WO2023019908A1

WO2023019908A1 - Method and apparatus for generating training sample set, and electronic device, storage medium and program

Info

Publication number: WO2023019908A1
Application number: PCT/CN2022/078350
Authority: WO
Inventors: 钟华平; 刘卓名; 何聪辉
Original assignee: 上海商汤智能科技有限公司
Priority date: 2021-08-19
Filing date: 2022-02-28
Publication date: 2023-02-23
Also published as: CN113642659B; CN113642659A

Abstract

Provided in the present disclosure are a method and apparatus for generating a training sample set, and an electronic device, a storage medium and a program. The method comprises: acquiring each unlabeled sample, and a target neural network obtained by performing training on the basis of a training sample set; on the basis of each unlabeled sample and the target neural network, determining an estimated effect degree value of each unlabeled sample regarding network training of the target neural network; selecting, from among unlabeled samples, a target unlabeled sample with the estimated effect degree value meeting a preset requirement; and when a target labeled sample is obtained by performing sample labeling on the selected target unlabeled sample, adding the target labeled sample to the training sample set, so as to obtain an updated training sample set, wherein the updated training sample set is used for performing network training again on the target neural network. By means of the present disclosure, automatic selection of an unlabeled sample is realized on the basis of an estimated effect degree value, which, compared with a manual selection scheme, saves time and labor, and also reduces subsequent labeling costs.

Description

A method, device, electronic device, storage medium and program for generating a training sample set

Cross References to Related Applications

This patent application requires that the Chinese patent application number submitted on August 19, 2021 is 202110953373.0, the applicant is: Shanghai Shangtang Technology Development Co., Ltd., and the application name is "a method, device, electronic equipment and storage medium for generating a training sample set ", which is incorporated into this disclosure by reference.

technical field

The present disclosure relates to the technical field of machine learning, and in particular to a method, device, electronic device and storage medium for generating a training sample set.

Background technique

With the continuous development of deep learning, based on the support of large-scale training data sets, various machine learning models have achieved more and more success in all walks of life. The training data set is a data set with rich annotation information. Collecting and annotating such a data set usually requires huge manpower and material resources.

In related technologies, when collecting and labeling the training data set, the training data can be screened manually to construct a better data set, which leads to high manpower and material cost.

Contents of the invention

Embodiments of the present disclosure at least provide a method, device, electronic device, storage medium, and program for generating a training sample set, so as to automatically select training samples, saving time and effort.

An embodiment of the present disclosure provides a method for generating a training sample set, the method comprising:

Obtain each unlabeled sample and the target neural network trained based on the training sample set;

Based on each of the unlabeled samples and the target neural network, determine the estimated influence degree value of each of the unlabeled samples on the network training of the target neural network;

Selecting target unlabeled samples whose estimated influence degree value meets the preset requirements from each of the unlabeled samples;

In the case of performing sample labeling on the selected target unlabeled samples to obtain target labeled samples, adding the target labeled samples to the training sample set to obtain an updated training sample set; the updated training samples The set is used to perform network training on the target neural network again.

Using the method of generating the training sample set above, in the case of obtaining each unlabeled sample and the target neural network, it is possible to first determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network, and then from each unlabeled sample The target unlabeled samples whose estimated impact value meets the preset requirements are selected from the labeled samples. In this way, after the target unlabeled samples are labeled, the target labeled samples can be obtained to update the training sample set. The disclosure realizes the automatic selection of unlabeled samples based on the estimated influence degree value, which is more time-saving and labor-saving than the manual selection scheme, and also reduces the subsequent labeling cost.

An embodiment of the present disclosure also provides a device for generating a training sample set, the device comprising:

An acquisition module configured to acquire each unlabeled sample and the target neural network trained based on the training sample set;

The determination module is configured to determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network based on the respective unlabeled samples and the target neural network;

The selection module is configured to select target unlabeled samples whose estimated influence degree value meets the preset requirements from each of the unlabeled samples;

The generation module is configured to add the target labeled samples to the training sample set to obtain an updated training sample set when the target unlabeled sample is sample labeled to obtain the target labeled sample; The updated training sample set is used to perform network training on the target neural network again.

An embodiment of the present disclosure also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the The memory communicates with each other through a bus, and when the machine-readable instructions are executed by the processor, the steps of the method for generating a training sample set described in any embodiment are executed.

An embodiment of the present disclosure also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method for generating a training sample set described in any embodiment is executed. step.

An embodiment of the present disclosure further provides a computer program, where the computer program includes computer readable codes, and when the computer readable codes run in an electronic device, a processor of the electronic device executes the program described in any embodiment. Steps of the method for generating the training sample set described above.

For the effect description of the device, electronic equipment, and computer-readable storage medium for generating the training sample set, please refer to the description of the method for generating the training sample set.

In order to make the above-mentioned objects, features and advantages of the present disclosure more comprehensible, preferred embodiments will be described in detail below together with the accompanying drawings.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present disclosure more clearly, the following will briefly introduce the accompanying drawings used in the embodiments. The accompanying drawings here are incorporated into the specification and constitute a part of the specification. The drawings show the embodiments consistent with the present disclosure, and are used together with the description to explain the technical solutions of the present disclosure. It should be understood that the following drawings only show some embodiments of the present disclosure, and therefore should not be regarded as limiting the scope. For those skilled in the art, they can also make From these drawings other related drawings are obtained.

FIG. 1 shows a schematic flowchart of a method for generating a training sample set provided by an embodiment of the present disclosure;

FIG. 2 shows a schematic diagram of a system architecture that can be applied to a method for generating a training sample set according to an embodiment of the present disclosure;

FIG. 3 shows a schematic structural diagram of an apparatus for generating a training sample set provided by an embodiment of the present disclosure;

Fig. 4 shows a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only It is a part of the embodiments of the present disclosure, but not all of them. The components of the disclosed embodiments generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations. Accordingly, the following detailed description of the embodiments of the present disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort shall fall within the protection scope of the present disclosure.

