A method, device and storage medium for automatically updating deep learning models

Abstract

The present invention proposes a method, device and storage medium for automatically updating a deep learning model, the method comprising: a server deploys a trained initial first deep learning model to a client, collects input data for a user as actual samples collected, and uses a second deep learning model on the server to machine-label the actual samples to obtain labeled actual samples; the server uses a first data set and the labeled actual samples to retrain the initial first deep learning model to update the initial first deep learning model. The present invention deploys multiple powerful algorithm models on a back-end server to collect and automatically label data in a real application environment, and then uses the labeled data set to retrain a relatively simple model on the client, and updates it to the front-end device, thereby solving the technical problem that the client learning model cannot be updated in time due to the slow speed of manual labeling of samples, and improving the user experience.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device and storage medium for automatically updating a deep learning model. Background Art

[0002] With the development of artificial intelligence, machine learning-based algorithms are increasingly being used. The main steps for training algorithm models are: manual data labeling, model training, and model deployment. Algorithm models need to be trained using a large number of training samples, and then tested using test samples. After the test passes, they are put into online use. Currently, data labeling still mainly relies on manual methods. As the amount of data increases, the cost of using manual labeling methods will increase significantly. For complex applications such as semantic segmentation, manually labeling data is inefficient.

[0003] Currently, the data sets used for model training mainly use open source public data sets. The models trained with these data sets may not belong to the same distribution as the actual scenario application data, and the trained models are not effective in actual scenario applications.

[0004] With the popularization of edge computing, more and more algorithm models are deployed in front-end devices. Due to the limitations of the computing performance of front-end devices and the rigid requirements of the business on the real-time performance of the algorithm, many relatively complex algorithm models with good effects cannot be deployed on the front end. We can only deploy some relatively small models with the required accuracy and speed.

[0005] It can be seen that in the prior art, it is impossible to use actual samples to retrain the learning model for specific scenarios, resulting in poor recognition performance of the learning model, affecting the user experience. In addition, in the prior art, training data sets (also called training samples) are manually labeled, and the labeling efficiency is low, which cannot adapt to the deployment speed of the learning model on the front end. Even if there are some machine-labeled examples, the standard accuracy is low due to the unreasonable setting of the model weight. Summary of the invention

[0006] The present invention aims at one or more defects in the above-mentioned prior art and proposes the following technical solutions.

[0007] A method for automatically updating a deep learning model, the method comprising:

[0008] A deployment step, in which the server trains the first deep learning model using the first data set to obtain a trained initial first deep learning model, and deploys the initial first deep learning model to the client;

[0009] In a collection step, a user uses the client to input data and sends the input data to the server as an actual sample collected;

[0010] A labeling step, using a second deep learning model on the server side to machine label the actual sample to obtain a labeled actual sample;

[0011] A retraining step, in which the server retrains the initial first deep learning model using the first data set and the annotated actual samples to obtain a retrained first deep learning model;

[0012] An updating step is performed to determine whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model. If so, the server sends the retrained first deep learning model to the client to update the initial first deep learning model.

[0013] Furthermore, the recognition performance of the first deep learning model is lower than that of the second deep learning model, and the system resources consumed when the first deep learning model is running are less than the system resources consumed when the second deep learning model is running.

[0014] Furthermore, the input data is face image data, fingerprint image data, iris image data or voiceprint data.

[0015] Furthermore, the operation of machine-labeling the actual samples using the second deep learning model on the server side to obtain the labeled actual samples is:

[0016] There are multiple second deep learning models, and the actual samples are identified by using the multiple second deep learning models. If the identification results of each second deep learning model are consistent, the actual samples are labeled by using the identification results;

[0017] Alternatively, there are multiple second deep learning models, and the multiple second deep learning models are used to identify the actual sample to obtain corresponding recognition results, and all recognition results are weightedly added to label the actual sample as the recognition result;

[0018] Alternatively, there are multiple second deep learning models, and the multiple second deep learning models are connected in series to identify the actual sample to obtain the final recognition result and label the actual sample.

