CN109615058A - A kind of training method of neural network model - Google Patents
A kind of training method of neural network model Download PDFInfo
- Publication number
- CN109615058A CN109615058A CN201811246828.XA CN201811246828A CN109615058A CN 109615058 A CN109615058 A CN 109615058A CN 201811246828 A CN201811246828 A CN 201811246828A CN 109615058 A CN109615058 A CN 109615058A
- Authority
- CN
- China
- Prior art keywords
- neural network
- network model
- training
- user terminal
- cloud server
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention relates to a kind of training methods of neural network model, belong to field of artificial intelligence.A kind of training method of neural network model, the user terminal that cloud server is set and is remotely connect with the cloud server；The training method specifically includes: step S1, and the cloud server is trained the neural network model using the first training dataset；Step S2, the cloud server judge the accuracy of the neural network model test: the neural network model is sent to the user terminal by step S3, the cloud server；Step S4, the user terminal continue to be trained the neural network model using the second training dataset；The neural network model is saved as available model by step S5, the user terminal, for subsequent use；In this way, neural network model can complete most training mission beyond the clouds, improve trained efficiency, also save the memory space of terminal device.
Description
Technical field
The present invention relates to a kind of field of artificial intelligence more particularly to a kind of training methods of neural network model.
Background technique
With artificial intelligence image recognition, in terms of application and its higher accuracy rate, by artificial intelligence
Can technology be applied to terminal device and have become the direction being concerned, but artificial intelligence technology, especially neural network mould
The operation of type has higher requirements for the performance and memory space of equipment, keeps it limited in the application of terminal device.Neural network
Model is generally divided into two processes, training process and reasoning process, training process be using established neural network model with
And the data set being collected into, model is trained, is made inferences later in reasoning process using trained model, it can
Enough carry out the identification or classification of image or sound.Since training process needs to constantly update data, i.e., train every time for same
One data need to be written and read at least once, therefore training process needs the higher processor of performance to read memory
It takes and is written, and reasoning process is calculated using data and trained model, is obtained a result and is made reasoning, and number need to be only read
According to, therefore reasoning process needs the higher processor of performance to be read out memory.Traditional application mode makes reasoning
All in same processing unit and storage unit, this is often required that between storage unit and processing unit for process and training process
Readwrite performance is all very high, brings challenges.The training process of the neural network model of terminal device is generally given at cloud at present simultaneously
Device is managed, model is sent to terminal device again using the data set completion training process in cloud and made inferences by cloud processor, when again
When needing training pattern, cloud processor is recycled to be trained, this often results in the loss of very big transmission data, to terminal
Become a main limiting factor for equipment.Meanwhile data set used in cloud processor is if it is from other terminals
It is collected in equipment, is trained then such data set is fully utilized, trained model is caused to be not necessarily suitable use
Family, but if the cloud processor data only collected using the terminal device of user come the model of training user, if user
Data set amount is insufficient, and the application precision that will also result in model is not high.
Summary of the invention
For the deficiency of technology, the problem to be solved in the present invention provides a kind of training method of neural network model.
Wherein, the user terminal that cloud server is set and is remotely connect with cloud server；
Training method specifically includes:
Step S1, cloud server are trained neural network model using the first training dataset；
Step S2, cloud server judge the accuracy of the test of neural network model:
If accuracy is greater than one first preset threshold, step S3 is turned to；
If accuracy is not more than the first preset threshold, return step S1, to continue to be trained neural network model；
Neural network model is sent to user terminal by step S3, cloud server；
Step S4, user terminal continues to be trained neural network model using the second training dataset, until nerve net
Until the accuracy of the test of network model is greater than one second preset threshold；
Neural network model is saved as available model by step S5, user terminal, for subsequent use；
Second preset threshold is greater than the first preset threshold.
During later using model, user terminal persistently detects the errors number of the reasoning of available model, and
Errors number re-execute the steps S4 to step S5 when being more than a third predetermined threshold value.
Cloud server is long-range respectively to connect a plurality of clients；
Cloud server collects the data that each user terminal uploads, as the first training dataset.
User terminal forms the second training dataset in such a way that data acquire.
The training system of the neural network model includes cloud server and user terminal, and cloud server remotely connects user
End；
Cloud server includes:
First training unit, for being trained using the first training dataset to neural network model；
First judging unit connects the first training unit, for judging the accuracy of the test of neural network model and defeated
First judging result out, when the first judging result indicates that the accuracy of the test of neural network model is not more than one first default threshold
When value, the first training unit continues to be trained neural network model；
Transmission unit is separately connected the first judging unit and the first training unit, for indicating mind in the first judging result
When the accuracy of test through network model is greater than first preset threshold, the neural network model that training is formed is sent to use
Family end；
User terminal includes:
Receiving unit, for receiving neural network model；
Second training unit, connect receiving unit, for using the second training dataset continue to neural network model into
Row training exports the mind that training is formed until the accuracy of the test of neural network model is greater than one second preset threshold
Through network model；
Storage unit connects the second training unit, and neural network model is saved as available model by storage unit, for rear
It is continuous to use；
Second preset threshold is greater than the first preset threshold.
User terminal further include:
Second judgment unit connects the second training unit, persistently detects available model during user is using model
Reasoning errors number, and errors number be more than a third predetermined threshold value when neural network model is sent back to the second instruction again
Practice and is trained in unit.
Cloud server is long-range respectively to connect a plurality of clients；
Cloud server collects the data that each user terminal uploads, as the first training dataset.
User terminal forms the second training dataset in such a way that data acquire.
Training method through the invention, neural network model can complete most training mission beyond the clouds, improve
The efficiency of training, also saves the memory space of terminal device, simultaneously as only using user terminal collection in the trained later period
Data set training neural network model rather than all using cloud collection data set, ensure that trained model more
Suitable for user itself, during use later, once the number that inference errors occurs in model has been more than certain threshold value,
It still then is trained model using the data set that terminal device and terminal device are collected, until the accuracy of model is higher than institute
It is required that value improve the service efficiency of equipment without being transmitted to cloud again.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached
Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair
The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this
A little attached drawings obtain other relevant attached drawings.
Fig. 1: the training method flow chart of neural network
Fig. 2: the system block diagram of neural network
Fig. 3: the accuracy threshold value of neural network
Fig. 4: the training and reasoning process of neural network model
Specific embodiment
A kind of training method of neural network model, specifically includes:
Step S1, cloud server 1 are trained neural network model using the first training dataset；
Step S2, cloud server 1 judge the accuracy of the test of neural network model:
If accuracy is greater than one first preset threshold, step S3 is turned to；
If accuracy is not more than the first preset threshold, return step S1, to continue to be trained neural network model；
Neural network model is sent to user terminal 2 by step S3, cloud server 1；
Step S4, user terminal 2 continues to be trained neural network model using the second training dataset, until nerve net
Until the accuracy of the test of network model is greater than one second preset threshold；
Neural network model is saved as available model by step S5, user terminal 2, for subsequent use；
Second preset threshold is greater than the first preset threshold.
In the accuracy of the test of neural network model, test refers to output data and positive exact figures to neural network model
According to being compared, correct data here refers to that the training data for being trained to neural network model concentrates corresponding mark
Quasi- output data.
During later using model, user terminal 2 persistently detects the errors number of the reasoning of available model, and
Errors number re-execute the steps S4 to step S5 when being more than a third predetermined threshold value.
Here reasoning refers to output prediction result, and whether mistake is judged by user feedback for reasoning, user
Reasoning is wrong is just denoted as primary mistake for feedback.
As shown in Figure 1 and Figure 4, it by doing repetition training to neural network model with a large amount of training data, ultimately forms
Neural network model the reasoning results meet result required by neural network model input data；
Cloud server 1 is long-range respectively to connect a plurality of clients 2；
Cloud server 1 collects the data that each user terminal 2 uploads, as the first training dataset.
The data set of cloud server 1 refers to the set for the data set that cloud server 1 is collected from a plurality of clients 2, such as
The set of data set A shown in Fig. 4, data set B, data set C, data set D.
User terminal 2 forms the second training dataset in such a way that data acquire.
Data acquisition may include the data directly inputted by the different collected data of sensor or user
Deng.
The training system of the neural network model includes cloud server 1 and user terminal 2, and cloud server 1 remotely connects
User terminal 2；
Cloud server 1 includes:
First training unit 11, for being trained using the first training dataset to neural network model；
First judging unit 12 connects the first training unit 11, the accuracy of the test for judging neural network model
And the first judging result is exported, when the first judging result indicates that the accuracy of the test of neural network model is pre- no more than one first
If when threshold value, the first training unit 11 continues to be trained neural network model；
Transmission unit 13 is separately connected the first judging unit 12 and the first training unit 11, in the first judging result
When indicating that the accuracy of the test of neural network model is greater than first preset threshold, the neural network model that training is formed is sent out
It send to user terminal 2；
User terminal 2 includes:
Receiving unit 21, for receiving neural network model；
Second training unit 22 connects receiving unit 21, for being continued using the second training dataset to neural network mould
Type is trained, and until the accuracy of the test of neural network model is greater than one second preset threshold, is exported training and is formed
Neural network model；
Storage unit 23 connects the second training unit 22, and neural network model is saved as available model by storage unit 23,
For subsequent use；
Second preset threshold is greater than the first preset threshold.
User terminal 2 trains accuracy threshold value, is determined according to the performance of user terminal 2 and the size of training dataset；
User terminal 2 can be mobile phone, tablet computer, computer；
User terminal 2 further include:
Second judgment unit 24 connects the second training unit 22, the mistake time of the reasoning for persistently detecting available model
Number, and neural network model is sent back to again in the second training unit 22 when errors number is more than a third predetermined threshold value and is carried out
Training.
As shown in Fig. 2,
Cloud server 1 includes the first training unit 11, the first judging unit 12, transmission unit 13；
User terminal 2 includes receiving unit 21, the second training unit 22, storage unit 23, second judgment unit 24；
In order to make it easy to understand, only showing a user terminal 2 in Fig. 2.
Cloud server 1 is long-range respectively to connect a plurality of clients 2；
Cloud server 1 collects the data that each user terminal 2 uploads, as the first training dataset.
User terminal 2 forms the second training dataset in such a way that data acquire.
The present invention is further illustrated in the following with reference to the drawings and specific embodiments, but not as the limitation of the invention.
Embodiment 1
For the training process of a user terminal neural network model, server 1 is trained model beyond the clouds first, institute
The data set that the data set used is collected from each user terminal 2 from cloud server 1, data set A as shown in Figure 4,
Data set B, data set C, data set D, when neural network model accuracy is more than the first preset threshold a% (as shown in Figure 3),
The parameter of neural network model is passed into user terminal 2, user terminal 2 utilizes the data set (data as shown in Figure 4 of themselves capture
Collection A or data set B or data set C or data set D) continue to train, until neural network model accuracy is higher than second
When preset threshold b% (as shown in Figure 3), user (user A or user B as shown in Figure 4 or user C or user D) starts to make
It is made inferences with neural network model, this process is as shown in Figure 4.Neural network model during use later, once
When inference errors number is more than third predetermined threshold value (both reasoning failure threshold values), illustrate that the generalization ability of neural network model (pushes away
Manage correct ability) decline, it needs to re-start training, until trained accuracy is more than user terminal training accuracy threshold value.
Training method through the invention, neural network model can complete most training mission beyond the clouds, improve
The efficiency of training, also saves the memory space of terminal device, simultaneously because the training later period only uses the number of user terminal collection
According to training neural network model is collected, so ensure that trained model is more applicable for user itself.
The above is only preferred embodiments of the present invention, are not intended to limit the implementation manners and the protection scope of the present invention, right
For those skilled in the art, it should can appreciate that and all replace with being equal made by description of the invention and diagramatic content
It changes and obviously changes obtained scheme, should all be included within the scope of the present invention.
Claims (8)
1. a kind of training method of neural network model, which is characterized in that setting cloud server and with the cloud service
The user terminal that device remotely connects；
The training method specifically includes:
Step S1, the cloud server are trained the neural network model using the first training dataset；
Step S2, the cloud server judge the accuracy of the neural network model test:
If the accuracy is greater than one first preset threshold, step S3 is turned to；
If the accuracy is not more than first preset threshold, the step S1 is returned to, to continue to the neural network
Model is trained；
The neural network model is sent to the user terminal by step S3, the cloud server；
Step S4, the user terminal continues to be trained the neural network model using the second training dataset, until institute
The accuracy of the test of neural network model is stated greater than until one second preset threshold；
The neural network model is saved as available model by step S5, the user terminal, for subsequent use；
Second preset threshold is greater than first preset threshold.
2. training method as described in claim 1, which is characterized in that the user terminal is during using the available model
The errors number of the reasoning of the available model is persistently detected, and again when the errors number is more than a third predetermined threshold value
The step S4 is executed to the step S5.
3. training method as described in claim 1, which is characterized in that the cloud server remotely connects multiple described respectively
User terminal；
The cloud server collects the data that each user terminal uploads, as first training dataset.
4. training method as described in claim 1, which is characterized in that the user terminal forms institute in such a way that data acquire
State the second training dataset.
5. a kind of training system of neural network model, which is characterized in that including cloud server and user terminal, the cloud clothes
Business device remotely connects the user terminal；
The cloud server includes:
First training unit, for being trained using the first training dataset to the neural network model；
First judging unit connects first training unit, the accuracy of the test for judging the neural network model
And the first judging result is exported, when first judging result indicates that the accuracy of the test of the neural network model is not more than
When one first preset threshold, first training unit continues to be trained the neural network model；
Transmission unit is separately connected first judging unit and first training unit, in the first judgement knot
When fruit indicates that the accuracy of the test of the neural network model is greater than first preset threshold, by the nerve of training formation
Network model is sent to the user terminal；
The user terminal includes:
Receiving unit, for receiving the neural network model；
Second training unit connects the receiving unit, for being continued using the second training dataset to the neural network mould
Type is trained, until the accuracy of the test of the neural network model is greater than one second preset threshold, output training
The neural network model formed；
Storage unit, connects second training unit, and the neural network model is saved as available mould by the storage unit
Type, for subsequent use；
Second preset threshold is greater than first preset threshold.
6. training system as claimed in claim 5, which is characterized in that the user terminal further include:
Second judgment unit connects second training unit, the mistake time of the reasoning for persistently detecting the available model
Number, and the neural network model is sent back into second training again when the errors number is more than a third predetermined threshold value
It is trained in unit.
7. training system as claimed in claim 5, which is characterized in that the cloud server remotely connects multiple described respectively
User terminal；
The cloud server collects the data that each user terminal uploads, as first training dataset.
8. training system as claimed in claim 5, which is characterized in that the user terminal forms institute in such a way that data acquire
State the second training dataset.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811246828.XA CN109615058A (en) | 2018-10-24 | 2018-10-24 | A kind of training method of neural network model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811246828.XA CN109615058A (en) | 2018-10-24 | 2018-10-24 | A kind of training method of neural network model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109615058A true CN109615058A (en) | 2019-04-12 |
Family
ID=66002142
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811246828.XA Pending CN109615058A (en) | 2018-10-24 | 2018-10-24 | A kind of training method of neural network model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109615058A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110415678A (en) * | 2019-06-13 | 2019-11-05 | 百度时代网络技术（北京）有限公司 | Customized voice broadcast client, server, system and method |
| CN110532445A (en) * | 2019-04-26 | 2019-12-03 | 长佳智能股份有限公司 | The cloud transaction system and its method of neural network training pattern are provided |
| CN110659732A (en) * | 2019-09-20 | 2020-01-07 | 上海新储集成电路有限公司 | Method for intelligently adjusting neural network model |
| CN110703994A (en) * | 2019-09-20 | 2020-01-17 | 上海新储集成电路有限公司 | Data storage system and method of neural network model |
| CN111598139A (en) * | 2020-04-24 | 2020-08-28 | 北京奇艺世纪科技有限公司 | Data processing method and system |
| CN112085208A (en) * | 2020-07-30 | 2020-12-15 | 北京聚云科技有限公司 | Method and device for model training by using cloud |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9336483B1 (en) * | 2015-04-03 | 2016-05-10 | Pearson Education, Inc. | Dynamically updated neural network structures for content distribution networks |
| CN106709917A (en) * | 2017-01-03 | 2017-05-24 | 青岛海信医疗设备股份有限公司 | Neural network model training method, device and system |
| CN107508866A (en) * | 2017-08-08 | 2017-12-22 | 重庆大学 | Reduce the method for the transmission consumption of mobile device end neural network model renewal |
| CN107871164A (en) * | 2017-11-17 | 2018-04-03 | 济南浪潮高新科技投资发展有限公司 | A kind of mist computing environment personalization deep learning method |
-
2018
- 2018-10-24 CN CN201811246828.XA patent/CN109615058A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9336483B1 (en) * | 2015-04-03 | 2016-05-10 | Pearson Education, Inc. | Dynamically updated neural network structures for content distribution networks |
| CN106709917A (en) * | 2017-01-03 | 2017-05-24 | 青岛海信医疗设备股份有限公司 | Neural network model training method, device and system |
| CN107508866A (en) * | 2017-08-08 | 2017-12-22 | 重庆大学 | Reduce the method for the transmission consumption of mobile device end neural network model renewal |
| CN107871164A (en) * | 2017-11-17 | 2018-04-03 | 济南浪潮高新科技投资发展有限公司 | A kind of mist computing environment personalization deep learning method |
Non-Patent Citations (2)
| Title |
|---|
| 刘庆等: "基于非监督预训练的结构优化卷积神经网络", 《工程科学与技术》, 30 June 2017 (2017-06-30), pages 213 - 218 * |
| 邵欣: "《物联网技术及应用》", 北京航空航天大学出版社, pages: 203 - 209 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110532445A (en) * | 2019-04-26 | 2019-12-03 | 长佳智能股份有限公司 | The cloud transaction system and its method of neural network training pattern are provided |
| CN110415678A (en) * | 2019-06-13 | 2019-11-05 | 百度时代网络技术（北京）有限公司 | Customized voice broadcast client, server, system and method |
| CN110659732A (en) * | 2019-09-20 | 2020-01-07 | 上海新储集成电路有限公司 | Method for intelligently adjusting neural network model |
| CN110703994A (en) * | 2019-09-20 | 2020-01-17 | 上海新储集成电路有限公司 | Data storage system and method of neural network model |
| CN110659732B (en) * | 2019-09-20 | 2023-07-04 | 上海新储集成电路有限公司 | Method for intelligently adjusting neural network model |
| CN111598139A (en) * | 2020-04-24 | 2020-08-28 | 北京奇艺世纪科技有限公司 | Data processing method and system |
| CN112085208A (en) * | 2020-07-30 | 2020-12-15 | 北京聚云科技有限公司 | Method and device for model training by using cloud |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109615058A (en) | A kind of training method of neural network model | |
| CN110263921B (en) | Method and device for training federated learning model | |
| CN112181666B (en) | Equipment assessment and federal learning importance aggregation method based on edge intelligence | |
| CN109948669B (en) | Abnormal data detection method and device | |
| US11176418B2 (en) | Model test methods and apparatuses | |
| US11379723B2 (en) | Method and apparatus for compressing neural network | |
| CN110473083B (en) | Tree risk account identification method, device, server and storage medium | |
| CN112748941B (en) | Method and device for updating target application program based on feedback information | |
| Li et al. | Online federated multitask learning | |
| KR20200145827A (en) | Facial feature extraction model learning method, facial feature extraction method, apparatus, device, and storage medium | |
| CN112149733B (en) | Model training method, model quality determining method, model training device, model quality determining device, electronic equipment and storage medium | |
| CN112001274B (en) | Crowd density determining method, device, storage medium and processor | |
| CN111915020A (en) | Method and device for updating detection model and storage medium | |
| CN110414581B (en) | Picture detection method and device, storage medium and electronic device | |
| CN111414540B (en) | Online learning recommendation method and device, online learning system and server | |
| CN114861172B (en) | Data processing method and system based on government affair service system | |
| CN111526213A (en) | Equipment control system based on block chain | |
| CN112817563B (en) | Target attribute configuration information determining method, computer device, and storage medium | |
| CN112861659A (en) | Image model training method and device, electronic equipment and storage medium | |
| CN113052025A (en) | Training method of image fusion model, image fusion method and electronic equipment | |
| CN110610140B (en) | Training method, device and equipment of face recognition model and readable storage medium | |
| CN115690592B (en) | Image processing method and model training method | |
| CN110472537B (en) | Self-adaptive identification method, device, equipment and medium | |
| CN116435995A (en) | Time series processing method, computer readable storage medium and electronic device | |
| CN115830709A (en) | Action recognition method based on federal edge learning, server and electronic equipment |
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 |