CN108304876A - Disaggregated model training method, device and sorting technique and device - Google Patents

Abstract

The present invention provides a kind of disaggregated model training method, device and sorting technique and devices, wherein the disaggregated model training method includes：Common characteristic capture is carried out to source domain data and target numeric field data using first nerves network, makes the common characteristic of first object characteristic of field vector study source domain data and target numeric field data；Otherness Feature capturing is carried out to source domain data and target numeric field data using nervus opticus network, makes the difference characteristic of the second target domain characterization vector study source domain data and target numeric field data；First object characteristic of field vector and the second target domain characterization vector are clustered respectively；According to the result of cluster and the first classification results, epicycle training is carried out to first nerves network and the first grader.This method can either use the same characteristic features between source domain and aiming field, can also utilize the difference characteristic between source domain and aiming field, and the disaggregated model of training gained can obtain more accurate classification results.

Description

Disaggregated model training method, device and sorting technique and device

Technical field

The present invention relates to depth learning technology field, in particular to a kind of disaggregated model training method, device and Sorting technique and device.

Background technology

Transfer learning can utilize training sample (can be described as source domain data) the training classification mould for having label in known art Type is demarcated come the data (can be described as target numeric field data) to target domain, and is not required for source domain data and target numeric field data Data distribution having the same.Transfer learning is indeed through the connection between looking for data and known label data to be calibrated Source domain and target numeric field data, are mapped in the same space, the source domain data under the space by system for example, by using the mode of kernel function Possess identical distribution with target numeric field data, so as to utilize the source domain sample data training point for having label of the space representation Class device demarcates target domain.

In existing transfer learning method, same neural network is trained usually using source domain data and target numeric field data, is obtained Obtain parameter sharing network.The network training process can find out the general character between source domain data and target numeric field data, usually by source Numeric field data and target numeric field data are mapped in the high bit space with high comparativity, to obtain the distribution characteristics in two domains.It is this Although training method can use the same characteristic features between source domain data and target numeric field data, make source domain and aiming field Between difference characteristic lose serious, cause the disaggregated model that training obtains to exist when classifying to target numeric field data certain Error.

Invention content

In view of this, the embodiment of the present invention is designed to provide a kind of disaggregated model training method, device and classification side Method and device can either use the same characteristic features between source domain and aiming field, also can will be between source domain and aiming field Difference characteristic utilized, training gained disaggregated model can obtain more accurate classification results.

In a first aspect, an embodiment of the present invention provides a kind of disaggregated model training method, this method includes：

Obtain the source domain data for carrying label and the target numeric field data for not carrying label；

The source domain data and the target numeric field data are inputted into first nerves network, for source domain data extraction the One source domain feature vector extracts first object characteristic of field vector for the target numeric field data；And

Common characteristic capture is carried out to the source domain data and the target numeric field data, make the first object characteristic of field to Amount learns the common characteristic of the source domain data and the target numeric field data；And

The first source domain feature vector is inputted into the first grader and obtains the first classification results；

The source domain data and the target numeric field data are inputted into nervus opticus network, are extracted for the target numeric field data Second target domain characterization vector；And

Otherness Feature capturing is carried out to the source domain data and the target numeric field data, makes second target domain characterization Vector learns the difference characteristic of the source domain data and the target numeric field data；

First object characteristic of field vector and the second target domain characterization vector are clustered respectively；

According to cluster as a result, and first classification results, to the first nerves network and first point described Class device carries out epicycle training；

By carrying out more wheel training to the first nerves network and first grader, disaggregated model is obtained.

Second aspect, the embodiment of the present invention also provide a kind of sorting technique, and this method includes：

Obtain data to be sorted；

The data to be sorted are input to point obtained by disaggregated model training method provided by the embodiments of the present application In class model, the classification results of the data to be sorted are obtained；

Wherein, the disaggregated model includes：The first nerves network and first grader.

Second aspect, the embodiment of the present invention also provide a kind of disaggregated model training device, including：

Acquisition module, for obtaining the source domain data for carrying label and the target numeric field data for not carrying label；

First processing module is source domain data for source domain data and target numeric field data to be inputted first nerves network The first source domain feature vector is extracted, first object characteristic of field vector is extracted for target numeric field data；And to source domain data and aiming field Data carry out common characteristic capture, make the common characteristic of first object characteristic of field vector study source domain data and target numeric field data； And the first source domain feature vector is inputted into the first grader and obtains the first classification results；

Second processing module is aiming field number for source domain data and target numeric field data to be inputted nervus opticus network According to extraction the second target domain characterization vector；And otherness Feature capturing is carried out to source domain data and target numeric field data, make the second mesh Mark the difference characteristic of characteristic of field vector study source domain data and target numeric field data；

Cluster module, for being clustered respectively to first object characteristic of field vector and the second target domain characterization vector；

Training module, for according to cluster as a result, and the first classification results, to first nerves network and first Grader carries out epicycle training；By carrying out more wheel training to first nerves network and the first grader, disaggregated model is obtained.

Fourth aspect, the embodiment of the present invention also provide a kind of sorter, including：Data acquisition module to be sorted, is used for Obtain data to be sorted；

Sort module, for being input to data to be sorted by disaggregated model training method provided by the embodiments of the present application In obtained disaggregated model, the classification results of data to be sorted are obtained；Wherein, disaggregated model includes：First nerves network and One grader.

5th aspect, the embodiment of the present invention also provide a kind of computer equipment, which includes memory, processor and deposit Store up the computer program that can be run on the memory and on the processor, wherein above-mentioned processor executes above computer The step of disaggregated model training method provided by the embodiments of the present application is realized when program.

6th aspect, the embodiment of the present invention also provide a kind of computer readable storage medium, the computer-readable storage medium It is stored with computer program in matter, the disaggregated model that our embodiment is provided is executed when which is run by processor The step of training method.

The embodiment of the present invention utilizes first nerves network, and the first source domain feature vector is extracted for source domain data, is aiming field When data extract first object characteristic of field vector, can source domain data and target numeric field data be carried out with the capture of common characteristic, made While the common characteristic of first object characteristic of field vector study source domain data and target numeric field data, nervus opticus network can be utilized The second target domain characterization vector is extracted for target numeric field data；Otherness feature can be carried out to source domain data and target numeric field data to catch It catches, makes the difference characteristic of the second target domain characterization vector study source domain data and target numeric field data, then according to the phase captured With feature and otherness feature, first nerves network and the first grader are trained, obtained disaggregated model, can either be incited somebody to action Same characteristic features between source domain and aiming field use, and also can the difference characteristic between source domain and aiming field be carried out profit With, and then the disaggregated model can obtain more accurate classification results.