It should be noted that like numerals and letters denote similar items in the following figures, therefore, once an item is defined in one figure, it does not require further definition and explanation in subsequent figures.

The term "and/or" in this article only describes an association relationship, which means that there can be three kinds of relationships, for example, A and/or B can mean: there is A alone, A and B exist at the same time, and B exists alone. situation. In addition, the term "at least one" herein means any one of a variety or any combination of at least two of the more, for example, including at least one of A, B, and C, which may mean including from A, Any one or more elements selected from the set formed by B and C.

After research, it is found that in related technologies, manual screening of training data can be used to construct a better data set, which leads to high cost of manpower and material resources.

In order to solve the above problems, the related technology also provides an active learning training data screening method, which considers the uncertainty of the model for unlabeled samples, or the impact of unlabeled samples on the diversity of labeled datasets. Measure the importance of samples, which can automatically find high-value data in unlabeled samples. Compared with manual operation, the above method only takes a fraction of the time to build a better dataset, and uses very little data to train an efficient model, thereby reducing the cost of labeling.

Uncertainty-based active learning algorithms and diversity-based active learning algorithms are two mainstream algorithms in related technologies, but both have their own defects. Uncertainty-based methods: Because neural networks often exhibit overconfidence in unfamiliar samples, such methods choose samples inaccurately. Methods based on sample diversity: The current state of the model is not considered to select samples, and the computational complexity is usually proportional to the square of the size of the data set.

Based on the above research, the embodiments of the present disclosure provide a method, device, electronic device, storage medium and program for generating a training sample set, so as to automatically realize the selection of training samples, saving time and effort.

In order to facilitate the understanding of this embodiment, a method for generating a training sample set disclosed in the embodiment of the present disclosure is first introduced in detail. The execution subject of the method for generating a training sample set provided by the embodiment of the present disclosure generally has a certain computing power Electronic equipment, the electronic equipment includes, for example: terminal equipment or server or other processing equipment, the terminal equipment can be user equipment (User Equipment, UE), mobile equipment, user terminal, cellular phone, cordless phone, personal digital assistant (Personal Digital Assistant) Assistant, PDA), handheld devices, computing devices, vehicle-mounted devices, wearable devices, etc. In some possible implementation manners, the method for generating a training sample set may be implemented by a processor invoking computer-readable instructions stored in a memory.

Referring to FIG. 1 , it is a schematic flowchart of a method for generating a training sample set provided by an embodiment of the present disclosure, the method is executed by an electronic device, and the method includes steps S101 to S104, wherein:

S101: Obtain each unlabeled sample and a target neural network trained based on the training sample set;

S102: Based on each unlabeled sample and the target neural network, determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network;

S103: Select target unlabeled samples whose estimated influence degree value meets the preset requirements from each unlabeled sample;

S104: In the case of sample labeling the selected target unlabeled sample to obtain the target labeled sample, add the target labeled sample to the training sample set to obtain the updated training sample set; the updated training sample set is used to train the target neuron The network goes through network training again.

Referring to FIG. 2, FIG. 2 shows a schematic diagram of a system architecture that can be used in a method for generating a training sample set in an embodiment of the present disclosure; as shown in FIG. 2, the system architecture includes: an acquisition terminal 201, a network 202 And control terminal 203. In order to support an exemplary application, the acquisition terminal 201 and the control terminal 203 establish a communication connection through the network 202, the acquisition terminal 201 reports the obtained unlabeled samples and the target neural network to the control terminal 203 through the network 202, and the control terminal 203 acquires After arriving at each unlabeled sample and target neural network, determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network; and select the target whose estimated influence degree value meets the preset requirements from each unlabeled sample Unlabeled samples; when the selected target unlabeled samples are labeled to obtain the target labeled samples, the target labeled samples are added to the training sample set to obtain an updated training sample set; the updated training sample set is used for The target neural network undergoes network training again. Finally, the control terminal 203 sends the updated training sample set to the acquisition terminal 201 .

As an example, the acquisition terminal 201 may be an image acquisition device, and the control terminal 203 may include a data processing device with data processing capability or a remote server. The network 202 may be connected in a wired or wireless manner. Wherein, when the control terminal 203 is a data processing device, the acquisition terminal 201 can communicate with the data processing device through a wired connection, such as performing data communication through a bus; Data exchange between the network and the remote server.

In order to facilitate the understanding of the method for generating the training sample set provided by the embodiment of the present disclosure, the application scenario of the method will be described in detail below. The method for generating the training sample set in the embodiments of the present disclosure can be applied to the training preparation process of the neural network in any application scenario. In order to better train the target neural network, it is necessary to prepare a richer training sample set before training the neural network, where the training sample set can be a set of marked samples. How to automatically select the target unlabeled samples that can adapt to the target neural network training from a large number of unlabeled samples has become a key task to update the training sample set.

The manual screening method provided in related technologies results in high manpower and material costs, while the screening method related to active learning either screens unlabeled samples that are not accurate enough, or takes a long time to achieve screening. Just to solve these problems, the embodiment of the present disclosure provides an unlabeled sample screening scheme based on estimated influence degree value estimation, so that the target neural network trained by the obtained updated training sample set is more accurate, and Automated screening, saving time and effort.

Among them, for different application scenarios, the target neural network here is also different. For example, in the target classification application scenario, the target neural network here can be a classification network that determines the target classification; for another example, in the target detection application scenario, the target neural network here can determine the target position, size and other information detection network. In addition, the target neural network here may also be other networks, which is not limited in this embodiment of the present disclosure.

