CN111612161A - Method, device and storage medium for automatically updating deep learning model - Google Patents
Method, device and storage medium for automatically updating deep learning model Download PDFInfo
- Publication number
- CN111612161A CN111612161A CN202010456082.6A CN202010456082A CN111612161A CN 111612161 A CN111612161 A CN 111612161A CN 202010456082 A CN202010456082 A CN 202010456082A CN 111612161 A CN111612161 A CN 111612161A
- Authority
- CN
- China
- Prior art keywords
- deep learning
- learning model
- initial
- actual
- retrained
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013136 deep learning model Methods 0.000 title claims abstract description 217
- 238000000034 method Methods 0.000 title claims abstract description 34
- 238000002372 labelling Methods 0.000 claims abstract description 53
- 238000012549 training Methods 0.000 claims description 17
- 238000004590 computer program Methods 0.000 claims description 3
- 230000001815 facial effect Effects 0.000 claims description 3
- 238000004422 calculation algorithm Methods 0.000 abstract description 31
- 238000012360 testing method Methods 0.000 description 12
- 238000001914 filtration Methods 0.000 description 10
- 230000000694 effects Effects 0.000 description 9
- 238000001514 detection method Methods 0.000 description 6
- 230000009466 transformation Effects 0.000 description 4
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a method, a device and a storage medium for automatically updating a deep learning model, wherein the method comprises the following steps: deploying the trained initial first deep learning model to the client by the server, collecting input data for a user as a collected actual sample, and performing machine labeling on the actual sample by using a second deep learning model at the server to obtain a labeled actual sample; and the server updates the initial first deep learning model by using the first data set and the labeled actual sample to the retrained first deep learning model. According to the invention, a plurality of powerful algorithm models are deployed at the back-end server to carry out data acquisition in a real application environment and automatic labeling is carried out, then a relatively simple model of the client is retrained by using the labeled data set and is updated to the front-end equipment, so that the technical problem that the learning model of the client cannot be updated in time due to the slow speed of manually labeling samples is solved, and the user experience is improved.
Description
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to a method and a device for automatically updating a deep learning model and a storage medium.
Background
With the development of artificial intelligence, the algorithm based on machine learning is applied more and more widely at present. The general training algorithm model mainly comprises the following steps: manual annotation of data, model training, and model deployment. The algorithm model needs to be trained by a large number of training samples, then is tested by using the test samples, and is put into online use after the test is passed. At present, data annotation mainly depends on a manual participation method. With the increasing of the data volume, the cost of the method using manual labeling is greatly increased, and the efficiency of manually labeling data is low for complex applications such as semantic segmentation and other tasks.
The data set used by the current model training mainly uses an open-source public data set, the models trained by the data sets may not belong to the same distribution with the actual scene application data, and the effect of the trained models is not good in the actual scene application.
With the popularization of edge computing, more and more algorithm models are deployed in front-end equipment, and due to the limitation of computing performance conditions of the front-end equipment and the hard requirement of business on algorithm real-time performance, many more complex algorithm models with good effects cannot be deployed at the front end, only some models can be deployed in a compromise mode, the accuracy can meet the requirement, and the speed can meet the required algorithm models.
It can be seen that, in the prior art, the learning model cannot be retrained by using actual samples for a specific scene, which results in poor recognition performance of the learning model and affects user experience, and in the prior art, training data sets (also referred to as training samples) are marked manually, so that marking efficiency is low, which cannot be applied to the deployment speed of the learning model at the front end, and even if there are some examples of machine marking, standard accuracy is low due to unreasonable weight setting of the model.
Disclosure of Invention
The present invention addresses one or more of the above-mentioned deficiencies in the prior art and proposes the following technical solutions.
A method of automatically updating a deep learning model, the method comprising:
a deployment step, wherein a server trains a first deep learning model by using a first data set to obtain a trained initial first deep learning model, and deploys the initial first deep learning model to a client;
a collection step, wherein a user uses the client to input data and sends the input data to the server as a collected actual sample;
a labeling step, in which a second deep learning model is used at the server end to perform machine labeling on the actual sample to obtain a labeled actual sample;
retraining, namely retraining the initial first deep learning model by using the first data set and the labeled actual sample by the server to obtain a retrained first deep learning model;
and an updating step, namely judging whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model, and if so, sending the retrained first deep learning model to the client by the server to update the initial first deep learning model.
