CN110009038B - Training method and device for screening model and storage medium - Google Patents
Training method and device for screening model and storage medium Download PDFInfo
- Publication number
- CN110009038B CN110009038B CN201910269536.6A CN201910269536A CN110009038B CN 110009038 B CN110009038 B CN 110009038B CN 201910269536 A CN201910269536 A CN 201910269536A CN 110009038 B CN110009038 B CN 110009038B
- Authority
- CN
- China
- Prior art keywords
- domain
- sub
- classifier
- target
- source
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
The embodiment of the invention provides a training device and method for screening models and a computer readable storage medium. Wherein the training device of screening model includes: the clustering unit is used for dividing the source field data by utilizing a clustering algorithm and dividing the source field into a plurality of sub-fields; the learning unit is used for respectively carrying out transfer learning on the plurality of sub-fields to obtain a plurality of sub-field classifiers used for classifying the target field data; and the integration unit is used for integrating the plurality of sub-field classifiers to obtain the target field classifier. According to the embodiment of the invention, because finer-grained knowledge transfer is carried out in a plurality of sub-fields, a classifier with better effect can be obtained, and meanwhile, the data annotation cost in the target field is reduced through multi-source transfer learning.
Description
Technical Field
The present invention relates to the field of information technology, and in particular, to a training method and apparatus for a screening model, and a computer-readable storage medium.
Background
When the conventional artificial intelligence screening system is deployed in a new application environment, because the screening models used in the new application environment are different, the data sources are different, for example, the brightness and the contrast of an image may be different, and the performance is affected by directly using source field data. Therefore, large-scale annotation data needs to be collected in a new application environment for training the screening model, so that the screening model with higher prediction performance is obtained. The data collection and annotation for this migration process is costly.
In addition, due to differences in data acquisition devices, such as differences in the variety of photographing devices and in photographing angles, lighting conditions, and the like, data of a source domain may be subject to different data distributions. The distribution difference of the data among each other can cause the classifier trained in the source domain to be difficult to obtain satisfactory effect in new application environment.
Current approaches to deploying screening systems for new application environments include the following:
(1) and (3) collecting a large amount of data, and then asking experts to manually label to obtain training data with labels, so as to learn the classifier.
(2) And training by directly utilizing the existing source field marking data to obtain a classifier, and classifying the data in the new application environment.
(3) And learning a classifier which can be used for classifying the target field data by using a single-source migration learning technology and using the source field data with labels and the target field data without labels.
The main drawbacks of the above deployment approaches include:
(1) the method for manually labeling the data in the new application environment needs a lot of manpower, material resources and time, and the cost of data labeling is high. Furthermore, this approach does not make efficient use of existing source domain data.
(2) The method for directly utilizing the existing source field labeling data has larger distribution difference between the existing source field labeling data and the data in the new environment because the existing source field labeling data and the data in the new environment come from different acquisition equipment and crowds. This difference often results in the source domain trained classifier failing to perform satisfactorily in new application environments.
(3) In the single-source migration learning method, because of the differences in the aspects of equipment, lighting conditions, crowds, operation methods of shooting personnel and the like in the data acquisition process of the source field, a plurality of sub-fields may exist in the source field, and the data of each sub-field has a certain distribution difference. By adopting the single-source migration learning algorithm, the distribution difference between the sub-fields in the field can be ignored, so that the performance of the source field classifier is restricted.
Disclosure of Invention
Embodiments of the present invention provide a training method and apparatus for a screening model, and a computer-readable storage medium, so as to solve one or more technical problems in the prior art.
In a first aspect, an embodiment of the present invention provides a training apparatus for screening a model, including:
the clustering unit is used for dividing the source field data by utilizing a clustering algorithm and dividing the source field into a plurality of sub-fields;
the learning unit is used for respectively carrying out transfer learning on the plurality of sub-fields to obtain a plurality of sub-field classifiers used for classifying the target field data;
and the integration unit is used for integrating the plurality of sub-field classifiers to obtain the target field classifier.
In one embodiment, the clustering unit is configured to:
extracting the characteristics of the source field data;
and performing clustering analysis on the features by using a clustering algorithm, thereby dividing the source field into a plurality of sub-fields.