In practical applications, the target neural network here can be trained based on a training sample set containing several labeled samples, for example, it can be a vehicle classification network trained based on multiple labeled vehicle pictures.

Here, in order to select the target unlabeled samples that are more suitable for the above-mentioned target neural network from each unlabeled sample, the estimated influence degree value of each unlabeled sample on the network training of the target neural network can be determined here.

In the embodiment of the present disclosure, the estimated influence degree value of each unlabeled sample can be determined, and the influence of each unlabeled sample on the estimated influence degree value of the network training of the target neural network can be converted into each first marked The estimated influence degree and value of the sample on the network training parameters, and the estimated influence value of each unlabeled sample on the network training parameters. This is considering that network training parameters are the most direct considerations for network training, and the degree of proximity between the estimated influence degree and value of each first labeled sample and the estimated influence degree value of each unlabeled sample can be determined to a certain extent. It reflects the role that unlabeled samples can play in the training of the target neural network.

In the case of determining the estimated influence degree value of each unlabeled sample on the network training of the target neural network, in the embodiment of the present disclosure, the target whose estimated influence degree value meets the preset requirements can be selected from each unlabeled sample Unlabeled samples. Here, unlabeled samples with estimated influence degree values greater than the second preset threshold may be selected as target unlabeled samples; or the unlabeled samples may be sorted in order of estimated influence degree values from large to small, Then determine the target unlabeled samples according to the ranking results; for example, select the unlabeled samples whose ranking results are in the top 10 as the target unlabeled samples. Here, the second preset threshold is set according to the type of the selected target neural network. Since the selected target neural network type to be trained is different, the estimated influence value of each unlabeled sample is also different. Therefore, the second preset The setting of the threshold is also different.

When the target unlabeled sample is selected, the target unlabeled sample can be firstly labeled, and then the labeled target labeled sample can be added to the training sample set to obtain an updated training sample set.

In practical applications, multiple rounds of target unlabeled samples can be selected and the training sample set updated to obtain a more accurate updated target neural network. Each round of update can be achieved through the following steps:

Step 1. Screen out the target unlabeled samples from each unlabeled sample to obtain each updated unlabeled sample; and determine the updated target neural network based on the training of the updated training sample set; the updated training sample set includes each the first labeled sample and the target labeled sample;

Step 2. Based on the updated unlabeled samples and the updated target neural network, determine the estimated influence degree value of the updated unlabeled samples on the network training of the updated target neural network;

Step 3. Select the target unlabeled samples whose estimated impact degree value meets the preset requirements from the updated unlabeled samples;

Step 4. In the case of labeling the selected target unlabeled samples to obtain the target labeled samples, add the target labeled samples to the updated training sample set, and obtain the updated target neural network for training. training sample set.

In some embodiments, the target non-concerned samples that meet the preset requirements are unlabeled samples whose estimated influence degree value is greater than the second preset threshold value as the target unlabeled sample; The small order sorts each unlabeled sample, and then determines the target unlabeled sample according to the sorting result.

Here, the target unlabeled samples can be screened out from each unlabeled sample to obtain the updated unlabeled samples, and at this time, the target labeled samples corresponding to the target unlabeled samples can be added to the updated training sample set, And the updated target neural network can be obtained by training based on the updated training sample set. In this way, the estimated influence degree value of each updated unlabeled sample can be determined based on each updated unlabeled sample and the updated target neural network, and then the target unlabeled sample is selected and the updated training sample set is updated again. , and so on, until the loop cut-off condition is reached, and the updated target neural network is obtained.

Wherein, the above-mentioned cycle cut-off condition may be that the number of cycles reaches the preset number of times, or that the relevant evaluation indicators of the updated target neural network obtained after training reach the preset indicators, for example, the cycle cut-off condition that the prediction accuracy reaches 0.75 .

Considering that the determination of the estimated influence degree value plays a key role in the selection of unlabeled samples, in some embodiments, the process of determining the estimated influence degree value can be described. Here, determining the estimated impact degree value may include the following steps:

Step 1. Based on each first labeled sample and the target neural network, determine the estimated influence degree of each first labeled sample on the network training parameters during the forward propagation process of the target neural network based on each first labeled sample and values; and, based on each unlabeled sample and the target neural network, determine the estimated influence degree value of each unlabeled sample on the network training parameters during the forward propagation of the target neural network based on each unlabeled sample;

Step 2. Based on the estimated influence degree and value of each first labeled sample on the network training parameters and the estimated influence degree value of each unlabeled sample on the network training parameters, determine the network effect of each unlabeled sample on the target neural network. The estimated impact value for training.

Here, firstly, the determination of the estimated influence degree and value of each first labeled sample on the network training parameters can be realized based on the forward propagation of the labeled samples; secondly, each unlabeled sample can be realized based on the forward propagation of the unlabeled samples Determination of the estimated influence degree value of the sample on the network training parameters; finally, the estimated influence degree value of each unlabeled sample on the network training of the target neural network can be determined based on the above estimated influence degree sum value and the estimated influence degree value .

Forward propagation in the embodiments of the present disclosure may refer to a process of inputting samples into a trained target neural network to obtain a gradient and a Hessian matrix corresponding to a relevant loss function. During the forward propagation, the network parameter values of the target neural network are not adjusted.

For unlabeled samples, the determination of the estimated influence degree of unlabeled samples on network training parameters can be realized based on the determination of pseudo-labeled information, which can be achieved by the following steps:

Step 1. For each unlabeled sample in each unlabeled sample, input each unlabeled sample into the target neural network, and determine the probability value for each candidate prediction result output by the target neural network;

Step 2: Based on the probability values for each candidate prediction result, determine the pseudo-label information of the unlabeled sample.

Step 3. Based on the pseudo-label information, determine the gradient value corresponding to the loss function of the target neural network in the case of forward propagation of unlabeled samples;

Step 4: The determined gradient value is used as the estimated influence degree value of the unlabeled samples on the network training parameters.