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 during the running of the first deep learning model are smaller than those consumed during the running of the second deep learning model.
Further, the input data is face image data, fingerprint image data, iris image data or voiceprint data.
Further, the operation of performing machine labeling on the actual sample by using a second deep learning model at the server end to obtain a labeled actual sample is as follows:
the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples, and if the identification results of the second deep learning models are consistent, the identification results are used for marking the actual samples;
or the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples to obtain corresponding identification results, and all the identification results are weighted and added to serve as identification results to label the actual samples;
or the number of the second deep learning models is multiple, and the actual samples are identified after the multiple second deep learning models are connected in series to obtain the final identification result, and then the actual samples are labeled.
Further, 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 as follows: and identifying a part of marked actual samples by using the retrained first deep learning model, wherein if the identification rate is greater than the accuracy rate of the part of marked actual samples by the initial first deep learning model, the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.
The invention also provides a device for automatically updating the deep learning model, which comprises:
the deployment unit is used for training the first deep learning model by using a first data set to obtain a trained initial first deep learning model and deploying the initial first deep learning model to the client;
the acquisition unit is used for inputting data by a user through the client and sending the input data to the server as an acquired actual sample;
the labeling unit is used for performing machine labeling on the actual sample by using a second deep learning model at the server end to obtain a labeled actual sample;
the retraining unit is used for retraining the initial first deep learning model by using the first data set and the labeled actual sample by the server to obtain a retrained first deep learning model;
and the updating unit is used for judging 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.
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 during the running of the first deep learning model are smaller than those consumed during the running of the second deep learning model.
Further, the input data is face image data, fingerprint image data, iris image data or voiceprint data.
Further, the operation of performing machine labeling on the actual sample by using a second deep learning model at the server end to obtain a labeled actual sample is as follows:
the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples, and if the identification results of the second deep learning models are consistent, the identification results are used for marking the actual samples;
or the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples to obtain corresponding identification results, and all the identification results are weighted and added to serve as identification results to label the actual samples;
or the number of the second deep learning models is multiple, and the actual samples are identified after the multiple second deep learning models are connected in series to obtain the final identification result, and then the actual samples are labeled.
Further, 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 as follows: and identifying a part of marked actual samples by using the retrained first deep learning model, wherein if the identification rate is greater than the accuracy rate of the part of marked actual samples by the initial first deep learning model, the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.
The present invention also proposes a computer-readable storage medium having stored thereon computer program code means for performing any of the above-mentioned means when said computer program code means is executed by a computer.
The invention has the technical effects that: the invention discloses a method, a device and a storage medium for automatically updating a deep learning model, wherein the method comprises the following steps: the server trains a first deep learning model by using a first data set to obtain a trained initial first deep learning model, and deploys the initial first deep learning model to the client; a user inputs data by using the client and sends the input data to the server as a collected actual sample; performing machine labeling on the actual sample by using a second deep learning model at the server side to obtain a labeled actual sample; the server retrains the initial first deep learning model by using the first data set and the labeled actual sample to obtain a retrained first deep learning model; and judging whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model, if so, sending the retrained first deep learning model to the client by the server to update the initial first deep learning model. According to the invention, a plurality of powerful algorithm models are deployed at the back-end server to carry out data acquisition and automatic labeling in a real application environment, and then the labeled data set is used for retraining a simpler model at the client and is updated into the front-end equipment, so that the technical problem that the learning model at the client cannot be updated in time due to the slow speed of manually labeling samples is solved, the user experience is improved, a weight calculation formula during multi-model labeling is provided, and the accuracy of using multi-model labeling is improved.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the following detailed description of non-limiting embodiments thereof, made with reference to the accompanying drawings.
FIG. 1 is a flow diagram of a method of automatically updating a deep learning model, according to an embodiment of the invention.
Fig. 2 is a block diagram of an apparatus for automatically updating a deep learning model according to an embodiment of the present invention.
Detailed Description
The present application will be described in further detail with reference to the following drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not restrictive of the invention. It should be noted that, for convenience of description, only the portions related to the related invention are shown in the drawings.
It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict. The present application will be described in detail below with reference to the embodiments with reference to the attached drawings.
FIG. 1 illustrates a method of the present invention for automatically updating a deep learning model, the method comprising:
and a deploying step S101, wherein the server trains the first deep learning model by 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. For example, for a currently existing data set, the data set is a general data set, such as a data set downloaded through a network, a model is trained for the first time through a back-end server (i.e., a first deep learning model, a deep neural network model, etc.), and the data set is deployed into a front-end device (i.e., a client). For example, the face detection algorithm Adaboost + LBP is a relatively simple face detection algorithm model, and the initial learning model has relatively low recognition performance in an actual environment due to the adoption of a universal data set for training.
And a collecting step S102, wherein a user inputs data by using the client and sends the input data to the server as a collected actual sample. For example, when a user uses a client to pay for an online scan, a first number of samples are collected from online input samples, for example, a first number of facial image samples are collected from users distributed across the country.
And a labeling step S103, performing machine labeling on the actual sample by using a second deep learning model at the server side to obtain a labeled actual sample. For example, the face detection algorithms MTCNN and Retinaface (i.e., the second deep learning model) are used, and the performance of the algorithms is stronger than that of the Adaboost + LBP algorithm. The multiple algorithms can be filtered in a cascading mode, because the accuracy of data labeling is concerned more, the final labeling result is considered to be accurate only when the labeling result of each algorithm at the rear end is consistent, and the filtering is performed once the labeling results of some samples are inconsistent. The filtering is performed in a cascading mode, which is equivalent to a vote rejection, namely, when a model considers that a detected sample is not a positive sample, the filtering is performed, so that the accuracy can be greatly improved. Other filtering means may of course be used, such as most obeying a minority, such as two models considered positive samples and one model considered negative samples, and the result still considered positive samples, but this may result in a loss of accuracy.
The method comprises the steps of deploying a plurality of models stronger than a front end at a rear end, and labeling actual scene sample data, wherein the computation capability of a rear end server is much better than that of a client, so that the rear end server is used for labeling the sample data in an actual environment by using a high-performance learning algorithm, the technical problem that the client learning model cannot be updated in time due to low manual labeling speed in the prior art is solved, and the method is an important invention point.
And a retraining step S104, in which the server retrains the initial first deep learning model by using the first data set and the labeled actual sample to obtain a retrained first deep learning model.
And an updating step S105, judging whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model, if so, sending the retrained first deep learning model to the client by the server to update the initial first deep learning model.
In one embodiment, the first deep learning model has lower recognition performance than the second deep learning model, and the first deep learning model consumes less system resources at runtime than the second deep learning model. The input data is face image data, fingerprint image data, iris image data or voiceprint data, and the data is collected through a camera, a fingerprint sensor, a voice sensor and the like.
In one embodiment, in order to prepare for machine labeling of sample data in an actual environment, performing, at the server side, machine labeling on the actual sample using a second deep learning model to obtain an operation of obtaining a labeled actual sample is:
the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples, and if the identification results of the second deep learning models are consistent, the identification results are used for marking the actual samples; this is the case where the determined sample set is optimal, i.e. it is guaranteed that each sample is correct.
Or the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples to obtain corresponding identification results, and all the identification results are weighted and added to serve as identification results to label the actual samples; in this case, the performance of the plurality of second deep learning models is taken into consideration, and different recognition weights may be given according to the second deep learning models having different performances.
Preferably, the recognition weights of the plurality of second deep learning models set a formula:
wherein y isiThe accuracy of the ith model. For example, if the back end deploys three second deep learning models with respective accuracies of 85%, 90%, and 95%, then the weights are respectively
The weight setting method is created by the exclusive research institute of the application, can realize the accuracy of the labeling of the sample machine to the maximum extent, and is the inventionThe method is one of important inventions.
Or the number of the second deep learning models is multiple, and the actual samples are identified after the multiple second deep learning models are connected in series to obtain the final identification result, and then the actual samples are labeled. In this case, the collected actual samples are identified by filtering, and when an actual sample is identified as not by a certain second deep learning model, the sample is discarded, and other actual samples are identified again.
Through the operation, the collected labeling samples are guaranteed to be positive samples, and due to the fact that in the training process, algorithm training convergence is guaranteed, negative samples are also collected, and a good model can be trained only when the positive samples and the negative samples are guaranteed to maintain a certain proportion. The negative sample can be automatically generated by a machine or downloaded through a network. The proportion of positive and negative samples during model training may be different according to different algorithms, and the collection of negative samples is simpler, for example: and randomly cutting the actual scene picture, including various subsequent operations, such as color transformation, mirror image transformation and the like.
