from sample data of which the confidence level of the training data set is higher than a first threshold interval comprises:

and dividing the sample data of which the confidence coefficient is higher than the first threshold interval of the training data set into at least two data intervals according to the confidence coefficient, and uniformly sampling the sample data meeting the preset quantity condition in each data interval.

6. The method of claim 4, wherein said extracting sample data in said training data set with a confidence level below a first threshold interval comprises:

and eliminating invalid training data with the confidence coefficient being lower than a first threshold interval from low to high in the training data set, and extracting sample data with the confidence coefficient being lower than the first threshold interval from the training data set with the invalid training data eliminated.

7. The method according to any one of claims 4-6, wherein the training data set comprises at least two preset categories of training data;

the extracting of the sample data with the confidence coefficient lower than a first threshold interval in the training data set comprises:

for sample data of a target preset category in the training data set, extracting a first preset number of sample data with a confidence coefficient lower than a first threshold interval as effective positive sample data of the target preset category according to the confidence coefficient classified as the target preset category by the current model to be optimized;

and for the sample data of the non-target preset type in the training data set, extracting second preset quantity of sample data with the highest confidence coefficient as effective negative sample data of the target preset type according to the confidence coefficient classified as the target preset type by the current model to be optimized.

8. The method of any of claims 1-6, wherein, in said two consecutive model optimization operations, the data size of the second valid sample data mined in the next model optimization operation is larger than the data size of the second valid sample data mined in the previous model optimization operation.

9. The method of claim 8, wherein the data volume of the second valid sample data increases with the number of times of performing the model optimization operation by a preset scaling factor.

10. A model optimization apparatus, comprising:

the acquisition unit is configured to acquire a training data set and a misjudgment data set of a model to be optimized;

the sample mining unit is configured to perform sample mining on the training data set based on a current model to be optimized to obtain first effective sample data;

the first training unit is configured to take the first effective sample data and the misjudgment data in the misjudgment data set as training samples, and perform iterative training on a model to be optimized to update the model to be optimized;

the first testing unit is configured to test the first performance of the model to be optimized after iterative training based on the training sample by adopting a preset testing data set;

a second training unit configured to repeatedly perform a model optimization operation until a first performance of the iteratively trained model to be optimized satisfies a first preset condition, wherein the model optimization operation includes: and carrying out sample mining on the training data set based on the current model to be optimized to obtain second effective sample data, taking the second effective sample data and misjudgment data in the misjudgment data set as new training samples, and carrying out iterative training on the current model to be optimized to update the model to be optimized.

11. The apparatus of claim 10, wherein the apparatus further comprises:

a second testing unit configured to determine, based on the misjudgment data set, a second performance of the model to be optimized after iterative optimization based on the current training sample in response to determining that the first performance of the model to be optimized after iterative training based on the current training sample satisfies a first preset condition;

and the third training unit is configured to copy and add the misjudged data set to the current training sample in response to the fact that the second performance of the model to be optimized after iterative optimization based on the current training sample does not meet a second preset condition, and conduct iterative training on the model to be optimized based on the current training sample after the misjudged data set is added.

12. The apparatus of claim 11, wherein the apparatus further comprises:

a determining unit configured to determine the model to be optimized after iterative optimization based on the current training sample as an optimized model in response to determining that a second performance of the model to be optimized after iterative optimization based on the current training sample satisfies a second preset condition.

13. The apparatus of claim 10, wherein the sample mining unit comprises:

the prediction module is configured to predict the confidence coefficient of each sample data in the training data set based on the current model to be optimized, and the confidence coefficient represents the reliability of the classification result of the sample data;

the first extraction module is configured to extract sample data with the confidence coefficient lower than a first threshold interval in the training data set as effective sample data;

and the second extraction module is configured to sample data meeting a preset quantity condition from the sample data of which the confidence coefficient of the training data set is higher than the first threshold interval, and the sample data is used as effective sample data.

14. The apparatus according to claim 13, wherein the second decimation module is further configured to sample data satisfying a preset number condition from sample data of which the confidence of the training data set is higher than a first threshold interval as follows:

15. The apparatus of claim 13, wherein the first extraction module is further configured to extract sample data in the training data set with a confidence level below a first threshold interval as follows:

16. The apparatus according to any one of claims 13-15, wherein the training data set comprises at least two preset categories of training data;

the first extraction module comprises:

the first extraction submodule is configured to extract, for sample data of a target preset category in the training data set, a first preset number of sample data with a confidence coefficient lower than a first threshold interval as effective positive sample data of the target preset category according to the confidence coefficient classified as the target preset category by the current model to be optimized;

and the second extraction sub-module is configured to extract a second preset number of sample data with the highest confidence as effective negative sample data of the target preset category according to the confidence of the sample data of the non-target preset category classified into the target preset category by the current model to be optimized.

17. The apparatus of any of claims 10-15, wherein, in said two consecutive model optimization operations, the data size of the second valid sample data mined in the next model optimization operation is larger than the data size of the second valid sample data mined in the previous model optimization operation.

18. The apparatus according to claim 17, wherein the data size of the second valid sample data increases with the number of times of performing the model optimization operation by a preset scaling factor.

19. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-9.

20. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-9.