In one embodiment, the learning unit includes:
a first training subunit, configured to train a source-domain encoder and the sub-domain classifier in the plurality of sub-domains, respectively;
a second training subunit, configured to train a domain classifier and a target domain encoder using a generative confrontation network in each of the plurality of sub-domains;
wherein the input information of the source-domain encoder comprises first samples derived from the sub-domain, and the output information of the source-domain encoder comprises feature vectors extracted from the first samples;
the sub-domain classifier is used for classifying the feature vectors;
the input information of the target domain encoder comprises second samples derived from the target domain, and the output information of the target domain encoder comprises feature vectors extracted from the second samples;
the domain classifier is configured to discriminate whether the input feature vector is derived from the sub-domain or the target domain.
In one embodiment, the second training subunit is configured to:
inputting the feature vector output by the source domain encoder and the feature vector output by the target domain encoder into the domain classifier, and judging whether the input feature vector is from the sub-domain or the target domain by the domain classifier;
and if the domain classifier can not judge that the input feature vector is from the sub-domain or the target domain, using the target domain encoder and the sub-domain classifier for classifying the target domain data.
In one embodiment, the integrated unit is configured to:
and taking the average value of the classification results of the plurality of sub-domain classifiers as the classification result of the target domain classifier.
In a second aspect, an embodiment of the present invention provides a training method for a screening model, including:
dividing the source field data by using a clustering algorithm, and dividing the source field into a plurality of sub-fields;
respectively carrying out transfer learning on the plurality of sub-fields to obtain a plurality of sub-field classifiers for classifying the target field data;
and integrating a plurality of the sub-field classifiers to obtain a target field classifier.
In one embodiment, the dividing the source domain data into a plurality of sub-domains by using a clustering algorithm includes:
extracting the characteristics of the source field data;
and performing clustering analysis on the features by using a clustering algorithm, thereby dividing the source field into a plurality of sub-fields.
In one embodiment, performing migration learning on a plurality of the sub-domains respectively to obtain a plurality of sub-domain classifiers for classifying target domain data includes:
training a source domain encoder and the sub-domain classifier in a plurality of the sub-domains respectively;
training a domain classifier and a target domain encoder by using a generative confrontation network in a plurality of the sub-domains respectively;
wherein the input information of the source-domain encoder comprises first samples derived from the sub-domain, and the output information of the source-domain encoder comprises feature vectors extracted from the first samples;
the sub-domain classifier is used for classifying the feature vectors;
the input information of the target domain encoder comprises second samples derived from the target domain, and the output information of the target domain encoder comprises feature vectors extracted from the second samples;
the domain classifier is configured to discriminate whether the input feature vector is derived from the sub-domain or the target domain.
In one embodiment, training a domain classifier and a target domain encoder using a generative confrontation network in a plurality of the sub-domains, respectively, comprises:
inputting the feature vector output by the source domain encoder and the feature vector output by the target domain encoder into the domain classifier, and judging whether the input feature vector is from the sub-domain or the target domain by the domain classifier;
and if the domain classifier can not judge that the input feature vector is from the sub-domain or the target domain, using the target domain encoder and the sub-domain classifier for classifying the target domain data.
In one embodiment, integrating a plurality of the sub-domain classifiers to obtain a target domain classifier includes:
and taking the average value of the classification results of the plurality of sub-domain classifiers as the classification result of the target domain classifier.
In a third aspect, an embodiment of the present invention provides a training device for screening models, where functions of the device may be implemented by hardware, or may be implemented by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above-described functions.
In one possible design, the apparatus includes a processor and a memory in a structure, the memory is used for storing a program for supporting the apparatus to execute the training method of the screening model, and the processor is configured to execute the program stored in the memory. The apparatus may also include a communication interface for communicating with other devices or a communication network.
In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium for storing computer software instructions for a training apparatus for screening models, which includes a program for executing the training method for screening models.
One of the above technical solutions has the following advantages or beneficial effects: because the knowledge transfer with finer granularity is carried out in a plurality of sub-fields, a classifier with better effect can be obtained, and meanwhile, the data labeling cost in the target field is reduced through multi-source transfer learning.
Another technical scheme in the above technical scheme has the following advantages or beneficial effects: by utilizing the generative countermeasure network and the multi-source transfer learning algorithm, the feature subspace shared among the fields can be learned, so that the difference among data distribution of different fields is reduced.
The foregoing summary is provided for the purpose of description only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the present invention will be readily apparent by reference to the drawings and following detailed description.
Drawings
In the drawings, like reference numerals refer to the same or similar parts or elements throughout the several views unless otherwise specified. The figures are not necessarily to scale. It is appreciated that these drawings depict only some embodiments in accordance with the disclosure and are therefore not to be considered limiting of its scope.
Fig. 1 shows a block diagram of a training apparatus for screening a model according to an embodiment of the present invention.
Fig. 2 shows a schematic fundus view of a normal eyeball.
Fig. 3 shows a schematic view of the fundus with diabetic retinopathy.
Fig. 4 shows a block diagram of a learning unit of a training apparatus for screening a model according to an embodiment of the present invention.
FIG. 5 shows a flow diagram of a training method of a screening model according to an embodiment of the present invention.
FIG. 6 illustrates a flow chart of a training method of a screening model according to another embodiment of the present invention.
FIG. 7 shows a flow chart of a training method of a screening model according to yet another embodiment of the present invention.
Fig. 8 is a block diagram illustrating a structure of a training apparatus for screening a model according to still another embodiment of the present invention.
Detailed Description
In the following, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments may be modified in various different ways, all without departing from the spirit or scope of the present invention. Accordingly, the drawings and description are to be regarded as illustrative in nature, and not as restrictive.
Fig. 1 shows a block diagram of a training apparatus for screening a model according to an embodiment of the present invention. As shown in fig. 1, the training device for screening a model according to an embodiment of the present invention includes:
a clustering unit 100, configured to divide source domain data by using a clustering algorithm, and divide a source domain into multiple sub-domains;
a learning unit 200, configured to perform migration learning on the multiple sub-fields respectively to obtain multiple sub-field classifiers for classifying target field data;
an integrating unit 300, configured to integrate the multiple sub-domain classifiers to obtain a target domain classifier.
In practical applications, due to the change of screening models, the change of data sources, and the like, the application environment of the artificial intelligence screening system may change. The original application environment may be generally referred to as a source domain and the new application environment may be generally referred to as a target domain. The embodiment of the invention utilizes the transfer learning to transfer the knowledge in the source field to the target field, so that the target field can obtain better learning effect. In general, the amount of data in the source domain is sufficient, while the amount of data in the target domain is small, and the migration learning requires that the knowledge learned in the case of sufficient amount of data be migrated to a new application environment with small amount of data.
In one example, an artificial intelligence screening system, when deployed into a new application environment, assists in resolving classification problems in the target domain using a migration learning technique with existing labeled training data (source domain data) in addition to data collected in the new application environment (target domain). With the help of the training data in the source domain, the requirement for the amount of training data in the target domain can be reduced.
Taking screening of diabetic retinopathy as an example, automatic screening of diabetic retinopathy can be realized by analyzing the shot fundus picture. Fig. 2 shows a schematic fundus view of a normal eyeball. Fig. 3 shows a schematic view of the fundus with diabetic retinopathy with bleeding spots due to diabetes.
As described above, in practical applications, due to differences in data acquisition apparatuses, such as various apparatuses for taking fundus images, and differences in the angles of shooting, lighting conditions, and the like, data in the source domain may be subject to different data distributions. These data have certain differences from each other. Therefore, in the embodiment of the invention, the data in the source field is divided by using a clustering-based method, the source field is divided into a plurality of sub-fields, and then the classification models in the plurality of sub-fields are used for assisting the training of the classification models in the target field, so that the performance of the classifier in the target field is improved.
Specifically, in the clustering unit 100, the source domain data is divided by using a clustering algorithm to obtain a plurality of sub-domains. Then, in the learning unit 200, single-source migration learning is performed in each sub-domain, and a plurality of sub-domain classifiers for the target domain are obtained. Finally, in the integration unit 300, a plurality of integration units 300 are integrated to obtain a final target domain classifier.
One of the above technical solutions has the following advantages or beneficial effects: because the knowledge transfer with finer granularity is carried out in a plurality of sub-fields, a classifier with better effect can be obtained, and meanwhile, the data labeling cost in the target field is reduced through multi-source transfer learning.
In one embodiment, the clustering unit is configured to:
extracting the characteristics of the source field data;
and performing clustering analysis on the features by using a clustering algorithm, thereby dividing the source field into a plurality of sub-fields.
In one example, a source domain dataset with a tag may be defined as X ═ X1,x2,...,xn}. Wherein x represents a training picture in the source domain dataset; n represents the number of training pictures in the source domain dataset. Firstly, extracting the characteristics of a source field data set, and then clustering the picture characteristics of the training pictures by using a clustering algorithm to obtain k clusters.
Wherein features may be extracted using a machine learning model provided in an image recognition database. For example, first, a feature extraction is performed on a data set using a pretrained ResNet model on ImageNet. ImageNet is a large visual database for visual object recognition software research. And then clustering the picture characteristics by using a clustering algorithm to obtain k clusters. Clustering is a process of categorically organizing data members of a data set that are similar in some way. In one example, the picture features may be clustered using a K-means clustering algorithm. The K-means clustering algorithm is to randomly select K objects as initial clustering centers. The distance between each object and the respective seed cluster center is then calculated, and each object is assigned to the cluster center closest to it. The cluster centers and the objects assigned to them represent a cluster. Once all objects are assigned, the cluster center for each cluster is recalculated based on the objects existing in the cluster. This process will be repeated until some termination condition is met. The termination condition may be that no (or minimum number) objects are reassigned to different clusters, no (or minimum number) cluster centers are changed again, and the sum of squared errors is locally minimal.
In particular, the process of ResNet extracting features can be defined as mapping fθWhere θ is the corresponding parameter. For example, a feature vector f of 2048 × 1 size may be extracted from each pictureθ(xn). The features of the n training pictures are divided into K sets by K-means (K ≦ n) so that the sum of the squares of the differences between the feature vectors in the sets and the mean is minimized, i.e. the following problem is solved:
wherein, muiIs the sub-field SiAverage of all points in, i.e.
Dividing the source domain into a plurality of sub-domains S through k-means clustering1,S2,...,Sk。
Fig. 4 shows a block diagram of a learning unit of a training apparatus for screening a model according to an embodiment of the present invention. As shown in fig. 4, in one embodiment, the learning unit 200 includes:
a first training subunit 210, configured to train a source-domain encoder and the sub-domain classifier in the plurality of sub-domains, respectively;
a second training subunit 220, configured to train a domain classifier and a target domain encoder using a generative confrontation network in a plurality of the sub-domains, respectively;
wherein the input information of the source-domain encoder comprises first samples derived from the sub-domain, and the output information of the source-domain encoder comprises feature vectors extracted from the first samples;
the sub-domain classifier is used for classifying the feature vectors;
the input information of the target domain encoder comprises second samples derived from the target domain, and the output information of the target domain encoder comprises feature vectors extracted from the second samples;
the domain classifier is configured to discriminate whether the input feature vector is derived from the sub-domain or the target domain.
In one embodiment, the second training subunit is configured to:
inputting the feature vector output by the source domain encoder and the feature vector output by the target domain encoder into the domain classifier, and judging whether the input feature vector is from the sub-domain or the target domain by the domain classifier;
and if the domain classifier can not judge that the input feature vector is from the sub-domain or the target domain, using the target domain encoder and the sub-domain classifier for classifying the target domain data.
In the learning unit 200, single-source migration learning is performed in each sub-domain, and a sub-domain classifier for the target domain corresponding to each sub-domain is obtained.
In one example, in the sub-domain S1,S2,...,SkAnd (4) training. And for each sub-field, obtaining a sub-field classifier corresponding to each sub-field and aiming at the target field by utilizing a generative confrontation network and a multi-source transfer learning algorithm.
Among them, the generation of a countermeasure network (GAN) is a deep learning model. The framework of such a model comprises: a Generative Model (G) and a discriminant Model (D). The game learning by the two models with each other yields a fairly good output. In one example, a deep neural network may be used as the generative model G and the discriminative model D.
In the embodiment of the invention, the generative model in the generative countermeasure network comprises 2 encoders which are a source domain encoder and a target domain encoder respectively. Wherein the input information of the source region encoder is derived fromSample pictures of the sub-fields; the input information of the target domain encoder is a sample picture from the target domain; the output information of these 2 encoders is a feature vector f extracted from the sample pictureθ(x) In that respect The discriminative model in the generative confrontation network includes a domain classifier. And inputting the feature vectors output by the 2 encoders into a domain classifier, and enabling the domain classifier to distinguish whether the input feature vectors are from the sub-domain or the target domain.
If the domain classifier cannot distinguish the sub-domain from the target domain, the source domain data and the target domain data are similar in distribution. That is, feature spaces constituted by feature vectors respectively extracted from each of the sub-fields in the source field and from the target field are similar.
In the method, the data of the sub-field and the data of the target field are used for training the field classifier, whether the data is from the sub-field or from the target field is identified, and the source field encoder and the target field encoder are trained to confuse the field classifier, so that the field classifier cannot distinguish whether the data is from the sub-field or from the target field, and the data of the source field and the data of the target field are distributed similarly.
In the first training subunit 210, the annotation data S of the sub-domain can be usediTraining source field encoderAnd sub-domain classifierThe corresponding optimization problem is as follows:
wherein, the sub-domain classifierFor classifying sample pictures, e.g. diagnostic sample picturesOr whether the eye is a fundus picture with diabetic retinopathy. Y in the formula represents the classification result, for example, y may represent the result category of diagnosis, 0 is a normal sample, and 1 is a diseased sample.
In the training process, the sub-field classifier is trained in each sub-fieldThe annotation data of the corresponding sub-domain in the source domain can be used. Because the source field is divided into a plurality of sub-fields through clustering, the labeled data of the corresponding sub-fields in the source field needs to be reused to train the classifier of the sub-fields after clustering division in the embodiment of the invention
In the second training subunit 220, the domain classifier D and the target domain encoder are trained using the generative confrontation network and the multi-source transfer learning algorithmThe corresponding optimization problem is as follows:
the domain classifier D is a discriminant model in the generative countermeasure network. As described above, the target domain encoder takes a sample picture of the target domain as an input, and extracts a feature vector from the sample picture. In order to make the feature space formed by the feature vectors as similar as possible to the corresponding sub-domain in the source domain, the target domain encoder needs to generate output results that are sufficient to "confuse" the discrimination model D, so that the domain classifier cannot distinguish whether the data originates from the sub-domain or from the target domain. This goal can be achieved by an antagonistic training between the generative model and the discriminative model, i.e. the discriminative model continuously increases the ability to distinguish data sources, while the generative model strives to generate results that are difficult to distinguish by discriminative model D, continuously increasing the ability to "confuse" the discriminative model. That is, on the one hand, discriminant model D is trained to maximize the probability of correctly distinguishing data sources; on the other hand, the training generation model minimizes the probability that the discrimination model D obtains a correct answer. The training process described above can be viewed as a very small game process with respect to the loss function.
After the training is completed, the characteristic distribution of the sub-domain data and the target domain data is basically consistent, so that the target domain encoder can be usedAnd sub-domain classifierAnd classifying the target field data.
The technical scheme has the following advantages or beneficial effects: by utilizing the generative countermeasure network and the multi-source transfer learning algorithm, the feature subspace shared among the fields can be learned, so that the difference among data distribution of different fields is reduced.
In one embodiment, the integrated unit 300 is configured to:
and taking the average value of the classification results of the plurality of sub-domain classifiers as the classification result of the target domain classifier.
Specifically, migration learning is performed based on a plurality of sub-fields to obtain a plurality of sub-field classifiers, and finally prediction results of the plurality of sub-field classifiers are integrated to obtain a final prediction value, namely a classification result of the target field classifier. For the target domain data x, the prediction formula is as follows:
wherein the content of the first and second substances,representing a classification result of the target domain classifier; k represents the number of sub-fields;representing target domain encoderSub-domain classifiers of corresponding sub-domainsThe predicted result of (1).
In practical application, the sample picture to be classified can be input into each sub-domain classifierAnd respectively obtaining a prediction result. And integrating the prediction results of the plurality of sub-field classifiers to obtain a final classification result.
FIG. 5 shows a flow diagram of a training method of a screening model according to an embodiment of the present invention. As shown in fig. 5, the training method of the screening model includes:
step S110, dividing source field data by using a clustering algorithm, and dividing the source field into a plurality of sub-fields;
step S120, respectively carrying out transfer learning on the plurality of sub-fields to obtain a plurality of sub-field classifiers for classifying the target field data;
and step S130, integrating a plurality of sub-field classifiers to obtain a target field classifier.
In an embodiment, in step S110 in fig. 5, dividing the source domain data by using a clustering algorithm, and dividing the source domain into a plurality of sub-domains may specifically include:
extracting the characteristics of the source field data;
and performing clustering analysis on the features by using a clustering algorithm, thereby dividing the source field into a plurality of sub-fields.
FIG. 6 illustrates a flow chart of a training method of a screening model according to another embodiment of the present invention. As shown in fig. 6, in an embodiment, in step S120 in fig. 5, performing migration learning on a plurality of sub-domains to obtain a plurality of sub-domain classifiers for classifying the target domain data, which may specifically include:
step S210, respectively training a source domain encoder and the sub-domain classifier in the plurality of sub-domains;
step S220, training a domain classifier and a target domain encoder by utilizing a generative confrontation network respectively in a plurality of sub-domains;
wherein the input information of the source-domain encoder comprises first samples derived from the sub-domain, and the output information of the source-domain encoder comprises feature vectors extracted from the first samples;
the sub-domain classifier is used for classifying the feature vectors;
the input information of the target domain encoder comprises second samples derived from the target domain, and the output information of the target domain encoder comprises feature vectors extracted from the second samples;
the domain classifier is configured to discriminate whether the input feature vector is derived from the sub-domain or the target domain.
FIG. 7 shows a flow chart of a training method of a screening model according to yet another embodiment of the present invention. As shown in fig. 7, in an embodiment, in step S220 in fig. 6, the training the domain classifier and the target domain encoder using the generative confrontation network in a plurality of the sub-domains respectively may specifically include:
step S310, inputting the feature vector output by the source domain encoder and the feature vector output by the target domain encoder into the domain classifier, and judging whether the input feature vector is from the sub-domain or the target domain by the domain classifier;
step S320, if the domain classifier cannot distinguish whether the input feature vector is from the sub-domain or the target domain, using the target domain encoder and the sub-domain classifier to classify the target domain data.
In an embodiment, in step S130 in fig. 5, integrating a plurality of the sub-domain classifiers to obtain a target domain classifier specifically includes:
and taking the average value of the classification results of the plurality of sub-domain classifiers as the classification result of the target domain classifier.
The implementation of each step in the training method for screening models in the embodiment of the present invention may refer to the corresponding description of the functions of each unit in the above apparatus, and is not described herein again.
Fig. 8 is a block diagram illustrating a structure of a training apparatus for screening a model according to still another embodiment of the present invention. As shown in fig. 8, the apparatus includes: a memory 910 and a processor 920, the memory 910 having stored therein computer programs operable on the processor 920. The processor 920, when executing the computer program, implements the training method of the screening model in the above embodiments. The number of the memory 910 and the processor 920 may be one or more.
The device also includes:
and a communication interface 930 for communicating with an external device to perform data interactive transmission.
If the memory 910, the processor 920 and the communication interface 930 are implemented independently, the memory 910, the processor 920 and the communication interface 930 may be connected to each other through a bus and perform communication with each other. The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended ISA (Extended Industry Standard Architecture) bus, or the like. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is shown in FIG. 8, but this is not intended to represent only one bus or type of bus.
Optionally, in an implementation, if the memory 910, the processor 920 and the communication interface 930 are integrated on a chip, the memory 910, the processor 920 and the communication interface 930 may complete communication with each other through an internal interface.
An embodiment of the present invention provides a computer-readable storage medium, which stores a computer program, and the computer program is used for implementing the method of any one of the above embodiments when being executed by a processor.
In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality" means two or more unless specifically defined otherwise.
Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps of the process, and alternate implementations are included within the scope of the preferred embodiment of the present invention in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the present invention.
The logic and/or steps represented in the flowcharts or otherwise described herein, e.g., an ordered listing of executable instructions that can be considered to implement logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or more wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable read-only memory (CDROM). Additionally, the computer-readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via for instance optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and then stored in a computer memory.
It should be understood that portions of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the various steps or methods may be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.
It will be understood by those skilled in the art that all or part of the steps carried by the method for implementing the above embodiments may be implemented by hardware related to instructions of a program, which may be stored in a computer readable storage medium, and when the program is executed, the program includes one or a combination of the steps of the method embodiments.
In addition, functional units in the embodiments of the present invention may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode. The integrated module, if implemented in the form of a software functional module and sold or used as a separate product, may also be stored in a computer readable storage medium. The storage medium may be a read-only memory, a magnetic or optical disk, or the like.
The above description is only for the specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive various changes or substitutions within the technical scope of the present invention, and these should be covered by the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.
Claims (10)
1. A training device for a diabetic retina screening model, comprising:
the clustering unit is used for dividing the image data of the source field by utilizing a clustering algorithm and dividing the source field into a plurality of sub-fields;
the learning unit is used for respectively carrying out transfer learning on the plurality of sub-fields to obtain a plurality of sub-field classifiers used for classifying the target field picture data;
the integration unit is used for integrating the plurality of sub-field classifiers to obtain a target field classifier;
the learning unit includes:
a first training subunit, configured to train a source-domain encoder and the sub-domain classifier in the plurality of sub-domains, respectively;
a second training subunit, configured to train a domain classifier and a target domain encoder using a generative confrontation network in each of the plurality of sub-domains;
wherein the input information of the source-domain encoder comprises first samples derived from the sub-domain, and the output information of the source-domain encoder comprises feature vectors extracted from the first samples;
the sub-domain classifier is used for classifying the feature vectors;
the input information of the target domain encoder comprises second samples derived from the target domain, and the output information of the target domain encoder comprises feature vectors extracted from the second samples;
the domain classifier is configured to discriminate whether the input feature vector is derived from the sub-domain or the target domain.
2. The apparatus of claim 1, wherein the clustering unit is configured to:
extracting the characteristics of the source field picture data;
and performing clustering analysis on the features by using a clustering algorithm, thereby dividing the source field into a plurality of sub-fields.
3. The apparatus of claim 1, wherein the second training subunit is configured to:
inputting the feature vector output by the source domain encoder and the feature vector output by the target domain encoder into the domain classifier, and judging whether the input feature vector is from the sub-domain or the target domain by the domain classifier;
and if the domain classifier can not judge that the input feature vector is from the sub-domain or the target domain, using the target domain encoder and the sub-domain classifier for classifying the target domain picture data.
4. The apparatus according to any one of claims 1 to 3, wherein the integrated unit is configured to:
and taking the average value of the classification results of the plurality of sub-domain classifiers as the classification result of the target domain classifier.
5. A method for training a diabetic retina screening model, comprising:
dividing the image data of the source field by using a clustering algorithm, and dividing the source field into a plurality of sub-fields;
respectively carrying out transfer learning on the plurality of sub-fields to obtain a plurality of sub-field classifiers for classifying the image data of the target field;
integrating a plurality of sub-field classifiers to obtain a target field classifier;
respectively carrying out transfer learning in a plurality of sub-fields to obtain a plurality of sub-field classifiers for classifying target field picture data, wherein the sub-field classifiers comprise:
training a source domain encoder and the sub-domain classifier in a plurality of the sub-domains respectively;
training a domain classifier and a target domain encoder by using a generative confrontation network in a plurality of the sub-domains respectively;
wherein the input information of the source-domain encoder comprises first samples derived from the sub-domain, and the output information of the source-domain encoder comprises feature vectors extracted from the first samples;
the sub-domain classifier is used for classifying the feature vectors;
the input information of the target domain encoder comprises second samples derived from the target domain, and the output information of the target domain encoder comprises feature vectors extracted from the second samples;
the domain classifier is configured to discriminate whether the input feature vector is derived from the sub-domain or the target domain.
6. The method of claim 5, wherein the dividing the source domain picture data into a plurality of sub-domains by using a clustering algorithm comprises:
extracting the characteristics of the source field picture data;
and performing clustering analysis on the features by using a clustering algorithm, thereby dividing the source field into a plurality of sub-fields.
7. The method of claim 5, wherein training a domain classifier and a target domain encoder using a generative confrontation network in a plurality of the sub-domains respectively comprises:
inputting the feature vector output by the source domain encoder and the feature vector output by the target domain encoder into the domain classifier, and judging whether the input feature vector is from the sub-domain or the target domain by the domain classifier;
and if the domain classifier can not judge that the input feature vector is from the sub-domain or the target domain, using the target domain encoder and the sub-domain classifier for classifying the target domain picture data.
8. The method according to any one of claims 5 to 7, wherein integrating the plurality of sub-domain classifiers to obtain a target domain classifier comprises:
and taking the average value of the classification results of the plurality of sub-domain classifiers as the classification result of the target domain classifier.
9. A training device for a diabetic retina screening model, comprising:
one or more processors;
storage means for storing one or more programs;
the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any of claims 5-8.
10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the method according to any one of claims 5 to 8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910269536.6A CN110009038B (en) | 2019-04-04 | 2019-04-04 | Training method and device for screening model and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910269536.6A CN110009038B (en) | 2019-04-04 | 2019-04-04 | Training method and device for screening model and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110009038A CN110009038A (en) | 2019-07-12 |
CN110009038B true CN110009038B (en) | 2022-02-22 |
Family
ID=67169913
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910269536.6A Active CN110009038B (en) | 2019-04-04 | 2019-04-04 | Training method and device for screening model and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110009038B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110399492A (en) * | 2019-07-22 | 2019-11-01 | 阿里巴巴集团控股有限公司 | The training method and device of disaggregated model aiming at the problem that user's question sentence |
CN112434746B (en) * | 2020-11-27 | 2023-10-27 | 平安科技(深圳)有限公司 | Pre-labeling method based on hierarchical migration learning and related equipment thereof |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2018061091A (en) * | 2016-10-03 | 2018-04-12 | 株式会社Preferred Networks | Data compression device, data reproduction device, data compression method, data reproduction method and data transfer method |
Family Cites Families (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103761311B (en) * | 2014-01-23 | 2016-11-02 | 中国矿业大学 | Sensibility classification method based on multi-source field instance migration |
US11205103B2 (en) * | 2016-12-09 | 2021-12-21 | The Research Foundation for the State University | Semisupervised autoencoder for sentiment analysis |
CN106650806B (en) * | 2016-12-16 | 2019-07-26 | 北京大学深圳研究生院 | A kind of cooperating type depth net model methodology for pedestrian detection |
CN108197643B (en) * | 2017-12-27 | 2021-11-30 | 佛山科学技术学院 | Transfer learning method based on unsupervised clustering and metric learning |
CN108710948B (en) * | 2018-04-25 | 2021-08-31 | 佛山科学技术学院 | Transfer learning method based on cluster balance and weight matrix optimization |
CN109190750B (en) * | 2018-07-06 | 2021-06-08 | 国家计算机网络与信息安全管理中心 | Small sample generation method and device based on countermeasure generation network |
CN108959265A (en) * | 2018-07-13 | 2018-12-07 | 深圳市牛鼎丰科技有限公司 | Cross-domain texts sensibility classification method, device, computer equipment and storage medium |
CN109308689B (en) * | 2018-10-15 | 2022-06-14 | 聚时科技(上海)有限公司 | Mask generation based unsupervised image restoration method for resisting network transfer learning |
CN109492099B (en) * | 2018-10-28 | 2022-03-15 | 北京工业大学 | Cross-domain text emotion classification method based on domain impedance self-adaption |
CN109523538A (en) * | 2018-11-21 | 2019-03-26 | 上海七牛信息技术有限公司 | A kind of people counting method and system based on generation confrontation neural network |
CN109522973A (en) * | 2019-01-17 | 2019-03-26 | 云南大学 | Medical big data classification method and system based on production confrontation network and semi-supervised learning |
-
2019
- 2019-04-04 CN CN201910269536.6A patent/CN110009038B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2018061091A (en) * | 2016-10-03 | 2018-04-12 | 株式会社Preferred Networks | Data compression device, data reproduction device, data compression method, data reproduction method and data transfer method |
Also Published As
Publication number | Publication date |
---|---|
CN110009038A (en) | 2019-07-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10410353B2 (en) | Multi-label semantic boundary detection system | |
JP5174040B2 (en) | Computer-implemented method for distinguishing between image components and background and system for distinguishing between image components and background | |
CN109978893A (en) | Training method, device, equipment and the storage medium of image, semantic segmentation network | |
US11334992B2 (en) | Computer-implemented method to increase the accuracy of machine learning models for medical image processing | |
Zhao et al. | Weakly-supervised simultaneous evidence identification and segmentation for automated glaucoma diagnosis | |
US9330336B2 (en) | Systems, methods, and media for on-line boosting of a classifier | |
JP2006252559A (en) | Method of specifying object position in image, and method of classifying images of objects in different image categories | |
Shenavarmasouleh et al. | Drdr: Automatic masking of exudates and microaneurysms caused by diabetic retinopathy using mask r-cnn and transfer learning | |
CN110009038B (en) | Training method and device for screening model and storage medium | |
CN110706200B (en) | Data prediction method and device | |
Di Cataldo et al. | ANAlyte: A modular image analysis tool for ANA testing with indirect immunofluorescence | |
KR101093107B1 (en) | Image information classification method and apparatus | |
CN109034218B (en) | Model training method, device, equipment and storage medium | |
JP7355299B2 (en) | Learning dataset generation system, learning server, and learning dataset generation program | |
CN116805522A (en) | Diagnostic report output method, device, terminal and storage medium | |
CN112819826B (en) | Spine image processing device based on artificial intelligence and computer equipment | |
Giordano et al. | An AI-based framework for supporting large scale automated analysis of video capsule endoscopy | |
Shen et al. | Explainable Hierarchical Semantic Convolutional Neural Network for Lung Cancer Diagnosis. | |
CN114283114A (en) | Image processing method, device, equipment and storage medium | |
CN109993191B (en) | Information processing method and device, electronic device and storage medium | |
US7634120B2 (en) | Incorporating spatial knowledge for classification | |
JP2020181265A (en) | Information processing device, system, information processing method, and program | |
CN111612021A (en) | Error sample identification method and device and terminal | |
Ajmal et al. | Covid-19 detection using deep convolutional neural networks from x-ray images | |
US20240037920A1 (en) | Continual-learning and transfer-learning based on-site adaptation of image classification and object localization modules |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |