CN107392125A - Training method/system, computer-readable recording medium and the terminal of model of mind - Google Patents
Training method/system, computer-readable recording medium and the terminal of model of mind Download PDFInfo
- Publication number
- CN107392125A CN107392125A CN201710561654.5A CN201710561654A CN107392125A CN 107392125 A CN107392125 A CN 107392125A CN 201710561654 A CN201710561654 A CN 201710561654A CN 107392125 A CN107392125 A CN 107392125A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- data set
- training
- mind
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
- G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of training method/system, computer-readable recording medium and the terminal of model of mind, and training method includes：The first data set and the markup information related to training mission to input carry out initial model training, to obtain benchmark model；New data is added, is incorporated in the first data set, forms the second data set；To the data test in the second data set and value assessment, to pick out the data that mark value is more than default mark value, the 3rd data set is formed；The data for not marking markup information in 3rd data set are labeled, are incorporated in the 3rd data set；Retraining, the benchmark model after being updated are carried out to benchmark model；It is the first new data set by the 3rd data set definition, adds new data, circulation performs above step, until the accuracy of the model after repetitive exercise is more than default accuracy.The present invention reduces the quantity manually marked, it is not necessary to marks total data, so as to save mark cost, and improves model training efficiency.
Description
Technical field
The invention belongs to field of artificial intelligence, is related to a kind of training method and system, more particularly to a kind of intelligence
Training method/system, computer-readable recording medium and the terminal of model.
Background technology
The development of artificial intelligence field is maked rapid progress, and in particular with the extensive use of depth learning technology, it is in object
The fields such as detection, identification achieve breakthrough progress.The feature of traditional engineer is different from, depth learning technology passes through
Substantial amounts of data are inputted, allow model itself to learn the character representation of object, the identification that can often match in excellence or beauty or even surmount the mankind is smart
Degree.
In general, the learning process of a complete artificial intelligence model include preparing two steps of data set and model training
Suddenly.Preparing data set includes data and the mark to data, for example, a recognition of face data set should include face picture with
To the markup information of the piece identity of every pictures.Model training is then input input data and markup information simultaneously, according to not
Same learning tasks carry out training pattern.Current most model learning flow is the learning process of one " static state ".I.e. once
Property collect and labeled data, then train model according to these data, then learning process terminates.However, it is somebody's turn to do of " static state "
Habit process has the disadvantage that in real application environment：
1) starting model training needs to wait all Data Collections and mark to finish.By taking depth learning technology as an example, model
Training needs substantial amounts of data, can be from pictures up to ten thousand to hundreds of thousands pictures, and collecting data and labeled data needs
Substantial amounts of man power and material is expended, cost is very high.
2) ability of model is shaped in an inquiry learning, and the data of disposable collecting are difficult complete adaptation practical application
Scene, so as to cause the performance of model in actual applications to have larger gap compared with experimental situation.
Therefore, how training method/system, computer-readable recording medium and the terminal of a kind of model of mind are provided, with
The problems such as solving Data Collection in the prior art and marking too high cost, model adaptability difference, the real this area that turned into are obtained employment
Personnel's technical problem urgently to be resolved hurrily.
The content of the invention
In view of the above the shortcomings that prior art, it is an object of the invention to provide a kind of training side of model of mind
Method/system, computer-readable recording medium and terminal, for solving Data Collection in the prior art and mark, cost is too high, mould
The problem of type adaptability difference.
In order to achieve the above objects and other related objects, one aspect of the present invention provides a kind of training method of model of mind,
The training method of the model of mind comprises the following steps：The first data set and the mark related to training mission letter to input
Breath carries out initial model training, to obtain a benchmark model；Addition newly counts with data attribute identical in first data set
According to, be incorporated in first data set, formed the second data set；The data tested in the second data set, to the second data set
Middle data carry out value assessment, and to pick out the data that mark value is more than default mark value, the data that will be singled out are formed
3rd data set；The data for not marking markup information in 3rd data set are labeled, the data of markup information will be labeled with
It is incorporated in the 3rd data set；Based on the 3rd data set after merging, retraining, the base after being updated are carried out to benchmark model
Quasi-mode type；It is the first new data set by the 3rd data set definition, adds new data, circulation performs above step, to be changed
Generation training, until the accuracy of the model after repetitive exercise is more than default accuracy.
In one embodiment of the invention, described pair of the first data set inputted and the markup information related to training mission
Initial model training is carried out, to obtain the loss letter related to training mission for referring to optimization the step of a benchmark model and pre-setting
Number, the functional value of the loss function is set constantly to reduce until preset function threshold value.
In one embodiment of the invention, after second data set is formed, the training method of the model of mind is also
Including being pre-processed to the data in second data set；The pretreatment includes carrying out examination, image to the data
Resolution ratio detection, image blur detection, and/or the adjustment of image balance degree.
In one embodiment of the invention, it is described test the second data set in data the step of include to described second count
Convolution, Chi Hua and more classification processing are carried out according to intensive data.
It is described that value assessment is carried out to the second data intensive data in one embodiment of the invention, to pick out
Mark value be more than the step of data of default mark value also include by it is predefined, for judging second data set
The cost function of the mark value of middle data, pick out the data that mark value is more than default mark value.
In one embodiment of the invention, the markup information for being directed to selected data is by artificial notation methods
The information of mark.
Another aspect of the present invention provides a kind of training system of model of mind, including：Initial training module, for input
The first data set and the markup information related to training mission carry out initial model training, to obtain a benchmark model；Merge
Module, for addition and data attribute identical new data in first data set, it is incorporated in first data set, shape
Into the second data set；Processing module, for testing the data in the second data set, value is carried out to the second data intensive data and commented
Estimate, to pick out the data that mark value is more than default mark value, the data that will be singled out form the 3rd data set；To the 3rd
The data for not marking markup information in data set are labeled, and the data for being labeled with markup information are incorporated in into the 3rd data set
In；Retraining module, for based on the 3rd data set after merging, retraining, the base after being updated to be carried out to benchmark model
Quasi-mode type；Loop module, for being the first new data set by the 3rd data set definition, new data is added, described in circular flow
Initial training module, merging module, processing module and retraining module, to be iterated training, until the mould after repetitive exercise
The accuracy of type is more than default accuracy.
In one embodiment of the invention, after second data set is formed, the training system of the model of mind is also
Including the pretreatment module coupled with the merging module, the pretreatment module is used for the data in second data set
Pre-processed；It is described pretreatment include to the data carry out examination, image resolution ratio detection, image blur detection and/
Or image balance degree adjustment.
In one embodiment of the invention, the data that the processing module is tested in the second data set refer to described second
Data intensive data carries out convolution, Chi Hua and more classification processing.
In one embodiment of the invention, the processing module be additionally operable to by it is predefined, for judging described second
The cost function of the mark value of data intensive data, pick out the data that mark value is more than default mark value.
Another aspect of the invention provides a kind of computer-readable recording medium, is stored thereon with computer program, its feature
It is, the program realizes the training method of the model of mind when being executed by processor.
Last aspect of the present invention provides a kind of terminal, including：Processor and memory；The memory is based on storing
Calculation machine program, the processor is used for the computer program for performing the memory storage, so that the terminal performs the intelligence
The training method of energy model.
As described above, training method/system, computer-readable recording medium and the terminal of the model of mind of the present invention, tool
There is following beneficial effect：
Training method/system, computer-readable recording medium and the terminal of model of mind of the present invention reduce people
The quantity of work mark, it is not necessary to mark total data, so as to save mark cost, and improve model training efficiency.
Brief description of the drawings
Fig. 1 is shown as schematic flow sheet of the training method of the model of mind of the present invention in an embodiment.
Fig. 2 is shown as theory structure schematic diagram of the training system of the model of mind of the present invention in an embodiment.
Component label instructions
The training system of 2 models of mind
21 initial training modules
22 merging modules
23 pretreatment modules
24 processing modules
25 retraining modules
26 loop modules
S11~S18 steps
Embodiment
Illustrate embodiments of the present invention below by way of specific instantiation, those skilled in the art can be by this specification
Disclosed content understands other advantages and effect of the present invention easily.The present invention can also pass through specific realities different in addition
The mode of applying is embodied or practiced, the various details in this specification can also be based on different viewpoints with application, without departing from
Various modifications or alterations are carried out under the spirit of the present invention.It should be noted that in the case where not conflicting, following examples and implementation
Feature in example can be mutually combined.
It should be noted that the diagram provided in following examples only illustrates the basic structure of the present invention in a schematic way
Think, only show the component relevant with the present invention in schema then rather than according to component count, shape and the size during actual implement
Draw, kenel, quantity and the ratio of each component can be a kind of random change during its actual implementation, and its assembly layout kenel
It is likely more complexity.
Training method/system, computer-readable recording medium and the terminal of model of mind of the present invention include two
Stage：Model initialization stage and model retraining stage.
The model initialization stage：In the starting stage of model training, prepare an initial data set and with training mission phase
The markup information of pass.The structure of model is designed, using data set and its markup information as information is inputted, carries out model training,
Obtain a benchmark model.Follow-up model carries out retraining on the basis of the model.
The model retraining stage：After getting out basic and training benchmark model.By retraining to model
Parameter is updated.Here retraining refers to be trained again by inputting new data and its mark to benchmark model, right
Model parameter is updated, and the model after renewal is typically higher than the model accuracy before renewal.Retraining needs the number inputted
According to there is two kinds of sources：First, being tested using benchmark model on former training dataset, difficult example therein is selected, such as identify
The relatively low data of the data and confidence level of mistake；Second, collect new data, by benchmark model and active Samples Estimates technology,
Value marking is carried out to data, picks out the data for wherein most marking value, then manually this partial data is labeled.
The data and its mark in two kinds of sources of merging collect to update the data.Retraining is carried out on new data set to benchmark model to obtain
Model after renewal.Above step can constantly be repeated according to actual conditions or iteration.
Embodiment one
The present embodiment provides a kind of training method of model of mind, including：
The first data set and the markup information related to training mission to input carry out initial model training, to obtain one
Benchmark model；
Addition and data attribute identical new data in first data set, are incorporated in first data set, shape
Into the second data set；
The data tested in the second data set, value assessment is carried out to the second data intensive data, to pick out mark valency
Value is more than the data of default mark value, and the data that will be singled out form the 3rd data set；
The data for not marking markup information in 3rd data set are labeled, the data for being labeled with markup information are merged
In the 3rd data set；
Based on the 3rd data set after merging, retraining, the benchmark model after being updated are carried out to benchmark model；
It is the first new data set by the 3rd data set definition, adds new data, circulation performs above step, to be changed
Generation training, until the accuracy of the model after repetitive exercise is more than default accuracy.
The training method of the model of mind provided below with reference to diagram the present embodiment is described in detail.Refer to
Fig. 1, it is shown as schematic flow sheet of the training method of model of mind in an embodiment.As shown in figure 1, the model of mind
Training method specifically includes following steps：
S11, the first data set and the markup information related to training mission to input carry out initial model training, to obtain
Take a benchmark model.For example, first data set is the first LFW face public data collection and the identity related to training mission
Markup information, or first data set are the positional information of the first vehicle detection public data collection and vehicle on picture.
Specific in practical application, to the first LFW face public data collection of input and the identity related to training mission
Markup information carries out initial model training, to obtain recognition of face benchmark model.
Or initial model is carried out to the positional information of the first vehicle detection public data collection and vehicle of input on picture
Training, to obtain vehicle detection benchmark model.
In the present embodiment, initial model training refers to the loss letter related to training mission that optimization is pre-set
Number, the functional value of the loss function is set constantly to reduce until preset function value threshold value.The preset function value threshold value levels off to 0.
In this process, the parameter of model is also continuously available renewal, so that the benchmark model preferably can appoint in adaptation training
Business.
S12, addition and data attribute identical new data in first data set, are incorporated in first data set
In, form the second data set.
In the present embodiment, the first LFW face public data collection is directed to, is added newly by modes such as web crawlers downloads
Face picture, formed the 2nd LFW face public data collection.
Or for the first vehicle detection public data collection, mode is downloaded etc. by web crawlers and adds new vehicle pictures,
Form the second vehicle detection public data collection.
S13, the data in second data set are pre-processed.The pretreatment includes sieving the data
Look into, image resolution ratio detects, image blur detection, and/or image balance degree adjust.In the present embodiment, climbed by network
The modes such as worm download collect the second data set.It is to be model retraining standard that data in second data set are carried out with pretreatment
The high data of standby quality.
Specifically, such as face picture is concentrated to carry out just the 2nd LFW faces public data using Dlib, OpenCV instrument
Walk the pretreatments such as face examination, image blur detection.
Or the vehicle figure concentrated using instruments such as Dlib, OpenCV to the second vehicle detection public data is carried out tentatively
The pretreatments such as vehicle examination, image resolution ratio, fuzziness detection, the adjustment of image balance degree.
S14, using the benchmark model, test pretreated second data set.In the present embodiment, institute is tested
The second data set is stated to refer to carry out the second data intensive data the processing such as convolution, Chi Hua and more classification processing.
Specifically, convolution, Chi Hua and more classification processing etc. are carried out to the 2nd LFW face public datas intensive data
Processing.
Or the processing such as convolution, Chi Hua and more classification processing are carried out to the second vehicle detection public data intensive data.
S15, value assessment is carried out to second data set, to pick out the number that mark value is more than default mark value
According to, and the costly data of mark that will be singled out form the 3rd data set.In the present embodiment, by it is predefined, be used for
Judge the cost function of the mark value of the second data intensive data, pick out mark value and be more than default mark value
Data.
Specifically, mark predefined, for judging the 2nd LFW faces public data concentration human body picture is passed through
The cost function of value, to pick out the face picture that mark value is more than default mark value, and the mark valency that will be singled out
It is worth high data and forms the 3rd LFW face public data collection.By it is predefined, for judging that second vehicle detection discloses
The cost function of the mark value of vehicle pictures in data set, to pick out the vehicle figure that mark value is more than default mark value
Piece, and the costly data of mark that will be singled out form the 3rd vehicle detection public data collection.
S16, the data for not marking markup information in the 3rd data set are labeled, the data of markup information will be labeled with
It is incorporated in the 3rd data set.
In the present embodiment, the markup information for being directed to selected data is the letter marked by artificial notation methods
Breath.For the face picture selected, face identity information is marked, and the face picture for being labeled with face identity information is merged
In the 3rd LFW face public data collection；For the vehicle pictures selected, the positional information of vehicle is marked, and will be marked
The vehicle pictures for having the positional information of vehicle are incorporated in the 3rd vehicle detection public data collection.The present embodiment is by selecting number
Reduce the quantity manually marked, it is not necessary to total data is marked, so as to save according to the data of most worthy are concentrated to be labeled
About mark cost.
S17, based on the 3rd data set after merging, retraining, the benchmark model after being updated are carried out to benchmark model.
In the present embodiment, training technique is identical with the training technique in step S11 used by retraining, i.e. optimization is pre-set
The loss function related to training mission, the functional value of the loss function is set constantly to reduce until preset function value threshold value.It is described
Preset function value threshold value levels off to 0.
S18, it is the first new data set by the 3rd data set definition after merging, and in the first new data set
New data is added, circulation performs above step S11 to S17, is trained to be iterated, until the model after repetitive exercise is accurate
Degree is more than default accuracy.In the present embodiment, by the way of training-mark iteration, partial data once can be only marked,
To partial data training pattern, renewal is then constantly iterated to model, training is more flexible, more rapidly.
Specifically, it is so that the accuracy of the model of recognition of face reaches more than 99% according to the demand of reality, works as model
Accuracy when being less than the value, with regard to continuous repeat step S11 to S17, until the accuracy of model reaches default accuracy 99%.
It is so that the accuracy of the model of vehicle identification reaches more than 99% according to the demand of reality, when the accuracy of model
During less than the value, with regard to continuous repeat step S11 to S17, until the accuracy of model reaches default accuracy 99%.
The present embodiment also provides a kind of computer-readable recording medium, is stored thereon with computer program, and the program is located
Reason device realizes the training method of the model of mind when performing.One of ordinary skill in the art will appreciate that：Realize above-mentioned each side
The all or part of step of method embodiment can be completed by the related hardware of computer program.Foregoing computer program can
To be stored in a computer-readable recording medium.The program upon execution, execution the step of including above-mentioned each method embodiment；
And foregoing storage medium includes：ROM, RAM, magnetic disc or CD etc. are various can be with the medium of store program codes.
The training method and computer-readable recording medium of model of mind described in the present embodiment reduce what is manually marked
Quantity, it is not necessary to mark total data, so as to save mark cost, and improve model training efficiency.
Embodiment two
The present embodiment provides a kind of training system of model of mind, including：
Initial training module, carried out for the first data set to input and the markup information related to training mission initial
Model training, to obtain a benchmark model；
Merging module, for add with data attribute identical new data in first data set, be incorporated in described the
In one data set, the second data set is formed；
Processing module, for testing the data in the second data set, value assessment is carried out to the second data intensive data, with
The data that mark value is more than default mark value are picked out, the data that will be singled out form the 3rd data set；To the 3rd data
Concentrate the data for not marking markup information to be labeled, the data for being labeled with markup information are incorporated in the 3rd data set；
Retraining module, for based on the 3rd data set after merging, retraining being carried out to benchmark model, after obtaining renewal
Benchmark model；
Loop module, for being the first new data set by the 3rd data set definition, new data is added, described in circular flow
Initial training module, merging module, processing module and retraining module, to be iterated training, until the mould after repetitive exercise
The accuracy of type is more than default accuracy.
The training system of the model of mind provided below with reference to diagram the present embodiment is described in detail.Need
Bright is, it should be understood that the division of the modules of apparatus above is only a kind of division of logic function, can be with when actually realizing
Completely or partially it is integrated on a physical entity, can also be physically separate.And these modules all can be passed through with software
The form that treatment element calls is realized；All it can also realize in the form of hardware；Treatment element can be passed through with part of module
The form of software is called to realize, part of module is realized by the form of hardware.For example, x modules can be the processing individually set up
Element, it can also be integrated in some chip of said apparatus and realize, in addition it is also possible to be stored in the form of program code
In the memory of said apparatus, called by some treatment element of said apparatus and perform the function of above x modules.Other moulds
The realization of block is similar therewith.In addition these modules can completely or partially integrate, and can also independently realize.It is described here
Treatment element can be a kind of integrated circuit, have signal disposal ability.In implementation process, each step of the above method
Or more modules can be completed by the instruction of the integrated logic circuit of the hardware in processor elements or software form.
For example, the above module can be arranged to implement one or more integrated circuits of above method, such as：
One or more specific integrated circuits (ApplicationSpecificIntegratedCircuit, abbreviation ASIC), or, one
Or multi-microprocessor (digitalsingnalprocessor, abbreviation DSP), or, one or more field-programmable gate array
Arrange (FieldProgrammableGateArray, abbreviation FPGA) etc..For another example, some module is dispatched by treatment element more than
When the form of program code is realized, the treatment element can be general processor, such as central processing unit
(CentralProcessingUnit, abbreviation CPU) or it is other can be with the processor of caller code.For another example, these modules can
To integrate, realized in the form of on-chip system (system-on-a-chip, abbreviation SOC).
Referring to Fig. 2, it is shown as theory structure schematic diagram of the training system of model of mind in an embodiment.Such as Fig. 2
Shown, the training system 2 of the model of mind includes initial training module 21, merging module 22, pretreatment module 23, processing mould
Block 24, retraining module 25 and loop module 26.
The initial training module 21 is used to enter the first data set of input and the markup information related to training mission
Row initial model training, to obtain a benchmark model.
Marked for example, first data set is the first LFW face public data collection and the identity related to training mission
Information, or first data set are the positional information of the first vehicle detection public data collection and vehicle on picture.
In specific practical application, the initial training module 21 to the first LFW face public data collection of input and with instruction
Practice the related identity markup information of task and carry out initial model training, to obtain recognition of face benchmark model.
Position of the initial training module 21 to the first vehicle detection public data collection and vehicle of input on picture
Information carries out initial model training, to obtain vehicle detection benchmark model.
In the present embodiment, initial model training refers to the loss letter related to training mission that optimization is pre-set
Number, the functional value of the loss function is set constantly to reduce until preset function value threshold value.The preset function value threshold value levels off to 0.
In this process, the parameter of model is also continuously available renewal, so that the benchmark model preferably can appoint in adaptation training
Business.
It is used to add and number in first data set with the merging module 22 of the initial training module 21 coupling
According to attribute identical new data, it is incorporated in first data set, forms the second data set.
In the present embodiment, the first LFW face public data collection is directed to, is added newly by modes such as web crawlers downloads
Face picture, formed the 2nd LFW face public data collection.
For the first vehicle detection public data collection, new vehicle pictures, shape are added by modes such as web crawlers downloads
Into the second vehicle detection public data collection.
It is used to carry out in advance the data in second data set with the pretreatment module 23 of the merging module 22 coupling
Processing.The pretreatment includes carrying out the data examination, image resolution ratio detection, image blur detection, and/or image
The degree of balance adjusts.In the present embodiment, the second data set is collected by modes such as web crawlers downloads.To in the second data set
It is to prepare the high data of quality for model retraining that data, which carry out pretreatment,.
Specifically, the pretreatment module 23 is concentrated using instruments such as Dlib, OpenCV to the 2nd LFW faces public data
Face picture carries out the pretreatments such as preliminary face examination, image blur detection.
Or the pretreatment module 23 is concentrated using instruments such as Dlib, OpenCV to the second vehicle detection public data
Vehicle figure carry out the pretreatments such as preliminary vehicle examination, image resolution ratio, fuzziness detection, the adjustment of image balance degree.
It is used to utilize the benchmark model with the processing module 24 of the pretreatment module 23 coupling, tests pretreated
Second data set.In the present embodiment, second data set is tested to refer to carry out the second data intensive data
The processing such as convolution, Chi Hua and more classification processing.
Specifically, processing module 24 carries out convolution, Chi Hua and more to the 2nd LFW face public datas intensive data
The processing such as classification processing.Or convolution, Chi Hua and more classification processing etc. are carried out to the second vehicle detection public data intensive data
Reason.
The processing module 24 carries out value after the data in testing the second data set, to second data set and commented
Estimate, be more than the data of default mark value to pick out mark value, and the costly data of mark that will be singled out form the
Three data sets.In the present embodiment, the processing module 24 by it is predefined, for judging the second data intensive data
Mark value cost function, pick out the data that mark value is more than default mark value.
Specifically, the processing module 24 by it is predefined, for judging that the 2nd LFW faces public data is concentrated
The cost function of the mark value of human body picture, to pick out the face picture that mark value is more than default mark value, and will
The costly data of mark picked out form the 3rd LFW face public data collection.
Or the processing module 24 by it is predefined, for judging that vehicle is concentrated in the second vehicle detection public data
The cost function of the mark value of picture, it is more than the vehicle pictures of default mark value to pick out mark value, and will selects
The costly data of mark gone out form the 3rd vehicle detection public data collection.
After the costly data of mark are picked out, the processing module 24 is additionally operable to not marked in the 3rd data set
The data of note information are labeled, and the data for being labeled with markup information are incorporated in the 3rd data set.
In the present embodiment, the markup information for being directed to selected data is the information marked by artificial notation methods.
For the face picture selected, face identity information, and the face figure that face identity information will be labeled with are marked
Piece is incorporated in the 3rd LFW face public data collection.
For the vehicle pictures selected, the positional information of vehicle is marked, and the car of the positional information of vehicle will be labeled with
Picture is incorporated in the 3rd vehicle detection public data collection.The present embodiment is by selecting the data of most worthy in data set
To be labeled, reduce the quantity manually marked, it is not necessary to total data is marked, so as to save mark cost.
It is used for the retraining module 25 of the processing module 24 coupling based on the 3rd data set after merging, to benchmark mould
Type carries out retraining, the benchmark model after being updated.In the present embodiment, training technique is same used by retraining module 25
Training technique in initial training module 21 is identical, that is, optimizes the loss function related to training mission pre-set, make this
The functional value of loss function is constantly reduced until preset function value threshold value.The preset function value threshold value levels off to 0.
With the initial training module 21, merging module 22, pretreatment module 23, processing module 24, retraining module 25
It is the first new data set that the loop module 26 of connection, which is used for the 3rd data set definition after merging, and described new first
New data is added in data set, the initial training module 21, merging module 22, pretreatment module 23, place described in circular flow
Reason module 24, retraining module 25 are to be iterated training, until the accuracy of the model after repetitive exercise is accurate more than default
Degree.In the present embodiment, by the way of training-mark iteration, partial data once can be only marked, partial data is trained
Model, renewal is then constantly iterated to model, training is more flexible, more rapidly.
Specifically, it is so that the accuracy of the model of recognition of face reaches more than 99% according to the demand of reality, works as model
Accuracy when being less than the value, just constantly rerun the initial training module 21, merging module 22, pretreatment module
23rd, processing module 24, retraining module 25, until the accuracy of model reaches default accuracy 99%.
It is so that the accuracy of the model of vehicle identification reaches more than 99% according to the demand of reality, when the accuracy of model
During less than the value, the initial training module 21 that just constantly reruns, merging module 22, pretreatment module 23, processing
Module 24, retraining module 25, until the accuracy of model reaches default accuracy 99%.
Embodiment three
The present embodiment provides a kind of terminal, including：Processor, memory, transceiver, communication interface and system bus；Deposit
Reservoir and communication interface are connected with processor and transceiver by system bus and complete mutual communication, and memory is used to deposit
Computer program is stored up, communication interface is used for and other equipment is communicated, and processor and transceiver are used to run computer program,
Make the step S11 to step S18 of the training system of model of mind of the terminal execution as described in embodiment one.
System bus mentioned above can be Peripheral Component Interconnect standard
(PeripheralPomponentInterconnect, abbreviation PCI) bus or EISA
(ExtendedIndustryStandardArchitecture, abbreviation EISA) bus etc..The system bus can be divided into address
Bus, data/address bus, controlling bus etc..For ease of representing, only represented in figure with a thick line, it is not intended that only one total
Line or a type of bus.Communication interface is used for accessing data base device and other equipment (such as client, read-write storehouse
And read-only storehouse) between communication.Memory may include random access memory (RandomAccessMemory, abbreviation RAM),
Nonvolatile memory (non-volatilememory), for example, at least a magnetic disk storage may also also be included.
Above-mentioned processor can be general processor, including central processing unit (CentralProcessingUnit, letter
Claim CPU), network processing unit (NetworkProcessor, abbreviation NP) etc.；It can also be digital signal processor
(DigitalSignalProcessing, abbreviation DSP), application specific integrated circuit
(ApplicationSpecificIntegratedCircuit, abbreviation ASIC), field programmable gate array (Field-
ProgrammableGateArray, abbreviation FPGA) either other PLDs, discrete gate or transistor logic device
Part, discrete hardware components.
In summary, training method/system, computer-readable recording medium and the terminal of model of mind of the present invention
Reduce the quantity manually marked, it is not necessary to mark total data, so as to save mark cost, and improve model training effect
Rate.So the present invention effectively overcomes various shortcoming of the prior art and has high industrial utilization.
The above-described embodiments merely illustrate the principles and effects of the present invention, not for the limitation present invention.It is any ripe
Know the personage of this technology all can carry out modifications and changes under the spirit and scope without prejudice to the present invention to above-described embodiment.Cause
This, those of ordinary skill in the art is complete without departing from disclosed spirit and institute under technological thought such as
Into all equivalent modifications or change, should by the present invention claim be covered.
Claims (12)
1. a kind of training method of model of mind, it is characterised in that the training method of the model of mind comprises the following steps：
The first data set and the markup information related to training mission to input carry out initial model training, to obtain a benchmark
Model；
Addition and data attribute identical new data in first data set, are incorporated in first data set, form the
Two data sets；
The data tested in the second data set, value assessment is carried out to the second data intensive data, it is big to pick out mark value
In the data of default mark value, the data that will be singled out form the 3rd data set；To mark letter is not marked in the 3rd data set
The data of breath are labeled, and the data for being labeled with markup information are incorporated in the 3rd data set；
Based on the 3rd data set after merging, retraining, the benchmark model after being updated are carried out to benchmark model；
It is the first new data set by the 3rd data set definition, adds new data, circulation performs above step, to be iterated instruction
Practice, until the accuracy of the model after repetitive exercise is more than default accuracy.
2. the training method of model of mind according to claim 1, it is characterised in that the first data set of described pair of input
And the markup information related to training mission carries out initial model training, set in advance with obtaining to refer to optimize the step of a benchmark model
The loss function related to training mission put, the functional value of the loss function is set constantly to reduce until preset function threshold value.
3. the training method of model of mind according to claim 1, it is characterised in that forming second data set
Afterwards, the training method of the model of mind also includes pre-processing the data in second data set；The pretreatment
Adjusted including the data are carried out with examination, image resolution ratio detection, image blur detection, and/or image balance degree.
4. the training method of model of mind according to claim 1, it is characterised in that in the second data set of the test
The step of data, includes carrying out the second data intensive data convolution, Chi Hua and more classification processing.
5. the training method of model of mind according to claim 1, it is characterised in that described in second data set
Data carry out value assessment, are also included by predefined with picking out the step of mark value is more than the data of default mark value
, the cost function of mark value for judging the second data intensive data, pick out mark value and be more than pre- bidding
Note the data of value.
6. the training method of model of mind according to claim 1, it is characterised in that described to be directed to selected data
Markup information is the information marked by artificial notation methods.
A kind of 7. training system of model of mind, it is characterised in that including：
Initial training module, initial model is carried out for the first data set to input and the markup information related to training mission
Training, to obtain a benchmark model；
Merging module, for addition and data attribute identical new data in first data set, it is incorporated in first number
According to concentration, the second data set is formed；
Processing module, for testing the data in the second data set, value assessment is carried out to the second data intensive data, to select
Go out the data that mark value is more than default mark value, the data that will be singled out form the 3rd data set；To in the 3rd data set
The data for not marking markup information are labeled, and the data for being labeled with markup information are incorporated in the 3rd data set；
Retraining module, for based on the 3rd data set after merging, retraining, the base after being updated to be carried out to benchmark model
Quasi-mode type；
Loop module, for being the first new data set by the 3rd data set definition, add new data, it is initial described in circular flow
Training module, merging module, processing module and retraining module, trained with being iterated, until the model after repetitive exercise
Accuracy is more than default accuracy.
8. the training system of model of mind according to claim 7, it is characterised in that forming second data set
Afterwards, the training system of the model of mind also includes the pretreatment module with merging module coupling, the pretreatment module
For being pre-processed to the data in second data set；The pretreatment includes carrying out examination, image to the data
Resolution ratio detection, image blur detection, and/or the adjustment of image balance degree.
9. the training system of model of mind according to claim 7, it is characterised in that the number of processing module test second
Refer to carry out the second data intensive data convolution, Chi Hua and more classification processing according to the data of concentration.
10. the training system of model of mind according to claim 7, it is characterised in that the processing module is additionally operable to lead to
The cost function of mark value predefined, for judging the second data intensive data is crossed, it is big to pick out mark value
In the data of default mark value.
11. a kind of computer-readable recording medium, is stored thereon with computer program, it is characterised in that the program is by processor
The training method of model of mind any one of claim 1 to 6 is realized during execution.
A kind of 12. terminal, it is characterised in that including：Processor and memory；
The memory is used to store computer program, and the processor is used for the computer journey for performing the memory storage
Sequence, so that the terminal performs the training method of the model of mind as any one of claim 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710561654.5A CN107392125A (en) | 2017-07-11 | 2017-07-11 | Training method/system, computer-readable recording medium and the terminal of model of mind |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710561654.5A CN107392125A (en) | 2017-07-11 | 2017-07-11 | Training method/system, computer-readable recording medium and the terminal of model of mind |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107392125A true CN107392125A (en) | 2017-11-24 |
Family
ID=60340400
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710561654.5A Pending CN107392125A (en) | 2017-07-11 | 2017-07-11 | Training method/system, computer-readable recording medium and the terminal of model of mind |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107392125A (en) |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108319938A (en) * | 2017-12-31 | 2018-07-24 | 奥瞳系统科技有限公司 | High quality training data preparation system for high-performance face identification system |
| CN108537289A (en) * | 2018-04-24 | 2018-09-14 | 百度在线网络技术（北京）有限公司 | Training method, device and the storage medium of data identification model |
| CN108537129A (en) * | 2018-03-14 | 2018-09-14 | 北京影谱科技股份有限公司 | The mask method of training sample, device and system |
| CN108564134A (en) * | 2018-04-27 | 2018-09-21 | 网易（杭州）网络有限公司 | Data processing method, device, computing device and medium |
| CN108664999A (en) * | 2018-05-03 | 2018-10-16 | 北京图森未来科技有限公司 | A kind of training method and its device, computer server of disaggregated model |
| CN108764372A (en) * | 2018-06-08 | 2018-11-06 | Oppo广东移动通信有限公司 | Construction method and device, mobile terminal, the readable storage medium storing program for executing of data set |
| CN108875963A (en) * | 2018-06-28 | 2018-11-23 | 北京字节跳动网络技术有限公司 | Optimization method, device, terminal device and the storage medium of machine learning model |
| CN108875821A (en) * | 2018-06-08 | 2018-11-23 | Oppo广东移动通信有限公司 | The training method and device of disaggregated model, mobile terminal, readable storage medium storing program for executing |
| CN108898162A (en) * | 2018-06-08 | 2018-11-27 | 东软集团股份有限公司 | A kind of data mask method, device, equipment and computer readable storage medium |
| CN108960409A (en) * | 2018-06-13 | 2018-12-07 | 南昌黑鲨科技有限公司 | Labeled data generation method, equipment and computer readable storage medium |
| CN109034397A (en) * | 2018-08-10 | 2018-12-18 | 腾讯科技（深圳）有限公司 | Model training method, device, computer equipment and storage medium |
| CN109359689A (en) * | 2018-10-19 | 2019-02-19 | 科大讯飞股份有限公司 | A kind of data identification method and device |
| CN109582793A (en) * | 2018-11-23 | 2019-04-05 | 深圳前海微众银行股份有限公司 | Model training method, customer service system and data labeling system, readable storage medium storing program for executing |
| CN109919317A (en) * | 2018-01-11 | 2019-06-21 | 华为技术有限公司 | A kind of machine learning model training method and device |
| CN109918444A (en) * | 2019-02-01 | 2019-06-21 | 上海尚阵智能科技有限公司 | Training/verifying/management method/system, medium and equipment of model result |
| CN110443141A (en) * | 2019-07-08 | 2019-11-12 | 深圳中兴网信科技有限公司 | Data set processing method, data set processing unit and storage medium |
| WO2020134533A1 (en) * | 2018-12-29 | 2020-07-02 | 北京市商汤科技开发有限公司 | Method and apparatus for training deep model, electronic device, and storage medium |
| WO2020206958A1 (en) * | 2019-04-09 | 2020-10-15 | 苏宁云计算有限公司 | Iterative learning-based smart aided labeling method and system |
| CN111815361A (en) * | 2020-07-10 | 2020-10-23 | 北京思特奇信息技术股份有限公司 | Region boundary calculation method and device, electronic equipment and storage medium |
| CN112036570A (en) * | 2020-07-30 | 2020-12-04 | 第四范式（北京）技术有限公司 | Method and device for scheduling annotation task and readable storage medium |
| CN112149701A (en) * | 2019-06-28 | 2020-12-29 | 杭州海康威视数字技术股份有限公司 | Image identification method, virtual sample data generation method and storage medium |
| CN112534367A (en) * | 2018-10-01 | 2021-03-19 | 欧姆龙株式会社 | Learning device, control device, learning method, and learning program |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090228411A1 (en) * | 2008-03-06 | 2009-09-10 | Kddi Corporation | Reducing method for support vector |
| CN103793510A (en) * | 2014-01-29 | 2014-05-14 | 苏州融希信息科技有限公司 | Classifier construction method based on active learning |
| CN105117429A (en) * | 2015-08-05 | 2015-12-02 | 广东工业大学 | Scenario image annotation method based on active learning and multi-label multi-instance learning |
-
2017
- 2017-07-11 CN CN201710561654.5A patent/CN107392125A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090228411A1 (en) * | 2008-03-06 | 2009-09-10 | Kddi Corporation | Reducing method for support vector |
| CN103793510A (en) * | 2014-01-29 | 2014-05-14 | 苏州融希信息科技有限公司 | Classifier construction method based on active learning |
| CN105117429A (en) * | 2015-08-05 | 2015-12-02 | 广东工业大学 | Scenario image annotation method based on active learning and multi-label multi-instance learning |
Cited By (38)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108319938A (en) * | 2017-12-31 | 2018-07-24 | 奥瞳系统科技有限公司 | High quality training data preparation system for high-performance face identification system |
| CN108319938B (en) * | 2017-12-31 | 2022-05-17 | 奥瞳系统科技有限公司 | High-quality training data preparation system for high-performance face recognition system |
| CN109919317A (en) * | 2018-01-11 | 2019-06-21 | 华为技术有限公司 | A kind of machine learning model training method and device |
| CN109919317B (en) * | 2018-01-11 | 2024-06-04 | 华为技术有限公司 | Machine learning model training method and device |
| CN108537129A (en) * | 2018-03-14 | 2018-09-14 | 北京影谱科技股份有限公司 | The mask method of training sample, device and system |
| CN108537129B (en) * | 2018-03-14 | 2021-01-08 | 北京影谱科技股份有限公司 | Method, device and system for marking training samples |
| CN108537289A (en) * | 2018-04-24 | 2018-09-14 | 百度在线网络技术（北京）有限公司 | Training method, device and the storage medium of data identification model |
| CN108537289B (en) * | 2018-04-24 | 2023-04-07 | 百度在线网络技术（北京）有限公司 | Training method and device of data recognition model and storage medium |
| CN108564134A (en) * | 2018-04-27 | 2018-09-21 | 网易（杭州）网络有限公司 | Data processing method, device, computing device and medium |
| CN108564134B (en) * | 2018-04-27 | 2021-07-06 | 网易（杭州）网络有限公司 | Data processing method, device, computing equipment and medium |
| CN108664999A (en) * | 2018-05-03 | 2018-10-16 | 北京图森未来科技有限公司 | A kind of training method and its device, computer server of disaggregated model |
| CN108664999B (en) * | 2018-05-03 | 2021-02-12 | 北京图森智途科技有限公司 | Training method and device of classification model and computer server |
| CN108898162A (en) * | 2018-06-08 | 2018-11-27 | 东软集团股份有限公司 | A kind of data mask method, device, equipment and computer readable storage medium |
| US11138478B2 (en) | 2018-06-08 | 2021-10-05 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method and apparatus for training, classification model, mobile terminal, and readable storage medium |
| CN108875821A (en) * | 2018-06-08 | 2018-11-23 | Oppo广东移动通信有限公司 | The training method and device of disaggregated model, mobile terminal, readable storage medium storing program for executing |
| CN108764372B (en) * | 2018-06-08 | 2019-07-16 | Oppo广东移动通信有限公司 | Construction method and device, mobile terminal, the readable storage medium storing program for executing of data set |
| CN108898162B (en) * | 2018-06-08 | 2021-03-30 | 东软集团股份有限公司 | Data annotation method, device and equipment and computer readable storage medium |
| WO2019233297A1 (en) * | 2018-06-08 | 2019-12-12 | Oppo广东移动通信有限公司 | Data set construction method, mobile terminal and readable storage medium |
| CN108764372A (en) * | 2018-06-08 | 2018-11-06 | Oppo广东移动通信有限公司 | Construction method and device, mobile terminal, the readable storage medium storing program for executing of data set |
| CN108960409A (en) * | 2018-06-13 | 2018-12-07 | 南昌黑鲨科技有限公司 | Labeled data generation method, equipment and computer readable storage medium |
| CN108960409B (en) * | 2018-06-13 | 2021-08-03 | 南昌黑鲨科技有限公司 | Method and device for generating annotation data and computer-readable storage medium |
| CN108875963A (en) * | 2018-06-28 | 2018-11-23 | 北京字节跳动网络技术有限公司 | Optimization method, device, terminal device and the storage medium of machine learning model |
| CN109034397B (en) * | 2018-08-10 | 2023-04-07 | 腾讯科技（深圳）有限公司 | Model training method and device, computer equipment and storage medium |
| CN109034397A (en) * | 2018-08-10 | 2018-12-18 | 腾讯科技（深圳）有限公司 | Model training method, device, computer equipment and storage medium |
| US11971709B2 (en) | 2018-10-01 | 2024-04-30 | Omron Corporation | Learning device, control device, learning method, and recording medium |
| CN112534367A (en) * | 2018-10-01 | 2021-03-19 | 欧姆龙株式会社 | Learning device, control device, learning method, and learning program |
| CN112534367B (en) * | 2018-10-01 | 2023-09-19 | 欧姆龙株式会社 | Learning device, control device, learning method, and computer-readable storage medium |
| CN109359689A (en) * | 2018-10-19 | 2019-02-19 | 科大讯飞股份有限公司 | A kind of data identification method and device |
| CN109582793A (en) * | 2018-11-23 | 2019-04-05 | 深圳前海微众银行股份有限公司 | Model training method, customer service system and data labeling system, readable storage medium storing program for executing |
| CN109582793B (en) * | 2018-11-23 | 2023-05-23 | 深圳前海微众银行股份有限公司 | Model training method, customer service system, data labeling system and readable storage medium |
| WO2020134533A1 (en) * | 2018-12-29 | 2020-07-02 | 北京市商汤科技开发有限公司 | Method and apparatus for training deep model, electronic device, and storage medium |
| CN109918444A (en) * | 2019-02-01 | 2019-06-21 | 上海尚阵智能科技有限公司 | Training/verifying/management method/system, medium and equipment of model result |
| WO2020206958A1 (en) * | 2019-04-09 | 2020-10-15 | 苏宁云计算有限公司 | Iterative learning-based smart aided labeling method and system |
| CN112149701A (en) * | 2019-06-28 | 2020-12-29 | 杭州海康威视数字技术股份有限公司 | Image identification method, virtual sample data generation method and storage medium |
| CN112149701B (en) * | 2019-06-28 | 2024-05-10 | 杭州海康威视数字技术股份有限公司 | Image recognition method, virtual sample data generation method and storage medium |
| CN110443141A (en) * | 2019-07-08 | 2019-11-12 | 深圳中兴网信科技有限公司 | Data set processing method, data set processing unit and storage medium |
| CN111815361A (en) * | 2020-07-10 | 2020-10-23 | 北京思特奇信息技术股份有限公司 | Region boundary calculation method and device, electronic equipment and storage medium |
| CN112036570A (en) * | 2020-07-30 | 2020-12-04 | 第四范式（北京）技术有限公司 | Method and device for scheduling annotation task and readable storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107392125A (en) | Training method/system, computer-readable recording medium and the terminal of model of mind | |
| CN110349147B (en) | Model training method, fundus macular region lesion recognition method, device and equipment | |
| CN109034159A (en) | image information extracting method and device | |
| CN109176512A (en) | A kind of method, robot and the control device of motion sensing control robot | |
| CN108229341A (en) | Sorting technique and device, electronic equipment, computer storage media, program | |
| CN109544598A (en) | Method for tracking target, device and readable storage medium storing program for executing | |
| JP2023044669A (en) | Graph model-based brain function registration method | |
| CN109685077A (en) | A kind of breast lump image-recognizing method and device | |
| CN109934227A (en) | System for recognizing characters from image and method | |
| CN110781976B (en) | Extension method of training image, training method and related device | |
| CN107845092A (en) | Clothes LOGO effect detection methods, device and electronic equipment | |
| CN111897985B (en) | Image multi-label classification method, system, equipment and readable storage medium | |
| CN109902548A (en) | A kind of object properties recognition methods, calculates equipment and system at device | |
| CN106383912A (en) | Picture retrieval method and apparatus | |
| CN108537168A (en) | Human facial expression recognition method based on transfer learning technology | |
| CN111915618B (en) | Peak response enhancement-based instance segmentation algorithm and computing device | |
| Shu et al. | LVC-Net: Medical image segmentation with noisy label based on local visual cues | |
| CN110222330A (en) | Method for recognizing semantics and device, storage medium, computer equipment | |
| CN110264274A (en) | Objective group's division methods, model generating method, device, equipment and storage medium | |
| CN115345938B (en) | Global-to-local-based head shadow mark point positioning method, equipment and medium | |
| CN109740752A (en) | Depth model training method and device, electronic equipment and storage medium | |
| CN112132812B (en) | Certificate verification method and device, electronic equipment and medium | |
| CN108595815A (en) | Artificial intelligence body training system and passive circuit Optimum Design System and method | |
| CN106874913A (en) | A kind of vegetable detection method | |
| CN110033012A (en) | A kind of production method for tracking target based on channel characteristics weighted convolution neural network |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171124 |