CN110503198A - Obtain method, apparatus, equipment and the storage medium of neural network test report - Google Patents
Obtain method, apparatus, equipment and the storage medium of neural network test report Download PDFInfo
- Publication number
- CN110503198A CN110503198A CN201910666325.6A CN201910666325A CN110503198A CN 110503198 A CN110503198 A CN 110503198A CN 201910666325 A CN201910666325 A CN 201910666325A CN 110503198 A CN110503198 A CN 110503198A
- Authority
- CN
- China
- Prior art keywords
- characteristic
- network
- parameter
- nervus opticus
- dimensionality reduction
- 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
- 238000000034 method Methods 0.000 title claims abstract description 70
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 54
- 238000012360 testing method Methods 0.000 title claims abstract description 41
- 230000009467 reduction Effects 0.000 claims abstract description 74
- 238000012549 training Methods 0.000 claims abstract description 66
- 210000005036 nerve Anatomy 0.000 claims abstract description 41
- 230000006870 function Effects 0.000 claims description 46
- 238000012545 processing Methods 0.000 claims description 24
- 230000008569 process Effects 0.000 claims description 23
- 230000004913 activation Effects 0.000 claims description 12
- 210000002569 neuron Anatomy 0.000 claims description 12
- 241000208340 Araliaceae Species 0.000 claims description 11
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 11
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 11
- 235000008434 ginseng Nutrition 0.000 claims description 11
- 238000010606 normalization Methods 0.000 claims description 11
- 238000012552 review Methods 0.000 claims description 10
- 238000006243 chemical reaction Methods 0.000 claims description 5
- 230000009466 transformation Effects 0.000 claims description 5
- 230000001537 neural effect Effects 0.000 claims description 3
- 238000003062 neural network model Methods 0.000 abstract description 15
- 238000001514 detection method Methods 0.000 abstract description 9
- 238000012544 monitoring process Methods 0.000 abstract description 5
- 238000004590 computer program Methods 0.000 description 5
- 230000000007 visual effect Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000009434 installation Methods 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 210000004218 nerve net Anatomy 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 239000002699 waste material Substances 0.000 description 2
- 230000008901 benefit Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 210000004027 cell Anatomy 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005538 encapsulation Methods 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 230000007257 malfunction Effects 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 230000003362 replicative effect Effects 0.000 description 1
- 230000000153 supplemental effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
This application involves page intelligent decision field, method, apparatus, equipment and the storage medium for obtaining neural network test report are provided, method includes: acquisition characteristic.The characteristic and feature are input to first nerves network, by the cost function training first nerves network model of first nerves network, to obtain model parameter w.Dimensionality reduction is carried out to the characteristic according to w, to obtain the characteristic after dimensionality reduction.Characteristic after the dimensionality reduction is input to nervus opticus network.Obtain the variable condition information and execution state information of the nervus opticus network parameter.The variable condition information and the execution state information are aggregated into log, test report is generated according to the log.The monitoring report of generation preferably monitors the variation of parameter and the link of detection training mistake and obtains the neural network model hyper parameter for meeting demand preferably according to the hyper parameter of the variation adjustment neural network model of parameter.
Description
Technical field
This application involves intelligent decision field, provides the method, apparatus for obtaining neural network test report, equipment and deposit
Storage media
Background technique
It is protected in loan field in letter, different letter guarantor loan platforms can having differences property and record in the data of preservation
There may be abnormal or missings for data.So causing the data of mode input to exist when data being caused to be trained model
Many abnormal data, so that the link to malfunction when neural network model training is relatively more, the test report matter that model is provided
Measure it is lower so that researcher can not know the error of which link and remove the hyper parameter of adjustment model according to test report, into
And researcher to waste a large amount of time go adjustment hyper parameter, so as to cause the waste of resource.Hyper parameter is to start to learn
The parameter of setting value before journey, rather than the supplemental characteristic obtained by training, such as learning rate (a variety of moulds of neural network
Formula), deep-neural-network hides the number of clusters in the number of plies and k mean cluster.
Summary of the invention
This application provides a kind of methods of page development function by encapsulation, are able to solve the page in the prior art and open
Hair efficiency it is too low, the problem of fast ripe iteration can not be done to version.
In a first aspect, the application provides a kind of acquisition neural network test report method, comprising:
Obtain characteristic.
The characteristic is defeated, the first nerves network model is trained by the cost function of first nerves network,
To obtain model parameter w.The feature refers at least one information in characteristic, the cost letter of the first nerves network
Number are as follows:Wherein y(i)Refer to the feature, w refers to first mind
Model parameter through network, x(i)Referring to the characteristic, λ is positive number, | | w | |1The L1 norm of expression parameter w.
Dimensionality reduction is carried out to the characteristic according to model parameter w, to obtain the characteristic after dimensionality reduction.
Characteristic after the dimensionality reduction is input to nervus opticus network.
The variable condition information of the nervus opticus network parameter in the nervus opticus network training process is obtained, with
And the execution state information for obtaining the nervus opticus network training, entering ginseng, going out ginseng and calculating process.
The variable condition information and the execution state information are aggregated into log, is generated and is tested according to the log
Report.