Here, when unlabeled samples are input into the target neural network, the probability values output by the target neural network for each candidate prediction result can be determined, and then the pseudo-label information of the unlabeled sample can be determined based on the probability values for each candidate prediction result .

Considering that the target neural networks corresponding to different application scenarios are different, the strategies for generating pseudo-label information for different target neural networks are also different. For example, when the target neural network is a classification network and the candidate prediction result is a candidate category, the candidate category with the highest probability value can be determined as the pseudo-label information of the unlabeled sample; When the prediction result is a candidate detection frame, the candidate detection frame whose probability value is greater than the first preset threshold is determined as the pseudo-label information of the unlabeled sample, that is, multiple candidate detection frames can be used as the pseudo-label information. In some embodiments, the first preset threshold can be set to 0.95, that is, the candidate detection frame with a probability value greater than 0.95 is determined as the pseudo-labeled information of the unlabeled sample. In addition, the corresponding pseudo-label information generation strategy can also be determined for other target neural networks.

The loss of unlabeled samples can be determined through pseudo-label information, and the gradient value corresponding to the loss function of the target neural network can be determined by backpropagating the loss to the target neural network, which can be used as the estimated influence degree of unlabeled samples on network training parameters value.

For the first marked sample, the first marked sample is input into the target neural network, and the gradient value corresponding to the loss function of the target neural network and the Hessian matrix corresponding to the loss function of the target neural network are obtained. The gradient value here is used to represent the degree of influence of each network parameter on the loss function in the case of forward propagation of the first labeled sample, and the Hessian matrix here is used to represent the forward propagation of the first labeled sample. , the degree to which each network parameter affects the loss function is affected by other network parameters. In the specific mathematical calculation process, the above gradient value corresponds to the first derivative of the loss function, and the Hessian matrix corresponds to the second derivative of the loss function.

In this way, for each labeled sample, the gradient value corresponding to the loss function of the target neural network obtained by each labeled sample and the Hessian matrix corresponding to the loss function of the target neural network can be superimposed to obtain the gradient sum value and Hessian The matrix and value, and then based on the product operation of the gradient sum value and the Hessian matrix sum value, determine the estimated influence degree and value.

In practical applications, the network performance test function of the training reference set can be combined to realize the determination of the estimated influence degree and value, and then determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network, which can be obtained by Follow these steps to achieve:

Step 1. Obtain each second labeled sample included in the training reference set; the training reference set and the training sample set do not have the same labeled samples;

Step 2. Based on each of the first labeled samples, each of the second labeled samples, and the target neural network, determine that in the process of forward propagation of the target neural network based on each of the first labeled samples and each of the second labeled samples, each The estimated influence degree and value of the first labeled sample and each second labeled sample on the network training parameters; and, based on each unlabeled sample and the target neural network, determine the forward propagation of the target neural network based on each unlabeled sample In the process of , each unlabeled sample has an estimated influence value on the network training parameters;

Step 3. Based on the estimated influence degree and value of each first labeled sample and each second labeled sample on the network training parameters and the estimated influence degree value of each unlabeled sample on the network training parameters, determine each unlabeled sample The value of the estimated influence of the sample on the network training of the target neural network.

See the above description for the process of determining the estimated influence value of each unlabeled sample on the network training parameters.

In the process of determining the above-mentioned estimated influence degree and value, firstly, multiple first labeled samples may be selected from each first labeled sample, and each first labeled sample in the multiple first labeled samples The marked samples are input into the target neural network, and the Hessian matrix and the value corresponding to the loss function of the target neural network are obtained; secondly, each second marked sample in each second marked sample can be input into the target neural network, Obtain the gradient and value corresponding to the loss function of the target neural network; finally, based on the product operation of the gradient sum value and the Hessian matrix sum value, determine the estimated influence degree and value.

It can be seen that here, the first labeled sample for training the target neural network and the second labeled sample for performance testing of the target neural network are distinguished, so that a more accurate network can be determined while valid network testing Estimated magnitude and value of impact.

In order to facilitate the understanding of the determination process of the estimated influence degree value of each unlabeled sample on the network training of the target neural network, it can be explained in conjunction with the formula.

In the embodiment of the present disclosure, it can be based on the gradient and value of the target neural network on the training reference set R

inverse matrix with Hessian and values

and the expected gradient value of the unlabeled sample z _i

to determine the impact of unlabeled samples on model performance, that is, to determine

To measure the importance of sample _zi , where the gradient and value can correspond to the degree of influence of each network parameter on the loss function in the case of forward propagation of each second labeled sample in each second labeled sample , the Hessian matrix and the value can correspond to the degree of influence of each network parameter on the loss function by the degree of influence of other network parameters in the case of forward propagation of each first labeled sample in each first labeled sample . Here the Hessian matrix and the values

can be defined as the sum of the Hessian matrices of all the first labeled samples, namely

In practical applications, considering the large amount of calculation of the Hessian matrix and value, the Hessian matrix and value will not be directly calculated, but the calculation

and

The product of , here the product can be recorded as s _test , where s _test can be determined by random estimation.

Here, the gradient and value of the network on the reference set can be first

Denote it as v, and then randomly select k samples {z ₁ ,z ₂ ,…,z _k } from the first labeled sample, and initialize

In the case of iterating through

Second, the obtained s _test results can be determined as the estimated impact degree and value.

In the embodiment of the present disclosure, in the process of determining the degree and value of the estimated impact, it can be realized through multiple rounds of iterative calculations. For the current round of iterative calculations, the marked samples pointed to by the current round of iterative calculations are determined, based on the determined marked samples corresponding to Hessian matrix, gradient and value, and the estimated influence degree and value corresponding to the previous round of iterative operation, determine the estimated influence degree and value corresponding to the current round of iterative operation, and through multiple rounds of iterations, the final estimated influence can be obtained degree and value.

Here, it is also possible to determine the estimated influence degree value of the unlabeled samples on the network training parameters, that is, it is necessary to determine the expected gradient of the unlabeled samples

For different target neural networks, the expected gradient determined here

Slightly different.

For example, for the classification network, the unlabeled picture z _i is forward propagated to the network, the category with the highest predicted score of the classifier is selected as the pseudo-labeling result p, and the pseudo-labeling result is used to determine the loss of the picture

in turn will lose

Backpropagation to the neural network to get the gradient

by

as the expected gradient for unlabeled samples

For another example, for the detection network, the unlabeled picture z _i is propagated forward to the network, and all detection frames P' of the picture z _i by the network are obtained. Here, you can use a threshold, such as 0.95, to filter out the detection frames with results lower than 0.95, that is, the untrusted detection frames, and use the remaining detection frames as the pseudo-labeling result P, and use the pseudo-labeling result to determine the loss of the picture

will lose

Backpropagation to the neural network to get the gradient

by

as the expected gradient for unlabeled samples

Here, after getting s _test we determine for each unlabeled sample z _i

by determining

To determine the impact of sample z _i on model performance, here, you can use

Denote as I(z _i ,R). The more negative the value of I(z _i ,R), the more positive impact the sample z _i can have on network performance. The N samples with the most negative values are selected for labeling and added to the first labeled sample to obtain an updated training sample set.

Those skilled in the art can understand that in the above method of specific implementation, the writing order of each step does not mean a strict execution order and constitutes any limitation on the implementation process. The specific execution order of each step should be based on its function and possible The inner logic is OK.

Based on the same inventive concept, the embodiment of the present disclosure also provides a device for generating a training sample set corresponding to the method for generating a training sample set. Since the problem-solving principle of the device in the embodiment of the present disclosure is the same as that of the above-mentioned training sample set in the embodiment of the present disclosure The generated method is similar, so the implementation of the device can refer to the implementation of the method.

Referring to FIG. 3 , it is a schematic structural diagram of a device for generating a training sample set provided by an embodiment of the present disclosure. The device includes: an acquisition module 301, a determination module 302, a selection module 303, and a generation module 304; wherein,

The obtaining module 301 is configured to obtain each unlabeled sample and the target neural network trained based on the training sample set;

The determination module 302 is configured to determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network based on each unlabeled sample and the target neural network;

The selection module 303 is configured to select target unlabeled samples whose estimated influence degree value meets the preset requirements from each unlabeled sample;

The generating module 304 is configured to add the target marked sample to the training sample set to obtain the updated training sample set when the target unlabeled sample is sample marked to obtain the target marked sample, and the updated training sample set is used In order to perform network training on the target neural network again.

Using the above-mentioned device for generating training sample sets, when each unlabeled sample and the target neural network are obtained, it is possible to first determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network, and then from each unlabeled sample. The target unlabeled samples whose estimated impact value meets the preset requirements are selected from the labeled samples. In this way, after the target unlabeled samples are labeled, the target labeled samples can be obtained to update the training sample set. The present disclosure realizes the automatic selection of unlabeled samples based on the estimated influence degree value, which is more time-saving and labor-saving than the manual selection scheme, and also reduces the subsequent labeling cost.

In a possible implementation manner, the training sample set includes each first labeled sample; the determining module 302 is configured to determine the contribution of each unlabeled sample to the target neural network based on the following steps: Estimated impact value for network training:

Based on each first marked sample and the target neural network, determine the estimated influence degree and value of each first marked sample on the network training parameters during the forward propagation of the target neural network based on each first marked sample; as well as,

Based on each unlabeled sample and the target neural network, determine the estimated influence degree value of each unlabeled sample on the network training parameters during the forward propagation of the target neural network based on each unlabeled sample;

Based on the estimated influence degree and value of each first labeled sample on the network training parameters and the estimated influence degree value of each unlabeled sample on the network training parameters, determine the network training prediction of each unlabeled sample on the target neural network. Estimate the degree of impact.

In a possible implementation manner, the determination module 302 is configured to determine the estimated impact degree and value according to the following steps:

Input each first marked sample in each first marked sample into the target neural network, and obtain the gradient sum value and the Hessian matrix sum value corresponding to the loss function of the target neural network; wherein, the gradient sum value is used to represent The summation result of the gradient value corresponding to each first labeled sample, the gradient value is used to represent the degree of influence of each network parameter on the loss function in the case of forward propagation of the first labeled sample; the Hessian matrix sum value is used In order to represent the summation result of the Hessian matrix corresponding to each first labeled sample, the Hessian matrix is used to indicate that in the case of forward propagation of the first labeled sample, the degree of influence of each network parameter on the loss function is affected by other The degree of influence of network parameters;

Based on the product operation of the gradient sum and the Hessian matrix sum, the estimated influence degree and value are determined.

In a possible implementation manner, the determination module 302 is configured to determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network based on each unlabeled sample and the target neural network according to the following steps:

Obtain each second labeled sample included in the training reference set; the training reference set and the training sample set do not have the same labeled sample;

Based on each first labeled sample, each second labeled sample, and the target neural network, it is determined that each first labeled sample and each second labeled sample are used for forward propagation of the target neural network. The degree and value of the estimated influence of the labeled sample and each second labeled sample on the network training parameters; and,

Based on the estimated influence degree and value of each first labeled sample and each second labeled sample on the network training parameters and the estimated influence degree value of each unlabeled sample on the network training parameters, determine the impact of each unlabeled sample on the target The estimated influence value of the network training of the neural network.

Select multiple first labeled samples from each of the first labeled samples, and input each first labeled sample in the multiple first labeled samples into the target neural network to obtain a loss function corresponding to the target neural network The Hessian matrix and value; the Hessian matrix and the value are used to represent the summation result of the Hessian matrix corresponding to each first labeled sample, and the Hessian matrix is used to represent the forward propagation of the first labeled sample. , the degree to which each network parameter affects the loss function is affected by other network parameters; and,

Input each second marked sample in each second marked sample into the target neural network to obtain the gradient and value corresponding to the loss function of the target neural network; the gradient and value are used to indicate that each second marked sample corresponds to The summation result of the gradient value of , the gradient value is used to represent the degree of influence of each network parameter on the loss function in the case of forward propagation of the second labeled sample;

In a possible implementation manner, the determination module 302 is configured to determine the estimated influence degree and value based on the product operation of the gradient sum value and the Hessian matrix sum value according to the following steps:

For the current round of iterative operations, determine the marked samples pointed to by the current round of iterative operations, and determine the current The estimated influence degree and value corresponding to the round iterative operation.

In a possible implementation, the determination module 302 is configured to determine, based on each unlabeled sample and the target neural network, in the process of forward propagation of the target neural network based on each unlabeled sample, each unlabeled The value of the estimated influence degree of the sample on the network training parameters:

For each unlabeled sample in each unlabeled sample, input each unlabeled sample into the target neural network, and determine the probability value for each candidate prediction result output by the target neural network;

Based on the probability value for each candidate prediction result, determine the pseudo-label information of the unlabeled sample;

Based on the pseudo-label information, determine the gradient value corresponding to the loss function of the target neural network in the case of forward propagation of unlabeled samples;

The determined gradient value is used as the estimated influence degree value of the unlabeled sample on the network training parameters.

In a possible implementation manner, the determining module 302 is configured to determine the pseudo-labeling information of the unlabeled samples based on the probability values for each candidate prediction result according to the following steps:

When the target neural network is a classification network and the candidate prediction result is a candidate category, determine the candidate category with the largest probability value as the pseudo-label information of the unlabeled sample; or,

When the target neural network includes a detection network and the candidate prediction result is a candidate detection frame, the candidate detection frame whose probability value is greater than the first preset threshold is determined as the pseudo-labeled information of the unlabeled sample.

In a possible implementation manner, the selection module 303 is configured to select target unlabeled samples whose estimated influence degree value meets the preset requirements from each unlabeled sample according to the following steps:

Selecting an unlabeled sample with an estimated influence degree value greater than a second preset threshold as a target unlabeled sample; or,

Sort each unlabeled sample in descending order of the estimated impact value, and determine the target unlabeled sample according to the sorting result.

In a possible implementation manner, the generating module 304 is further configured to:

After obtaining the updated training sample set, perform the following steps in a loop until the loop cut-off condition is reached, and the updated target neural network is obtained:

Screen out the target unlabeled samples from each unlabeled sample to obtain each updated unlabeled sample; and determine the updated target neural network based on the updated training sample set training; the updated training sample set includes each first already Labeled samples and target labeled samples;

Based on the updated unlabeled samples and the updated target neural network, determine the estimated influence degree value of the updated unlabeled samples on the network training of the updated target neural network;

From the updated unlabeled samples, select the target unlabeled samples whose estimated influence degree value meets the preset requirements;

In the case of sample labeling the selected target unlabeled samples to obtain the target labeled samples, add the target labeled samples to the updated training sample set to obtain an updated training sample set for training the updated target neural network .

For the description of the processing flow of each module in the device and the interaction flow between the modules, reference may be made to the relevant description in the above method embodiment, and details will not be described here.

The embodiment of the present disclosure also provides an electronic device, as shown in FIG. 4 , which is a schematic structural diagram of the electronic device provided by the embodiment of the present disclosure, including: a processor 401 , a memory 402 , and a bus 403 . The memory 402 stores machine-readable instructions executable by the processor 401 (for example, the execution instructions corresponding to the acquisition module 301, the determination module 302, the selection module 303, and the generation module 304 in the device in FIG. , the processor 401 communicates with the memory 402 through the bus 403, and when the machine-readable instructions are executed by the processor 401, the following processing is performed:

Based on each unlabeled sample and the target neural network, determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network;

Select the target unlabeled samples whose estimated influence degree value meets the preset requirements from each unlabeled sample;

In the case of labeling the selected target unlabeled samples to obtain the target labeled samples, the target labeled samples are added to the training sample set to obtain the updated training sample set; the updated training sample set is used to retrain the target neural network Do network training.

An embodiment of the present disclosure also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the method for generating the training sample set described in the above-mentioned method embodiment is executed. step. Wherein, the storage medium may be a volatile or non-volatile computer-readable storage medium.

An embodiment of the present disclosure also provides a computer program. The computer program includes computer readable codes. When the computer readable codes run in the electronic device, the processor of the electronic device executes the training samples as described in any of the above embodiments. The steps of the method for set generation.

Embodiments of the present disclosure also provide a computer program product, the computer program product carries program code, and the instructions included in the program code can be configured to execute the steps of the method for generating the training sample set described in the method embodiment above, which can be See the method example above.

Wherein, the above-mentioned computer program product may be realized by hardware, software or a combination thereof. In some embodiments, the computer program product may be embodied as a computer storage medium, and in other embodiments, the computer program product may be embodied as a software product, such as a software development kit (Software Development Kit, SDK) and the like.

The device involved in the embodiments of the present disclosure may be at least one of a system, a method, and a computer program product. A computer program product may include a computer readable storage medium having computer readable program instructions thereon for causing a processor to implement various aspects of the present disclosure.

A computer readable storage medium may be a tangible device that can retain and store instructions for use by an instruction execution device. A computer readable storage medium may be, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. Examples of computer-readable storage media (a non-exhaustive list) include: portable computer disks, hard disks, Random Access Memory (RAM), Read-Only Memory (ROM), erasable Electrical Programmable Read Only Memory (EPROM) or flash memory, Static Random-Access Memory (Static Random-Access Memory, SRAM), Portable Compact Disc Read-Only Memory (CD-ROM), Digital Video Discs (DVDs), memory sticks, floppy disks, mechanically encoded devices such as punched cards or raised structures in grooves with instructions stored thereon, and any suitable combination of the foregoing. As used herein, computer-readable storage media are not to be construed as transient signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., pulses of light through fiber optic cables), or transmitted electrical signals.

The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to a respective computing/processing device, or to an external computer or external storage device over at least one of a network, such as the Internet, a local area network, a wide area network, and a wireless network. . The network may include at least one of copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and edge servers. A network adapter card or a network interface in each computing/processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing/processing device .

Computer program instructions for performing the operations of the present disclosure may be assembly instructions, Industry Standard Architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or in one or more source or object code written in any combination of programming languages, including object-oriented programming languages—such as Smalltalk, C++, etc., and conventional procedural programming languages, such as the “C” language or similar programming languages. Computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server implement. In cases involving a remote computer, the remote computer may be connected to the user's computer through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or it may be connected to an external computer (for example, using Internet Service Provider to connect via the Internet). In some embodiments, electronic circuits, such as programmable logic circuits, FPGAs, or programmable logic arrays (Programmable Logic Arrays, PLAs), can be customized by using state information of computer-readable program instructions, which can execute computer-readable Read program instructions, thereby implementing various aspects of the present disclosure.

Those skilled in the art can clearly understand that for the convenience and brevity of description, for the specific working process of the system and device described above, reference can be made to the corresponding process in the foregoing method embodiments. In the several embodiments provided in the present disclosure, it should be understood that the disclosed systems, devices and methods may be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or May be integrated into another system, or some features may be ignored, or not implemented. In another point, the mutual coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit.

If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium executable by a processor. Based on this understanding, the essence of the technical solution of the present disclosure or the part that contributes to the related technology or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several The instructions are used to make an electronic device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present disclosure. The aforementioned storage medium includes: various media capable of storing program codes such as U disk, mobile hard disk, ROM, random access memory RAM, magnetic disk or optical disk.

Finally, it should be noted that: the above-mentioned embodiments are only specific implementations of the present disclosure, and are used to illustrate the technical solutions of the present disclosure, rather than limit them, and the protection scope of the present disclosure is not limited thereto, although referring to the aforementioned The embodiments have described the present disclosure in detail, and those skilled in the art should understand that any person familiar with the technical field can still modify the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present disclosure Changes can be easily imagined, or equivalent replacements can be made to some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in this disclosure. within the scope of protection. Therefore, the protection scope of the present disclosure should be defined by the protection scope of the claims.

Industrial Applicability

The present disclosure provides a method, device, electronic device, storage medium, and program for generating a training sample set, wherein the method includes: obtaining each unlabeled sample and the target neural network trained based on the training sample set; Label the samples and the target neural network, and determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network; select the target unlabeled sample whose estimated influence degree value meets the preset requirements from each unlabeled sample; When the selected target unlabeled sample is sample-labeled to obtain the target labeled sample, the target labeled sample is added to the training sample set to obtain the updated training sample set; the updated training sample set is used to re-train the target neural network. network training.

Claims

A method for generating a training sample set, the method being executed by an electronic device, the method comprising:

Obtain each unlabeled sample and the target neural network trained based on the training sample set;

Based on each of the unlabeled samples and the target neural network, determine the estimated influence degree value of each of the unlabeled samples on the network training of the target neural network;

Selecting target unlabeled samples whose estimated influence degree value meets the preset requirements from each of the unlabeled samples;

In the case of performing sample labeling on the selected target unlabeled samples to obtain target labeled samples, adding the target labeled samples to the training sample set to obtain an updated training sample set; the updated training samples The set is used to perform network training on the target neural network again.
The method according to claim 1, wherein the training sample set includes each first labeled sample; and based on the each unlabeled sample and the target neural network, determining that each unlabeled sample has a corresponding The estimated influence degree value of the network training of the target neural network, including:

Based on the respective first marked samples and the target neural network, determine the network training parameters of the respective first marked samples during the forward propagation of the target neural network based on the respective first marked samples The magnitude and value of the estimated impact of ; and,

Based on each of the unlabeled samples and the target neural network, determine the estimated influence degree of each of the unlabeled samples on the network training parameters during the forward propagation of the target neural network based on the respective unlabeled samples. value;

Based on the estimated influence degree and value of each of the first labeled samples on the network training parameters and the estimated influence degree value of each of the unlabeled samples on the network training parameters, determine the impact of each of the unlabeled samples on the The estimated influence value of network training for the target neural network.
The method according to claim 2, wherein the estimated influence degree and value are determined according to the following steps:

Input each of the first labeled samples in the first labeled samples into the target neural network to obtain a gradient sum and a Hessian matrix sum corresponding to a loss function of the target neural network; Wherein, the gradient sum value is used to represent the summation result of the gradient value corresponding to each of the first labeled samples, and the gradient value is used to represent that in the case of forward propagation of the first labeled sample, The degree of influence of each network parameter on the loss function; the Hessian matrix and the value are used to represent the summation result of the Hessian matrix corresponding to each of the first marked samples, and the Hessian matrix is used to represent the sum of the Hessian matrix in In the case of forward propagation of the first labeled sample, the degree of influence of each network parameter on the loss function is affected by the degree of influence of other network parameters;

The estimated influence degree and value are determined based on a product operation of the gradient sum value and the Hessian matrix sum value.
The method according to claim 2, wherein, based on the each unlabeled sample and the target neural network, determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network ,include:

Obtaining each second labeled sample included in the training reference set; the training reference set does not have the same labeled sample as the training sample set;

Based on the respective first marked samples, the respective second marked samples and the target neural network, determine the target neural network based on the respective first marked samples and the respective second marked samples before performing the target neural network In the process of forward propagation, the estimated influence degree and value of each of the first marked samples and each of the second marked samples on the network training parameters; and,

Based on each of the unlabeled samples and the target neural network, determine the estimated influence degree of each of the unlabeled samples on the network training parameters during the forward propagation of the target neural network based on the respective unlabeled samples. value;

Determine each The estimated influence degree value of the unlabeled samples on the network training of the target neural network.
The method according to claim 4, wherein the estimated influence degree and value are determined according to the following steps:

selecting a plurality of first labeled samples from the respective first labeled samples, and inputting each of the first labeled samples in the plurality of first labeled samples into the target neural network, Obtain the Hessian matrix and value corresponding to the loss function of the target neural network; the Hessian matrix and value are used to represent the summation result of the Hessian matrix corresponding to each of the first marked samples, and the Hessian The matrix is used to represent the degree of influence of each network parameter on the loss function by the degree of influence of other network parameters in the case of forward propagation of the first labeled sample; and,

Input each of the second marked samples in the respective second marked samples into the target neural network to obtain the gradient and value corresponding to the loss function of the target neural network; the gradient and value are used to represent A summation result of gradient values corresponding to each of the second labeled samples, where the gradient values are used to represent the influence of each network parameter on the loss function in the case of forward propagation of the second labeled samples degree;

The estimated influence degree and value are determined based on a product operation of the gradient sum value and the Hessian matrix sum value.
The method according to claim 3 or 5, wherein said determining the estimated influence degree and value based on the product operation of said gradient sum value and said Hessian matrix sum value comprises:

For the current round of iterative operation, determine the marked samples pointed to by the current round of iterative operation;

Based on the determined Hessian matrix corresponding to the marked sample, the gradient sum value, and the estimated influence degree and value corresponding to the previous round of iterative operation, determine the estimated influence degree and value corresponding to the current round of iterative operation .
The method according to any one of claims 2 to 6, wherein, based on the respective unlabeled samples and the target neural network, it is determined that in the process of forward propagation of the target neural network based on the respective unlabeled samples , each of the unlabeled samples has an estimated degree of influence on the network training parameters, including:

For each of the unlabeled samples in each unlabeled sample, input each of the unlabeled samples into the target neural network, and determine the probability value for each candidate prediction result output by the target neural network;

Based on the probability value for each candidate prediction result, determine the pseudo-label information of the unlabeled sample;

Based on the pseudo-label information, determine the gradient value corresponding to the loss function of the target neural network in the case of forward propagation of the unlabeled sample;

The determined gradient value is used as an estimated influence degree value of the unlabeled sample on the network training parameters.
The method according to claim 7, wherein the determining the pseudo-labeling information of the unlabeled samples based on the probability values for each candidate prediction result comprises:

When the target neural network is a classification network and the candidate prediction result is a candidate category, determining the candidate category with the largest probability value as the pseudo-label information of the unlabeled sample; or,

When the target neural network includes a detection network and the candidate prediction result is a candidate detection frame, determine a candidate detection frame with a probability value greater than a first preset threshold as the pseudo-label information of the unlabeled sample.
The method according to any one of claims 1 to 8, wherein the selection of target unlabeled samples whose estimated impact degree values meet the preset requirements from the various unlabeled samples includes:

Selecting an unlabeled sample with an estimated influence degree value greater than a second preset threshold as the target unlabeled sample; or,

The unlabeled samples are sorted in descending order of estimated influence degree values, and the target unlabeled samples are determined according to the sorting results.
The method according to any one of claims 1 to 9, wherein, after obtaining the updated training sample set, the method further comprises:

Perform the following steps in a loop until the loop cut-off condition is reached, and the updated target neural network is obtained:

Screen out the target unlabeled samples from each unlabeled sample to obtain each updated unlabeled sample; and determine an updated target neural network trained based on the updated training sample set; wherein, the updated The training sample set includes each first labeled sample and the target labeled sample;

Based on each of the updated unlabeled samples and the updated target neural network, determine an estimated influence degree value of each of the updated unlabeled samples on the network training of the updated target neural network;

Select the target unlabeled samples whose estimated impact value meets the preset requirements from each of the updated unlabeled samples;

In the case of performing sample labeling on the selected target unlabeled samples to obtain target labeled samples, adding the target labeled samples to the updated training sample set to obtain a neural network for the updated target neural network The updated training sample set for training.
A device for generating a training sample set, the device comprising:

An acquisition module configured to acquire each unlabeled sample and the target neural network trained based on the training sample set;

The determination module is configured to determine the estimated influence degree value of each unlabeled sample on the network training of the target neural network based on the respective unlabeled samples and the target neural network;

The selection module is configured to select target unlabeled samples whose estimated influence degree value meets the preset requirements from each of the unlabeled samples;

The generation module is configured to add the target labeled samples to the training sample set to obtain an updated training sample set when the target unlabeled sample is sample labeled to obtain the target labeled sample; The updated training sample set is used to perform network training on the target neural network again.
An electronic device, comprising: a processor, a memory, and a bus, the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory through the bus , when the machine-readable instructions are executed by the processor, the steps of the method for generating the training sample set according to any one of claims 1 to 10 are executed.
A computer-readable storage medium, on which a computer program is stored, and when the computer program is run by a processor, the steps of the method for generating a training sample set according to any one of claims 1 to 10 are executed.
A computer program, the computer program comprising computer readable code, in the case of the computer readable code running in an electronic device, the processor of the electronic device executes to implement any one of claims 1 to 10 The steps of the method for generating the training sample set.