[0019] Furthermore, the operation of judging whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model is: using the retrained first deep learning model to identify partially labeled actual samples, if the recognition rate is greater than the accuracy rate of the initial first deep learning model for the partially labeled actual samples, then the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.

[0020] The present invention also proposes a device for automatically updating a deep learning model, the device comprising:

[0021] A deployment unit, configured for the server to train the first deep learning model using the first data set to obtain a trained initial first deep learning model, and deploy the initial first deep learning model to the client;

[0022] A collection unit, used for a user to input data using the client and send the input data to the server as an actual sample collected;

[0023] A labeling unit, configured to use a second deep learning model on the server side to machine label the actual sample to obtain a labeled actual sample;

[0024] A retraining unit, configured for the server to retrain the initial first deep learning model using the first data set and the annotated actual samples to obtain a retrained first deep learning model;

[0025] An updating unit is used to determine whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model. If so, the server sends the retrained first deep learning model to the client to update the initial first deep learning model.

[0026] Furthermore, the recognition performance of the first deep learning model is lower than that of the second deep learning model, and the system resources consumed when the first deep learning model is running are less than the system resources consumed when the second deep learning model is running.

[0027] Furthermore, the input data is face image data, fingerprint image data, iris image data or voiceprint data.

[0028] Furthermore, the operation of machine-labeling the actual samples using the second deep learning model on the server side to obtain the labeled actual samples is:

[0029] There are multiple second deep learning models, and the actual samples are identified by using the multiple second deep learning models. If the identification results of each second deep learning model are consistent, the actual samples are labeled by using the identification results;

[0030] Alternatively, there are multiple second deep learning models, and the multiple second deep learning models are used to identify the actual sample to obtain corresponding recognition results, and all recognition results are weightedly added to label the actual sample as the recognition result;

[0031] Alternatively, there are multiple second deep learning models, and the multiple second deep learning models are connected in series to identify the actual sample to obtain the final recognition result and label the actual sample.

[0032] Furthermore, the operation of judging whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model is: using the retrained first deep learning model to identify partially labeled actual samples, if the recognition rate is greater than the accuracy rate of the initial first deep learning model for the partially labeled actual samples, then the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.

[0033] The present invention also provides a computer-readable storage medium, on which a computer program code is stored. When the computer program code is executed by a computer, any of the above-mentioned devices is executed.

[0034] The technical effect of the present invention is: a method, device and storage medium for automatically updating a deep learning model of the present invention, the method comprising: a server uses a first data set to train a first deep learning model to obtain an initial first deep learning model after training, and deploys the initial first deep learning model to a client; a user uses the client to input data and sends the input data to the server as an actual sample collected; the server uses a second deep learning model to machine-label the actual sample to obtain an annotated actual sample; the server uses the first data set and the annotated actual sample to retrain the initial first deep learning model to obtain a retrained first deep learning model; it is determined whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model, and if so, the server sends the retrained first deep learning model to the client to update the initial first deep learning model. The present invention deploys multiple powerful algorithm models on a back-end server to collect and automatically annotate data in a real application environment, and then uses the annotated data set to retrain a relatively simple model on the client, and updates it to the front-end device, thereby solving the technical problem that the client learning model cannot be updated in time due to the slow speed of manual sample labeling, improving the user experience, and proposing a weight calculation formula for multi-model labeling, thereby improving the accuracy of multi-model labeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Other features, objects and advantages of the present application will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings.

[0036] Figure 1 is a flowchart of a method for automatically updating a deep learning model according to an embodiment of the present invention.