To enable the above objects, features and advantages of the present invention to be clearer and more comprehensible, preferred embodiment cited below particularly, and coordinate Appended attached drawing, is described in detail below.

Description of the drawings

In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this A little attached drawings obtain other relevant attached drawings.

Fig. 1 shows a kind of flow chart for disaggregated model training method that the embodiment of the present invention is provided；

In the disaggregated model training method provided Fig. 2 shows the embodiment of the present invention, to source domain data and aiming field number According to the flow chart for the specific method for carrying out common characteristic capture；

Fig. 3 is shown in the disaggregated model training method that the embodiment of the present invention is provided, according to domain classification results, to first Neural network carries out the flow chart of the specific method of parameter adjustment；

Fig. 4 is shown in the disaggregated model training method that the embodiment of the present invention is provided, to source domain data and aiming field number According to the flow chart for the specific method for carrying out otherness Feature capturing；

Fig. 5 is shown in the disaggregated model training method that the embodiment of the present invention is provided, according to domain classification results, to second Neural network carries out the flow chart of the specific method of parameter adjustment；

Fig. 6 is shown in the disaggregated model training method that the embodiment of the present invention is provided, according to the first god of cluster result pair The flow chart of the specific method of epicycle training is carried out through network；

Fig. 7 shows in the disaggregated model training method that the embodiment of the present invention is provided that similarity calculation operates specific The flow chart of method；

Fig. 8 shows a kind of flow chart for sorting technique that the embodiment of the present invention is provided；

Fig. 9 shows a kind of structural schematic diagram for disaggregated model training device that the embodiment of the present invention is provided；

Figure 10 shows a kind of structural schematic diagram for sorter that the embodiment of the present invention is provided；

Figure 11 shows a kind of structural schematic diagram for computer equipment that the embodiment of the present invention is provided

Specific implementation mode

Unlike the prior art, the embodiment of the present invention utilizes first nerves network, and it is special to extract the first source domain for source domain data Sign vector can be total to source domain data and target numeric field data when extracting first object characteristic of field vector for target numeric field data The capture for having feature, while making the common characteristic of first object characteristic of field vector study source domain data and target numeric field data, meeting It is that target numeric field data extracts the second target domain characterization vector using nervus opticus network；It can be to source domain data and the aiming field Data carry out otherness Feature capturing, keep the difference of the second target domain characterization vector study source domain data and target numeric field data special Sign is trained first nerves network and the first grader, obtains then according to the same characteristic features and otherness feature captured The disaggregated model arrived can either use the same characteristic features between source domain and aiming field, also can be by source domain and aiming field Between difference characteristic utilized, and then the disaggregated model can obtain more accurate classification results.

For ease of understanding the present embodiment, first to a kind of disaggregated model training side disclosed in the embodiment of the present invention Method describes in detail, and this method is used for the training to the disaggregated model of various data, and the disaggregated model of gained can be to correspondence Data are classified.

Shown in Figure 1, disaggregated model training method provided in an embodiment of the present invention includes：

S101：Obtain the source domain data for carrying label and the target numeric field data for not carrying label.

When specific implementation, source domain data are the data with label, and target numeric field data is the number without label According to.Source domain data and target numeric field data have certain general character, and have certain otherness.

It is considered that source domain data are sufficient as training samples number, and current demand is needed to the mesh with default feature Mark numeric field data is classified, and the second target numeric field data with default feature is as training samples number deficiency, or training It is difficult larger in the process, then, it needs through transfer learning, during learning to source domain data, while learning default spy Sign, to merge default feature with source domain data characteristics；It is also desirable to by transfer learning, learn to source domain data During, while learning to the difference characteristic between source domain data and target numeric field data, thus by default feature, source domain data It is merged with source domain data characteristics with the difference characteristic of target numeric field data, abundant learning objective characteristic of field space, to aiming field number According to classification it is more accurate.

Herein, source domain data and target numeric field data can be that neural network learning may be used in image, video, language etc. The data classified.

Such as when source domain data and target numeric field data are image data, source domain data can be that quality preferably be schemed Picture, such as using the higher image acquisition equipment of resolution ratio under the conditions of uniform illumination, the facial have no occluder that is obtained it is clear Facial image.Face in source domain data can be the facial image of multiple angles, for example, face face image, side elevation image, Oblique-view image looks up image, overhead view image etc..

It such as the image that picture quality is poor can use resolution ratio that target numeric field data, which is the image with default feature, Lower image acquisition equipment unsharp facial image acquired under a variety of different illumination conditions heterogeneous.Aiming field Face in data can also be the facial image of multiple angles.

In another example when source domain data and target numeric field data are language data, source domain data are French vocabulary, aiming field Data are Spanish vocabulary, since French and Spanish belong to Romance, have part common between the two Feature；But the two belongs to two different language, therefore has certain difference again.Using the French that can be identified to western class The feature of tooth language is learnt, so as to identify Spanish.

In another example when source domain data and target numeric field data are language data, emotion is carried out to some vocabulary or words art Analysis；Source domain data are to be labelled with the vocabulary of affective tag, and target numeric field data is not mark art if affective tag.

S102：Source domain data and target numeric field data are inputted into first nerves network, the first source domain is extracted for source domain data Feature vector extracts first object characteristic of field vector for target numeric field data；And source domain data and target numeric field data are shared Feature capturing makes the common characteristic of first object characteristic of field vector study source domain data and target numeric field data；And by the first source domain Feature vector inputs the first grader and obtains the first classification results.

When specific implementation, convolutional neural networks (Convolutional may be used in first nerves network Neural Network, CNN) to extract the first source domain feature vector for source domain data, extract first object for target numeric field data Characteristic of field vector.

Source domain data are to carry the data of label, which is used to indicate the correct classification results of source domain data；Target Numeric field data is not carry the data of label.After by source domain data and target numeric field data input first nerves network, first nerves Network carries out source domain data and target numeric field data the feature learning of shared parameter.In this process, due to first nerves Network carries out supervised learning to source domain data, and carries out unsupervised learning to target numeric field data, is using same first nerves Network carries out source domain data and target numeric field data in the learning process of shared parameter, can constantly adjust institute in first nerves network The parameter used, to during to first nerves network training, allow the parameter of first nerves network by aiming field number According to influence so that first nerves network is to source domain data and target numeric field data after carrying out feature learning, to each source Numeric field data can be interfered by target numeric field data in the obtained first source domain feature vector of feature extraction, realizes source domain number It is mixed according between the domain of target numeric field data.

Meanwhile the first source domain feature vector is being extracted for source domain data, and extract first object domain spy for target numeric field data After sign vector, can also source domain data and target numeric field data be carried out with the capture of common characteristic, this makes to first nerves network When training, target numeric field data that the first source domain feature vector for extracting for source domain data is subject to is interfered, be exactly by To the interference of source domain data and target numeric field data common characteristic, to be merged between realization source domain data and the domain of target numeric field data When, the shared domain of source domain data and target numeric field data is merged, finally so that the study of the first source domain feature vector is arrived The common characteristic of source domain data and target numeric field data.

Shown in Figure 2, the embodiment of the present invention also provides a kind of to source domain data and target numeric field data progress common characteristic The specific method of capture.This method includes：

S201：The first source domain feature vector is being extracted for source domain data, and first object domain spy is extracted for target numeric field data After sign vector, the first source domain feature vector and first object characteristic of field vector are subjected to gradient reverse process.

S202：It will carry out the first source domain feature vector and the first object characteristic of field vector input first of gradient reverse process Domain grader.

S203：It is characterized respectively according to first domain grader pair the first source domain feature vector and first object characteristic of field vector Source domain data and target numeric field data domain classification results, to first nerves network carry out parameter adjustment.

When specific implementation, due to being trained to first nerves network using source domain data and target numeric field data Process, source domain data and target numeric field data are actually carried out to the process of domain mixing.Use first nerves network pair Source domain data, which carry out the first source domain feature vector that feature extraction is obtained, to be influenced by feature in target numeric field data, I.e. so that the first source domain feature vector is close to the feature of target numeric field data；Meanwhile using first nerves network to source domain data Carry out the first object characteristic of field vector acquired in feature extraction is influenced by feature in source domain data, namely so that first Target domain characterization vector is close to the feature of source domain data.Therefore, mixed to the domain of source domain data and target numeric field data in order to realize It closes, is extracting first object characteristic of field vector for each target numeric field data in target numeric field data, and be every in source domain data After a source domain data extract the first source domain feature vector, first object characteristic of field vector sum the first source domain feature vector is carried out Then gradient reverse process inputs first object characteristic of field vector sum the first source domain feature vector Jing Guo gradient reverse process To the first domain grader, domain point is carried out to first object characteristic of field vector sum the first source domain feature vector using the first domain grader Class.

The result of domain classification is correct namely the first domain grader can be correctly to the first source domain feature vector and first object The probability that characteristic of field vector is correctly classified is bigger, then illustrates that the degree of domain mixing is smaller；Domain classification result mistake it is general Rate is bigger namely domain grader pair the first source domain feature vector and first object characteristic of field vector classify, and correct probability is got over It is small, illustrate that the degree of domain mixing is bigger, therefore, to be based on the first domain grader to first object characteristic of field the first source domain of vector sum It is that the source domain data and target numeric field data that feature vector characterizes respectively are classified as a result, to first nerves network carry out parameter tune It is whole.

Herein, shown in Figure 3, it can determine operation by executing following domain Classification Loss, be classified according to domain to realize As a result, carrying out parameter adjustment to first nerves network：

S301：Determine current first source domain feature vector and the source domain data that first object characteristic of field vector characterizes respectively The domain Classification Loss classified with this domain of target numeric field data.

Herein, the degree of domain mixing is characterized by domain Classification Loss.The domain Classification Loss of source domain data refer to based on During first source domain feature vector and first object characteristic of field vector classify to source domain data and target numeric field data, point Class result is the quantity for the source domain data being assigned in aiming field.The domain Classification Loss of target numeric field data refers to based on the first source During characteristic of field vector sum first object characteristic of field vector classifies to source domain data and target numeric field data, classification results To be assigned to the quantity of the target numeric field data in source domain.Using first domain grader pair the first source domain feature vector and first After the source domain data and target numeric field data that target domain characterization vector characterizes respectively carry out domain classification, it will be able to obtain domain classification knot Fruit determines that source domain data close the corresponding domain Classification Loss of target numeric field data then according to domain classification results.

S302：It is not less than default differential threshold for the difference between the domain Classification Loss of nearest preset times, generates the Three feedback informations, and parameter tune is carried out to first nerves network based on the third feedback information.

Herein, it is constrained come the degree mixed to domain using default differential threshold.It is prestored in first domain grader The distribution for having domain belonging to the first source domain feature vector and first object characteristic of field vector difference, when the domain of nearest preset times is classified When difference between loss is not less than default differential threshold, then it is assumed that the also not up to stable state of domain classification, that is to say, that In the classification of certain domain, the first domain grader can correctly distinguish the first source domain feature vector and first object characteristic of field vector difference institute The domain of category, certain domain classification in, domain grader cannot correctly distinguish again the first source domain feature vector and first object characteristic of field to Domain belonging to amount difference, domain mixability is also unstable, then, need the parameter to first nerves network to be adjusted, therefore The excessive third feedback information of domain Classification Loss difference can be generated, and feeds back to first nerves network.First nerves network is connecing After receiving the excessive third feedback information of the domain Classification Loss difference, the parameter of itself can be adjusted.

S303：The use of first nerves network is that source domain data extract the first new source domain feature based on the parameter after adjustment Vector, and new first object characteristic of field vector is extracted for target numeric field data, and execute domain Classification Loss and determine operation, until poor Differential threshold great Yu not be preset, completes to train the epicycle of first nerves network based on the first domain grader.

Training based on the first domain grader to first nerves network, being will be according to the first domain grader to the first source domain Domain Classification Loss maintains in certain value determined by the classification results of feature vector and first object characteristic of field vector, to the greatest extent may be used It is to belong to source domain to still fall within aiming field on earth that Chu's target numeric field data and source domain data, which can be hard to tell, extracts the public spy of the two Sign.

Herein, it should be noted that when the difference between the domain Classification Loss of nearest preset times is less than default difference threshold When value, the suitable feedback information of domain Classification Loss can be also generated, and it is fed back into first nerves network.First nerves network exists After receiving the suitable feedback information of domain Classification Loss, also ladder can be made every effort to the parameter of itself into the smaller adjustment of line amplitude Degree drops to local optimum.

S103：Source domain data and target numeric field data are inputted into nervus opticus network, the second mesh is extracted for target numeric field data Mark characteristic of field vector；And otherness Feature capturing is carried out to source domain data and target numeric field data, keep the second target domain characterization vectorial Learn the difference characteristic of source domain data and target numeric field data.

When specific implementation, convolutional neural networks (Convolutional may be used in nervus opticus network Neural Network, CNN) to extract the second source domain feature vector for source domain data, extract the second target for target numeric field data Characteristic of field vector.

Optionally, first nerves network and nervus opticus network are the identical neural network of structure.

Source domain data are to carry the data of label, which is used to indicate the correct classification results of source domain data；Target Numeric field data is not carry the data of label.After by source domain data and target numeric field data input nervus opticus network, nervus opticus Network carries out source domain data and target numeric field data the feature learning of shared parameter.In this process, due to nervus opticus Network carries out supervised learning to source domain data, and carries out unsupervised learning to target numeric field data, is using same nervus opticus Network carries out source domain data and target numeric field data in the learning process of shared parameter, can constantly adjust institute in first nerves network The parameter used, to during to nervus opticus network training, allow the parameter of nervus opticus network by aiming field number According to influence so that nervus opticus network is to source domain data and target numeric field data after carrying out feature learning, to each source Numeric field data can be interfered by target numeric field data in the obtained second source domain feature vector of feature extraction, realizes source domain number It is mixed according between the domain of target numeric field data.

Meanwhile the second source domain feature vector is being extracted for source domain data, and extract the second aiming field spy for target numeric field data After sign vector, can also source domain data and target numeric field data be carried out with the capture of otherness feature, this makes to nervus opticus net When network training, target numeric field data that the second source domain feature vector for extracting for source domain data is subject to is interfered, that is, It is interfered by source domain data and aiming field data variance feature, to realize between source domain data and the domain of target numeric field data When fusion, the otherness domain of source domain data and target numeric field data is merged, finally so that the second source domain feature vector Learn the otherness feature to source domain data and target numeric field data.

To be that source domain data extract the second source before carrying out otherness Feature capturing to source domain data and target numeric field data Characteristic of field vector.

Shown in Figure 4, the embodiment of the present invention also provides a kind of special to source domain data and target numeric field data progress otherness Levy the specific method captured.This method includes：

S401：The second source domain feature vector of institute and the second target domain characterization vector are inputted into the second domain grader.

S402：Table is distinguished to the second source domain feature vector and the second target domain characterization vector according to the second domain grader The source domain data of sign and the domain classification results of target numeric field data carry out parameter adjustment to nervus opticus network.

It is anti-not carrying out gradient to the second source domain feature vector and the second target domain characterization vector when specific implementation To processing, but the two is directly inputted into the second domain grader, using second domain grader pair the second source domain feature vector and The source domain data and target numeric field data that second target domain characterization vector characterizes respectively carry out domain classification, obtained domain Classification Loss It is smaller, the domain belonging to source domain data and target numeric field data can be distinguished as far as possible.Nervus opticus network is captured To the otherness feature between source domain data and target numeric field data, distance between the two is zoomed out.

Specifically, shown in Figure 5, operation can be determined by executing following domain Classification Loss, to realize according to domain point Class to nervus opticus network as a result, carry out parameter adjustment：

S501：Determine current second source domain feature vector and the source domain data that the second target domain characterization vector characterizes respectively The domain Classification Loss classified with this domain of target numeric field data.

Herein, it carries out domain to characterize the second source domain feature vector and the second target domain characterization vector by domain Classification Loss and mixes The degree of conjunction.The domain Classification Loss of source domain data refers to right based on the second source domain feature vector and the second target domain characterization vector During source domain data and target numeric field data are classified, classification results are the quantity of the source domain data of target numeric field data.Mesh Mark numeric field data domain Classification Loss refer to based on the second source domain feature vector and the second target domain characterization vector to source domain data During being classified with target numeric field data, classification results are the quantity of the target numeric field data of source domain data.Using region The source domain data and target numeric field data that grader pair the second source domain feature vector and the second target domain characterization vector characterize respectively After carrying out domain classification, it will be able to obtain domain classification results, then according to domain classification results, determine that source domain data close target numeric field data Corresponding domain Classification Loss.

S502：For the situation of domain classification results mistake, the 4th feedback information is generated, and based on the 4th feedback information pair the Two neural networks carry out parameter adjustment.

Herein, due to the correctness of domain classification results to be ensured, only domain classification results are correct, can just zoom out source domain The distance between data and target numeric field data, namely the otherness data between source domain data and target numeric field data are extracted, because This will generate the 4th feedback information, and based on the 4th feedback information to nervus opticus network when the classification results mistake of domain Carry out parameter adjustment.

S503：The use of nervus opticus network is that source domain data extract the second new source domain feature based on the parameter after adjustment Vector, and be new the second target domain characterization vector of target numeric field data extraction, and execute domain Classification Loss and determine operation.

Up to domain classification results are correct or the accuracy of domain classification results reaches preset threshold value.

S104：First object characteristic of field vector and the second target domain characterization vector are clustered respectively.

S105：According to cluster as a result, and the first classification results, to first nerves network and the first grader into Row epicycle is trained.

When specific implementation, according to cluster as a result, and the first classification results, to first nerves network and First grader carry out epicycle training, be will according to cluster as a result, to the parameter of first nerves network in the training process into Row adjustment, and according to the first classification results, the parameter of degree first nerves network and the first grader in the training process is adjusted It is whole.

One：According to cluster as a result, when being adjusted to the parameter of first nerves network in the training process, due to being To use first nerves network is that target numeric field data extracts first object characteristic of field vector, and it is target to use nervus opticus network Numeric field data extracts the second target domain characterization vector, and the final training result of model is to be carried out as far as possible just to target numeric field data True classification, therefore during being trained to first nerves network and nervus opticus network, it is ensured that the two carries respectively The similitude of the distribution of first object characteristic of field vector sum the second target domain characterization vector taken is within limits.

Here the similitude of the distribution of first object characteristic of field vector sum the second target domain characterization vector, refers to target Numeric field data includes multiple data.This multiple data belongs to target numeric field data.To each target numeric field data extract feature to It is to extract feature vector respectively for each of this multiple data when amount；The first object characteristic of field vector of multiple data It is similar with the distribution of the second target domain characterization vector spatially.

Therefore, first object characteristic of field vector sum the second target domain characterization vector is clustered respectively herein.

It is shown in Figure 6, when carrying out epicycle training to first nerves network according to cluster result, including：

S601：According to first object characteristic of field vector clustered as a result, generate the first adjacency matrix.

It specifically, can be by each first object domain due to when being clustered first object characteristic of field vector Data regard the point being mapped in higher dimensional space as, and according to distance between points, cluster operation is carried out to these points, will be away from It is divided into same class from the point within predetermined threshold value.Then according to cluster operation as a result, forming distance between point-to-point The first adjacency matrix.

In the first adjacency matrix, if two points belong to same class in cluster, distance between the two is 1；Such as Two points of fruit are not belonging to same class in cluster, then the distance between 2 points are 0.

For example, target numeric field data has 5, obtained first object characteristic of field vector is respectively：1、2、3、4、5.Wherein, right The result that first object characteristic of field vector is clustered is：{ 1,3 }, { 2 }, { 4,5 }, the then adjacency matrix formed are：

S602：According to the second target domain characterization vector clustered as a result, generate the second adjacency matrix.

Herein, the method for generating the second adjacency matrix is similar with the method for the first adjacency matrix is generated, and details are not described herein.

S603：According to the similarity between the first adjacency matrix and the second adjacency matrix, first nerves network is carried out Training.

Herein, according to the similarity between the first adjacency matrix and the second adjacency matrix, to first nerves network into When row training, following similarity calculation operation is executed, until the similarity between the first adjacency matrix and the second adjacency matrix is small In preset similarity threshold.

Shown in Figure 7, similarity calculation operation includes：

S701：Calculate the similarity between currently available the first adjacency matrix and the second adjacency matrix.

It is similar between currently available the first adjacency matrix and the second adjacency matrix calculating when specific implementation When spending, the mark of the mark and the second adjacency matrix of the first adjacency matrix is calculated, the mark of the first adjacency matrix and the second adjacency matrix The distance between mark is closer, then the similarity between the first adjacency matrix and the second adjacency matrix is higher.To the first adjoining square When the mark of battle array and the distance between the mark of the second adjacency matrix are solved, the mark of the first adjacency matrix and second can be abutted Difference between the mark of matrix is as the similarity namely the first adjacency matrix between the first adjacency matrix and the second adjacency matrix Absolute value of the difference between mark and the mark of the second adjacency matrix is bigger, and the similarity of the first adjacency matrix and the second adjacency matrix is got over It is low.

S702：The case where being not less than preset similarity threshold for similarity, generates the first feedback information, and based on the One feedback information carries out parameter adjustment to first nerves network and nervus opticus network.

S703：The use of first nerves network is that aiming field number extracts new first object domain spy based on the parameter after adjustment Sign vector, and new the second target domain characterization vector is extracted for target numeric field data using nervus opticus network；

S704：New first object characteristic of field vector is clustered, the first new adjacency matrix is generated, and, to new The second target domain characterization vector clustered, generate the second new adjacency matrix, and execute similarity calculation operation again.

Since the similarity between the first adjacency matrix and the second adjacency matrix is higher, then the first adjacency matrix characterization pair The classification results and the second adjacency matrix that first object characteristic of field vector is classified characterize vectorial to the second target domain characterization The classification results classified are more similar, therefore right according to the similarity between the first adjacency matrix and the second adjacency matrix First nerves network carries out parameter adjustment so that first nerves network is special in the difference between source domain data and target numeric field data Sign is limited, and model convergence is accelerated.

Second, according to the first classification results, to the parameter of first nerves network and the first grader in the training process into Row adjustment.

Specifically, according to the first classification results, to first nerves network and the first grader in the training process When parameter is adjusted, following sort operation is executed, it is known that the first obtained classification results are correct, then complete to be based on nervus opticus Network trains the epicycle of first nerves network and the first grader.

The sort operation includes：

Classified to the first source domain feature vector currently extracted using the first grader；

For the situation of classification results mistake, the second feedback information is generated, and will be in the second feedback information, to first nerves The parameter of network and the first grader is adjusted；

The use of first nerves network is that source domain data extract the first new source domain feature vector based on the parameter after adjustment, And sort operation is executed again.

Herein, since disaggregated model will classify to target numeric field data, to ensure the premise of the correctness of classification as possible Under, while using the otherness feature of source domain and aiming field holder, it is ensured that the classification results to source domain data are correct , therefore the parameter of first nerves network and the first grader is constrained using the first classification results.

S106：By carrying out more wheel training to first nerves network and the first grader, disaggregated model is obtained.

When specific implementation, a wheel is carried out to first nerves network and the first grader and is trained, refers to using one group Source domain data and target numeric field data are trained first nerves network and the first grader.Later, it also will continue to input multigroup Numeric field data and target numeric field data are trained first nerves network and the first grader, until the first god met the requirements Through network and the first grader, and using obtained first nerves network and the first grader as obtained disaggregated model.

Priority, which is had no, in the above process, between S102 and S103 executes sequence.

In disaggregated model training method provided in an embodiment of the present invention, using first nerves network, extracted for source domain data First source domain feature vector can be to source domain data and aiming field when extracting first object characteristic of field vector for target numeric field data Data carry out the capture of common characteristic, make the common characteristic of first object characteristic of field vector study source domain data and target numeric field data While, it is that target numeric field data extracts the second target domain characterization vector that can utilize nervus opticus network；Can to source domain data and Target numeric field data carries out otherness Feature capturing, makes the difference of the second target domain characterization vector study source domain data and target numeric field data Different feature instructs first nerves network and the first grader then according to the same characteristic features and otherness feature captured Practice, obtained disaggregated model can either use the same characteristic features between source domain and aiming field, also can be by source domain and mesh Difference characteristic between mark domain is utilized, and then the disaggregated model can obtain more accurate classification results.

In another embodiment, since nervus opticus network to be based on is to first nerves network and the first grader, While to first nerves network training, can also training be synchronized to nervus opticus network.To nervus opticus network into When row training, for a moment according to cluster as a result, being adjusted to the parameter of nervus opticus network.Second, also to ensure Nervus opticus network is accurate to the classification results of source domain data.

Therefore after extracting the second source domain feature vector for source domain data, also by the second source domain feature vector input the Two graders obtain the second classification results；

According to the second classification results of the source domain data to the second source domain feature vector characterization of the second grader, to second Neural network carries out parameter adjustment.

Herein, according to the second grader to the second classification results of the source domain data of second source domain feature vector characterization, Parameter adjustment is carried out to nervus opticus network, is specifically included：

Following sort operations are executed, until obtained the second classification results are correct, are then completed to nervus opticus network and the The epicycle of two graders is trained；

Sort operation includes：

Classified to the second source domain feature vector currently extracted using the second grader；

For the situation of classification results mistake, the 5th feedback information is generated, and will be in the 5th feedback information, to nervus opticus The parameter of network and the second grader is adjusted；

The use of nervus opticus network is that source domain data extract the second new source domain feature vector based on the parameter after adjustment, And sort operation is executed again.

The process for carrying out parameter adjustment to nervus opticus network and the second grader using the second classification results, and uses the The process that one classification results carry out first nerves network and the first grader parameter adjustment is similar, and details are not described herein.

Shown in Figure 8, the embodiment of the present invention also provides a kind of sorting technique, and this method includes：

S801：Obtain data to be sorted；

S802：Data to be sorted are input to the classification that the disaggregated model training method that the embodiment of the present application is provided obtains In model, the classification results of data to be sorted are obtained；

Wherein, disaggregated model includes：First nerves network and the first grader.

Based on same inventive concept, classification mould corresponding with disaggregated model training method is additionally provided in the embodiment of the present invention Type training device, since the principle that the device in the embodiment of the present invention solves the problems, such as is instructed with the above-mentioned disaggregated model of the embodiment of the present invention It is similar to practice method, therefore the implementation of device may refer to the implementation of method, overlaps will not be repeated.

It is shown in Figure 9, disaggregated model training device provided in an embodiment of the present invention, including：

Acquisition module 901, for obtaining the source domain data for carrying label and the target numeric field data for not carrying label；

First processing module 902 is source domain number for source domain data and target numeric field data to be inputted first nerves network According to the first source domain feature vector of extraction, first object characteristic of field vector is extracted for target numeric field data；And to source domain data and target Numeric field data carries out common characteristic capture, makes the shared spy of first object characteristic of field vector study source domain data and target numeric field data Sign；And the first source domain feature vector is inputted into the first grader and obtains the first classification results；

Second processing module 903 is aiming field for source domain data and target numeric field data to be inputted nervus opticus network Data extract the second target domain characterization vector；And otherness Feature capturing is carried out to source domain data and target numeric field data, make second Target domain characterization vector learns the difference characteristic of source domain data and target numeric field data；

Cluster module 904, for gathering respectively to first object characteristic of field vector and the second target domain characterization vector Class；

Training module 905, for according to cluster as a result, and the first classification results, to first nerves network and First grader carries out epicycle training；By carrying out more wheel training to first nerves network and the first grader, classification mould is obtained Type.

The embodiment of the present invention utilizes first nerves network, and the first source domain feature vector is extracted for source domain data, is aiming field When data extract first object characteristic of field vector, can source domain data and target numeric field data be carried out with the capture of common characteristic, made While the common characteristic of first object characteristic of field vector study source domain data and target numeric field data, nervus opticus network can be utilized The second target domain characterization vector is extracted for target numeric field data；Otherness feature can be carried out to source domain data and target numeric field data to catch It catches, makes the difference characteristic of the second target domain characterization vector study source domain data and target numeric field data, then according to the phase captured With feature and otherness feature, first nerves network and the first grader are trained, obtained disaggregated model, can either be incited somebody to action Same characteristic features between source domain and aiming field use, and also can the difference characteristic between source domain and aiming field be carried out profit With, and then the disaggregated model can obtain more accurate classification results.

Optionally, first processing module 902 is used to carry out common characteristic to source domain data and target numeric field data such as under type It captures：

The first source domain feature vector is being extracted for source domain data, and first object characteristic of field vector is extracted for target numeric field data Later, the first source domain feature vector and first object characteristic of field vector are subjected to gradient reverse process；

The the first source domain feature vector for carrying out gradient reverse process and first object characteristic of field vector are inputted into the first domain point Class device；

The source characterized respectively according to first domain grader pair the first source domain feature vector and first object characteristic of field vector The domain classification results of numeric field data and target numeric field data carry out parameter adjustment to first nerves network.

Optionally, source domain data and target numeric field data are being inputted nervus opticus network by Second processing module 903, to source After numeric field data and target numeric field data carry out feature learning, the second source domain feature vector is extracted for source domain data；

Second processing module 903 is caught for carrying out otherness feature to source domain data and target numeric field data in the following way It catches：

Second source domain feature vector and the second target domain characterization vector are inputted into the second domain grader；

According to characterizing respectively to the second source domain feature vector and the second target domain characterization vector for the second domain grader The domain classification results of source domain data and target numeric field data carry out parameter adjustment to nervus opticus network.

Optionally, further include the second training module, after extracting the second source domain feature vector for source domain data, by the Two source domain feature vectors input the second grader and obtain the second classification results；

According to the second classification results of the source domain data to the second source domain feature vector characterization of the second grader, to second Neural network carries out parameter adjustment.

Optionally, training module 905, specifically for according to first object characteristic of field vector clustered as a result, raw At the first adjacency matrix；

According to the second target domain characterization vector clustered as a result, generate the second adjacency matrix；

According to the similarity between the first adjacency matrix and the second adjacency matrix, and the first classification results, to One neural network and the first grader carry out epicycle training.

Optionally, training module 905 are specifically used for：Following similarity calculation operation and sort operation are executed, until the Similarity between one adjacency matrix and the second adjacency matrix is less than preset similarity threshold, and the first obtained classification knot Fruit is correct, then completes to train the epicycle of first nerves network and the first grader based on nervus opticus network；

Similarity calculation operates：

Calculate the similarity between currently available the first adjacency matrix and the second adjacency matrix；

The case where being not less than preset similarity threshold for similarity, generates the first feedback information, and anti-based on first Feedforward information carries out parameter adjustment to first nerves network and nervus opticus network；

Based on the parameter after adjustment, using first nerves network be aiming field number extract new first object characteristic of field to Amount, and new the second target domain characterization vector is extracted for target numeric field data using nervus opticus network；

New first object characteristic of field vector is clustered, the first new adjacency matrix is generated, and, to new second Target domain characterization vector is clustered, and the second new adjacency matrix is generated, and executes similarity calculation operation again；

Sort operation includes：

Classified to the first source domain feature vector currently extracted using the first grader；

For the situation of classification results mistake, the second feedback information is generated, and will be in the second feedback information, to first nerves The parameter of network and the first grader is adjusted；

The use of first nerves network is that source domain data extract the first new source domain feature vector based on the parameter after adjustment, And sort operation is executed again.

Optionally, training module 905, specifically for calculating the first currently available adjacency matrix and according to following manner Similarity between two adjacency matrix：

Calculate the mark of the mark and the second adjacency matrix of the first adjacency matrix；

Difference between the mark of first adjacency matrix and the mark of the second adjacency matrix is adjacent as the first adjacency matrix and second Connect the similarity between matrix.

Optionally, the second training module is additionally operable to carry out parameter adjustment to nervus opticus network using following manner：

It executes following domain Classification Loss and determines operation：

Determine the source domain data and mesh that current second source domain feature vector and the second target domain characterization vector characterize respectively Mark the domain Classification Loss of this domain classification of numeric field data；

For the situation of domain classification results mistake, the 4th feedback information is generated, and based on the 4th the second god of feedback information pair Parameter adjustment is carried out through network；

The use of nervus opticus network is that source domain data extract the second new source domain feature vector based on the parameter after adjustment, And be new the second target domain characterization vector of target numeric field data extraction, and execute domain Classification Loss and determine operation.

Further embodiment of this invention also provides a kind of sorter, shown in Figure 10, what the embodiment of the present invention was provided Sorter includes：

Data acquisition module 1001 to be sorted, for obtaining data to be sorted；

Sort module 1002, for being input to data to be sorted by disaggregated model training provided by the embodiments of the present application In the disaggregated model that method obtains, the classification results of data to be sorted are obtained；Wherein, disaggregated model includes：First nerves network With the first grader.

Corresponding to the disaggregated model training method in Fig. 1, the embodiment of the present invention additionally provides a kind of computer equipment, such as schemes Shown in 11, which includes memory 1000, processor 2000 and is stored on the memory 1000 and can be in the processor The computer program run on 2000, wherein above-mentioned processor 2000 realizes above-mentioned classification mould when executing above computer program The step of type training method.

Specifically, above-mentioned memory 1000 and processor 2000 can be general memory and processor, not do here It is specific to limit, when the computer program of 2000 run memory 1000 of processor storage, it is able to carry out above-mentioned disaggregated model instruction Practice method, to solve in disaggregated model training process be can utilize it is identical between source domain data and target numeric field data Feature, and can not utilize source domain data and not disaggregated model caused by the variance data between poster data to target numeric field data Existing error problem when being classified, and then reach and the same characteristic features between source domain and aiming field can either be used, Also the difference characteristic between source domain and aiming field can be utilized, the disaggregated model of training gained can obtain more accurately Classification results effect.

Corresponding to the disaggregated model training method in Fig. 1, the embodiment of the present invention additionally provides a kind of computer-readable storage Medium is stored with computer program on the computer readable storage medium, is executed when which is run by processor The step of stating disaggregated model training method.

Specifically, which can be general storage medium, such as mobile disk, hard disk, on the storage medium Computer program when being run, above-mentioned disaggregated model training method is able to carry out, to solve to disaggregated model training It is the same characteristic features that can be utilized between source domain data and target numeric field data in journey, and source domain data and not poster number can not be utilized The existing error problem when classifying to target numeric field data of disaggregated model caused by variance data between, and then reach Same characteristic features between source domain and aiming field can either be used, it also can be by the difference characteristic between source domain and aiming field It is utilized, the disaggregated model of training gained can obtain the effect of more accurate classification results.

The computer journey for disaggregated model training method, device and the sorting technique and device that the embodiment of the present invention is provided Sequence product, including the computer readable storage medium of program code is stored, the instruction that said program code includes can be used for holding Method described in row previous methods embodiment, specific implementation can be found in embodiment of the method, and details are not described herein.

It is apparent to those skilled in the art that for convenience and simplicity of description, the system of foregoing description It with the specific work process of device, can refer to corresponding processes in the foregoing method embodiment, details are not described herein.

It, can be with if the function is realized in the form of SFU software functional unit and when sold or used as an independent product It is stored in a computer read/write memory medium.Based on this understanding, technical scheme of the present invention is substantially in other words The part of the part that contributes to existing technology or the technical solution can be expressed in the form of software products, the meter Calculation machine software product is stored in a storage medium, including some instructions are used so that a computer equipment (can be People's computer, server or network equipment etc.) it performs all or part of the steps of the method described in the various embodiments of the present invention. And storage medium above-mentioned includes：USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), arbitrary access are deposited The various media that can store program code such as reservoir (RAM, Random Access Memory), magnetic disc or CD.

The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any Those familiar with the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all contain Lid is within protection scope of the present invention.Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (10)

1. a kind of disaggregated model training method, which is characterized in that this method includes：

Obtain the source domain data for carrying label and the target numeric field data for not carrying label；

The source domain data and the target numeric field data are inputted into first nerves network, the first source is extracted for the source domain data Characteristic of field vector extracts first object characteristic of field vector for the target numeric field data；And

Common characteristic capture is carried out to the source domain data and the target numeric field data, makes the first object characteristic of field vector Practise the common characteristic of the source domain data and the target numeric field data；And

The first source domain feature vector is inputted into the first grader and obtains the first classification results；

The source domain data and the target numeric field data are inputted into nervus opticus network, second is extracted for the target numeric field data Target domain characterization vector；And

Otherness Feature capturing is carried out to the source domain data and the target numeric field data, makes the second target domain characterization vector Learn the difference characteristic of the source domain data and the target numeric field data；

First object characteristic of field vector and the second target domain characterization vector are clustered respectively；

According to cluster as a result, and first classification results, to the first nerves network and first grader Carry out epicycle training；

By carrying out more wheel training to the first nerves network and first grader, disaggregated model is obtained.

2. according to the method described in claim 1, it is characterized in that, in the following way to the source domain data and the target Numeric field data carries out common characteristic capture：

The first source domain feature vector is being extracted for the source domain data, and first object characteristic of field is extracted for the target numeric field data After vector, the first source domain feature vector and the first object characteristic of field vector are subjected to gradient reverse process；

The the first source domain feature vector for carrying out gradient reverse process and first object characteristic of field vector are inputted into the first domain grader；

The institute characterized respectively according to first domain grader pair the first source domain feature vector and first object characteristic of field vector The domain classification results for stating source domain data and the target numeric field data carry out parameter adjustment to the first nerves network.

3. according to the method described in claim 1, it is characterized in that, described by the source domain data and the target numeric field data Nervus opticus network is inputted, after carrying out feature learning to the source domain data and the target numeric field data, further includes：

The second source domain feature vector is extracted for the source domain data；

Otherness Feature capturing is carried out to the source domain data and the target numeric field data in the following way：

The second source domain feature vector and the second target domain characterization vector are inputted into the second domain grader；

According to characterizing respectively to the second source domain feature vector and the second target domain characterization vector for second domain grader The domain classification results of the source domain data and the target numeric field data carry out parameter adjustment to the nervus opticus network.

4. according to the method described in claim 3, it is characterized in that, for the source domain data extract the second source domain feature vector it Afterwards, further include：

The second source domain feature vector is inputted into the second grader and obtains the second classification results；

According to the second classification results of the source domain data to the second source domain feature vector characterization of second grader, to described Nervus opticus network carries out parameter adjustment.

5. according to the method described in any of claim 1 to 4, which is characterized in that it is described according to cluster as a result, and institute The first classification results stated carry out epicycle training to the first nerves network and first grader, specifically include：

According to the first object characteristic of field vector clustered as a result, generate the first adjacency matrix；

According to the second target domain characterization vector clustered as a result, generate the second adjacency matrix；

According between first adjacency matrix and the second adjacency matrix similarity and first classification results, Epicycle training is carried out to the first nerves network and first grader.

6. according to the method described in claim 5, it is characterized in that, described abut according to first adjacency matrix and second Similarity between matrix and first classification results, to the first nerves network and first grader Epicycle training is carried out, is specifically included：

Following similarity calculation operation and sort operation are executed, until the phase between the first adjacency matrix and the second adjacency matrix It is correct less than preset similarity threshold, and the first obtained classification results like degree, then it completes to be based on nervus opticus network pair The epicycle of the first nerves network and first grader is trained；

The similarity calculation operates：

Calculate the similarity between currently available the first adjacency matrix and the second adjacency matrix；

The case where being not less than preset similarity threshold for similarity, generates the first feedback information, and anti-based on described first Feedforward information carries out parameter adjustment to the first nerves network and the nervus opticus network；

Based on the parameter after adjustment, using first nerves network be the aiming field number extract new first object characteristic of field to Amount, and new the second target domain characterization vector is extracted for the target numeric field data using nervus opticus network；

New first object characteristic of field vector is clustered, the first new adjacency matrix is generated, and, to the second new target Characteristic of field vector is clustered, and the second new adjacency matrix is generated, and executes the similarity calculation operation again；

The sort operation includes：

Classified to the first source domain feature vector currently extracted using first grader；

For the situation of classification results mistake, the second feedback information is generated, and will be in second feedback information, to described first Neural network and the parameter of first grader are adjusted；

The use of first nerves network is that the source domain data extract the first new source domain feature vector based on the parameter after adjustment, And the sort operation is executed again.

7. according to the method described in claim 6, it is characterized in that, described calculate currently available the first adjacency matrix and second Similarity between adjacency matrix, specifically includes：

Calculate the mark of the mark and second adjacency matrix of the first adjacency matrix；

Using the difference between the mark of the first adjacency matrix and the mark of second adjacency matrix as the first adjacency matrix and described Similarity between two adjacency matrix.

8. according to the method described in claim 2, it is characterized in that, described according to first domain the first source domain of grader pair spy The domain classification knot for the source domain data and the target numeric field data that sign vector and first object characteristic of field vector characterize respectively Fruit carries out parameter adjustment to the first nerves network, specifically includes：

It executes following domain Classification Loss and determines operation：

Determine the source domain data and mesh that current first source domain feature vector and the first object characteristic of field vector characterize respectively Mark the domain Classification Loss of this domain classification of numeric field data；

It is not less than default differential threshold for the difference between the domain Classification Loss of nearest preset times, generates third feedback letter Breath, and parameter adjustment is carried out to the first nerves network based on the third feedback information；

The use of first nerves network is that the source domain data extract the first new source domain feature vector based on the parameter after adjustment, And new first object characteristic of field vector is extracted for the target numeric field data, and execute the domain Classification Loss and determine operation, directly It is more than default differential threshold to difference, completes to train the epicycle of the first nerves network based on first domain grader.

9. according to the method described in claim 3, it is characterized in that, it is described according to second domain grader to the second source domain Classify in the domain for the source domain data and the target numeric field data that feature vector and the second target domain characterization vector characterize respectively As a result, carrying out parameter adjustment to the nervus opticus network, specifically include：

It executes following domain Classification Loss and determines operation：

Determine the source domain data and mesh that current second source domain feature vector and the second target domain characterization vector characterize respectively Mark the domain Classification Loss of this domain classification of numeric field data；

For the situation of domain classification results mistake, the 4th feedback information is generated, and based on the 4th feedback information to described the Two neural networks carry out parameter adjustment；

The use of nervus opticus network is that the source domain data extract the second new source domain feature vector based on the parameter after adjustment, And be new the second target domain characterization vector of target numeric field data extraction, and execute the domain Classification Loss and determine operation.

10. a kind of sorting technique, which is characterized in that this method includes：

Obtain data to be sorted；

The data to be sorted are input to the classification obtained by the disaggregated model training method of claim 1-9 any one In model, the classification results of the data to be sorted are obtained；

Wherein, the disaggregated model includes：The first nerves network and first grader.