Embodiment
Here exemplary embodiment will be illustrated in detail, its example is illustrated in the accompanying drawings.Following description is related to
During accompanying drawing, unless otherwise indicated, the same numbers in different accompanying drawings represent same or analogous key element.Following exemplary embodiment
Described in embodiment do not represent all embodiments consistent with this specification.On the contrary, they are only and such as institute
The example of the consistent apparatus and method of some aspects be described in detail in attached claims, this specification.
It is only merely for the purpose of description specific embodiment in the term that this specification uses, and is not intended to be limiting this explanation
Book." one kind " of used singulative, " described " and "the" are also intended to bag in this specification and in the appended claims
Most forms are included, unless context clearly shows that other implications.It is also understood that term "and/or" used herein is
Refer to and any or all may be combined comprising the associated list items purpose of one or more.
It will be appreciated that though various information may be described using term first, second, third, etc. in this specification, but
These information should not necessarily be limited by these terms.These terms are only used for same type of information being distinguished from each other out.For example, do not taking off
In the case of this specification scope, the first information can also be referred to as the second information, and similarly, the second information can also be claimed
For the first information.Depending on linguistic context, word as used in this " if " can be construed to " ... when " or
" when ... " or " in response to determining ".
As previously mentioned, for data prediction, such as linear regression traditionally can be individually used, or individually use GBDT
(Gradient Boosting Decision Tree, gradient lifting decision tree) returns.
However, individually by the way of linear regression prediction data, when returning because practically sample data is usual
In the presence of it is a large amount of it is nonlinear in the case of, it is therefore desirable to complex curve could be finally fitted by doing substantial amounts of non-linear conversion.The opposing party
Face, linear regression with can not solving level nonlinear problem, such as this logical combination of "AND", "or", distance are non-
Linear problem.For example, in trading volume is predicted, the feature of two time serieses be present, whether it " is activity that a feature is
Day ", whether another be characterized in " being weekend ";Under normal circumstances, the trading volume of active day can show as rising, inactive day
Trading volume can show as declining;The trading volume at weekend can be shown as declining, and the trading volume at non-weekend can be shown as
Rise;But when be active day be weekend again in the case of (such as because the public holiday will be adjusted to working day at weekend), transaction
It is bigger to measure the degree risen;Such case is that linear regression is unpredictable, so situation as once occurring, using linear
The result meeting deviation of regression forecasting is larger.
Individually returned to using GBDT, it is necessary to which smoothly sample data could fitted trend.It is described that smoothly sample data can
To refer to, the degree that sample data deviates standard deviation is the smaller the better.Actual sample data is usually jiggly, show as by
Year inflationary spiral, sample data increase year by year, cause sample data to deviate the degree of standard deviation and also increase year by year.Although can profit
This jiggly trend is made up with the mode such as difference processing, logization processing, it can only accomplishing to mitigate and can not do
To trend is completely eliminated.Therefore, it is also undesirable using the result precision of GBDT regression forecastings.
In order to provide data prediction scheme more precisely, a kind of data prediction of this specification shown in Fig. 1 refer to
Embodiment of the method, as shown in figure 1, this method may comprise steps of:
Step 110:Obtain the time factor of time series.
In this explanation, the time factor can be expressed as some day in a certain year.
To predict in financial market exemplified by the trading volume of stock market, it is assumed that be currently on July 1st, 2017, it is necessary to predict tomorrow
The trading volume on July 2nd, 1, then user can input on July 2nd, 2017;Correspondingly, for carrying out the clothes of data prediction
The time factor that business end is got can be on July 2nd, 2017.
It is noted that this specification embodiment can also disposably predict data corresponding to multiple time factors.Example
As, it is necessary to predict following 1 week data, then service end can get 7 time factors corresponding to 7 days futures.
Step 120:The time factor is input to gradient lifting decision-tree model.
Step 130:Obtain the first prediction data of the gradient lifting decision-tree model input.
In this specification, GBDT (Gradient Boosting Decision Tree) is also known as MART
(Multiple Additive Regression Tree)、GBT(Gradient Boosting Tree)、GTB
(Gradient Tree Boosting) or GBRT (Gradient Boosting Regression Tree), this specification
In be collectively referred to as GBDT.The GBDT is a kind of decision Tree algorithms of iteration, and the algorithm can be made up of more decision trees, is owned
The result of these decision trees can obtain final result after adding up.
In general, GBDT iteration can use it is preceding to Distribution Algorithm (Forward Stagewise Algorithm), and
And weak learner can use CART regression tree models.In GBDT iteration, it is assumed that the strong learner that previous round iteration obtains
It is ft-1(x), loss function is L (y, ft-1(x)), then the target of epicycle iteration can find a CART regression tree model
Weak learner ht(x), so that loss L (y, the f of epicyclet(x))=L (y, ft-1(x))+ht(x) it is minimum.It is to be understood that
The decision tree that epicycle iteration is found, it is that the loss of sample to be allowed becomes smaller as far as possible.
In this specification, the gradient lifts decision-making number (GBDT) model, can train to obtain in the following way:
A1:Obtain the sample set for training;The sample set is by some time factor, data to forming;
Such as sample set Z as follows:
Z={ (x1,y1),(x2,y2),(x3,y3),...,(xi,yi),...,(xn,yn), wherein, xiIt can represent i-th
Time factor, yiThe data of i-th of time factor can be represented, sample set Z has n group samples.
A2:Train to obtain gradient lifting decision tree mould with reference to the sample set based on gradient lifting decision tree regression algorithm
Type.
Generally, gradient lifting decision tree regression algorithm also needs to configure maximum iteration T and loss function L.
The maximum iteration T can be the empirical value artificially set.
The loss function L can use the loss function that calculation is returned for GBDT commonly used in the trade:
The first:Mean square deviation, L (y, f (x))=(y-f (x))2
Second:Definitely loss, L (y, f (x))=| y-f (x) |
Corresponding to negative gradient error is:
sign(yi-f(xi));
The sign functions are sign function.
The third:Huber loses:
Wherein, δ represents it can is the empirical value artificially set;
| y-f (x) | during≤δ,
| y-f (x) | during > δ,
Corresponding to negative gradient error is
Huber losses are a kind of mean square deviations and the compromise mode definitely lost, can be adopted for deep abnormity point
With absolute loss, mean square deviation can be used for the point of immediate vicinity.
4th kind:Quantile is lost:
Wherein, θ represents quantile;The quantile can be an empirical value being manually set.
Corresponding to negative gradient error is
It should be noted that above-mentioned loss function is merely illustrative, the not loss letter to specifically using in this specification
Number is defined.
Training process introduced below:
First, it is as follows to initialize a weak learner:
Then:Carry out the iterative calculation of T (maximum iteration) wheels;
Iterative process is as follows each time:
B1:According to i=1 in sample set Z, 2 ..., n data, negative gradient is calculated.
Equation below can be passed through by calculating negative gradient:
Wherein, rtiFor i-th of negative gradient,For local derviation symbol, L (y, f (xi)) it is loss function.That is, each
X are all corresponding with a negative gradient, i.e. (xi,rti), i=1,2 ..., n.
B2:According to the negative gradient calculated, a CART regression tree is fitted.
As it was previously stated, according to negative gradient (xi,rti), i=1,2 ..., n;Can is fitted a CART regression tree, wherein,
Corresponding leaf node region is Rtj, j=1,2 ..., J, J is the number of the CART regression trees leaf node.
B3:According to the area foliage of the CART regression trees, best-fit values are calculated.
Equation below can be passed through by calculating best-fit values:
Wherein, ctjFor best-fit values, RtjFor the area foliage of CART regression trees, and j=1,2 ..., J, J are that CART is returned
Return the number of leaf child node.
B4:Update strong learner
B5:Weak learner using the strong learner after renewal as next round iteration, carry out next round iteration;Until reaching
Maximum iteration.
Finally, the expression formula (as follows) for iterating to calculate the strong learner drawn for the last time is defined as GBDT moulds
Type.
So far, GBDT model constructions are completed.
In this specification, the sample set of input can be with as follows:
Z=
(x11,y11),(x12,y12),(x13,y13),...,(x1n,y1n)
(x21,y21),(x22,y22),(x23,y23),...,(x2n,y2n)
...
(xm1,ym1),(xm2,ym2),(xm3,ym3),...,(xmn,ymn)
}
Wherein, xmnIt can represent that n-th day of m, such as the January 2 of the 1st year can be designated as x12;ymnCan be represented
The m data of n-th day.
The sample set Z inputted in one specifically embodiment, during the structure GBDT models, it is also necessary to smoothly located
Reason, so as to eliminate average trend and inflationary spiral.
The average trend can refer to, Change in Mean situation year by year, such as increases year by year, reduces year by year
Or increase sometimes and reduce sometimes.There is average trend in sample data, easily cause the unstable of sample data, and then
GBDT is unable to fitted trend.
For example, the data mean value that the data mean value that the data mean value of the 1st year is the 1, the 2nd year is the 2, the 3rd year is 3, then
It can increase year by year taking human as average trend.
The training sample set can eliminate average trend in the following way:
C1:Equation 1 below is fitted using one-variable linear regression combination sample set, calculates slope and the intercept of each year.
ymn=xmnwm+bmFormula 1
Wherein, xmn(independent variable) is n-th day of m in sample set;ymn(dependent variable) is n-th of m in sample set
It data;wmFor m slope, bmFor m intercept.
C2:According to the slope and intercept, the average trend based on equation 2 below elimination sample intensive data.
Ymn=ymn-(bm-b1)-xmn*wmFormula 2
Wherein, YmnTo eliminate the m data of n-th day after average trend, ymnTo eliminate m n-th day before average trend
Data, bmFor m intercept, b1For the intercept of the 1st year, xmnFor m n-th day, wmFor m slope.
By above-mentioned formula 2, the sample set eliminated after average trend can be with as follows:
Z=
(x11,Y11),(x12,Y12),(x13,Y13),...,(x1n,Y1n)
(x21,Y21),(x22,Y22),(x23,Y23),...,(x2n,Y2n)
...
(xm1,Ym1),(xm2,Ym2),(xm3,Ym3),...,(xmn,Ymn)
}
The inflationary spiral can refer to, standard deviation intensity of variation, such as increase year by year, reduce year by year year by year
Or increase sometimes and reduce sometimes.There is inflationary spiral in sample data, easily cause the unstable of sample data, and then
GBDT is unable to fitted trend.
It should be noted that eliminating inflationary spiral is performed on the basis of the sample set of average trend is eliminated.
The training sample set can eliminate inflationary spiral in the following way:
D1:In the case of the average trend of the elimination sample intensive data, each year is calculated according to equation below 3
According to arithmetic mean of instantaneous value.
Arithmetic mean of instantaneous value (Arithmetic Mean), also known as average.
Shown in the formula equation below 3 for calculating arithmetic mean of instantaneous value:
Wherein, μmFor m arithmetic mean of instantaneous value, YmnTo eliminate the m data of n-th day after average trend.
Below with the data instance of the 1st year in sample set,
The data of the 1st year include (Y11,Y12,Y13,...,Y1n)
Then, the arithmetic mean of instantaneous value of the 1st year
D2:According to formula 4, the standard deviation of each annual data is calculated.
Standard deviation (Standard Deviation), also known as mean square deviation or standard deviation.Standard deviation can reflect one
The dispersion degree of data intensive data.
Shown in the formula equation below 4 for calculating standard deviation:
Wherein, YmiFor the m data of i-th day, i=1,2 ..., n, σmFor m standard deviation, μmFor m calculation
Art average value.
Still with the data instance of the 1st year in sample set,
The standard deviation of the 1st year
D3:According to formula 5, the coefficient of expansion of each year of calculating.
Shown in the formula equation below 5 for calculating the coefficient of expansion:
Wherein, PmFor the m coefficient of expansion, σmFor m standard deviation, σ1For the standard deviation of the 1st year.
D4:According to the coefficient of expansion, the inflationary spiral of sample intensive data is eliminated based on equation 6 below.
Ymn'=μm+(Ymn-μm)*PmFormula 6
Wherein, Ymn' it is to eliminate the m data of n-th day after inflationary spiral, μmFor m arithmetic mean of instantaneous value, YmnTo disappear
Except the m data of n-th day, P before inflationary spiralmFor the m coefficient of expansion.
By above-mentioned formula 6, the sample set eliminated after inflationary spiral can be with as follows:
Z=
(x11,Y11'),(x12,Y12'),(x13,Y13'),...,(x1n,Y1n')
(x21,Y21'),(x22,Y22'),(x23,Y23'),...,(x2n,Y2n')
...
(xm1,Ym1'),(xm2,Ym2'),(xm3,Ym3'),...,(xmn,Ymn')
}
By said process, the average trend for the GBDT sample intensive datas modeled and inflationary spiral are eliminated, is made
It is steady to obtain sample intensive data;So, it is easy to GBDT fitted trends, so as to find periodicity, seasonal, festivals or holidays data
Rule.
Step 140:First prediction data is input to linear regression model (LRM).
Step 150:Obtain the second prediction data of the linear regression model (LRM) input.
The linear regression model (LRM) trains to obtain in the following way:
E1:The sample set is inputted into the gradient lifting decision-tree model.
E2:Obtain the recurrence sample set of the gradient lifting decision-tree model input.
As it was previously stated, after GBDT models are constructed, can continue according to the GBDT models to the x in sample set ZmnEnter
Line number is it was predicted that generation recurrence sample set is as follows:
Z=
(x11,z11),(x12,z12),(x13,z13),...,(x1n,z1n)
(x21,z21),(x22,z22),(x23,z23),...,(x2n,z2n)
...
(xm1,zm1),(xm2,zm2),(xm3,zm3),...,(xmn,zmn)
}
Wherein, xmnN-th day of m, z can be representedmnN-th day of the m that GBDT model predictions obtain can be represented
Data.
E3:Train to obtain linear regression model (LRM) with reference to the recurrence sample set based on linear regression algorithm.
Combined based on linear regression algorithm and thus return sample set training linear regression model (LRM).
Linear regression (Linear Regression), be it is a kind of using linear regression side into least square function pair one
The regression analysis that relation is modeled between individual or multiple independents variable and dependent variable.This function generally can be one or more
The referred to as linear combination of the model parameter of regression coefficient.
In regression analysis, an independent variable and a dependent variable are only included, and the relation of the two can be near with straight line
Like expression, this regression analysis is referred to as simple linear regression analysis.If regression analysis include it is two or more from
Variable, and be linear relationship between dependent variable and independent variable, then referred to as multiple linear regression analysis.In this specification, use
It is simple linear regression analysis.
Based on linear regression algorithm with reference to the recurrence sample set, can train to obtain linear regression model (LRM).That is, by GBDT
The input of the prediction data that model obtains linear regression the most.
Shown in the expression-form equation below 7 of linear regression, it can typically be expressed as:
zmn=xmnW+b formula 7
Wherein, w is slope, and b is intercept.It should be noted that the linear regression can be multiple linear regression, with preceding public affairs
One-variable linear regression shown in formula 1 is different.
After optimal w and b is found, the linear regression model (LRM), which is just built, to be completed.
The linear regression expression formula finally given is the linear regression model (LRM).
This specification is namely based on the model (GBDT regression models and linear regression model (LRM)) of the two structures, is getting
After the time factor of time series;Decision-tree model is first lifted based on gradient, calculates the first prediction data of the time factor;
Linear regression model (LRM) is based on again, calculates the second prediction data of the time factor, second prediction data is final
Prediction data.
This specification is returned and linear regression progress data prediction by combining GBDT, you can to play linear regression fit
The characteristics of trend, the problem of avoiding being not easy fitted trend when individually using GBDT to return;Can also play that GBDT returns can be with
Fit non-linear solves such as logical problem with level, and avoiding can not solve with level when individually using linear regression
Property logical problem.In this way, the accuracy of data prediction can be improved.
Further, using one-variable linear regression fitted trend, slope and intercept are obtained, and according to the slope and intercept
Eliminate the average trend of data;Further, the coefficient of expansion is obtained by calculating standard deviation, and according to the coefficient of expansion
Eliminate the inflationary spiral of data so that the sample data region of training pattern is steady, the data that the model prediction so built goes out
It is more accurate.
Corresponding with foregoing data predication method embodiment, this specification additionally provides the implementation of data prediction device
Example.Described device embodiment can be realized by software, can also be realized by way of hardware or software and hardware combining.With soft
It is by nonvolatile memory by the processor of equipment where it as the device on a logical meaning exemplified by part is realized
In corresponding computer program instructions read in internal memory what operation was formed.For hardware view, as shown in Fig. 2 being this theory
A kind of hardware structure diagram of equipment where bright book data prediction device, except the processor shown in Fig. 2, network interface, internal memory with
And outside nonvolatile memory, the equipment in embodiment where device may be used also generally according to the actual functional capability of the data prediction
Including other hardware, to be repeated no more to this.
Data predication method embodiment shown in corresponding diagram 1, the data that the embodiment of this specification one shown in Figure 3 provides
The module map of prediction meanss, described device include:
Acquiring unit 310, obtain the time factor of time series;
First input block 320, the time factor is input to gradient lifting decision-tree model;
First predicting unit 330, obtain the first prediction data of the gradient lifting decision-tree model input;
Second input block 340, first prediction data is input to linear regression model (LRM);
Second predicting unit 350, obtain the second prediction data of the linear regression model (LRM) input.
In an optional embodiment:
The gradient lifting decision-tree model is obtained by following subelement training:
Subelement is obtained, obtains the sample set for training;The sample set is by some time factor, data to forming;
First training subelement, train to obtain gradient carrying with reference to the sample set based on gradient lifting decision tree regression algorithm
Rise decision-tree model;
The linear regression model (LRM) is obtained by following subelement training:
Subelement is inputted, the sample set is inputted into the gradient lifts decision-tree model;
Subelement is exported, obtains the recurrence sample set of the gradient lifting decision-tree model input;
Second training subelement, train to obtain linear regression mould with reference to the recurrence sample set based on linear regression algorithm
Type.
In an optional embodiment:
Before the described first training subelement, described device also includes:
Subelement is eliminated, eliminates average trend and the inflationary spiral of the sample intensive data.
In an optional embodiment:
The sample set eliminates average trend by following subelement:
First computation subunit, using one-variable linear regression with reference to the sample set fitting formula ymn=xmnwm+bm, calculate
The slope and intercept of each year;Wherein, xmnFor n-th day of m in sample set;ymnFor n-th day of m in sample set
Data;wmFor m slope, bmFor m intercept;
Second computation subunit, according to the slope and intercept, and it is based on formula Ymn=ymn-(bm-b1)-xmn*wmEliminate
The average trend of sample intensive data;Wherein, YmnTo eliminate the m data of n-th day after average trend, ymnTo eliminate average
The m data of n-th day before trend, bmFor m intercept, b1For the intercept of the 1st year, xmnFor m n-th day, wmFor m
The slope in year.
In an optional embodiment:
The sample set eliminates inflationary spiral by following subelement:
3rd computation subunit, in the case of the average trend of the elimination sample intensive data, according to formulaCalculate the arithmetic mean of instantaneous value of each annual data;Wherein, μmFor m arithmetic
Average value, YmnTo eliminate the m data of n-th day after average trend;
4th computation subunit, according to formulaCalculate the mark of each annual data
It is accurate poor;Wherein, YmiFor the m data of i-th day, i=1,2 ..., n, σmFor m standard deviation, μmFor m arithmetic
Average value;
5th computation subunit, according to formulaCalculate the coefficient of expansion of each year;Wherein, PmFor m
The coefficient of expansion, σmFor m standard deviation, σ1For the standard deviation of the 1st year;
6th computation subunit, according to the coefficient of expansion, based on formula Ymn'=μm+(Ymn-μm)*PmEliminate sample set
The inflationary spiral of middle data;Wherein, Ymn' it is to eliminate the m data of n-th day after inflationary spiral, μmFor m arithmetic average
Value, YmnTo eliminate the m data of n-th day before inflationary spiral, PmFor the m coefficient of expansion.
System, device, module or the unit that above-described embodiment illustrates, it can specifically be realized by computer chip or entity,
Or realized by the product with certain function.One kind typically realizes that equipment is computer, and the concrete form of computer can
To be personal computer, laptop computer, cell phone, camera phone, smart phone, personal digital assistant, media play
In device, navigation equipment, E-mail receiver/send equipment, game console, tablet PC, wearable device or these equipment
The combination of any several equipment.
The function of unit and the implementation process of effect specifically refer to and step are corresponded in the above method in said apparatus
Implementation process, it will not be repeated here.
For device embodiment, because it corresponds essentially to embodiment of the method, so related part is real referring to method
Apply the part explanation of example.Device embodiment described above is only schematical, wherein described be used as separating component
The unit of explanation can be or may not be physically separate, can be as the part that unit is shown or can also
It is not physical location, you can with positioned at a place, or can also be distributed on multiple NEs.Can be according to reality
Need to select some or all of module therein to realize the purpose of this specification scheme.Those of ordinary skill in the art are not
In the case of paying creative work, you can to understand and implement.
Figure 3 above describes inner function module and the structural representation of data prediction device, its substantial executive agent
Can be a kind of electronic equipment, such as server, Fig. 4 are a kind of hardware knots of server according to an exemplary embodiment
The schematic diagram of structure, reference picture 4, the server can include:
Processor;
For storing the memory of processor-executable instruction;
Wherein, the processor is configured as:
Obtain the time factor of time series;
The time factor is input to gradient lifting decision-tree model;
Obtain the first prediction data of the gradient lifting decision-tree model input;
First prediction data is input to linear regression model (LRM);
Obtain the second prediction data of the linear regression model (LRM) input.
Preferably, the gradient lifting decision-tree model trains to obtain in the following way:
Obtain the sample set for training;The sample set is by some time factor, data to forming;
Train to obtain gradient lifting decision-tree model with reference to the sample set based on gradient lifting decision tree regression algorithm;
The linear regression model (LRM) trains to obtain in the following way:
The sample set is inputted into the gradient lifting decision-tree model;
Obtain the recurrence sample set of the gradient lifting decision-tree model input;
Train to obtain linear regression model (LRM) with reference to the recurrence sample set based on linear regression algorithm.
Preferably, train to obtain gradient lifting with reference to the sample set based on gradient lifting decision tree regression algorithm described
Before decision-tree model, methods described also includes:
Eliminate average trend and the inflationary spiral of the sample intensive data.
Preferably, the sample set eliminates average trend in the following way:
Using one-variable linear regression with reference to the sample set fitting formula ymn=xmnwm+bm, calculate each year slope and
Intercept;Wherein, xmnFor n-th day of m in sample set;ymnFor the data of n-th day of m in sample set;wmFor m's
Slope, bmFor m intercept;
According to the slope and intercept, and it is based on formula Ymn=ymn-(bm-b1)-xmn*wmEliminate the equal of sample intensive data
Value trend;Wherein, YmnTo eliminate the m data of n-th day after average trend, ymnTo eliminate m n-th day before average trend
Data, bmFor m intercept, b1For the intercept of the 1st year, xmnFor m n-th day, wmFor m slope.
Preferably, the sample set eliminates inflationary spiral in the following way:
In the case of the average trend of the elimination sample intensive data, according to formulaCalculate the arithmetic mean of instantaneous value of each annual data;Wherein, μmFor m arithmetic
Average value, YmnTo eliminate the m data of n-th day after average trend;
According to formulaCalculate the standard deviation of each annual data;Wherein, YmiFor
The m data of i-th day, i=1,2 ..., n, σmFor m standard deviation, μmFor m arithmetic mean of instantaneous value;
According to formulaCalculate the coefficient of expansion of each year;Wherein, PmFor the m coefficient of expansion, σmFor
M standard deviation, σ1For the standard deviation of the 1st year;
According to the coefficient of expansion, based on formula Ymn'=μm+(Ymn-μm)*PmThe expansion for eliminating sample intensive data becomes
Gesture;Wherein, Ymn' it is to eliminate the m data of n-th day after inflationary spiral, μmFor m arithmetic mean of instantaneous value, YmnIt is swollen to eliminate
The m data of n-th day, P before swollen trendmFor the m coefficient of expansion.
In the embodiment of above-mentioned electronic equipment, it should be appreciated that the processor can be CPU (English:
Central Processing Unit, referred to as:CPU), it can also be other general processors, digital signal processor (English:
Digital Signal Processor, referred to as:DSP), application specific integrated circuit (English:Application Specific
Integrated Circuit, referred to as:ASIC) etc..General processor can be microprocessor or the processor can also be
Any conventional processor etc., and foregoing memory can be read-only storage (English:Read-only memory, abbreviation:
ROM), random access memory (English:Random access memory, referred to as:RAM), flash memory, hard disk or solid
State hard disk.The step of method with reference to disclosed in the embodiment of the present invention, can be embodied directly in hardware processor and perform completion, or
Hardware and software module combination in person's processor perform completion.
Fig. 5 is a kind of schematic diagram of server 1000 according to an exemplary embodiment.Reference picture 5, server
1000 include processing component 1022, and it further comprises one or more processors, and as depositing representated by memory 1032
Memory resource, can be by the instruction of the execution of processing component 1022, such as application program for storing.Stored in memory 1032
Application program can include it is one or more each correspond to the module of one group of instruction.In addition, processing component 1022
Execute instruction is configured as, to perform all or part of step of the above-mentioned picture retrieval method based on convolutional neural networks.
Server 1000 can also include the power management that a power supply module 1026 is configured as execute server 1000,
One wired or wireless network interface 1050 is configured as server 1000 being connected to network, and an input and output (I/O)
Interface 1058.Server 1000 can be operated based on the operating system for being stored in memory 1032, such as Windows
ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or similar.
Each embodiment in this specification is described by the way of progressive, identical similar portion between each embodiment
Divide mutually referring to what each embodiment stressed is the difference with other embodiment.Set especially for electronics
For standby embodiment, because it is substantially similar to embodiment of the method, so description is fairly simple, related part is real referring to method
Apply the part explanation of example.
Those skilled in the art will readily occur to this specification after considering specification and putting into practice invention disclosed herein
Other embodiments.This specification is intended to any modification, purposes or adaptations of this specification, these modifications,
Purposes or adaptations follow the general principle of this specification and undocumented in the art including this specification
Common knowledge or conventional techniques.Description and embodiments be considered only as it is exemplary, the true scope of this specification and
Spirit is pointed out by following claim.
It should be appreciated that the precision architecture that this specification is not limited to be described above and is shown in the drawings,
And various modifications and changes can be being carried out without departing from the scope.The scope of this specification is only limited by appended claim
System.