[0037] Figure 2 is a structural diagram of a device for automatically updating a deep learning model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0039] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0040] Figure 1A method for automatically updating a deep learning model of the present invention is shown, the method comprising:

[0041] In the deployment step S101, the server uses the first data set to train the first deep learning model to obtain the trained initial first deep learning model, and deploys the initial first deep learning model to the client. For example, for the currently available data set, which is a general data set, such as a data set downloaded from the network, the model (i.e., the first deep learning model, which can be a deep neural network model, etc.) is trained for the first time by the back-end server and deployed to the front-end device (i.e., the client). For example, the face detection algorithm Adaboost+LBP is a relatively simple face detection algorithm model. Due to the use of a general data set for training, the recognition performance of the initial learning model is relatively low in the actual environment.

[0042] In the collection step S102, the user uses the client to input data and sends the input data to the server as the collected actual samples. For example, when the user uses the client to perform online face scanning payment, a first number of samples are collected from online input samples, such as collecting a first number of face image samples from users distributed across the country.

[0043] In the labeling step S103, the actual sample is machine labeled using the second deep learning model on the server side to obtain the labeled actual sample. For example, the face detection algorithm MTCNN and Retinaface (i.e., the second deep learning model) are used, and the performance of these algorithms is stronger than the Adaboost+LBP algorithm. Multiple algorithms can be filtered in a cascade manner, because they pay more attention to the accuracy of data labeling, so only when the results of each algorithm labeling in the back end are consistent, the final labeling result is considered accurate. Once there are inconsistencies in the labeling results of some samples, filtering is performed. Filtering in a cascade manner is equivalent to a veto, that is, when a model believes that the detected sample is not a positive sample, it is filtered out, which can greatly improve the accuracy. Of course, other filtering methods can also be used, such as the majority obeys the minority, such as two models believe that it is a positive sample, and one model believes that it is a negative sample. The result is still considered to be a positive sample, but this will lead to a decrease in accuracy.

[0044] By using the backend to deploy multiple models that are more powerful than the frontend, the actual scene sample data is labeled. This is because the computing power of the backend server is much better than that of the client. Therefore, the backend server is used to use a high-performance learning algorithm to label the sample data in the actual environment, which solves the technical problem in the prior art that the slow manual labeling speed makes it impossible to update the client learning model in time. This is an important invention point of the present invention.

[0045] In the retraining step S104, the server uses the first data set and the annotated actual samples to retrain the initial first deep learning model to obtain a retrained first deep learning model.

[0046] In an updating step S105 , it is determined whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model. If so, the server sends the retrained first deep learning model to the client to update the initial first deep learning model.

[0047] In one embodiment, the recognition performance of the first deep learning model is lower than that of the second deep learning model, and the system resources consumed by the first deep learning model when running are less than the system resources consumed by the second deep learning model when running. The input data is face image data, fingerprint image data, iris image data or voiceprint data, which are collected through a camera, fingerprint sensor, sound sensor, etc.

[0048] In one embodiment, in order to prepare to use machine to annotate sample data in an actual environment, the operation of machine annotating the actual samples using the second deep learning model on the server to obtain the annotated actual samples is:

[0049] There are multiple second deep learning models, and the actual samples are identified using the multiple second deep learning models. If the identification results of each of the second deep learning models are consistent, the actual samples are labeled using the identification results. This situation determines that the sample set is optimal, which ensures that each sample is correct.

[0050] Alternatively, there are multiple second deep learning models, and the actual samples are identified using the multiple second deep learning models to obtain corresponding recognition results, and all recognition results are weighted added together to label the actual samples as the recognition results; this case takes into account the performance of multiple second deep learning models, and different recognition weights can be given according to second deep learning models with different performances.

[0051] Preferably, the recognition weight setting formulas of the multiple second deep learning models are: where y i is the accuracy of the i-th model. For example, the backend deploys three second-order deep learning models with accuracy rates of 85%, 90%, and 95%, respectively. Then their weights are This weight setting method is created by the exclusive research institute of this application, which can maximize the accuracy of sample machine labeling and is one of the important invention points of the present invention.

[0052] Alternatively, there are multiple second deep learning models, and the multiple second deep learning models are connected in series to identify the actual sample to obtain the final recognition result and label the actual sample. In this case, the collected actual samples are identified by filtering. When an actual sample is identified as negative by a second deep learning model, the sample is discarded and other actual samples are re-identified.

[0053] Through this operation, it is ensured that the collected labeled samples are all positive samples. During the training process, in order to ensure the convergence of the algorithm training, negative samples must also be collected. Only when the positive and negative samples maintain a certain ratio can a better model be trained. Negative samples can be automatically generated by a machine or downloaded from the Internet. The ratio of positive and negative samples during model training may vary according to different algorithms. The collection of negative samples is relatively simple, for example: random cropping of actual scene pictures, including various subsequent operations, such as color transformation, mirror transformation, etc.

[0054] In one embodiment, the operation of determining whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model is: using the retrained first deep learning model to identify partially labeled actual samples, if the recognition rate is greater than the accuracy rate of the initial first deep learning model for the partially labeled actual samples, then the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.

[0055] A portion of the annotated actual sample data is used as a test set. Since the first deep learning should be as effective as possible in actual scenarios, the test set must contain only actual scenario data, rather than standard online data sets. The retrained first deep learning model is tested and verified using the test set. When the test result is better than the online model (i.e., the initial first deep learning model), the model is updated. After a period of time, when the actual samples reach a certain level, the effect of the first deep learning model tends to be stable. After that, the first deep learning model can no longer be retrained, freeing up the resources of the backend server and reducing the pressure on the server.

[0056] This method deploys multiple powerful algorithm models on the back-end server to collect and automatically annotate data in a real application environment, and then uses the annotated data set to retrain a relatively simple model on the client side and update it to the front-end device, thereby solving the technical problem that the client learning model cannot be updated in time due to the slow speed of manual annotation of samples, and improving the user experience. The present invention can train the front-end learning model with data from the actual environment using automatic annotation in the background until the performance of the learning model tends to be stable, which is another important inventive point of the present invention.

[0057] Figure 2 A device for automatically updating a deep learning model of the present invention is shown, the device comprising:

[0058] The deployment unit 201 is used for the server to train the first deep learning model using the first data set to obtain the trained initial first deep learning model, and deploy the initial first deep learning model to the client. For example, for the currently available data set, which is a general data set, such as a data set downloaded from the network, the model (i.e., the first deep learning model, which can be a deep neural network model, etc.) is trained for the first time by the back-end server and deployed to the front-end device (i.e., the client). For example, the face detection algorithm Adaboost+LBP is a relatively simple face detection algorithm model. Due to the use of a general data set for training, the recognition performance of the initial learning model is relatively low in the actual environment.

[0059] The collection unit 202 is used for the user to input data using the client and send the input data to the server as the collected actual samples. For example, when the user uses the client to perform online face scanning payment, a first number of samples are collected from online input samples, such as collecting a first number of face image samples from users distributed across the country.

[0060] The labeling unit 203 is used to use the second deep learning model to machine label the actual sample on the server side to obtain the labeled actual sample. For example, the face detection algorithm MTCNN and Retinaface (i.e., the second deep learning model) are used, and the performance of these algorithms is stronger than the Adaboost+LBP algorithm. Multiple algorithms can be filtered in a cascade manner, because they pay more attention to the accuracy of data labeling, so only when the results of each algorithm labeling at the back end are consistent, the final labeling result is considered accurate. Once there are inconsistencies in the labeling results of some samples, filtering is performed. Filtering in a cascade manner is equivalent to a veto, that is, when a model believes that the detected sample is not a positive sample, it is filtered out, which can greatly improve the accuracy. Of course, other filtering methods can also be used, such as the majority obeys the minority, such as two models believe that it is a positive sample, and one model believes that it is a negative sample. The result is still considered to be a positive sample, but this will lead to a decrease in accuracy.

[0061] By using the backend to deploy multiple models that are more powerful than the frontend, the actual scene sample data is labeled. This is because the computing power of the backend server is much better than that of the client. Therefore, the backend server is used to use a high-performance learning algorithm to label the sample data in the actual environment, which solves the technical problem in the prior art that the slow manual labeling speed makes it impossible to update the client learning model in time. This is an important invention point of the present invention.

[0062] The retraining unit 204 is used for the server to retrain the initial first deep learning model using the first data set and the annotated actual samples to obtain a retrained first deep learning model.

[0063] The updating unit 205 is used to determine whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model. If so, the server sends the retrained first deep learning model to the client to update the initial first deep learning model.

[0064] In one embodiment, the recognition performance of the first deep learning model is lower than that of the second deep learning model, and the system resources consumed by the first deep learning model when running are less than the system resources consumed by the second deep learning model when running. The input data is face image data, fingerprint image data, iris image data or voiceprint data, which are collected through a camera, fingerprint sensor, sound sensor, etc.

[0065] In one embodiment, in order to prepare to use machine to annotate sample data in an actual environment, the operation of machine annotating the actual samples using the second deep learning model on the server to obtain the annotated actual samples is:

[0066] There are multiple second deep learning models, and the actual samples are identified using the multiple second deep learning models. If the identification results of each of the second deep learning models are consistent, the actual samples are labeled using the identification results. This situation determines that the sample set is optimal, which ensures that each sample is correct.

[0067] Alternatively, there are multiple second deep learning models, and the actual samples are identified using the multiple second deep learning models to obtain corresponding recognition results, and all recognition results are weighted added together to label the actual samples as the recognition results; this case takes into account the performance of multiple second deep learning models, and different recognition weights can be given according to second deep learning models with different performances.

[0068] Preferably, the recognition weight setting formulas of the multiple second deep learning models are: where y iis the accuracy of the i-th model. For example, the backend deploys three second-order deep learning models with accuracy rates of 85%, 90%, and 95%, respectively. Then their weights are This weight setting method is created by the exclusive research institute of this application, which can maximize the accuracy of sample machine labeling and is one of the important invention points of the present invention.

[0069] Alternatively, there are multiple second deep learning models, and the multiple second deep learning models are connected in series to identify the actual sample to obtain the final recognition result and label the actual sample. In this case, the collected actual samples are identified by filtering. When an actual sample is identified as negative by a second deep learning model, the sample is discarded and other actual samples are re-identified.

[0070] Through this operation, it is ensured that the collected labeled samples are all positive samples. During the training process, in order to ensure the convergence of the algorithm training, negative samples must also be collected. Only when the positive and negative samples maintain a certain ratio can a better model be trained. Negative samples can be automatically generated by a machine or downloaded from the Internet. The ratio of positive and negative samples during model training may vary according to different algorithms. The collection of negative samples is relatively simple, for example: random cropping of actual scene pictures, including various subsequent operations, such as color transformation, mirror transformation, etc.

[0071] In one embodiment, the operation of determining whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model is: using the retrained first deep learning model to identify partially labeled actual samples, if the recognition rate is greater than the accuracy rate of the initial first deep learning model for the partially labeled actual samples, then the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.

[0072] A portion of the annotated actual sample data is used as a test set. Since the first deep learning should be as effective as possible in actual scenarios, the test set must contain only actual scenario data, rather than standard online data sets. The retrained first deep learning model is tested and verified using the test set. When the test result is better than the online model (i.e., the initial first deep learning model), the model is updated. After a period of time, when the actual samples reach a certain level, the effect of the first deep learning model tends to be stable. After that, the first deep learning model can no longer be retrained, freeing up the resources of the backend server and reducing the pressure on the server.

[0073] The present invention is used to collect and automatically label data in a real application environment by deploying multiple powerful algorithm models on the back-end server, and then use the labeled data set to retrain a relatively simple model on the client side and update it to the front-end device, thereby solving the technical problem that the client learning model cannot be updated in time due to the slow speed of manual labeling of samples, and improving the user experience. The present invention can train the front-end learning model with data from the actual environment using automatic labeling in the background until the performance of the learning model tends to be stable, which is another important inventive point of the present invention.

[0074] For the convenience of description, the above device is described in various units according to their functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0075] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the devices described in the various embodiments of the present application or certain parts of the embodiments.

[0076] Finally, it should be noted that the above embodiments are only intended to illustrate rather than limit the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the present invention can still be modified or replaced by equivalents. Any modification or partial replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for automatically updating a deep learning model, characterized in that: The method includes: A deployment step, in which the server trains the first deep learning model using the first data set to obtain a trained initial first deep learning model, and deploys the initial first deep learning model to the client; In a collection step, a user uses the client to input data and sends the input data to the server as an actual sample collected; A labeling step, using a second deep learning model on the server side to machine label the actual sample to obtain a labeled actual sample; The operation of machine-labeling the actual samples using the second deep learning model on the server side to obtain the labeled actual samples is: There are multiple second deep learning models, and the actual samples are identified by using the multiple second deep learning models to obtain corresponding identification results, and all the identification results are weightedly added to label the actual samples as the identification results; The recognition weight setting formula of multiple second deep learning models is: Among them, y i Represents the accuracy of the i-th model; A retraining step, in which the server retrains the initial first deep learning model using the first data set and the annotated actual samples to obtain a retrained first deep learning model; An updating step is performed to determine whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model. If so, the server sends the retrained first deep learning model to the client to update the initial first deep learning model.

2. The method according to claim 1, characterized in that The recognition performance of the first deep learning model is lower than that of the second deep learning model, and the system resources consumed by the first deep learning model when running are less than the system resources consumed by the second deep learning model when running.

3. The method according to claim 2, characterized in that The input data is face image data, fingerprint image data, iris image data or voiceprint data.

4. The method according to claim 1, characterized in that: The operation of judging whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model is: using the retrained first deep learning model to identify partially labeled actual samples, if the recognition rate is greater than the accuracy rate of the initial first deep learning model for the partially labeled actual samples, then the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.

5. A device for automatically updating a deep learning model, characterized in that: The device includes: A deployment unit, configured for the server to train the first deep learning model using the first data set to obtain a trained initial first deep learning model, and deploy the initial first deep learning model to the client; A collection unit, used for a user to input data using the client and send the input data to the server as an actual sample collected; A labeling unit, configured to use a second deep learning model on the server side to machine label the actual sample to obtain a labeled actual sample; The operation of machine-labeling the actual samples using the second deep learning model on the server side to obtain the labeled actual samples is: There are multiple second deep learning models, and the actual samples are identified by using the multiple second deep learning models to obtain corresponding identification results, and all the identification results are weightedly added to label the actual samples as the identification results; The recognition weight setting formula of multiple second deep learning models is: Among them, y i Represents the accuracy of the i-th model; A retraining unit, configured for the server to retrain the initial first deep learning model using the first data set and the annotated actual samples to obtain a retrained first deep learning model; An updating unit is used to determine whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model. If so, the server sends the retrained first deep learning model to the client to update the initial first deep learning model.

6. The device according to claim 5, characterized in that The recognition performance of the first deep learning model is lower than that of the second deep learning model, and the system resources consumed by the first deep learning model when running are less than the system resources consumed by the second deep learning model when running.

7. The device according to claim 6, characterized in that The input data is face image data, fingerprint image data, iris image data or voiceprint data.

8. The device according to claim 5, characterized in that The operation of judging whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model is: using the retrained first deep learning model to identify partially labeled actual samples, if the recognition rate is greater than the accuracy rate of the initial first deep learning model for the partially labeled actual samples, then the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.

9. A computer-readable storage medium, characterized in that: The storage medium stores computer program codes, and when the computer program codes are executed by a computer, the method according to any one of claims 1 to 4 is performed.

Images