CN109754068A - Transfer learning method and terminal device based on deep learning pre-training model - Google Patents
Transfer learning method and terminal device based on deep learning pre-training model Download PDFInfo
- Publication number
- CN109754068A CN109754068A CN201811473650.2A CN201811473650A CN109754068A CN 109754068 A CN109754068 A CN 109754068A CN 201811473650 A CN201811473650 A CN 201811473650A CN 109754068 A CN109754068 A CN 109754068A
- Authority
- CN
- China
- Prior art keywords
- training
- data
- model
- data set
- new
- 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
Landscapes
- Electrically Operated Instructional Devices (AREA)
Abstract
The present invention is suitable for model construction techniques field, a kind of transfer learning method and terminal device based on deep learning pre-training model is provided, this method comprises: data set is divided into training dataset and test data set;New model is obtained to the pre-training model re -training of acquisition according to the data that the training data is concentrated;According to the data that the test data is concentrated, detection is carried out to the Generalization Capability of the new model and obtains testing result;When the testing result reaches pre-set level value, it is determined that the new model is the model for meeting application.The program can solve the data of scale needed for when facing the particular problem in a certain field, may being usually unable to get building model in the prior art, and construct the new model problem that time-consuming, resource consumption is larger.
Description
Technical field
The invention belongs to model construction techniques field more particularly to a kind of migrations based on deep learning pre-training model
Learning method and terminal device.
Background technique
Under traditional machine learning frame, the task of study is exactly to be constructed on the basis of data according to given train up
A new model.However, following key is had in machine learning in current research:
1, large-scale data needed for training new model.By the large-scale data for a certain specific field and bad obtain
It takes.
2, time-consuming.Deep learning is a large-scale neural network, and the number of plies is relatively more, training needed for expend the time compared with
It is long, and neural network is more complicated, and data are more, it would be desirable to which the time spent in training process is also more.
3, resource is consumed.Neural network usually requires a large amount of marker samples, usually a large amount of data and neural network
In the response of each layer can consume a large amount of memories.It is identical that traditional machine learning usually assumes that training data is obeyed with test data
Data distribution.However, in many cases, this same distributional assumption is simultaneously unsatisfactory for, such as training data is expired, results in the need for me
Remove to mark a large amount of training data again to meet the needs of we train, but mark new data and need a large amount of manpower and object
Power, even and if we have a large amount of, training data under different distributions, abandoning these data completely is also to waste very much
's.
Summary of the invention
In view of this, the embodiment of the invention provides a kind of transfer learning method based on deep learning pre-training model and
Terminal device, can solve in the prior art face a certain field particular problem when, may usually be unable to get building mould
The data of scale needed for type, and construct the new model problem that time-consuming, resource consumption is larger.
The first aspect of the embodiment of the present invention provides a kind of transfer learning method based on deep learning pre-training model,
Include:
Data set is divided into training dataset and test data set;
New model is obtained to the pre-training model re -training of acquisition according to the data that the training data is concentrated;
According to the data that the test data is concentrated, detection is carried out to the Generalization Capability of the new model and obtains detection knot
Fruit;
When the testing result reaches pre-set level value, it is determined that the new model is the model for meeting application.
In one embodiment, the training dataset and the test data set are that two data distributions are consistent and mutual exclusion
Data acquisition system.
In one embodiment, the test data set is carried out sampling in the data set by way of stratified sampling and be obtained
, the data in the data set in addition to the test data set are the training dataset.
In one embodiment, the data that the training data is concentrated are more than the data that the test data is concentrated.
In one embodiment, the ratio section that the data that the training data is concentrated account for the data of the data set is [2/
3,4/5]。
It is in one embodiment, described that data set is divided into training dataset and test data set, comprising:
Data set is subjected to n times random division, N group training dataset is obtained and corresponding test data set, the N is greater than
Equal to 1.
In one embodiment, the data concentrated according to the training data, instruct the pre-training model of acquisition again
Practice, obtain new model, comprising:
According to the data that the training data is concentrated, to the training in the level of model rear end in the pre-training model of acquisition
Weight re -training obtains new weight;
According to the training data concentrate data, to the parameter in the level of model rear end in the pre-training model into
Row adjustment, obtains new parameter;
According to remained unchanged in the level of model front end in the pre-training model training weight, the new weight with
And the new parameter, obtain new model.
The second aspect of the embodiment of the present invention provides a kind of transfer learning device based on deep learning pre-training model,
Include:
Division module, for data set to be divided into training dataset and test data set;
Training module, the data for being concentrated according to the training data obtain the pre-training model re -training of acquisition
Obtain new model;
Test module, the data for being concentrated according to the test data, examines the Generalization Capability of the new model
It surveys and obtains testing result;
Determining module, for when the testing result reaches pre-set level value, it is determined that the new model is to meet to answer
Model.
The third aspect of the embodiment of the present invention provides a kind of terminal device, including memory, processor and is stored in
In the memory and the computer program that can run on the processor, which is characterized in that described in the processor executes
Step described in the transfer learning method based on deep learning pre-training model realized when computer program.
The fourth aspect of the embodiment of the present invention provides a kind of computer readable storage medium, the computer-readable storage
Media storage has computer program, which is characterized in that realizes when the computer program is executed by processor and is based on deep learning
Step described in the transfer learning method of pre-training model.
Existing beneficial effect is the embodiment of the present invention compared with prior art: scheme provided in an embodiment of the present invention is led to
It crosses and is adjusted the part layer in the pre-training model in deep learning to obtain new model according to the data of frontier, then to new
Model is assessed, adjusts to be applied to practical problem, trains up data to solve and need to give in the prior art
On the basis of learn a new model, when the new model of needs is a large-scale neural network, time-consuming and consumes
Problem more than resource.
Detailed description of the invention
It to describe the technical solutions in the embodiments of the present invention more clearly, below will be to embodiment or description of the prior art
Needed in attached drawing be briefly described, it should be apparent that, the accompanying drawings in the following description is only of the invention some
Embodiment for those of ordinary skill in the art without any creative labor, can also be according to these
Attached drawing obtains other attached drawings.
Fig. 1 is a kind of process of transfer learning method based on deep learning pre-training model provided in an embodiment of the present invention
Schematic diagram;
Fig. 2 is the stream of another transfer learning method based on deep learning pre-training model provided in an embodiment of the present invention
Journey schematic diagram;
Fig. 3 is a kind of example of transfer learning device based on deep learning pre-training model provided in an embodiment of the present invention
Figure;
Fig. 4 is the schematic diagram of terminal device provided in an embodiment of the present invention.
Specific embodiment
In being described below, for illustration and not for limitation, the tool of such as particular system structure, technology etc is proposed
Body details, to understand thoroughly the embodiment of the present invention.However, it will be clear to one skilled in the art that there is no these specific
The present invention also may be implemented in the other embodiments of details.In other situations, it omits to well-known system, device, electricity
The detailed description of road and method, in case unnecessary details interferes description of the invention.
In order to illustrate technical solutions according to the invention, the following is a description of specific embodiments.
The embodiment of the present invention provides a kind of transfer learning method based on deep learning pre-training model, as shown in Figure 1, should
Method the following steps are included:
Step 101, data set is divided into training dataset and test data set.
Optionally, the training dataset and the test data set are that two data distributions are consistent and the data of mutual exclusion
Set.For example, data set D is divided into two data distributions unanimously and the data acquisition system of mutual exclusion, one is training dataset
S, it be the intersection of D, S and T is empty that one, which is the union of test data set T, S and T,.
Optionally, S is consistent with the data distribution of T, be in order to avoid because data partition process introduces additional deviation, thus
Influence final result.
Further, in order to guarantee the consistency of data distribution, in the application by the way of stratified sampling to data into
Row sampling.Specifically, the test data set carries out sampling acquisition by way of stratified sampling in the data set, described
Data in data set in addition to the test data set are the training dataset.For example, having m in data set D1A positive sample
This, there is m2A negative sample, and the ratio that S accounts for D is p, then the ratio that T accounts for D is (1-p), then it can be by m1A positive sample
Middle acquisition (m1* the positive sample that p) a sample is concentrated as training data, and by m2(m is acquired in a negative sample2* p) a sample
This negative sample concentrated as training data, remaining sample concentrated as test data.
Optionally, the data that the training data is concentrated are more than the data that the test data is concentrated.
After dividing to data set D, there are many sample in training dataset S, close to D, train the new model and D come
Itself training the model come may be very close to, but since T is smaller, it is not accurate enough at this time to may result in assessment result
Stablize;If S sample is seldom, and the new model for training can be made to train the model come with D and differed greatly.Therefore,
Specifically, the ratio section that the data that the training data is concentrated account for the data of the data set is [2/3,4/5], the then survey
The ratio section that data in examination data set account for the data of the data set is [1/3,1/5].
Further, when being divided to data set, data set can be subjected to n times random division, obtains the training of N group
Data set and corresponding test data set, the N are more than or equal to 1, the N group testing result of acquisition can be averaged work in this way
For final assessment result, the assessment result for being is more acurrate, more meets the quick application of different field.
Step 102, the data concentrated according to the training data obtain new the pre-training model re -training of acquisition
Model.
Optionally, as shown in Fig. 2, the step includes following sub-step:
Step 1021, the data concentrated according to the training data, to the level of model rear end in the pre-training model
In training weight re -training, obtain new weight.
It optionally, can also include: to obtain pre-training model before this step, the pre-training model includes training power
Weight.
During deep learning, since computing resource is limited or training dataset is smaller, but we want again obtain compared with
It is well more stable as a result, still can obtain some trained models, i.e. pre-training model first, directly to pre-training
Model carries out re -training acquisition new model can save a large amount of people without trained a new model of starting from scratch in this way
Power material resources.
The source model of one pre-training be it is select from available model, many research institutions have all issued based on super
The model of large data sets, these all can serve as the choice of source model.The pre-training model that this programme obtains is with training
The pre-training model of weight.
Further alternative, deep learning passes through forward calculation and backpropagation, continuous adjusting parameter, to extract optimal spy
Sign, to achieve the purpose that prediction.Advanced connection of the level of model front end commonly used to capture input data, such as image border
With main body etc.;The level of model rear end helps to make the information finally determined commonly used to capture, such as distinguishing target
The detailed information of output.
After obtaining pre-training model, re -training total is not needed, it is only necessary to several layers of be trained for therein
?.Some layers of the weight that model originates is remained unchanged, the subsequent layer of re -training obtains new weight.That is basis
The data that the training data is concentrated, instruct the training weight in the level of model rear end in the pre-training model of acquisition again
Practice, obtains new weight.
During adjusting model, we can repeatedly be attempted, the number concentrated according to the different training data of N group
It is adjusted according to pre-training model, freezes layer frozen layers and retraining layer retrain so as to find according to result
Best collocation between layers.
Step 1022, the data concentrated according to the training data, to the level of model rear end in the pre-training model
In parameter be adjusted, obtain new parameter.
Optionally, it is finely adjusted, trained model is applied to similar or only by the parameter to pre-training model
Have in the different task of nuance.
Step 1023, according to the training weight, described remained unchanged in the level of model front end in the pre-training model
New weight and the new parameter obtains new model.
It should be understood that the application of new model is the process of loop iteration, only by the lasting adjustment of model and tune
It is excellent just to adapt to online data and business objective, it can just find most effective new model.
Step 103, the data concentrated according to the test data, carry out detection acquisition to the Generalization Capability of the new model
Testing result.
After obtaining new model, need to assess the performance of new model, the application passes through the generalization to new model
It can be carried out detection, to obtain the result to new model assessment.
Generalization Capability, that is, generalization ability, generalization ability (generalization ability) refer to machine learning algorithm pair
The adaptability of fresh sample.The destination of study is to acquire the rule for lying in data to behind, to same rule
The data other than collection are practised, trained network can also provide suitable output, which is known as generalization ability.
Optionally, in order to obtain more accurate testing result, the data that N group test data is concentrated, to the new model
Generalization Capability detected, obtain N number of testing result, take the average value of N number of testing result as final testing result.
Step 104, when the testing result reaches pre-set level value, it is determined that the new model is the mould for meeting application
Type.
, can be online by new model when new model reaches the pre-set level value of setting, it puts into production, obtains for enterprise
The transfer learning method model used with user.
Optionally, as shown in Fig. 2, before this step further include:
Step 105, detect whether the testing result reaches pre-set level value.
Step 106, when the testing result is not up to the pre-set level value, it is determined that the new model is unsatisfactory for answering
With, continue adjust model, i.e., execution step 1021, until assessment result meet apply.
Transfer learning method provided in an embodiment of the present invention based on deep learning pre-training model, by being drawn to data set
It is divided into the training dataset and test data set of independent same distribution and mutual exclusion, according to the data that the training data is concentrated, to obtaining
The pre-training model re -training taken obtains new model;According to the data that the test data is concentrated, to the general of the new model
Change performance and carries out detection acquisition testing result;When the testing result reaches pre-set level value, it is determined that the new model is
Meet the model of application.Can solve in the prior art face a certain field particular problem when, may usually be unable to get
The problem of large-scale data needed for constructing model, and if it is given train up data on the basis of it is new to learn one
Model, when the new model of needs is a large-scale neural network, time-consuming and consumes the problem more than resource.This programme is logical
Re -training, the new model of acquisition can be carried out to the part layer in pre-training model for certain categorical data by crossing transfer learning
In relationship can also be easily applied to the different problems in same field.
It should be understood that the size of the serial number of each step is not meant that the order of the execution order in above-described embodiment, each process
Execution sequence should be determined by its function and internal logic, the implementation process without coping with the embodiment of the present invention constitutes any limit
It is fixed.
The embodiment of the present invention provides a kind of transfer learning device based on deep learning pre-training model, as shown in figure 3, should
Device includes: division module 301, training module 302, test module 303, determining module 304.
Division module 301, for data set to be divided into training dataset and test data set.
Optionally, the training dataset and the test data set are that two data distributions are consistent and the data of mutual exclusion
Set.For example, data set D is divided into two data distributions unanimously and the data acquisition system of mutual exclusion, one is training dataset
S, it be the intersection of D, S and T is empty that one, which is the union of test data set T, S and T,.
Optionally, S is consistent with the data distribution of T, be in order to avoid because data partition process introduces additional deviation, thus
Influence final result.
Further, in order to guarantee the consistency of data distribution, in the application by the way of stratified sampling to data into
Row sampling.Specifically, the test data set carries out sampling acquisition by way of stratified sampling in the data set, described
Data in data set in addition to the test data set are the training dataset.For example, having m in data set D1A positive sample
This, there is m2A negative sample, and the ratio that S accounts for D is p, then the ratio that T accounts for D is (1-p), then it can be by m1A positive sample
Middle acquisition (m1* the positive sample that p) a sample is concentrated as training data, and by m2(m is acquired in a negative sample2* p) a sample
This negative sample concentrated as training data, remaining sample concentrated as test data.
Optionally, the data that the training data is concentrated are more than the data that the test data is concentrated.
After dividing to data set D, there are many sample in training dataset S, close to D, train the new model and D come
Itself training the model come may be very close to, but since T is smaller, it is not accurate enough at this time to may result in assessment result
Stablize;If S sample is seldom, and the new model for training can be made to train the model come with D and differed greatly.Therefore,
Specifically, the ratio section that the data that the training data is concentrated account for the data of the data set is [2/3,4/5], the then survey
The ratio section that data in examination data set account for the data of the data set is [1/3,1/5].
Further, when being divided to data set, data set can be subjected to n times random division, obtains the training of N group
Data set and corresponding test data set, the N are more than or equal to 1, the N group testing result of acquisition can be averaged work in this way
For final assessment result, the assessment result for being is more acurrate, more meets the quick application of different field.
Training module 302, the data for being concentrated according to the training data, instructs the pre-training model of acquisition again
Practice, obtains new model.
Optionally, training module 302, specifically for the data concentrated according to the training data, to the pre-training of acquisition
Training weight re -training in model in the level of model rear end, obtains new weight;It is concentrated according to the training data
Data are adjusted the parameter in the level of model rear end in the pre-training model, obtain new parameter;According to described pre-
Training weight, the new weight and the new parameter remained unchanged in the level of model front end in training pattern, is obtained
Obtain new model.
During deep learning, since computing resource is limited or training dataset is smaller, but we want again obtain compared with
It is well more stable as a result, still can obtain some trained models, i.e. pre-training model first, directly to pre-training
Model carries out re -training acquisition new model can save a large amount of people without trained a new model of starting from scratch in this way
Power material resources.
The source model of one pre-training be it is select from available model, many research institutions have all issued based on super
The model of large data sets, these all can serve as the choice of source model.The pre-training model that this programme obtains is with training
The pre-training model of weight.
After obtaining pre-training model, re -training total is not needed, it is only necessary to several layers of be trained for therein
?.Some layers of the weight that model originates is remained unchanged, the subsequent layer of re -training obtains new weight.That is basis
The data that the training data is concentrated, instruct the training weight in the level of model rear end in the pre-training model of acquisition again
Practice, obtains new weight.
During adjusting model, we can repeatedly be attempted, the number concentrated according to the different training data of N group
It is adjusted according to pre-training model, freezes layer frozen layers and retraining layer retrain so as to find according to result
Best collocation between layers.
Test module 303, the data for being concentrated according to the test data carry out the Generalization Capability of the new model
Detection obtains testing result.
Optionally, in order to obtain more accurate testing result, the data that test module 303 concentrates N group test data are right
The Generalization Capability of the new model is detected, and N number of testing result is obtained, and takes the average value of N number of testing result as final
Testing result.
Determining module 304, for when the testing result reaches pre-set level value, it is determined that the new model is to meet
The model of application.
When the testing result is not up to the pre-set level value, it is determined that the new model is unsatisfactory for applying, then by
Training module 302 continues to adjust model, until the assessment result of the new model of acquisition meets application.
The embodiment of the present invention provides a kind of transfer learning device based on deep learning pre-training model, passes through transfer learning
Pre-training model in deep learning is adjusted to obtain new model, re-test mould by training module according to the data of frontier
Block assesses the new model of acquisition, so that the new model for meeting application that determining module determines can also be applied easily
In the different problems in same field.
Fig. 4 is the schematic diagram for the terminal device that one embodiment of the invention provides.As shown in figure 4, the terminal of the embodiment is set
Standby 4 include: processor 401, memory 402 and are stored in the memory 402 and can run on the processor 401
Computer program 403, such as the transfer learning program based on deep learning pre-training model.The processor 401 executes institute
The step in the above-mentioned transfer learning embodiment of the method based on deep learning pre-training model is realized when stating computer program 403,
Such as step 101 shown in FIG. 1, to 104 or step 101 shown in Fig. 2 to step 106, the processor 401 executes described
The function of each module in above-mentioned each Installation practice, such as the function of module 301 to 304 shown in Fig. 3 are realized when computer program 403
Energy.
Illustratively, the computer program 403 can be divided into one or more modules, one or more of
Module is stored in the memory 402, and is executed by the processor 401, to complete the present invention.It is one or more of
Module can be the series of computation machine program instruction section that can complete specific function, and the instruction segment is for describing the computer
Implementation procedure of the program 403 in the device or terminal device 4 of the transfer learning based on deep learning pre-training model.
For example, the computer program 403 can be divided into division module 301, training module 302, test module 303 determines mould
Block 304, each module concrete function is as shown in figure 3, this is no longer going to repeat them.
The terminal device 4 can be the calculating such as desktop PC, notebook, palm PC and cloud server and set
It is standby.The terminal device may include, but be not limited only to, processor 401, memory 402.It will be understood by those skilled in the art that
Fig. 4 is only the example of terminal device 4, does not constitute the restriction to terminal device 4, may include more more or fewer than illustrating
Component, perhaps combines certain components or different components, for example, the terminal device can also include input-output equipment,
Network access equipment, bus etc..
Alleged processor 401 can be central processing unit (Central Processing Unit, CPU), can also be
Other general processors, digital signal processor (Digital Signal Processor, DSP), specific integrated circuit
(Application Specific Integrated Circuit, ASIC), ready-made programmable gate array (Field-
Programmable Gate Array, FPGA) either other programmable logic device, discrete gate or transistor logic,
Discrete hardware components etc..General processor can be microprocessor or the processor is also possible to any conventional processor
Deng.
The memory 402 can be the internal storage unit of the terminal device 4, for example, terminal device 4 hard disk or
Memory.The memory 402 is also possible to the External memory equipment of the terminal device 4, such as is equipped on the terminal device 4
Plug-in type hard disk, intelligent memory card (Smart Media Card, SMC), secure digital (Secure Digital, SD) card,
Flash card (Flash Card) etc..Further, the memory 402 can also have been deposited both the inside including the terminal device 4
Storage unit also includes External memory equipment.The memory 402 is for storing the computer program and the terminal device 4
Other required programs and data.The memory 402, which can be also used for temporarily storing, have been exported or will export
Data.
It is apparent to those skilled in the art that for convenience of description and succinctly, only with above-mentioned each function
Can unit, module division progress for example, in practical application, can according to need and by above-mentioned function distribution by different
Functional unit, module are completed, i.e., the internal structure of described device is divided into different functional unit or module, more than completing
The all or part of function of description.Each functional unit in embodiment, module can integrate in one processing unit, can also
To be that each unit physically exists alone, can also be integrated in one unit with two or more units, it is above-mentioned integrated
Unit both can take the form of hardware realization, can also realize in the form of software functional units.In addition, each function list
Member, the specific name of module are also only for convenience of distinguishing each other, the protection scope being not intended to limit this application.Above system
The specific work process of middle unit, module, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.
In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, is not described in detail or remembers in some embodiment
The part of load may refer to the associated description of other embodiments.
Those of ordinary skill in the art may be aware that list described in conjunction with the examples disclosed in the embodiments of the present disclosure
Member and algorithm steps can be realized with the combination of electronic hardware or computer software and electronic hardware.These functions are actually
It is implemented in hardware or software, the specific application and design constraint depending on technical solution.Professional technician
Each specific application can be used different methods to achieve the described function, but this realization is it is not considered that exceed
The scope of the present invention.
In embodiment provided by the present invention, it should be understood that disclosed device/terminal device and method, it can be with
It realizes by another way.For example, device described above/terminal device embodiment is only schematical, for example, institute
The division of module or unit is stated, only a kind of logical function partition, there may be another division manner in actual implementation, such as
Multiple units or components can be combined or can be integrated into another system, or some features can be ignored or not executed.Separately
A bit, shown or discussed mutual coupling or direct-coupling or communication connection can be through some interfaces, device
Or the INDIRECT COUPLING or communication connection of unit, it can be electrical property, mechanical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme
's.
It, can also be in addition, the functional units in various embodiments of the present invention may be integrated into one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also realize in the form of software functional units.
If the integrated module/unit be realized in the form of SFU software functional unit and as independent product sale or
In use, can store in a computer readable storage medium.Based on this understanding, the present invention realizes above-mentioned implementation
All or part of the process in example method, can also instruct relevant hardware to complete, the meter by computer program
Calculation machine program can be stored in a computer readable storage medium, the computer program when being executed by processor, it can be achieved that on
The step of stating each embodiment of the method.Wherein, the computer program includes computer program code, the computer program
Code can be source code form, object identification code form, executable file or certain intermediate forms etc..Computer-readable Jie
Matter may include: can carry the computer program code any entity or device, recording medium, USB flash disk, mobile hard disk,
Magnetic disk, CD, computer storage, read-only memory (ROM, Read-Only Memory), random access memory (RAM,
Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium etc..It should be noted that described
The content that computer-readable medium includes can carry out increasing appropriate according to the requirement made laws in jurisdiction with patent practice
Subtract, such as does not include electric carrier signal and electricity according to legislation and patent practice, computer-readable medium in certain jurisdictions
Believe signal.
Embodiment described above is merely illustrative of the technical solution of the present invention, rather than its limitations;Although referring to aforementioned reality
Applying example, invention is explained in detail, those skilled in the art should understand that: it still can be to aforementioned each
Technical solution documented by embodiment is modified or equivalent replacement of some of the technical features;And these are modified
Or replacement, the spirit and scope for technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution should all
It is included within protection scope of the present invention.
Claims (10)
1. a kind of transfer learning method based on deep learning pre-training model characterized by comprising
Data set is divided into training dataset and test data set;
New model is obtained to the pre-training model re -training of acquisition according to the data that the training data is concentrated;
According to the data that the test data is concentrated, detection is carried out to the Generalization Capability of the new model and obtains testing result;
When the testing result reaches pre-set level value, it is determined that the new model is the model for meeting application.
2. the transfer learning method as described in claim 1 based on deep learning pre-training model, which is characterized in that the instruction
Practicing data set and the test data set is that two data distributions are consistent and the data acquisition system of mutual exclusion.
3. the transfer learning method as claimed in claim 2 based on deep learning pre-training model, which is characterized in that the survey
Examination data set carries out sampling acquisition by way of stratified sampling in the data set, removes the test number in the data set
It is the training dataset according to the data except collection.
4. the transfer learning method as claimed in claim 3 based on deep learning pre-training model, which is characterized in that the instruction
The data practiced in data set are more than the data that the test data is concentrated.
5. the transfer learning method as claimed in claim 4 based on deep learning pre-training model, which is characterized in that the instruction
The ratio section for practicing the data that the data in data set account for the data set is [2/3,4/5].
6. the transfer learning method according to any one of claims 1 to 5 based on deep learning pre-training model, feature
It is, it is described that data set is divided into training dataset and test data set, comprising:
Data set is subjected to n times random division, N group training dataset is obtained and corresponding test data set, the N is more than or equal to
1。
7. the transfer learning method as claimed in claim 6 based on deep learning pre-training model, which is characterized in that described
New model is obtained to the pre-training model re -training of acquisition according to the data that the training data is concentrated, comprising:
According to the data that the training data is concentrated, to the training weight in the level of model rear end in the pre-training model of acquisition
Re -training obtains new weight;
According to the data that the training data is concentrated, the parameter in the level of model rear end in the pre-training model is adjusted
It is whole, obtain new parameter;
According to the training weight, the new weight and institute remained unchanged in the level of model front end in the pre-training model
New parameter is stated, new model is obtained.
8. a kind of transfer learning device based on deep learning pre-training model characterized by comprising
Division module, for data set to be divided into training dataset and test data set;
Training module, the data for being concentrated according to the training data obtain new the pre-training model re -training of acquisition
Model;
Test module, the data for being concentrated according to the test data carry out detection to the Generalization Capability of the new model and obtain
Take testing result;
Determining module, for when the testing result reaches pre-set level value, it is determined that the new model is to meet application
Model.
9. a kind of terminal device, including memory, processor and storage are in the memory and can be on the processor
The computer program of operation, which is characterized in that the processor realizes such as claim 1 to 7 when executing the computer program
The step of any one the method.
10. a kind of computer readable storage medium, the computer-readable recording medium storage has computer program, and feature exists
In when the computer program is executed by processor the step of any one of such as claim 1 to 7 of realization the method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811473650.2A CN109754068A (en) | 2018-12-04 | 2018-12-04 | Transfer learning method and terminal device based on deep learning pre-training model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811473650.2A CN109754068A (en) | 2018-12-04 | 2018-12-04 | Transfer learning method and terminal device based on deep learning pre-training model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109754068A true CN109754068A (en) | 2019-05-14 |
Family
ID=66403533
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811473650.2A Pending CN109754068A (en) | 2018-12-04 | 2018-12-04 | Transfer learning method and terminal device based on deep learning pre-training model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109754068A (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110532314A (en) * | 2019-08-30 | 2019-12-03 | 国家电网有限公司 | The method and terminal device of High-Voltage Electrical Appliances quality testing |
CN110598737A (en) * | 2019-08-06 | 2019-12-20 | 深圳大学 | Online learning method, device, equipment and medium of deep learning model |
CN110688288A (en) * | 2019-09-09 | 2020-01-14 | 平安普惠企业管理有限公司 | Automatic testing method, device, equipment and storage medium based on artificial intelligence |
CN110929877A (en) * | 2019-10-18 | 2020-03-27 | 平安科技(深圳)有限公司 | Model establishing method, device, equipment and storage medium based on transfer learning |
CN111191558A (en) * | 2019-12-25 | 2020-05-22 | 深圳市优必选科技股份有限公司 | Robot and face recognition teaching method and storage medium thereof |
CN111898650A (en) * | 2020-07-08 | 2020-11-06 | 国网浙江省电力有限公司杭州供电公司 | Marketing and distribution data automatic clustering analysis equipment and method based on deep learning |
CN112434717A (en) * | 2019-08-26 | 2021-03-02 | 杭州海康威视数字技术股份有限公司 | Model training method and device |
CN112712213A (en) * | 2021-01-15 | 2021-04-27 | 上海交通大学 | Method and system for predicting energy consumption of deep migration learning of centralized air-conditioning house |
CN113094994A (en) * | 2021-04-12 | 2021-07-09 | 上海电享信息科技有限公司 | Power battery prediction method based on big data migration learning |
CN113127614A (en) * | 2020-01-16 | 2021-07-16 | 微软技术许可有限责任公司 | Providing QA training data and training QA model based on implicit relevance feedback |
CN114121161A (en) * | 2021-06-04 | 2022-03-01 | 东莞太力生物工程有限公司 | Culture medium formula development method and system based on transfer learning |
CN115114863A (en) * | 2022-08-23 | 2022-09-27 | 苏州清研精准汽车科技有限公司 | Battery pack Y capacitance prediction method and device, computer equipment and storage medium |
WO2024016637A1 (en) * | 2022-07-22 | 2024-01-25 | 中控技术股份有限公司 | Method for constructing parameter setting model and industrial process control method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108399420A (en) * | 2018-01-30 | 2018-08-14 | 北京理工雷科电子信息技术有限公司 | A kind of visible light naval vessel false-alarm elimination method based on depth convolutional network |
CN108805137A (en) * | 2018-04-17 | 2018-11-13 | 平安科技(深圳)有限公司 | Extracting method, device, computer equipment and the storage medium of livestock feature vector |
CN108875590A (en) * | 2018-05-25 | 2018-11-23 | 平安科技(深圳)有限公司 | BMI prediction technique, device, computer equipment and storage medium |
-
2018
- 2018-12-04 CN CN201811473650.2A patent/CN109754068A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108399420A (en) * | 2018-01-30 | 2018-08-14 | 北京理工雷科电子信息技术有限公司 | A kind of visible light naval vessel false-alarm elimination method based on depth convolutional network |
CN108805137A (en) * | 2018-04-17 | 2018-11-13 | 平安科技(深圳)有限公司 | Extracting method, device, computer equipment and the storage medium of livestock feature vector |
CN108875590A (en) * | 2018-05-25 | 2018-11-23 | 平安科技(深圳)有限公司 | BMI prediction technique, device, computer equipment and storage medium |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110598737A (en) * | 2019-08-06 | 2019-12-20 | 深圳大学 | Online learning method, device, equipment and medium of deep learning model |
CN112434717A (en) * | 2019-08-26 | 2021-03-02 | 杭州海康威视数字技术股份有限公司 | Model training method and device |
CN112434717B (en) * | 2019-08-26 | 2024-03-08 | 杭州海康威视数字技术股份有限公司 | Model training method and device |
CN110532314A (en) * | 2019-08-30 | 2019-12-03 | 国家电网有限公司 | The method and terminal device of High-Voltage Electrical Appliances quality testing |
CN110688288A (en) * | 2019-09-09 | 2020-01-14 | 平安普惠企业管理有限公司 | Automatic testing method, device, equipment and storage medium based on artificial intelligence |
CN110688288B (en) * | 2019-09-09 | 2023-11-07 | 新疆北斗同创信息科技有限公司 | Automatic test method, device, equipment and storage medium based on artificial intelligence |
CN110929877A (en) * | 2019-10-18 | 2020-03-27 | 平安科技(深圳)有限公司 | Model establishing method, device, equipment and storage medium based on transfer learning |
CN110929877B (en) * | 2019-10-18 | 2023-09-15 | 平安科技(深圳)有限公司 | Model building method, device, equipment and storage medium based on transfer learning |
CN111191558A (en) * | 2019-12-25 | 2020-05-22 | 深圳市优必选科技股份有限公司 | Robot and face recognition teaching method and storage medium thereof |
CN111191558B (en) * | 2019-12-25 | 2024-02-02 | 深圳市优必选科技股份有限公司 | Robot and face recognition teaching method and storage medium thereof |
CN113127614A (en) * | 2020-01-16 | 2021-07-16 | 微软技术许可有限责任公司 | Providing QA training data and training QA model based on implicit relevance feedback |
CN111898650A (en) * | 2020-07-08 | 2020-11-06 | 国网浙江省电力有限公司杭州供电公司 | Marketing and distribution data automatic clustering analysis equipment and method based on deep learning |
CN112712213A (en) * | 2021-01-15 | 2021-04-27 | 上海交通大学 | Method and system for predicting energy consumption of deep migration learning of centralized air-conditioning house |
CN112712213B (en) * | 2021-01-15 | 2023-07-04 | 上海交通大学 | Method and system for predicting deep migration learning energy consumption of concentrated air conditioning house |
CN113094994A (en) * | 2021-04-12 | 2021-07-09 | 上海电享信息科技有限公司 | Power battery prediction method based on big data migration learning |
CN114121161B (en) * | 2021-06-04 | 2022-08-05 | 深圳太力生物技术有限责任公司 | Culture medium formula development method and system based on transfer learning |
CN114121161A (en) * | 2021-06-04 | 2022-03-01 | 东莞太力生物工程有限公司 | Culture medium formula development method and system based on transfer learning |
WO2024016637A1 (en) * | 2022-07-22 | 2024-01-25 | 中控技术股份有限公司 | Method for constructing parameter setting model and industrial process control method |
CN115114863A (en) * | 2022-08-23 | 2022-09-27 | 苏州清研精准汽车科技有限公司 | Battery pack Y capacitance prediction method and device, computer equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109754068A (en) | Transfer learning method and terminal device based on deep learning pre-training model | |
CN112163465B (en) | Fine-grained image classification method, fine-grained image classification system, computer equipment and storage medium | |
US20190130266A1 (en) | System and method for improved neural network training | |
CN110968701A (en) | Relationship map establishing method, device and equipment for graph neural network | |
CN110147456A (en) | A kind of image classification method, device, readable storage medium storing program for executing and terminal device | |
CN109919316A (en) | The method, apparatus and equipment and storage medium of acquisition network representation study vector | |
CN107480789A (en) | The efficient conversion method and device of a kind of deep learning model | |
CN106951825A (en) | A kind of quality of human face image assessment system and implementation method | |
CN107871166A (en) | For the characteristic processing method and characteristics processing system of machine learning | |
US11423307B2 (en) | Taxonomy construction via graph-based cross-domain knowledge transfer | |
Yang et al. | Associative memory optimized method on deep neural networks for image classification | |
CN110288007A (en) | The method, apparatus and electronic equipment of data mark | |
CN110399487A (en) | A kind of file classification method, device, electronic equipment and storage medium | |
CN110008961A (en) | Text real-time identification method, device, computer equipment and storage medium | |
CN106682702A (en) | Deep learning method and system | |
CN106778910A (en) | Deep learning system and method based on local training | |
CN111523324A (en) | Training method and device for named entity recognition model | |
Liu et al. | Research of animals image semantic segmentation based on deep learning | |
CN108228684A (en) | Training method, device, electronic equipment and the computer storage media of Clustering Model | |
CN113761250A (en) | Model training method, merchant classification method and device | |
Zhou et al. | Classroom learning status assessment based on deep learning | |
CN109214444B (en) | Game anti-addiction determination system and method based on twin neural network and GMM | |
CN108537270A (en) | Image labeling method, terminal device and storage medium based on multi-tag study | |
CN109670927A (en) | The method of adjustment and its device of credit line, equipment, storage medium | |
CN112035671A (en) | State detection method and device, computer equipment and storage medium |
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 |