Compared to the prior art, in scheme provided by the present application, by pre-processing number unrelated with neural network model
According to the error link rejected and detection neural network model is trained, complete to independent parameter in neural network model training process
Detection and reduce training parameter.Due to the reduction of training parameter and the detection of training link, monitoring report can be more
The link of the good variation for monitoring parameter and detection training mistake, preferably can adjust nerve net according to the variation of parameter
The hyper parameter of network model obtains the neural network model for more meeting demand.
In some possible designs, after the generation test report according to the log, the method also includes:
Whether succeeded according to nervus opticus network training described in the report review.
Whether the nervus opticus network training according to the report review succeeds, comprising:
Whether there is mistake according to the execution state information training of judgement process, if there is not mistake, described in judgement
The success of nervus opticus network training.
Alternatively, determining the range of parameter according to the upper limit and lower limit of the nervus opticus network parameter;The parameter is extremely
It less include the value of learning rate η, accuracy rate and loss function;
Nervus opticus network parameter is judged according to the variable condition information and the nervus opticus network parameter range
Whether fall in the parameter area, if nervus opticus network parameter falls in the parameter area, judges the nervus opticus
Network training success.
Alternatively, judging whether the value of the loss function of nervus opticus network is higher than threshold value, institute is judged if not being higher than threshold value
State the success of nervus opticus network training.The loss function refers to, wherein described refer to loss function, after referring to the dimensionality reduction
Characteristic is input to the desired output of the nervus opticus network, and the characteristic after referring to the dimensionality reduction is input to described
The reality output of two neural networks.
It is described that dimensionality reduction is carried out to the characteristic according to model parameter w in some possible designs, to obtain dimensionality reduction
After characteristic afterwards, the method also includes:
Normalized is done to the characteristic after the dimensionality reduction.The normalization refers to the characteristic after the dimensionality reduction
According to linear transformation is carried out, the characteristic after making dimensionality reduction is mapped between [0,1].The normalization is held by following mathematical formulae
Row:
xi*=(xi-mini)/(maxi-mini)
Wherein maxiThe maximum value of the ith feature of characteristic after referring to the dimensionality reduction, miniAfter referring to the dimensionality reduction
Characteristic ith feature minimum value, xiThe arbitrary number of the ith feature of characteristic after referring to the dimensionality reduction
According to xi* the corresponding data of characteristic ith feature feature after dimensionality reduction after referring to conversion.
In some possible designs, before the acquisition characteristic, the method also includes:
Handle the missing values of the characteristic.
In some possible designs, the missing values of the processing characteristic include at least following implementations it
One:
The missing values are filled by average value, mode and median.
Alternatively, being recorded the missing values as a kind of state.
Alternatively, the record of missing values is deleted.
In some possible designs, after the acquisition characteristic, the method also includes:
It does feature to the characteristic to derive, to obtain new feature.The feature derivative refers to the characteristic
Derived according to the stability bandwidth for counting, sum, seeking ratio, doing the time difference and find the characteristic was passed through.
In some possible designs, which is characterized in that the first nerves network and the nervus opticus network are logical
Cross the output that activation primitive carries out map neural network;The mathematics form of expression of the activation primitive is as follows:
Wherein y refers to the output of the neuron of the first nerves network and nervus opticus network, and x refers to described first
The input of the neuron of neural network and nervus opticus network, a are non-zero constant.
Second aspect, the application provide a kind of acquisition neural network test report device, have and realize and correspond to above-mentioned the
On the one hand the function of the method for the acquisition neural network test report provided.The function can be by hardware realization, can also be with
Corresponding software realization is executed by hardware.Hardware or software include one or more modules corresponding with above-mentioned function, institute
Stating module can be software and/or hardware.
The acquisition neural network test report device includes:
Input/output module module, for obtaining characteristic.
Processing module passes through the generation of first nerves network for that the characteristic will be input to first nerves network
The valence function training first nerves network model, to obtain model parameter w.The feature refers at least one in characteristic
Item information, the cost function of the first nerves network are as follows:Wherein y(i)Refer to the feature, w refers to the model parameter of the first nerves network, x(i)Referring to the characteristic, λ is positive number, | |
w||1The L1 norm of expression parameter w.Dimensionality reduction is carried out to the characteristic according to model parameter w, to obtain the feature after dimensionality reduction
Data.Characteristic after the dimensionality reduction is input to nervus opticus network.It obtains in the nervus opticus network training process
The nervus opticus network parameter variable condition information, and obtain the nervus opticus network training, enter ginseng, go out ginseng with
And the execution state information of calculating process.The variable condition information and the execution state information are aggregated into log, root
Test report is generated according to the log.
In some possible designs, the processing module is also used to:
Whether succeeded according to nervus opticus network training described in the report review.
Whether the nervus opticus network training according to the report review succeeds, comprising:
Whether there is mistake according to the execution state information training of judgement process, if there is not mistake, described in judgement
The success of nervus opticus network training.
Alternatively, determining the range of parameter according to the upper limit and lower limit of the nervus opticus network parameter;The parameter is extremely
It less include the value of learning rate η, accuracy rate and loss function;
Nervus opticus network parameter is judged according to the variable condition information and the nervus opticus network parameter range
Whether fall in the parameter area, if nervus opticus network parameter falls in the parameter area, judges the nervus opticus
Network training success.
Alternatively, judging whether the value of the loss function of nervus opticus network is higher than threshold value, institute is judged if not being higher than threshold value
State the success of nervus opticus network training.The loss function refers to, wherein described refer to loss function, after referring to the dimensionality reduction
Characteristic is input to the desired output of the nervus opticus network, and the characteristic after referring to the dimensionality reduction is input to described
The reality output of two neural networks.
In some possible designs, the processing module is also used to:
Normalized is done to the characteristic after the dimensionality reduction.The normalization refers to the characteristic after the dimensionality reduction
According to linear transformation is carried out, the characteristic after making dimensionality reduction is mapped between [0,1].The normalization is held by following mathematical formulae
Row:
xi*=(xi-mini)/(maxi-mini)
Wherein maxiThe maximum value of the ith feature of characteristic after referring to the dimensionality reduction, miniAfter referring to the dimensionality reduction
Characteristic ith feature minimum value, xiThe arbitrary number of the ith feature of characteristic after referring to the dimensionality reduction
According to xi* the corresponding data of characteristic ith feature feature after dimensionality reduction after referring to conversion.
In some possible designs, the processing module is also used to:
Handle the missing values of the characteristic.
In some possible designs, the processing module is also used to:
The missing values are filled by average value, mode and median.
Alternatively, being recorded the missing values as a kind of state.
Alternatively, the record of missing values is deleted.
In some possible designs, the processing module is also used to:
It does feature to the characteristic to derive, to obtain new feature.The feature derivative refers to the characteristic
Derived according to the stability bandwidth for counting, sum, seeking ratio, doing the time difference and find the characteristic was passed through.
In some possible designs, the processing module is also used to:
The first nerves network and the nervus opticus network pass through the defeated of activation primitive progress map neural network
Out;The mathematics form of expression of the activation primitive is as follows:
Wherein y refers to the output of the neuron of the first nerves network and nervus opticus network, and x refers to described first
The input of the neuron of neural network and nervus opticus network, a are non-zero constant.
The another aspect of the application provides a kind of equipment for obtaining neural network test report comprising at least one connection
Processor, memory, input-output unit, wherein the memory is for storing program code, and the processor is for adjusting
Method described in above-mentioned various aspects is executed with the program code in the memory.
The another aspect of the application provides a kind of computer storage medium comprising instruction, when it runs on computers
When, so that computer executes method described in above-mentioned various aspects.
Detailed description of the invention
Fig. 1 is a kind of flow diagram that neural network test report method is obtained in the embodiment of the present application;
Fig. 2 is the flow diagram that the embodiment of the present application obtains neural network test report;
Fig. 3 is the functional block diagram that the embodiment of the present application obtains neural network test report.
The embodiments will be further described with reference to the accompanying drawings for realization, functional characteristics and the advantage of the application purpose.
Specific embodiment
It should be appreciated that specific embodiment described herein is not used to limit the application only to explain the application.This
The specification and claims of application and term " first " in above-mentioned attached drawing, " second " etc. are for distinguishing similar right
As without being used to describe a particular order or precedence order.It should be understood that the data used in this way in the appropriate case can be with
It exchanges, so that the embodiments described herein can be implemented with the sequence other than the content for illustrating or describing herein.In addition,
Term " includes " and " having " and their any deformation, it is intended that cover it is non-exclusive include, for example, containing a system
The process, method, system, product or equipment of column step or module those of are not necessarily limited to be clearly listed step or module, and
Being may include other steps or module being not clearly listed or intrinsic for these process, methods, product or equipment, this
The division of module appeared in application, only a kind of division in logic can have other when realizing in practical application
Division mode, such as multiple modules can be combined into or are integrated in another system, or some features can be ignored, or not held
Row.
Fig. 1 is please referred to, a kind of method for providing acquisition neural network test report to the application below is illustrated,
The described method includes:
101, characteristic is obtained.
The characteristic includes at least user's gender, user's monthly income, user month expenditure, user's reference record, user
Address and user month premium.
The characteristic can be protected by cell formats, xml format, propretis format and yml format
It deposits.The characteristic can be obtained by network, can also be obtained by database.Such as network acquisition can pass through
The crawler of python is obtained.
102, the characteristic is input to first nerves network, passes through the cost function training institute of first nerves network
First nerves network model is stated, to obtain model parameter w.
The feature refers at least one information in characteristic, the cost function of the first nerves network are as follows:Wherein y(i)Refer to the feature, w refers to the first nerves network
Model parameter, x(i)Referring to the characteristic, λ is positive number, | | w | |1The L1 norm of expression parameter w.
The neural network refers to that developer establishes the neural network model refunded after prediction client borrows.Wherein, neural
Network refers to a kind of method for replicating this intensive neuroid.By the multiple data flows of single treatment, computer can
Time needed for substantially reducing processing data.This technology is applied to deep learning and has produced artificial neural network.This
A little artificial neural networks are made of input node, output node and node layer.
Input node, input node for receiving data.
Output node, for exporting result data.
Node layer, for the data inputted from input node to be converted to the content that output node can be used.Node layer
Refer to multiple concealed nodes between input node and output node, node layer can also become hidden layer.When data pass through
When these concealed nodes advance, neural network determines to pass data to next concealed nodes using logic.
103, dimensionality reduction is carried out to the characteristic according to model parameter w, to obtain the characteristic after dimensionality reduction.
By w 0 can will be set to the weight for the characteristic number for not generating effect in training process.Wherein client gender is believed
Breath does not generate any effect in realistic model training it is necessary to mention client gender information deletion or to other information
It takes, to accelerate the training speed of AI model.
104, the characteristic after the dimensionality reduction is input to nervus opticus network.
The nervus opticus network can be any neural network model.Such as to can be K equal for the neural network model
It is worth clustering algorithm (K-means), support vector machines (Support Vector Machine, SVM) and linear regression.It is described poly-
Class is one by similar data member carries out the process of taxonomic organization in some aspects in data set, and cluster is exactly a kind of discovery
The technology of this immanent structure, clustering technique are commonly referred to as unsupervised learning.
105, the variable condition letter of the nervus opticus network parameter in the nervus opticus network training process is obtained
Breath, and the execution state information for obtaining the nervus opticus network training, entering ginseng, going out ginseng and calculating process.
Such as 2 the 0th of system number represent the execution state for passing the 1st layer of neural network.After having executed whole flow process, return
One 32 2 system numbers execute state, if the 0th is 1, the 1st layer of neural network as Indexs measure neural network node
It runs succeeded, if the 0th is 0, executes failure, gone wrong neural network node is judged by the method.In another example
When passing through neural metwork training, the parameter for extracting setting node layer carrys out the training condition of feedback node layer as index.
106, the variable condition information and the execution state information are aggregated into log, are generated according to the log
Test report.
The test report is counted and is shown the information of test report by Visual Chart.The Visual Chart is at least
Including bar chart, histogram, line chart and pie chart.The Visual Chart is selected by selection Visual Chart type,
It can save and show any one or multinomial Visual Chart.
In some embodiments, after the generation test report according to the log, the method also includes:
Whether succeeded according to nervus opticus network training described in the report review.
Whether the nervus opticus network training according to the report review succeeds, comprising:
Whether there is mistake according to the execution state information training of judgement process, if there is not mistake, described in judgement
The success of nervus opticus network training.
Alternatively, determining the range of parameter according to the upper limit and lower limit of the nervus opticus network parameter;The parameter is extremely
It less include the value of learning rate η, accuracy rate and loss function;
Nervus opticus network parameter is judged according to the variable condition information and the nervus opticus network parameter range
Whether fall in the parameter area, if nervus opticus network parameter falls in the parameter area, judges the nervus opticus
Network training success.
Alternatively, judging whether the value of the loss function of nervus opticus network is higher than threshold value, institute is judged if not being higher than threshold value
State the success of nervus opticus network training.The loss function refers to, wherein described refer to loss function, after referring to the dimensionality reduction
Characteristic is input to the desired output of the nervus opticus network, and the characteristic after referring to the dimensionality reduction is input to described
The reality output of two neural networks.
In above embodiment, the accuracy of the model under hyper parameter different characteristic combination is extracted, according to summarizing out
The test report come, selects optimal feature to combine.
It is described that dimensionality reduction is carried out to the characteristic according to model parameter w in some embodiments, after obtaining dimensionality reduction
After characteristic, the method also includes:
Normalized is done to the characteristic after the dimensionality reduction.The normalization refers to the characteristic after the dimensionality reduction
According to linear transformation is carried out, the characteristic after making dimensionality reduction is mapped between [0,1].The normalization is held by following mathematical formulae
Row:
xi*=(xi-mini)/(maxi-mini)
Wherein maxiThe maximum value of the ith feature of characteristic after referring to the dimensionality reduction, miniAfter referring to the dimensionality reduction
Characteristic ith feature minimum value, xiThe arbitrary number of the ith feature of characteristic after referring to the dimensionality reduction
According to xi* the corresponding data of characteristic ith feature feature after dimensionality reduction after referring to conversion.
In above embodiment, the training of neural network can be accelerated by data normalization and increase neural network
Accuracy.
In some embodiments, before the acquisition characteristic, the method also includes:
Handle the missing values of the characteristic.
In above embodiment, by the missing values of processing feature data, missing values is allowed to be loaded into neural network
Model.
In some embodiments, the missing values of the processing characteristic include at least one of following implementations:
The missing values are filled by average value, mode and median.
Alternatively, being recorded the missing values as a kind of state.
Alternatively, the record of missing values is deleted.
In above embodiment, missing values can be filled by above-mentioned.
In some embodiments, after the acquisition characteristic, the method also includes:
It does feature to the characteristic to derive, to obtain new feature.The feature derivative refers to the characteristic
Derived according to the stability bandwidth for counting, sum, seeking ratio, doing the time difference and find the characteristic was passed through.
In above embodiment, the feature is total including at least the number of the loan of preset time, the consumption of preset time
The accounting of volume, loan application amount and preset time.
In some embodiments, the first nerves network and the nervus opticus network are reflected by activation primitive
Penetrate the output of neural network;The mathematics form of expression of the activation primitive is as follows:
Wherein y refers to the output of the neuron of the first nerves network and nervus opticus network, and x refers to described first
The input of the neuron of neural network and nervus opticus network, a are non-zero constant.
In above embodiment, the activation primitive (Activation Function), exactly in artificial neural network
The function run on neuron is responsible for the input of neuron being mapped to output end.
Compared to the prior art, in scheme provided by the present application, by pre-processing number unrelated with neural network model
According to the error link rejected and detection neural network model is trained, complete to independent parameter in neural network model training process
Detection and reduce training parameter.Due to the reduction of training parameter and the detection of training link, monitoring report can be more
The link of the good variation for monitoring parameter and detection training mistake, preferably can adjust nerve net according to the variation of parameter
The hyper parameter of network model obtains the neural network model for more meeting demand.
A kind of structural schematic diagram of the device 20 of acquisition neural network test report as shown in Figure 2, can be applied to obtain
Take neural network test report.The device of acquisition neural network test report in the embodiment of the present application can be realized corresponding to upper
The step of stating the method for acquisition neural network test report performed in embodiment corresponding to Fig. 1-1.Obtain neural network
The function that the device 20 of test report is realized can also execute corresponding software realization by hardware realization by hardware.
Hardware or software include one or more modules corresponding with above-mentioned function, and the module can be software and/or hardware.Institute
Stating and obtaining the device of neural network test report may include input/output module 201 and processing module 202, the processing module
202 and the function of input/output module 201 realize and can refer in embodiment corresponding to Fig. 1-1 performed operation, herein not
It repeats.Input/output module 201 can be used for controlling the input, output and acquisition operation of the input/output module 201.
In some embodiments, input/output module module, for obtaining characteristic.
Processing module, for passing through the cost function of first nerves network for the characteristic to first nerves network
The training first nerves network model, to obtain model parameter w.The feature refers at least one letter in characteristic
Breath, the cost function of the first nerves network are as follows:Wherein y(i)It is
Refer to the feature, w refers to the model parameter of the first nerves network, x(i)Referring to the characteristic, λ is positive number, | | w | |1
The L1 norm of expression parameter w.Dimensionality reduction is carried out to the characteristic according to model parameter w, to obtain the characteristic after dimensionality reduction.
Characteristic after the dimensionality reduction is input to nervus opticus network.It obtains described in the nervus opticus network training process
The variable condition information of nervus opticus network parameter, and obtain the nervus opticus network training, enter ginseng, go out to join and calculate
The execution state information of process.The variable condition information and the execution state information are aggregated into log, according to described
Log generates test report.
In some possible designs, the processing module 202 is also used to:
Whether succeeded according to nervus opticus network training described in the report review.
Whether the nervus opticus network training according to the report review succeeds, comprising:
Whether there is mistake according to the execution state information training of judgement process, if there is not mistake, described in judgement
The success of nervus opticus network training.
Alternatively, determining the range of parameter according to the upper limit and lower limit of the nervus opticus network parameter;The parameter is extremely
It less include the value of learning rate η, accuracy rate and loss function;
Nervus opticus network parameter is judged according to the variable condition information and the nervus opticus network parameter range
Whether fall in the parameter area, if nervus opticus network parameter falls in the parameter area, judges the nervus opticus
Network training success.
Alternatively, judging whether the value of the loss function of nervus opticus network is higher than threshold value, institute is judged if not being higher than threshold value
State the success of nervus opticus network training.The loss function refers to, wherein described refer to loss function, after referring to the dimensionality reduction
Characteristic is input to the desired output of the nervus opticus network, and the characteristic after referring to the dimensionality reduction is input to described
The reality output of two neural networks.
In some possible designs, the processing module 202 is also used to:
Normalized is done to the characteristic after the dimensionality reduction.The normalization refers to the characteristic after the dimensionality reduction
According to linear transformation is carried out, the characteristic after making dimensionality reduction is mapped between [0,1].The normalization is held by following mathematical formulae
Row:
xi*=(xi-mini)/(maxi-mini)
Wherein maxiThe maximum value of the ith feature of characteristic after referring to the dimensionality reduction, miniAfter referring to the dimensionality reduction
Characteristic ith feature minimum value, xiThe arbitrary number of the ith feature of characteristic after referring to the dimensionality reduction
According to xi* the corresponding data of characteristic ith feature feature after dimensionality reduction after referring to conversion.
In some possible designs, the processing module 202 is also used to:
Handle the missing values of the characteristic.
In some possible designs, the processing module 202 is also used to:
The missing values are filled by average value, mode and median.
Alternatively, being recorded the missing values as a kind of state.
Alternatively, the record of missing values is deleted.
In some possible designs, the processing module 202 is also used to:
It does feature to the characteristic to derive, to obtain new feature.The feature derivative refers to the characteristic
Derived according to the stability bandwidth for counting, sum, seeking ratio, doing the time difference and find the characteristic was passed through.
In some possible designs, the processing module 202 is also used to:
The first nerves network and the nervus opticus network pass through the defeated of activation primitive progress map neural network
Out;The mathematics form of expression of the activation primitive is as follows:
Wherein y refers to the output of the neuron of the first nerves network and nervus opticus network, and x refers to described first
The input of the neuron of neural network and nervus opticus network, a are non-zero constant.
The creating device in the embodiment of the present application is described respectively from the angle of modular functionality entity above, below from hard
Part angle introduces a kind of computer equipment, as shown in figure 3, comprising: processor, memory, input-output unit (are also possible to
Transceiver does not identify in Fig. 3) and storage is in the memory and the computer journey that can run on the processor
Sequence.For example, the computer program can be corresponding to obtain the method for neural network test report in embodiment corresponding to Fig. 1
Program.For example, when computer equipment realizes the function of the device 20 of acquisition neural network test report as shown in Figure 2, institute
It states and is realized when processor executes the computer program in embodiment corresponding to above-mentioned Fig. 2 by acquisition neural network test report
Device 20 execute acquisition neural network test report method in each step.Alternatively, the processor executes the meter
The function of each module in the device 20 of the acquisition neural network test report of embodiment corresponding to above-mentioned Fig. 2 is realized when calculation machine program
Energy.In another example the computer program can be corresponding to obtain the method for neural network test report in embodiment corresponding to Fig. 1
Program.
Alleged processor can be central processing unit (Central Processing Unit, CPU), can also be it
His general processor, 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 processor is the control centre of the computer installation, utilizes various interfaces and the entire computer installation of connection
Various pieces.
The memory can be used for storing the computer program and/or module, and the processor is by operation or executes
Computer program in the memory and/or module are stored, and calls the data being stored in memory, described in realization
The various functions of computer installation.The memory can mainly include storing program area and storage data area, wherein storage program
It area can application program (such as sound-playing function, image player function etc.) needed for storage program area, at least one function
Deng.Storage data area, which can be stored, uses created data (such as audio data, video data etc.) etc. according to mobile phone.This
Outside, memory may include high-speed random access memory, can also include nonvolatile memory, such as hard disk, memory, insert
Connect formula hard disk, intelligent memory card (Smart Media Card, SMC), secure digital (Secure Digital, SD) card, flash memory
Block (Flash Card), at least one disk memory, flush memory device or other volatile solid-state parts.
The input-output unit can also be replaced with receiver and transmitter, can be real for same or different physics
Body.When for identical physical entity, input-output unit may be collectively referred to as.The input and output can be transceiver.
The memory can integrate in the processor, can also be provided separately with the processor.
Through the above description of the embodiments, those skilled in the art can be understood that above-described embodiment side
Method can be realized by means of software and necessary general hardware platform, naturally it is also possible to by hardware, but in many cases
The former is more preferably embodiment.Based on this understanding, the technical solution of the application substantially in other words does the prior art
The part contributed out can be embodied in the form of software products, which is stored in a storage medium
In (such as ROM/RAM), including some instructions are used so that a terminal (can be mobile phone, computer, server or network are set
It is standby etc.) execute method described in each embodiment of the application.
Embodiments herein is described above in conjunction with attached drawing, but the application be not limited to it is above-mentioned specific
Embodiment, the above mentioned embodiment is only schematical, rather than restrictive, those skilled in the art
Under the enlightenment of the application, when not departing from the application objective and scope of the claimed protection, can also it make very much
Form, it is all using equivalent structure or equivalent flow shift made by present specification and accompanying drawing content, directly or indirectly
Other related technical areas are used in, these are belonged within the protection of the application.
Claims (10)
1. a kind of method for obtaining neural network test report, which is characterized in that the described method includes:
Obtain characteristic;
The characteristic is input to first nerves network, passes through the cost function training of first nerves network first mind
Through network model, to obtain model parameter w;, the cost function of the first nerves network are as follows:
Wherein y(i)Refer to the feature, w refers to the first nerves
The model parameter of network, x(i)Referring to the characteristic, λ is positive number, | | w | |1The L1 norm of expression parameter w;
Dimensionality reduction is carried out to the characteristic according to model parameter w, to obtain the characteristic after dimensionality reduction;
Characteristic after the dimensionality reduction is input to nervus opticus network;
The variable condition information of the nervus opticus network parameter in the nervus opticus network training process is obtained, and is obtained
The execution state information for taking the nervus opticus network training, entering ginseng, going out ginseng and calculating process;
The variable condition information and the execution state information are aggregated into log, test report is generated according to the log
It accuses.
2. the method according to claim 1, wherein it is described according to the log generate test report after, institute
State method further include:
Whether succeeded according to nervus opticus network training described in the report review;
Whether the nervus opticus network training according to the report review succeeds, comprising:
Whether there is mistake according to the execution state information training of judgement process, if there is not mistake, judges described second
Neural metwork training success;
Alternatively, determining the range of parameter according to the upper limit and lower limit of the nervus opticus network parameter;The parameter is at least wrapped
Include the value of learning rate η, accuracy rate and loss function;
Whether nervus opticus network parameter is judged according to the variable condition information and the nervus opticus network parameter range
It falls in the parameter area, if nervus opticus network parameter falls in the parameter area, judges the nervus opticus network
It trains successfully;
Alternatively, judge whether the value of the loss function of nervus opticus network is higher than threshold value, described the is judged if not being higher than threshold value
The success of two neural metwork trainings;The loss function refers to L [(Y, f (X)]=[Y-f (X)]2, wherein the L [(Y, f (X)] be
Refer to that loss function, Y refer to that the characteristic after the dimensionality reduction is input to the desired output of the nervus opticus network, f (X) refers to
Characteristic after the dimensionality reduction is input to the reality output of the nervus opticus network.
3. the method according to claim 1, wherein described carry out the characteristic according to model parameter w
Dimensionality reduction, after obtaining the characteristic after dimensionality reduction, the method also includes:
Normalized is done to the characteristic after the dimensionality reduction;It is described normalization refer to the characteristic after the dimensionality reduction into
Row linear transformation, the characteristic after making dimensionality reduction are mapped between [0,1];The normalization is executed by following mathematical formulae:
xi*=(xi-mini)/(maxi-mini)
Wherein maxiThe maximum value of the ith feature of characteristic after referring to the dimensionality reduction, miniSpy after referring to the dimensionality reduction
Levy the minimum value of the ith feature of data, xiThe arbitrary data of the ith feature of characteristic after referring to the dimensionality reduction, xi*
The corresponding data of characteristic ith feature after the dimensionality reduction after referring to conversion.
4. the method according to claim 1, wherein before the acquisition characteristic, the method also includes:
Handle the missing values of the characteristic.
5. according to the method described in claim 4, it is characterized in that, the missing values of the processing characteristic include at least
One of following implementations:
The missing values are filled by average value, mode and median;
Alternatively, being recorded the missing values as a kind of state;
Alternatively, the record of missing values is deleted.
6. the method according to claim 1, wherein after the acquisition characteristic, the method also includes:
It does feature to the characteristic to derive, to obtain new feature;The feature derivative refers to logical with the characteristic
The stability bandwidth count, sum, seeking ratio, doing the time difference and find the characteristic was crossed to be derived.
7. the method according to claim 1, wherein the first nerves network and the nervus opticus network
The output of map neural network is carried out by activation primitive;The mathematics form of expression of the activation primitive is as follows:
Wherein y refers to the output of the neuron of the first nerves network and nervus opticus network, and x refers to the first nerves
The input of the neuron of network and nervus opticus network, a are non-zero constant.
8. a kind of device for obtaining neural network test report, which is characterized in that described device includes:
Input/output module obtains characteristic;
The characteristic is input to first nerves network by processing module, passes through the cost function training of first nerves network
The first nerves network model, to obtain model parameter w;The feature refers at least one information in characteristic, institute
State the cost function of first nerves network are as follows:Wherein y(i)Refer to described
Feature, w refer to the model parameter of the first nerves network, x(i)Referring to the characteristic, λ is positive number, | | w | |1Indicate ginseng
The L1 norm of number w;Dimensionality reduction is carried out to the characteristic according to model parameter w, to obtain the characteristic after dimensionality reduction;It will be described
Characteristic after dimensionality reduction is input to nervus opticus network;Obtain second mind in the nervus opticus network training process
Variable condition information through network parameter, and obtain the nervus opticus network model training, enter ginseng, go out to join and calculated
The execution state information of journey;The variable condition information and the execution state information are aggregated into log, according to the day
Will generates test report.
9. a kind of computer equipment, which is characterized in that the computer equipment includes:
At least one processor, memory and input-output unit;
Wherein, the memory is for storing program code, and the processor is for calling the program stored in the memory
Code is executed such as method of any of claims 1-7.
10. a kind of computer storage medium, which is characterized in that it includes instruction, when run on a computer, so that calculating
Machine executes such as method of any of claims 1-7.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910666325.6A CN110503198A (en) | 2019-07-23 | 2019-07-23 | Obtain method, apparatus, equipment and the storage medium of neural network test report |
PCT/CN2020/098841 WO2021012894A1 (en) | 2019-07-23 | 2020-06-29 | Method and apparatus for obtaining neural network test report, device, and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910666325.6A CN110503198A (en) | 2019-07-23 | 2019-07-23 | Obtain method, apparatus, equipment and the storage medium of neural network test report |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110503198A true CN110503198A (en) | 2019-11-26 |
Family
ID=68586705
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910666325.6A Pending CN110503198A (en) | 2019-07-23 | 2019-07-23 | Obtain method, apparatus, equipment and the storage medium of neural network test report |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN110503198A (en) |
WO (1) | WO2021012894A1 (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111738886A (en) * | 2020-06-30 | 2020-10-02 | 上海松鼠课堂人工智能科技有限公司 | Learning report analysis system based on neural network |
CN111898742A (en) * | 2020-08-05 | 2020-11-06 | 上海眼控科技股份有限公司 | Method and equipment for monitoring training state of neural network model |
WO2021012894A1 (en) * | 2019-07-23 | 2021-01-28 | 平安科技(深圳)有限公司 | Method and apparatus for obtaining neural network test report, device, and storage medium |
CN114338248A (en) * | 2022-03-15 | 2022-04-12 | 北京大学 | User abnormal behavior detection method and device based on machine learning |
CN115261963A (en) * | 2022-09-27 | 2022-11-01 | 南通如东依航电子研发有限公司 | Method for improving deep plating capability of PCB |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108711100A (en) * | 2018-05-20 | 2018-10-26 | 冯世程 | A kind of system of the P2P platform operation risk assessment based on neural network |
CN109443382A (en) * | 2018-10-22 | 2019-03-08 | 北京工业大学 | Vision SLAM closed loop detection method based on feature extraction Yu dimensionality reduction neural network |
CN109544190A (en) * | 2018-11-28 | 2019-03-29 | 北京芯盾时代科技有限公司 | A kind of fraud identification model training method, fraud recognition methods and device |
US20190156211A1 (en) * | 2017-11-21 | 2019-05-23 | International Business Machines Corporation | Feature extraction using multi-task learning |
US20190164057A1 (en) * | 2019-01-30 | 2019-05-30 | Intel Corporation | Mapping and quantification of influence of neural network features for explainable artificial intelligence |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6954744B2 (en) * | 2001-08-29 | 2005-10-11 | Honeywell International, Inc. | Combinatorial approach for supervised neural network learning |
CN105787500A (en) * | 2014-12-26 | 2016-07-20 | 日本电气株式会社 | Characteristic selecting method and characteristic selecting device based on artificial neural network |
CN106779051A (en) * | 2016-11-24 | 2017-05-31 | 厦门中控生物识别信息技术有限公司 | A kind of convolutional neural networks model parameter processing method and system |
CN108491927A (en) * | 2018-03-16 | 2018-09-04 | 新智认知数据服务有限公司 | A kind of data processing method and device based on neural network |
CN109344968A (en) * | 2018-10-10 | 2019-02-15 | 郑州云海信息技术有限公司 | A kind of method and device of the hyper parameter processing of neural network |
CN110503198A (en) * | 2019-07-23 | 2019-11-26 | 平安科技(深圳)有限公司 | Obtain method, apparatus, equipment and the storage medium of neural network test report |
-
2019
- 2019-07-23 CN CN201910666325.6A patent/CN110503198A/en active Pending
-
2020
- 2020-06-29 WO PCT/CN2020/098841 patent/WO2021012894A1/en active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190156211A1 (en) * | 2017-11-21 | 2019-05-23 | International Business Machines Corporation | Feature extraction using multi-task learning |
CN108711100A (en) * | 2018-05-20 | 2018-10-26 | 冯世程 | A kind of system of the P2P platform operation risk assessment based on neural network |
CN109443382A (en) * | 2018-10-22 | 2019-03-08 | 北京工业大学 | Vision SLAM closed loop detection method based on feature extraction Yu dimensionality reduction neural network |
CN109544190A (en) * | 2018-11-28 | 2019-03-29 | 北京芯盾时代科技有限公司 | A kind of fraud identification model training method, fraud recognition methods and device |
US20190164057A1 (en) * | 2019-01-30 | 2019-05-30 | Intel Corporation | Mapping and quantification of influence of neural network features for explainable artificial intelligence |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2021012894A1 (en) * | 2019-07-23 | 2021-01-28 | 平安科技(深圳)有限公司 | Method and apparatus for obtaining neural network test report, device, and storage medium |
CN111738886A (en) * | 2020-06-30 | 2020-10-02 | 上海松鼠课堂人工智能科技有限公司 | Learning report analysis system based on neural network |
CN111738886B (en) * | 2020-06-30 | 2021-04-30 | 上海松鼠课堂人工智能科技有限公司 | Learning report analysis system based on neural network |
CN111898742A (en) * | 2020-08-05 | 2020-11-06 | 上海眼控科技股份有限公司 | Method and equipment for monitoring training state of neural network model |
CN114338248A (en) * | 2022-03-15 | 2022-04-12 | 北京大学 | User abnormal behavior detection method and device based on machine learning |
CN114338248B (en) * | 2022-03-15 | 2022-08-05 | 北京大学 | User abnormal behavior detection method and device based on machine learning |
CN115261963A (en) * | 2022-09-27 | 2022-11-01 | 南通如东依航电子研发有限公司 | Method for improving deep plating capability of PCB |
Also Published As
Publication number | Publication date |
---|---|
WO2021012894A1 (en) | 2021-01-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110503198A (en) | Obtain method, apparatus, equipment and the storage medium of neural network test report | |
CN102075352B (en) | Method and device for predicting network user behavior | |
CN110009479A (en) | Credit assessment method and device, storage medium, computer equipment | |
CN106777024A (en) | Recognize the method and device of malicious user | |
CN109035003A (en) | Anti- fraud model modelling approach and anti-fraud monitoring method based on machine learning | |
CN108898154A (en) | A kind of electric load SOM-FCM Hierarchical clustering methods | |
CN108416663A (en) | The method and device of the financial default risk of assessment | |
Frye et al. | Credit loss and systematic LGD | |
CN108427720A (en) | System log sorting technique | |
CN108898476A (en) | A kind of loan customer credit-graded approach and device | |
TWI752349B (en) | Risk identification method and device | |
CN110276679A (en) | A kind of network individual credit fraud detection method towards deep learning | |
CN105354210A (en) | Mobile game payment account behavior data processing method and apparatus | |
CN109034194A (en) | Transaction swindling behavior depth detection method based on feature differentiation | |
CN115510042A (en) | Power system load data filling method and device based on generation countermeasure network | |
CN111368926B (en) | Image screening method, device and computer readable storage medium | |
CN107133874A (en) | A kind of quantization strategy generation system based on genetic planning | |
CN107992978A (en) | It is a kind of to net the method for prewarning risk and relevant apparatus for borrowing platform | |
CN107862423A (en) | System evaluation method, intelligent evaluation system and computer-readable recording medium | |
CN109711707A (en) | A kind of Ship Power Equipment synthetical condition assessment method | |
CN109669935A (en) | Check data screening method, apparatus, equipment and storage medium | |
CN112561176A (en) | Early warning method for online running state of electric power metering device | |
CN103281555B (en) | Half reference assessment-based quality of experience (QoE) objective assessment method for video streaming service | |
CN114154672A (en) | Data mining method for customer churn prediction | |
CN108765133A (en) | The demand for loan and matched method, apparatus of loan product, system 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 |