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 as follows: and identifying a part of marked actual samples by using the retrained first deep learning model, wherein if the identification rate is greater than the accuracy rate of the part of marked actual samples by the initial first deep learning model, the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.
And (3) taking part of data of the labeled actual sample as a test set, wherein the test set is required to be completely actual scene data instead of a standard data set on the network because the first deep learning is ensured to have the best effect in an actual scene as far as possible. And (3) carrying out test verification on the retrained first deep learning model by using the test set, and updating the model when the test effect is better than that of the on-line model (namely the initial first deep learning model). After a period of time, when the actual sample reaches a certain magnitude, the effect of the first deep learning model tends to be stable, and then the first deep learning model can not be retrained any more, so that the resources of the back-end server are released, and the pressure of the server is reduced.
According to the method, a plurality of powerful algorithm models are deployed at a back-end server to carry out data acquisition and automatic labeling in a real application environment, then a labeled data set is used for retraining a simpler model of a client and is updated into front-end equipment, the technical problem that the learning model of the client cannot be updated in time due to low speed of manual labeling of samples is solved, user experience is improved, the method can be used for training the learning model of the front end by using data in the automatic labeling actual environment at the background until the performance of the learning model tends to be stable, and the method is another important invention point of the method.
FIG. 2 illustrates an apparatus for automatically updating a deep learning model of the present invention, the apparatus comprising:
the deployment unit 201 is configured to train the first deep learning model by using the first data set by the server to obtain a trained initial first deep learning model, and deploy the initial first deep learning model to the client. For example, for a currently existing data set, the data set is a general data set, such as a data set downloaded through a network, a model is trained for the first time through a back-end server (i.e., a first deep learning model, a deep neural network model, etc.), and the data set is deployed into a front-end device (i.e., a client). For example, the face detection algorithm Adaboost + LBP is a relatively simple face detection algorithm model, and the initial learning model has relatively low recognition performance in an actual environment due to the adoption of a universal data set for training.
And the acquisition unit 202 is used for inputting data by a user through the client and sending the input data to the server as an acquired actual sample. For example, when a user uses a client to pay for an online scan, a first number of samples are collected from online input samples, for example, a first number of facial image samples are collected from users distributed across the country.
And the labeling unit 203 is configured to perform machine labeling on the actual sample at the server end by using a second deep learning model to obtain a labeled actual sample. For example, the face detection algorithms MTCNN and Retinaface (i.e., the second deep learning model) are used, and the performance of the algorithms is stronger than that of the Adaboost + LBP algorithm. The multiple algorithms can be filtered in a cascading mode, because the accuracy of data labeling is concerned more, the final labeling result is considered to be accurate only when the labeling result of each algorithm at the rear end is consistent, and the filtering is performed once the labeling results of some samples are inconsistent. The filtering is performed in a cascading mode, which is equivalent to a vote rejection, namely, when a model considers that a detected sample is not a positive sample, the filtering is performed, so that the accuracy can be greatly improved. Other filtering means may of course be used, such as most obeying a minority, such as two models considered positive samples and one model considered negative samples, and the result still considered positive samples, but this may result in a loss of accuracy.
The method comprises the steps of deploying a plurality of models stronger than a front end at a rear end, and labeling actual scene sample data, wherein the computation capability of a rear end server is much better than that of a client, so that the rear end server is used for labeling the sample data in an actual environment by using a high-performance learning algorithm, the technical problem that the client learning model cannot be updated in time due to low manual labeling speed in the prior art is solved, and the method is an important invention point.
And the retraining unit 204 is configured to retrain the initial first deep learning model by using the first data set and the labeled actual sample by the server to obtain a retrained first deep learning model.
An updating unit 205, configured to determine 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.
In one embodiment, the first deep learning model has lower recognition performance than the second deep learning model, and the first deep learning model consumes less system resources at runtime than the second deep learning model. The input data is face image data, fingerprint image data, iris image data or voiceprint data, and the data is collected through a camera, a fingerprint sensor, a voice sensor and the like.
In one embodiment, in order to prepare for machine labeling of sample data in an actual environment, performing, at the server side, machine labeling on the actual sample using a second deep learning model to obtain an operation of obtaining a labeled actual sample is:
the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples, and if the identification results of the second deep learning models are consistent, the identification results are used for marking the actual samples; this is the case where the determined sample set is optimal, i.e. it is guaranteed that each sample is correct.
Or the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples to obtain corresponding identification results, and all the identification results are weighted and added to serve as identification results to label the actual samples; in this case, the performance of the plurality of second deep learning models is taken into consideration, and different recognition weights may be given according to the second deep learning models having different performances.
Preferably, the recognition weights of the plurality of second deep learning models set a formula:
wherein y isiThe accuracy of the ith model. For example, if the back end deploys three second deep learning models with respective accuracies of 85%, 90%, and 95%, then the weights are respectively
The weight setting method is created by the exclusive research institute of the application, can realize the accuracy of the labeling of the sample machine to the maximum degree, and is one of the important invention points of the invention.
Or the number of the second deep learning models is multiple, and the actual samples are identified after the multiple second deep learning models are connected in series to obtain the final identification result, and then the actual samples are labeled. In this case, the collected actual samples are identified by filtering, and when an actual sample is identified as not by a certain second deep learning model, the sample is discarded, and other actual samples are identified again.
Through the operation, the collected labeling samples are guaranteed to be positive samples, and due to the fact that in the training process, algorithm training convergence is guaranteed, negative samples are also collected, and a good model can be trained only when the positive samples and the negative samples are guaranteed to maintain a certain proportion. The negative sample can be automatically generated by a machine or downloaded through a network. The proportion of positive and negative samples during model training may be different according to different algorithms, and the collection of negative samples is simpler, for example: and randomly cutting the actual scene picture, including various subsequent operations, such as color transformation, mirror image transformation and the like.
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 as follows: and identifying a part of marked actual samples by using the retrained first deep learning model, wherein if the identification rate is greater than the accuracy rate of the part of marked actual samples by the initial first deep learning model, the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.
And (3) taking part of data of the labeled actual sample as a test set, wherein the test set is required to be completely actual scene data instead of a standard data set on the network because the first deep learning is ensured to have the best effect in an actual scene as far as possible. And (3) carrying out test verification on the retrained first deep learning model by using the test set, and updating the model when the test effect is better than that of the on-line model (namely the initial first deep learning model). After a period of time, when the actual sample reaches a certain magnitude, the effect of the first deep learning model tends to be stable, and then the first deep learning model can not be retrained any more, so that the resources of the back-end server are released, and the pressure of the server is reduced.
The method is used for acquiring and automatically labeling data in a real application environment by deploying a plurality of powerful algorithm models at a back-end server, then retraining a simpler model at a client by using a labeled data set, and updating the simpler model into front-end equipment, so that the technical problem that a client learning model cannot be updated in time due to low speed of manually labeling samples is solved, and the user experience is improved.
For convenience of description, the above devices are described as being divided into various units by function, and are described separately. Of course, the functionality of the units may be implemented in one or more software and/or hardware when implementing the present application.
From the above description of the embodiments, it is clear to those skilled in the art that the present application can be implemented by software plus necessary general hardware platform. Based on such understanding, the technical solutions of the present application may be essentially implemented or the portions that contribute to the prior art may be embodied in the form of a software product, which may 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 may be a personal computer, a server, or a network device, etc.) to execute the apparatuses described in the embodiments or some portions of the embodiments of the present application.
Finally, it should be noted that: although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that: modifications and equivalents may be made thereto without departing from the spirit and scope of the invention and it is intended to cover in the claims the invention as defined in the appended claims.
Claims (11)
1. A method for automatically updating a deep learning model, the method comprising:
a deployment step, wherein a server trains a first deep learning model by using a first data set to obtain a trained initial first deep learning model, and deploys the initial first deep learning model to a client;
a collection step, wherein a user uses the client to input data and sends the input data to the server as a collected actual sample;
a labeling step, in which a second deep learning model is used at the server end to perform machine labeling on the actual sample to obtain a labeled actual sample;
retraining, namely retraining the initial first deep learning model by using the first data set and the labeled actual sample by the server to obtain a retrained first deep learning model;
and an updating step, namely judging whether the performance of the retrained first deep learning model is greater than that of the initial first deep learning model, and if so, sending the retrained first deep learning model to the client by the server to update the initial first deep learning model.
2. The method of claim 1, wherein the first deep learning model has lower recognition performance than the second deep learning model, and wherein the first deep learning model consumes less system resources at runtime than the second deep learning model.
3. The method of claim 2, wherein the input data is facial image data, fingerprint image data, iris image data, or voice print data.
4. The method according to claim 3, wherein the operation of performing machine labeling on the actual samples by using the second deep learning model at the server end to obtain labeled actual samples is:
the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples, and if the identification results of the second deep learning models are consistent, the identification results are used for marking the actual samples;
or the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples to obtain corresponding identification results, and all the identification results are weighted and added to serve as identification results to label the actual samples;
or the number of the second deep learning models is multiple, and the actual samples are identified after the multiple second deep learning models are connected in series to obtain the final identification result, and then the actual samples are labeled.
5. The method of claim 4, wherein the operation of determining whether the performance of the retrained first deep learning model is greater than the initial first deep learning model is: and identifying a part of marked actual samples by using the retrained first deep learning model, wherein if the identification rate is greater than the accuracy rate of the part of marked actual samples by the initial first deep learning model, the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.
6. An apparatus for automatically updating a deep learning model, the apparatus comprising:
the deployment unit is used for training the first deep learning model by using a first data set to obtain a trained initial first deep learning model and deploying the initial first deep learning model to the client;
the acquisition unit is used for inputting data by a user through the client and sending the input data to the server as an acquired actual sample;
the labeling unit is used for performing machine labeling on the actual sample by using a second deep learning model at the server end to obtain a labeled actual sample;
the retraining unit is used for retraining the initial first deep learning model by using the first data set and the labeled actual sample by the server to obtain a retrained first deep learning model;
and the updating unit is used for judging 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.
7. The apparatus of claim 6, wherein the first deep learning model has lower recognition performance than the second deep learning model, and wherein the first deep learning model consumes less system resources at runtime than the second deep learning model.
8. The apparatus of claim 7, wherein the input data is face image data, fingerprint image data, iris image data, or voiceprint data.
9. The apparatus of claim 8, wherein the operation of performing machine labeling on the actual samples using the second deep learning model at the server end to obtain labeled actual samples is:
the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples, and if the identification results of the second deep learning models are consistent, the identification results are used for marking the actual samples;
or the number of the second deep learning models is multiple, the multiple second deep learning models are used for identifying the actual samples to obtain corresponding identification results, and all the identification results are weighted and added to serve as identification results to label the actual samples;
or the number of the second deep learning models is multiple, and the actual samples are identified after the multiple second deep learning models are connected in series to obtain the final identification result, and then the actual samples are labeled.
10. The apparatus of claim 9, wherein the operation of determining whether the performance of the retrained first deep learning model is greater than the initial first deep learning model is: and identifying a part of marked actual samples by using the retrained first deep learning model, wherein if the identification rate is greater than the accuracy rate of the part of marked actual samples by the initial first deep learning model, the performance of the retrained first deep learning model is greater than that of the initial first deep learning model.
11. A computer-readable storage medium, characterized in that the storage medium has stored thereon computer program code which, when executed by a computer, performs the apparatus of any of claims 1-5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010456082.6A CN111612161A (en) | 2020-05-26 | 2020-05-26 | Method, device and storage medium for automatically updating deep learning model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010456082.6A CN111612161A (en) | 2020-05-26 | 2020-05-26 | Method, device and storage medium for automatically updating deep learning model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111612161A true CN111612161A (en) | 2020-09-01 |
Family
ID=72200451
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010456082.6A Pending CN111612161A (en) | 2020-05-26 | 2020-05-26 | Method, device and storage medium for automatically updating deep learning model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111612161A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112347947A (en) * | 2020-11-10 | 2021-02-09 | 厦门长江电子科技有限公司 | Image data processing system and method integrating intelligent detection and automatic test |
| CN112598058A (en) * | 2020-12-22 | 2021-04-02 | 深延科技(北京)有限公司 | Automatic deep learning method and device and storage medium |
| CN113344916A (en) * | 2021-07-21 | 2021-09-03 | 上海媒智科技有限公司 | Method, system, terminal, medium and application for acquiring machine learning model capability |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107967491A (en) * | 2017-12-14 | 2018-04-27 | 北京木业邦科技有限公司 | Machine learning method, device, electronic equipment and the storage medium again of plank identification |
| CN110674295A (en) * | 2019-09-11 | 2020-01-10 | 成都数之联科技有限公司 | Data labeling system based on deep learning |
| CN110705717A (en) * | 2019-09-30 | 2020-01-17 | 支付宝(杭州)信息技术有限公司 | Training method, device and equipment of machine learning model executed by computer |
| CN110928553A (en) * | 2019-10-16 | 2020-03-27 | 中国平安人寿保险股份有限公司 | Deployment method, device and system of deep learning model |
| CN111126574A (en) * | 2019-12-30 | 2020-05-08 | 腾讯科技（深圳）有限公司 | Method and device for training machine learning model based on endoscopic image and storage medium |
-
2020
- 2020-05-26 CN CN202010456082.6A patent/CN111612161A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107967491A (en) * | 2017-12-14 | 2018-04-27 | 北京木业邦科技有限公司 | Machine learning method, device, electronic equipment and the storage medium again of plank identification |
| CN110674295A (en) * | 2019-09-11 | 2020-01-10 | 成都数之联科技有限公司 | Data labeling system based on deep learning |
| CN110705717A (en) * | 2019-09-30 | 2020-01-17 | 支付宝(杭州)信息技术有限公司 | Training method, device and equipment of machine learning model executed by computer |
| CN110928553A (en) * | 2019-10-16 | 2020-03-27 | 中国平安人寿保险股份有限公司 | Deployment method, device and system of deep learning model |
| CN111126574A (en) * | 2019-12-30 | 2020-05-08 | 腾讯科技（深圳）有限公司 | Method and device for training machine learning model based on endoscopic image and storage medium |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112347947A (en) * | 2020-11-10 | 2021-02-09 | 厦门长江电子科技有限公司 | Image data processing system and method integrating intelligent detection and automatic test |
| CN112598058A (en) * | 2020-12-22 | 2021-04-02 | 深延科技(北京)有限公司 | Automatic deep learning method and device and storage medium |
| CN113344916A (en) * | 2021-07-21 | 2021-09-03 | 上海媒智科技有限公司 | Method, system, terminal, medium and application for acquiring machine learning model capability |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104239858B (en) | A kind of method and apparatus of face characteristic checking | |
| CN107169049B (en) | Application tag information generation method and device | |
| Darlow et al. | Fingerprint minutiae extraction using deep learning | |
| CN111612161A (en) | Method, device and storage medium for automatically updating deep learning model | |
| CN110781829A (en) | Light-weight deep learning intelligent business hall face recognition method | |
| CN104143079A (en) | Method and system for face attribute recognition | |
| CN109145717A (en) | A kind of face identification method of on-line study | |
| US10423817B2 (en) | Latent fingerprint ridge flow map improvement | |
| CN111291773A (en) | Feature identification method and device | |
| CN111401105B (en) | Video expression recognition method, device and equipment | |
| CN108960412A (en) | Image-recognizing method, device and computer readable storage medium | |
| CN113870254B (en) | Target object detection method and device, electronic equipment and storage medium | |
| CN110288085A (en) | A kind of data processing method, device, system and storage medium | |
| CN110991346A (en) | Suspected drug addict identification method and device and storage medium | |
| CN111401343B (en) | Method for identifying attributes of people in image and training method and device for identification model | |
| CN108875559A (en) | The face identification method and system shone based on certificate photo and scene | |
| CN110443577A (en) | A kind of campus attendance checking system based on recognition of face | |
| CN116386118B (en) | Drama matching cosmetic system and method based on human image recognition | |
| CN111652242A (en) | Image processing method, image processing device, electronic equipment and storage medium | |
| CN111611917A (en) | Model training method, feature point detection device, feature point detection equipment and storage medium | |
| Aboalhsan et al. | Deep learning technique for fingerprint recognition | |
| CN116206373A (en) | Living body detection method, electronic device and storage medium | |
| CN116246303A (en) | Sample construction method, device, equipment and medium for model cross-domain training | |
| CN115116117A (en) | Learning input data acquisition method based on multi-mode fusion network | |
| CN114566184A (en) | Audio recognition method and related device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |