WO2022242459A1 - 数据分类识别方法、装置、设备、介质及程序产品 - Google Patents

数据分类识别方法、装置、设备、介质及程序产品 Download PDF

Info

Publication number
WO2022242459A1
WO2022242459A1 PCT/CN2022/090902 CN2022090902W WO2022242459A1 WO 2022242459 A1 WO2022242459 A1 WO 2022242459A1 CN 2022090902 W CN2022090902 W CN 2022090902W WO 2022242459 A1 WO2022242459 A1 WO 2022242459A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
classification
classification model
model
prediction result
Prior art date
Application number
PCT/CN2022/090902
Other languages
English (en)
French (fr)
Inventor
魏东
孙镜涵
马锴
王连生
郑冶枫
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2022242459A1 publication Critical patent/WO2022242459A1/zh
Priority to US18/077,709 priority Critical patent/US20230105590A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • 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
    • 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
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/40—Software arrangements specially adapted for pattern recognition, e.g. user interfaces or toolboxes therefor
    • G06F18/41—Interactive pattern learning with a human teacher
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/7715—Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06V10/7753—Incorporation of unlabelled data, e.g. multiple instance learning [MIL]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/94—Hardware or software architectures specially adapted for image or video understanding
    • G06V10/95—Hardware or software architectures specially adapted for image or video understanding structured as a network, e.g. client-server architectures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images

Definitions

  • the embodiments of the present application relate to the field of machine learning, and in particular to a data classification and identification method, device, equipment, medium and program product.
  • the diagnosis of diseases based on medical images usually includes the diagnosis of rare diseases and the diagnosis of common diseases, that is, after inputting medical images into the machine learning model, the machine learning model analyzes the medical images to determine the corresponding Physical abnormalities.
  • rare diseases themselves are diseases with a low probability of occurrence, and it is difficult to collect image data of rare diseases and label rare disease information, resulting in low training efficiency of classification models, and thus low classification accuracy of classification models.
  • the embodiment of the present application provides a data classification and recognition method, device, equipment, medium and program product, which can improve the training efficiency of a recognition model for recognition and classification of rare diseases.
  • the technical scheme is as follows.
  • a data classification and recognition method is provided, which is applied to a computer device, and the method includes:
  • the first data set includes first data
  • the second data set includes second data marked with a sample label
  • the second classification model is a classification model whose model parameters are to be adjusted
  • a data classification identification device includes:
  • An acquisition module configured to acquire a first data set and a second data set, the first data set includes first data, and the second data set includes second data marked with a sample label;
  • a training module configured to use the first data in an unsupervised training mode and use the second data in a supervised training mode to train candidate classification models to obtain a first classification model
  • the obtaining module is also used to obtain a second classification model, and the second classification model is a classification model whose model parameters are to be adjusted;
  • the training module is further configured to use the first prediction result of the first classification model for the first data as a reference, and based on the second prediction result of the second classification model for the first data to Adjusting the model parameters of the second classification model to obtain a data classification model;
  • the prediction module is used to classify and predict the target data through the data classification model, and obtain the classification result of the target data.
  • a computer device in another aspect, includes a processor and a memory, at least one instruction, at least one program, code set or instruction set are stored in the memory, the at least one instruction, the at least A program, the code set or instruction set is loaded and executed by the processor to implement the data classification identification method described in any one of the above-mentioned embodiments of the present application.
  • a computer-readable storage medium wherein at least one instruction, at least one program, code set or instruction set are stored in the storage medium, the at least one instruction, the at least one program, the code
  • the set or instruction set is loaded and executed by the processor to implement the data classification identification method described in any one of the above-mentioned embodiments of the present application.
  • a computer program product or computer program comprising computer instructions stored in a computer readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the data classification identification method described in any one of the above embodiments.
  • the first classification model After performing unsupervised training on unlabeled first data and supervised training on labeled second data, the first classification model is obtained, so that on the basis of the first classification model, a second classification model is created for knowledge distillation training, The teacher model is used for supervised training to achieve the purpose of distillation, and finally a student model with higher performance and accuracy is obtained.
  • the training mainly relies on a large amount of first data, while the data volume requirement for the labeled second data is small, avoiding the need for The cumbersome process of labeling a large number of sample data improves the training efficiency and accuracy of the data classification model.
  • Figure 1 is a schematic diagram of the implementation process of the overall solution provided by an exemplary embodiment of the present application
  • Fig. 2 is a schematic diagram of an implementation environment provided by an exemplary embodiment of the present application.
  • Fig. 3 is a flowchart of a data classification identification method provided by an exemplary embodiment of the present application.
  • Fig. 4 is a flowchart of a data classification identification method provided by another exemplary embodiment of the present application.
  • Fig. 5 is a data classification identification method provided by another exemplary embodiment of the present application.
  • Fig. 6 is an overall schematic diagram of the training process of the rare disease classification and identification model provided by an exemplary embodiment of the present application.
  • Fig. 7 is a structural block diagram of a data classification identification device provided by an exemplary embodiment of the present application.
  • Fig. 8 is a structural block diagram of a data classification identification device provided by another exemplary embodiment of the present application.
  • Fig. 9 is a structural block diagram of a server provided by an exemplary embodiment of the present application.
  • Pseudo-label It refers to the label that predicts the unlabeled data through the trained model to obtain the prediction result, and labels the data based on the prediction result. That is to say, the pseudo-label is not a label manually labeled according to the actual situation of the data, but a label with a certain error tolerance rate labeled by the trained model.
  • the classification prediction result obtained after classifying and predicting the data with respect to the classification model obtained through training is the pseudo-label corresponding to the data.
  • the diagnosis of rare diseases requires a classification model for the diagnosis of rare diseases
  • the training of the classification model requires a large number of sample image data marked with rare disease information
  • the classification model is used to classify and identify the sample image data
  • the recognition result is obtained, and the classification model is trained by the difference between the marked rare disease information and the recognition result.
  • the rarity of the rare disease itself it is difficult to obtain the sample image data, which requires a lot of manpower to collect the sample image data and label the sample image data with rare disease information, and the training efficiency of the classification model is low. It is precisely because of the difficulty in obtaining sample image data of rare diseases that it will lead to insufficient training sample data, and the accuracy of the trained classification model is low.
  • a data classification and recognition method which improves the training efficiency and accuracy of the data classification model when the number of labeled samples is small.
  • FIG. 1 is a schematic diagram of the implementation flow of the overall solution provided by an exemplary embodiment of the present application, taking the training process of a classification model for rare diseases as an example, as shown in FIG. 1 .
  • the first image data set 110 and the second image data set 120 are obtained, wherein the first image data set 110 includes medical images of common diseases, and the medical images in the first image data set 110 do not use label information or are not labeled Label information: the second image data set 120 includes a small number of medical images of rare diseases, and the medical images in the second image data set 120 include labeled label information, which is used to indicate the rare disease information corresponding to the medical images.
  • knowledge distillation training refers to guiding the classification ability of the untrained model based on the classification ability of the trained model.
  • the prediction result of the first classification model F on the data is used as a reference to guide the prediction ability of the second classification model F' on the data, that is, the data a
  • the prediction result predicted by the first classification model F is used as a reference to guide the prediction ability of the second classification model F' on the data, that is, the data a
  • the difference of pseudo labels is used to train the second classification model F', so as to guide the classification accuracy of the second classification model F' to approach the first classification model F.
  • the terminal 210 includes a first terminal 211 and a second terminal 212 .
  • the first terminal 211 is used to send medical images to the server 220 .
  • the first terminal 211 is a terminal used by doctors.
  • doctors use classification models to assist diagnosis, thereby improving the accuracy of diagnosis;
  • the first terminal 211 is a user application terminal, such as: the interviewer himself, or relatives of the interviewer, etc., the user sends the medical image to the server, so as to obtain the reference diagnosis result; or, the first terminal 211 is a terminal connected to the medical image scanning device, and the medical image
  • the scanning device transmits the scanned medical image to the first terminal 211.
  • the first terminal 211 forwards the medical image to the server 220 for auxiliary diagnosis.
  • the server 220 trains the rare disease classification model 221 in the manner shown in FIG. 1 above. After obtaining the rare disease classification model 221, it receives the medical images uploaded by the first terminal 211, and classifies and recognizes the medical images through the rare disease classification model 221. , to obtain the classification and diagnosis results of medical images in the rare disease classification set. The server 220 feeds back the classification diagnosis result to the first terminal 211 or sends the classification diagnosis result to the second terminal 212 .
  • the server 220 sends the classification and diagnosis results to the second terminal 212, and the second terminal 212 is implemented as a terminal for doctor applications or a terminal for user applications.
  • the foregoing terminal may be various forms of terminal equipment such as a mobile phone, a tablet computer, a desktop computer, and a portable notebook computer, which is not limited in this embodiment of the present application.
  • server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, and can also provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network Cloud servers for basic cloud computing services such as services, cloud communications, middleware services, domain name services, security services, content delivery network (Content Delivery Network, CDN), and big data and artificial intelligence platforms.
  • cloud services cloud databases, cloud computing, cloud functions, cloud storage, network Cloud servers for basic cloud computing services such as services, cloud communications, middleware services, domain name services, security services, content delivery network (Content Delivery Network, CDN), and big data and artificial intelligence platforms.
  • Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and network in a wide area network or a local area network to realize data calculation, storage, processing, and sharing.
  • the above server can also be implemented as a node in the blockchain system.
  • the information including but not limited to user equipment information, user personal information, etc.
  • data including but not limited to data used for analysis, stored data, displayed data, etc.
  • signals involved in this application All are individually authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.
  • the medical images mentioned in this application were all obtained with full authorization.
  • the doctor sends the medical image to the server through the terminal, and the server classifies and recognizes the medical image through the trained classification model, obtains the classification diagnosis result corresponding to the medical image, and feeds back the classification diagnosis result to the terminal used by the doctor for further processing. Display, so that the doctor can make an auxiliary diagnosis through the classified diagnosis results, and get the final diagnosis result.
  • the user sends the medical images to the server, and the server classifies and recognizes the medical images through the trained classification model, obtains the classification diagnosis results corresponding to the medical images, and feeds back the classification diagnosis results to
  • the terminal of the user application is displayed, and the user first has a preliminary understanding of the abnormal life state according to the classified diagnosis results, and then obtains the detailed diagnosis results through the doctor's diagnosis.
  • the classification model can also be applied to other classification scenarios.
  • the classification model can also be applied to object recognition scenarios, speech recognition scenarios, handwriting recognition scenarios, etc., which is not limited in this embodiment of the present application.
  • the data classification and identification method provided by this application is described. Taking the application of the method in the server as an example, as shown in FIG. 3 , the method includes the following steps.
  • Step 301 acquiring a first data set and a second data set.
  • the first data set includes first data
  • the second data set includes second data labeled with sample labels.
  • the first data in the first data set is unlabeled data
  • the second data in the second data set is data labeled with sample labels; or, the first data in the first data set is labeled There are labels, but the label labeled by the first data is not used in this embodiment.
  • the first data belongs to the first classification set
  • the second data belongs to the target classification set, that is, the first data and the second data belong to data corresponding to different classification sets.
  • the category set is used to represent a set containing multiple subordinate categories, that is, the category set belongs to the upper-level induction concept of the classification, and the upper-level induction includes the lower-level categories belonging to the induction characteristics.
  • the first-level inductive concept is disease, and those belonging to the concept of disease are divided into two lower-level categories, namely common diseases and rare diseases.
  • the first classification set corresponds to the common disease classification set.
  • the first classification set includes common eye diseases such as myopia, hyperopia, and conjunctivitis
  • the target classification set corresponds to the rare disease classification set, taking eye diseases as an example, for example: the target classification set includes dry eye syndrome, visual snow syndrome, genetic disease, etc.
  • Rare eye diseases such as optic neuropathy.
  • common diseases and rare diseases refer to diseases corresponding to the same organ or the same body part, or common diseases and rare diseases belong to the same disease type, such as: body pain type, vomiting type.
  • the first data included in the first data set are medical images corresponding to common diseases, such as: images in the form of computer tomography (Computed Tomography, CT) images, X-ray images, ultrasonic images, etc.;
  • the second data included in the second data set are medical images corresponding to rare diseases, such as images in the form of CT images, X-ray images, and ultrasonic images.
  • first data and second data in the embodiment of the present application can also be implemented as other types of data, such as: voice data, text data, etc., the embodiment of the present application This is not limited.
  • the data amount of the first data in the first data set (that is, the number of medical images) is greater than the data amount of the second data in the second data set (that is, the number of medical images), and optionally, the second data in the second data set
  • the quantity of data is within the required quantity range, such as: less than the preset quantity.
  • the first data in the first data set is data randomly sampled from the basic data set, which includes common disease data;
  • the second data in the second data set is data randomly sampled from the rare disease data set, rare
  • the rare disease data set includes rare disease data, and the second data is marked with rare disease information, that is, the type of rare disease corresponding to each medical image.
  • Step 302 train the candidate classification model by using the first data in an unsupervised training mode and by using the second data in a supervised training mode to obtain a first classification model.
  • unsupervised training is performed on the feature extraction network in the candidate classification model based on the first data
  • the classification regression network in the candidate classification model is combined with the unsupervised trained feature extraction network to obtain a classification model
  • the classification and regression network is used for data classification.
  • the classification and regression network is used for data classification in the target classification set (that is, the classification set where the second data is located).
  • Supervised training is performed on the classification model by using the second data and sample labels in the second data set to obtain the first classification model.
  • the first data in the first data set has no corresponding label
  • the first data can only be used for unsupervised training of the feature extraction network.
  • the second data in the second data set has correspondingly labeled sample labels, so the second data can be used for supervised training of the classification model.
  • Step 303 acquiring a second classification model.
  • the second classification model is a classification model whose model parameters are to be adjusted.
  • the second classification model is a randomly initialized classification model
  • the second classification model includes model parameters
  • the second classification model is used to perform distillation training according to the knowledge output by the first classification model.
  • knowledge distillation refers to the process of using the supervisory information output by the teacher model as knowledge
  • the student model learns the supervisory information transferred from the teacher model, and uses the teacher model to conduct supervised training to achieve the purpose of distillation, and finally obtain higher performance and Accurate student model.
  • Step 304 based on the first prediction result of the first classification model for the first data, and based on the second prediction result of the second classification model for the first data, the model parameters of the second classification model are adjusted to obtain a data classification model.
  • the first classification model is used to classify and predict the first data to obtain the first prediction result, and optionally, the first prediction result is used as the category pseudo-label in the corresponding target classification set of the first data; through the second classification
  • the model performs classification prediction on the first data to obtain a second prediction result, and adjusts model parameters of the second classification model based on the difference between the first prediction result and the second prediction result to obtain a data classification model.
  • the first classification model and the second classification model perform prediction on the same first data, obtain a first prediction result output by the first classification model, and obtain a second prediction result output by the second classification model.
  • the first data set includes data A, data B and data C, then firstly predict data A through the first classification model to obtain the first prediction result a, and predict data A through the second classification model to obtain For the second prediction result b, the model parameters of the second classification model are adjusted according to the difference between the first prediction result a and the second prediction result b.
  • the pseudo-label output by the first classification model after classifying and predicting the first data is used as knowledge, and the pseudo-label is transferred by the second classification model for distillation, thereby realizing the knowledge distillation training of the second classification model.
  • the second classification model migrates the pseudo-label to perform distillation means that after the second classification model classifies and predicts the first data, it evaluates the prediction result based on the pseudo-label, thereby adjusting the model parameters, and also That is, the pseudo-labels identified by the first classification model are used as data labels to guide the classification prediction results of the second classification model to align with the pseudo-labels, thereby improving the classification prediction accuracy of the second classification model.
  • Step 305 classify and predict the target data through the data classification model, and obtain the classification result of the target data.
  • the target data is classified and predicted by the data classification model, and the classification result of the target data in the target classification set is obtained, wherein the target classification set is the classification set to which the second data labeled with the sample label belongs.
  • the target data is classified through the data classification model to obtain a classification result of the target data in the target classification set.
  • the target data may be data in actual application, such as: medical images in actual application; or, the target data may also be data in the test set for testing the data classification model.
  • the data classification recognition method obtains the first classification model after performing unsupervised training on unlabeled first data and supervised training on labeled second data. Based on the model, create a second classification model for knowledge distillation training, use the teacher model for supervised training to achieve the purpose of distillation, and finally obtain a student model with higher performance and accuracy.
  • the training mainly relies on a large amount of first data, and for some The data volume requirement of the second data of the label is small, avoiding the cumbersome process of labeling a large amount of sample data, and improving the training efficiency and accuracy of the data classification model.
  • FIG. 4 is provided by another exemplary embodiment of the present application
  • the flow chart of the data classification and identification method is illustrated by taking the application of the method in the server as an example, as shown in FIG. 4 , the method includes the following steps.
  • Step 401 acquiring a first data set and a second data set.
  • the first data set includes first data
  • the second data set includes second data labeled with sample labels.
  • the first data in the first dataset is unlabeled data
  • the second data in the second dataset is data labeled with sample labels.
  • the first data belongs to the first classification set
  • the second data belongs to the target classification set, that is, the first data and the second data belong to data corresponding to different classification sets.
  • the first classification set corresponds to the common disease classification set
  • the target classification set corresponds to the rare disease classification set.
  • Step 402 train the candidate classification model by using the first data in an unsupervised training mode and the second data in a supervised training mode to obtain a first classification model.
  • unsupervised training is performed on the feature extraction network in the candidate classification model based on the first data in the first data set, and the classification regression network in the candidate classification model is combined with the unsupervised trained feature extraction network, A classification model is obtained, wherein the classification regression network is used to classify data in the target classification set, and the classification model is supervised and trained by the second data and sample labels in the second data set to obtain the first classification model.
  • the first classification model has better classification performance, but in the process of representation learning, the knowledge related to the target classification set is ignored. Therefore, in the embodiment of this application, the first classification model is used as the benchmark model, and the first classification The knowledge output by the model performs distillation training on the second classification model.
  • the second classification model is a model whose model parameters are to be adjusted for classification in the target classification set.
  • Step 403 perform classification prediction on the first data through the first classification model to obtain a first prediction result.
  • the first data set is medical images of common diseases
  • the second data set is medical images of rare diseases
  • the data in the first dataset and the second dataset have similar features in color, texture or shape. Therefore, the first classification model is used as the reference model to predict the probability that the images in the first data set belong to each classification in the target classification set.
  • the first data in the first data set is classified and predicted by the first classification model, the probability value corresponding to the classification in the target classification set of the first data is obtained, and the corresponding value of the first data is determined from the target classification set based on the probability value.
  • the pseudo-label of is used as the first prediction result.
  • the hard label of the category in the target category set corresponding to the first data is further determined according to the soft label determined above.
  • the category with the highest probability is marked as the pseudo-label corresponding to the first data, that is, the hard label of the category with the highest probability is 1, the hard label of other categories is 0, and the hard label is 1
  • the category is the first prediction result corresponding to the first data, that is, the pseudo-label corresponding to the first data.
  • a first prediction result is obtained by classification prediction for a first data x, that is, m first prediction results are obtained by classification prediction for m first data, m is a positive integer, and each first data can be obtained by Classification prediction obtains a first prediction result.
  • step 404 the second classification model is obtained, and the first data is classified and predicted by the second classification model to obtain a second prediction result corresponding to the first data.
  • the second classification model is a model whose model parameters are to be adjusted, and the second classification model is used to classify data corresponding to the target classification set.
  • the first prediction result and the second prediction result are obtained respectively, then the first prediction result and the second prediction result corresponding to the same first data There is a comparison between the results.
  • Step 405 Adjust the model parameters of the second classification model based on the difference between the first prediction result and the second prediction result to obtain a data classification model.
  • the second classification model includes the first query encoder and the first key-value encoder, then the first data is encoded by the first query encoder to obtain the first encoding result, and the first key-value encoder
  • the second encoding result of encoding the first data and the data in the first preset dynamic dictionary is used to train the second classification model based on the difference between the first encoding result and the second encoding result to obtain a data classification model.
  • the hybrid distillation loss is determined in combination with a pseudo-label supervision method and a contrastive discriminant method, wherein the pseudo-label supervision method is a model of the second classification model based on the difference between the first prediction result and the second prediction result The parameters are adjusted, and the comparison and discrimination method is to train the second classification model through the first query encoder and the first key-value encoder.
  • x is the first data in the first data set
  • ⁇ ' k is the parameter of the first key-value encoder f' k
  • the momentum is updated along with ⁇ ' q
  • the first key-value encoder f' k corresponds to The first preset dynamic dictionary, comparing the first encoding result extracted by the first query encoder f'q with the second encoding result of the data encoding in the first preset dynamic dictionary by the first key-value encoder, to obtain Contrast the loss L con in the discriminative method.
  • y represents the first prediction result identified by the first classification model, the first prediction result is compared with the second prediction result of the second classification model, and the loss L cls corresponding to the pseudo-label supervision method is obtained. In this way, the two parts of the loss are added to update the model parameter ⁇ ' q .
  • f'c is implemented with a fully connected layer (followed by a softmax operation), which facilitates end-to-end model training.
  • the first prediction result also corresponds to a confidence parameter
  • the confidence parameter of the first prediction result is obtained
  • the difference between the first prediction result and the second prediction result under the confidence parameter is determined
  • the model parameters of the second classification model are adjusted to obtain the data classification model.
  • the prediction value p' of the second classification model is combined with the pseudo-label y (that is, the first prediction result y) as the training target, please refer to the following formula 2.
  • ⁇ is a confidence parameter, which controls the proportion of the pseudo-label y generated by the first classification model to the training target.
  • ⁇ T is the confidence parameter value of the last training round, and schematically, ⁇ T is set to 0.7.
  • T is the total number of training rounds.
  • step 406 the target data is classified and predicted by the data classification model, and the classification result of the target data in the target classification set is obtained.
  • a data classification model is obtained, and the target data is classified through the data classification model to obtain a classification result of the target data in the target classification set.
  • the test data set is obtained, the test data in the test data set is used to test the training effect of the data classification model, the target data is obtained from the test data set, the target data is marked with reference classification information, and the data classification model is used to After the target data is classified and predicted to obtain the classification result, the training effect data of the data classification model is obtained based on the reference classification information and the classification result.
  • multiple target data in the test data set are obtained, classification predictions are performed respectively, and they are compared with reference classification information, and the training effect is determined according to the ratio of the correct target data to the total number of tested target data. That is to determine the prediction accuracy of the data classification model.
  • the data classification recognition method obtains the first classification model after performing unsupervised training on unlabeled first data and supervised training on labeled second data. Based on the model, create a second classification model for knowledge distillation training, use the teacher model for supervised training to achieve the purpose of distillation, and finally obtain a student model with higher performance and accuracy.
  • the training mainly relies on a large amount of first data, and for some The data volume requirement of the second data of the label is small, avoiding the cumbersome process of labeling a large amount of sample data, and improving the training efficiency and accuracy of the data classification model.
  • the method provided in this embodiment combines the pseudo-label supervision method and the contrastive discrimination method to determine the mixed distillation loss, and avoids the feature extraction of the data by the second classification model while performing distillation training on the second classification model through the first classification model Affected by the distillation training process, the training efficiency and accuracy of the second classification model are improved.
  • the first The confidence parameter of the prediction result avoids the deviation of the accuracy of the first prediction result from affecting the training effect of the second classification model, and improves the prediction accuracy of the classification model.
  • the method provided in this embodiment really introduces a linear growth method to the confidence parameter, and gradually adjusts the adjustment of the confidence parameter to the first prediction result, so as to avoid excessive intervention in the first prediction result caused by the subjective setting of the confidence parameter Or it is too low, which improves the calculation accuracy of the loss value.
  • the first classification model is obtained through unsupervised training on the first data and supervised training on the second data.
  • Fig. 5 is a flowchart of a data classification identification method provided by another exemplary embodiment of the present application. As shown in Fig. 5, taking the method applied to a server as an example, the method includes the following steps.
  • Step 501 acquire a first data set and a second data set.
  • the first data set includes first data
  • the second data set includes second data marked with sample labels
  • the second data belongs to the target classification set.
  • the first data belongs to the first classification set
  • the second data belongs to the target classification set, that is, the first data and the second data belong to data corresponding to different classification sets.
  • Step 502 perform unsupervised training on the feature extraction network in the candidate classification model based on the first data.
  • the feature extraction network includes a second query encoder and a second key-value encoder
  • the first data is encoded by the second query encoder to obtain a third encoding result
  • the second key-value encoder is obtained
  • unsupervised training is performed on the feature extraction network based on the difference between the third encoding result and the fourth encoding result.
  • the contrastive loss is used as the optimization function of the feature extraction network.
  • image enhancement is performed on the medical images in the first data set.
  • Image enhancement includes at least one of enhancement processing methods such as contrast enhancement, brightness enhancement, and sharpening enhancement.
  • the number of times of image enhancement is twice, so as to input the second query encoder and the second key-value encoder respectively, wherein, two image enhancements are to carry out two different degrees of enhancements for the same enhancement direction; or, two images
  • the enhancement is two enhancements of the same or different degrees for different enhancement directions; or, the two image enhancements are two superimposed enhancements of the same or different degrees for the same enhancement direction.
  • two image enhancements are performed on each image in the first data set to obtain and in, is the image obtained after the first image enhancement, is the image obtained after the second image enhancement.
  • the second query encoder and the second key-value encoder Perform feature extraction to obtain the corresponding features, where f q and f k are the second query encoder and the second key-value encoder composed of parameters ⁇ q and ⁇ k respectively.
  • f q and f k are the second query encoder and the second key-value encoder composed of parameters ⁇ q and ⁇ k respectively.
  • xi is the first data in the first data set
  • l is the number of key-value images stored in the second preset dynamic dictionary
  • ⁇ is the hyperparameter of the smoothing label.
  • the parameters ⁇ q are frozen.
  • step 503 the classification regression network in the candidate classification model is combined with the feature extraction network that has undergone unsupervised training to obtain a classification model.
  • a classification regression network is used to classify data within a target classification set.
  • the classification regression network and the unsupervised The trained second query encoder is concatenated to obtain the classification model.
  • Step 504 supervise the classification model by using the second data and sample labels in the second data set to obtain the first classification model.
  • the second data when using the second data to supervise the training of the classification model, the second data is input into the classification model for classification prediction, and the prediction result is obtained, and the second data itself is marked with a sample label, which is used to indicate the second data Actual classification, so as to reversely adjust the model parameters of the classification model according to the difference between the sample label and the predicted result to obtain the first classification model.
  • the loss value of the prediction result is calculated according to the sample label and the prediction result, so that the model parameters of the classification model are reversely adjusted according to the loss value until the loss value corresponding to the prediction result converges, such as: the loss value corresponding to the prediction result less than the preset threshold; or, the difference between the loss value of the predicted result in the qth iterative training and the loss value of the predicted result in the q-1th iterative training is less than the preset difference threshold, and q is an integer greater than 1.
  • Step 505 acquiring a second classification model.
  • the second classification model is a classification model whose model parameters are to be adjusted.
  • the second classification model is a randomly initialized classification model, and the second classification model includes model parameters. During random initialization, the initial model parameters of the second classification model are obtained randomly.
  • the second classification model is used for distillation training according to the knowledge output by the first classification model.
  • knowledge distillation refers to the supervision information output by the first classification model, that is, the first prediction result output by the first classification model as knowledge, and the second classification model learns the supervision information transferred from the first classification model as a distillation process.
  • the first classification model is used for supervised training to achieve the purpose of distillation, and finally a student model with higher performance and accuracy is obtained.
  • Step 506 based on the first prediction result of the first classification model for the first data, and based on the second prediction result of the second classification model for the first data, the model parameters of the second classification model are adjusted to obtain a data classification model.
  • the pseudo-label output by the first classification model after classifying and predicting the first data is used as knowledge, and the pseudo-label is transferred by the second classification model for distillation, thereby realizing distillation training of the second classification model.
  • Step 507 classify and predict the target data through the data classification model, and obtain the classification result of the target data.
  • the target data is classified through the data classification model to obtain a classification result of the target data in the target classification set.
  • the target data may be data in actual application, such as: medical images in actual application; or, the target data may also be data in the test set for testing the data classification model.
  • the data classification recognition method obtains the first classification model after performing unsupervised training on unlabeled first data and supervised training on labeled second data. Based on the model, create a second classification model for knowledge distillation training, use the teacher model for supervised training to achieve the purpose of distillation, and finally obtain a student model with higher performance and accuracy.
  • the training mainly relies on a large amount of first data, and for some The data volume requirement of the second data of the label is small, avoiding the cumbersome process of labeling a large amount of sample data, and improving the training efficiency and accuracy of the data classification model.
  • unsupervised training is performed on the feature extraction network through the unlabeled first data in the first data set
  • the supervised training is performed on the classification model through the second data with labels in the second data set, so that in the second
  • the data collection process is cumbersome, or the second data collection is difficult, only a small amount of second data needs to be collected to realize effective training of the first classification model and improve the training efficiency of the model.
  • the data classification model is tested through the test data in the test data set to determine the training effect of the data classification model, thereby assisting the further training or application of the data classification model, and improving the performance of the data classification model.
  • the classification accuracy of the data classification model is determined by the data classification model.
  • Figure 6 is a An overall schematic diagram of the training process of the rare disease classification and recognition model provided in the exemplary embodiment.
  • the process includes an unsupervised training phase 610 , a supervised training phase 620 , a pseudo-label generation phase 630 and a second classification model training phase 640 .
  • the unsupervised training stage 610 the unlabeled common disease medical image 611 is subjected to two image enhancements to obtain x q and x k , and the loss value is determined through the query encoder 612 and the key value encoder 613, Thus, the training of the query encoder 612 is completed, the parameters of the query encoder 612 are frozen, and the unsupervised trained query encoder 612 is applied to the connection with the classification regression model 621 in the supervised training stage 620 .
  • the first classification model 622 to be trained is obtained, and the first classification model 622 is supervised by medical images 623 of rare diseases marked with sample labels During training, the loss value is determined according to the labeled sample labels corresponding to the rare disease medical images 623 and the classification results of the first classification model 622 , and supervised training of the first classification model 622 is implemented.
  • the pseudo-label generation stage 630 After the training of the first classification model 622 is completed, in the pseudo-label generation stage 630, the medical images 611 of common diseases are classified and recognized by the first classification model 622, and pseudo-labels corresponding to the medical images 611 of common diseases are obtained.
  • the first loss value is obtained according to the pseudo-label corresponding to the medical image 611 of common diseases and the prediction result of the second classification model 641, and according to the query encoder 642 and the key in the second classification model 641
  • the second loss value is obtained from the encoding result of the value encoder 643 , so that the total loss value is determined according to the first loss value and the second loss value to train the second classification model 641 to obtain a rare disease classification recognition model.
  • Table 1 shows the comparison of the results of the technical solution of the present application on the skin lesion classification data set.
  • This data set contains 7 categories, the data sets of the four categories with the largest number of cases are used as the first data set, and the data sets of the remaining three categories are used as the second data set.
  • As the evaluation index Accuracy and F1score, an index used to measure the accuracy of the binary classification model in statistics, are selected.
  • N represents the number of test categories
  • K represents the number of labeled pictures provided by each test category. This technical solution compares the results of K being 1, 3, and 5 respectively.
  • the remaining image composition Q in the rare disease dataset is used as the test set for performance evaluation.
  • Fig. 7 is a schematic structural diagram of a data classification identification device provided by an exemplary embodiment of the present application. As shown in Fig. 7, the device includes the following parts:
  • An acquisition module 710 configured to acquire a first data set and a second data set, the first data set includes first data, and the second data set includes second data marked with a sample label;
  • a training module 720 configured to use the first data in an unsupervised training mode and use the second data in a supervised training mode to train candidate classification models to obtain a first classification model;
  • the obtaining module 710 is also used to obtain a second classification model, where the second classification model is a classification model whose model parameters are to be adjusted;
  • the training module 720 is further configured to use the first prediction result of the first classification model for the first data as a reference, and to perform the calculation based on the second prediction result of the second classification model for the first data. Adjusting the model parameters of the second classification model to obtain a data classification model;
  • the prediction module 730 is configured to perform classification prediction on the target data through the data classification model, and obtain a classification result of the target data.
  • the prediction module 730 is further configured to use the first classification model to classify and predict the first data to obtain a first prediction result;
  • the prediction module 730 is further configured to classify and predict the first data through the second classification model to obtain a second prediction result
  • the training module 720 also includes:
  • An adjustment unit 721 configured to adjust the model parameters of the second classification model based on the difference between the first prediction result and the second prediction result.
  • the obtaining module 710 is further configured to obtain a confidence parameter of the first prediction result
  • the adjustment unit 721 is further configured to determine a difference between the first prediction result under the confidence parameter and the second prediction result, and adjust the value of the second classification model based on the difference.
  • the model parameters are adjusted.
  • the prediction module 730 is further configured to use the first classification model to perform classification prediction on the first data, and obtain the probability value of the classification in the target classification set corresponding to the first data;
  • the first predictor is determined from the set of target categories based on the probability value.
  • the second classification model includes a first query encoder and a first key-value encoder
  • the device also includes:
  • An encoding module 740 configured to encode the first data through the first query encoder to obtain a first encoding result
  • the acquiring module 710 is further configured to acquire a second encoding result of encoding the first data and the data in the first preset dynamic dictionary by the first key-value encoder;
  • the training module 720 is further configured to train the second classification model based on the difference between the first encoding result and the second encoding result.
  • the training module 720 is further configured to perform unsupervised training on the feature extraction network in the candidate classification model based on the first data;
  • the feature extraction network is combined to obtain a classification model, and the classification regression network is used to classify data in the target classification set;
  • the training module 720 is further configured to perform supervised training on the classification model by using the second data in the second data set and the sample label to obtain the first classification model.
  • the feature extraction network includes a second query encoder and a second key-value encoder
  • the device also includes:
  • An encoding module 740 configured to encode the first data through the second query encoder to obtain a third encoding result
  • the acquiring module 710 is further configured to acquire a fourth encoding result of encoding the first data and the data in the second preset dynamic dictionary by the second key-value encoder;
  • the training module 720 is further configured to perform unsupervised training on the feature extraction network based on the difference between the third encoding result and the fourth encoding result.
  • the training module 720 is further configured to connect the classification regression network with the second query encoder that has undergone unsupervised training to obtain the classification model.
  • the obtaining module 710 is also used to obtain a test data set, the test data in the test data set is used to test the training effect of the data classification model; from the test data Centrally acquiring the target data, where the target data is marked with reference classification information;
  • the prediction module 730 is further configured to classify and predict the target data through the data classification model to obtain the classification result;
  • the obtaining module 710 is further configured to obtain training effect data of the data classification model based on the reference classification information and the classification result.
  • the data classification and identification device obtains the first classification model after performing unsupervised training on unlabeled first data and supervised training on labeled second data, so that in the first classification Based on the model, create a second classification model for knowledge distillation training, use the teacher model for supervised training to achieve the purpose of distillation, and finally obtain a student model with higher performance and accuracy.
  • the training mainly relies on a large amount of first data, and for some The data volume requirement of the second data of the label is small, avoiding the cumbersome process of labeling a large amount of sample data, and improving the training efficiency and accuracy of the data classification model.
  • the data classification identification device provided by the above-mentioned embodiment is only illustrated by the division of the above-mentioned functional modules. The structure is divided into different functional modules to complete all or part of the functions described above.
  • the data classification and recognition device provided in the above embodiments and the data classification and recognition method embodiment belong to the same concept, and the specific implementation process thereof is detailed in the method embodiment, and will not be repeated here.
  • Fig. 9 shows a schematic structural diagram of a server provided by an exemplary embodiment of the present application.
  • the server 900 includes a central processing unit (Central Processing Unit, CPU) 901, a system memory 904 including a random access memory (Random Access Memory, RAM) 902 and a read only memory (Read Only Memory, ROM) 903, and A system bus 905 that connects the system memory 904 and the central processing unit 901 .
  • Server 900 also includes mass storage device 906 for storing operating system 913 , application programs 914 and other program modules 915 .
  • Mass storage device 906 is connected to central processing unit 901 through a mass storage controller (not shown) connected to system bus 905 .
  • Mass storage device 906 and its associated computer-readable media provide non-volatile storage for server 900 . That is, mass storage device 906 may include a computer-readable medium (not shown) such as a hard disk or a Compact Disc Read Only Memory (CD-ROM) drive.
  • CD-ROM Compact Disc Read Only Memory
  • Computer-readable media may comprise computer storage media and communication media.
  • Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.
  • Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (Erasable Programmable Read Only Memory, EPROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other solid-state storage Its technology, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, tape cartridge, tape, disk storage or other magnetic storage device.
  • the computer storage medium is not limited to the above-mentioned ones.
  • the above-mentioned system memory 904 and mass storage device 906 may be collectively referred to as memory.
  • the server 900 can also run on a remote computer connected to the network through a network such as the Internet. That is to say, the server 900 can be connected to the network 912 through the network interface unit 911 connected to the system bus 905, or can use the network interface unit 911 to connect to other types of networks or remote computer systems (not shown).
  • the above-mentioned memory also includes one or more programs, one or more programs are stored in the memory and configured to be executed by the CPU.
  • the embodiment of the present application also provides a computer device, the computer device includes a processor and a memory, at least one instruction, at least one section of program, code set or instruction set are stored in the memory, at least one instruction, at least one section of program, code The set or instruction set is loaded and executed by the processor to implement the data classification and identification methods provided by the above method embodiments.
  • Embodiments of the present application also provide a computer-readable storage medium, on which at least one instruction, at least one program, code set or instruction set is stored, at least one instruction, at least one program, code set or The instruction set is loaded and executed by the processor, so as to implement the data classification and identification methods provided by the above method embodiments.
  • the computer-readable storage medium may include: a read-only memory (ROM, Read Only Memory), a random access memory (RAM, Random Access Memory), a solid-state hard drive (SSD, Solid State Drives) or an optical disc, etc.
  • random access memory may include resistive random access memory (ReRAM, Resistance Random Access Memory) and dynamic random access memory (DRAM, Dynamic Random Access Memory).
  • ReRAM resistive random access memory
  • DRAM Dynamic Random Access Memory
  • Embodiments of the present application also provide a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the data classification identification method described in any one of the above embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Databases & Information Systems (AREA)
  • Computing Systems (AREA)
  • Medical Informatics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Mathematical Physics (AREA)
  • Human Computer Interaction (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种数据分类识别方法、装置、设备及可读存储介质，涉及机器学习领域。该方法包括：获取第一数据集和第二数据集(301)，第二数据集中包括标注有样本标签的第二数据；通过第一数据以无监督训练模式，以及第二数据以监督训练模式训练得到第一分类模型(302)；获取第二分类模型(303)；对第二分类模型的模型参数进行蒸馏训练，得到数据分类模型(304)；通过数据分类模型对目标数据进行分类预测(305)。

Description

数据分类识别方法、装置、设备、介质及程序产品
本申请要求于2021年05月17日提交的申请号为202110532246.3、发明名称为“数据分类识别方法、装置、设备及可读存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及机器学习领域，特别涉及一种数据分类识别方法、装置、设备、介质及程序产品。
背景技术
在基于医学影像的疾病诊断方面通常包括罕见病的诊断和常见病的诊断，也即，将医学影像输入至机器学习模型后，由机器学习模型对医学影像进行分析，从而判断医学影像所对应的身体异常情况。
相关技术中，在针对罕见病进行诊断时，将医学影像输入至罕见病的分类模型中，由分类模型对医学影像进行分析诊断，从而确定医学影像所表达的图像特征属于哪一种罕见病。其中，分类模型在训练过程中，需要大量标注有罕见病信息的图像数据集，从而确保模型准确率。
然而，罕见病本身属于出现几率较低的病症，收集罕见病的图像数据以及对罕见病信息进行标注的难度较大，导致分类模型的训练效率较低，从而分类模型的分类准确率较低。
发明内容
本申请实施例提供了一种数据分类识别方法、装置、设备、介质及程序产品，能够提高对针对罕见病进行识别分类的识别模型的训练效率。所述技术方案如下。
一方面，提供了一种数据分类识别方法，应用于计算机设备，所述方法包括：
获取第一数据集和第二数据集，所述第一数据集中包括第一数据，所述第二数据集中包括标注有样本标签的第二数据；
通过所述第一数据以无监督训练模式，以及通过所述第二数据以监督训练模式对候选分类模型进行训练，得到第一分类模型；
获取第二分类模型，所述第二分类模型为模型参数待调整的分类模型；
以所述第一分类模型对所述第一数据的第一预测结果为基准，基于所述第二分类模型对所述第一数据的第二预测结果对所述第二分类模型的所述模型参数进行调整，得到数据分类模型；
通过所述数据分类模型对目标数据进行分类预测，得到所述目标数据的分类结果。另一方面，提供了一种数据分类识别装置，所述装置包括：
获取模块，用于获取第一数据集和第二数据集，所述第一数据集中包括第一数据，所述第二数据集中包括标注有样本标签的第二数据；
训练模块，用于通过所述第一数据以无监督训练模式，以及通过所述第二数据以监督训练模式对候选分类模型进行训练，得到第一分类模型；
所述获取模块，还用于获取第二分类模型，所述第二分类模型为模型参数待调整的分类模型；
所述训练模块，还用于以所述第一分类模型对所述第一数据的第一预测结果为基准，基于所述第二分类模型对所述第一数据的第二预测结果对所述第二分类模型的所述模型参数进行调整，得到数据分类模型；
预测模块，用于通过所述数据分类模型对目标数据进行分类预测，得到所述目标数据的分类结果。
另一方面，提供了一种计算机设备，所述计算机设备包括处理器和存储器，所述存储器中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由所述处理器加载并执行以实现如上述本申请实施例中任一所述数据分类识别方法。
另一方面，提供了一种计算机可读存储介质，所述存储介质中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由处理器加载并执行以实现如上述本申请实施例中任一所述的数据分类识别方法。
另一方面，提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述实施例中任一所述的数据分类识别方法。
本申请实施例提供的技术方案带来的有益效果至少包括：
在通过无标签的第一数据进行无监督训练以及有标签的第二数据进行监督训练后，得到第一分类模型，从而在第一分类模型的基础上，创建第二分类模型进行知识蒸馏训练，利用教师模型进行监督训练来达到蒸馏的目的，最终得到更高性能和精度的学生模型，训练主要依赖大量的第一数据，而对有标签的第二数据的数据量要求较小，避免了对样本数据进行大量标注的繁琐过程，提高了数据分类模型的训练效率以及准确率。
附图说明
图1是本申请一个示例性实施例提供的整体方案实施流程示意图；
图2是本申请一个示例性实施例提供的实施环境示意图；
图3是本申请一个示例性实施例提供的数据分类识别方法的流程图；
图4是本申请另一个示例性实施例提供的数据分类识别方法的流程图；
图5是本申请另一个示例性实施例提供的数据分类识别方法；
图6是本申请一个示例性实施例提供的罕见病分类识别模型的训练过程整体示意图；
图7是本申请一个示例性实施例提供的数据分类识别装置的结构框图；
图8是本申请另一个示例性实施例提供的数据分类识别装置的结构框图；
图9是本申请一个示例性实施例提供的服务器的结构框图。
具体实施方式
首先，针对本申请实施例中涉及的名词进行简单介绍。
伪标签：是指通过经过训练的模型对未标注的数据进行预测后得到预测结果，并基于预测结果对数据进行标注的标签。也即，伪标签并非根据数据的实际情况人工标注的标签，而是由训练好的模型标注的存在一定容错率的标签。在一些实施例中，针对训练得到的分类模型对数据进行分类预测后所得到的分类预测结果即为该数据对应的伪标签。
相关技术中，针对罕见病的诊断，需要通过用于罕见病诊断的分类模型，而分类模型的训练则需要通过标注有罕见病信息的大量样本图像数据，通过分类模型对样本图像数据进行分类识别后，得到识别结果，通过标注的罕见病信息与识别结果之间的差异对分类模型进行训练。然而，由于罕见病本身的罕见性，导致样本图像数据的获取难度较大，需要大量的人力采集样本图像数据，并对样本图像数据进行罕见病信息的标注，分类模型的训练效率较低。而正是由于罕见病的样本图像数据获取的难度大，会导致训练样本数据不足的情况，所训练得到的分类模型的准确率较低。
本申请实施例中，提供了一种数据分类识别方法，在标注有标签的样本数量较少的情况 下提高了数据分类模型的训练效率和准确率。
示意性的，图1是本申请一个示例性实施例提供的整体方案实施流程示意图，以罕见病的分类模型训练过程为例，如图1所示。
首先获取第一图像数据集110和第二图像数据集120，其中，第一图像数据集110中包括常见病的医学影像，且第一图像数据集110中的医学影像不采用标签信息或者未标注标签信息；第二图像数据集120中包括少量罕见病的医学影像，且第二图像数据集120中的医学影像包括标注的标签信息，该标签信息用于表示医学影像对应的罕见病信息。
通过第一图像数据集110对特征提取网络f q进行无监督训练后，将经过无监督训练的f q与分类网络f c连接，得到第一分类模型F，通过第二图像数据集120对第一分类模型F进行监督训练，并基于训练后的第一分类模型F对第二分类模型F’进行知识蒸馏训练，从而得到罕见病的分类模型(也即训练后的第二分类模型F’)。
其中，知识蒸馏训练是指由训练好的模型的分类能力为基础，引导未训练好的模型的分类能力。本实施例中，知识蒸馏训练在实现过程中，主要是由第一分类模型F对数据的预测结果为基准，对第二分类模型F’对数据的预测能力进行引导，也即，将数据a输入第一分类模型F后，输出第一分类模型F预测得到的伪标签，将数据a输入第二分类模型F’后，输出第二分类模型F’预测得到的分类结果，根据该分类结果和伪标签的差异，对第二分类模型F’进行训练，从而引导第二分类模型F’的分类准确率向第一分类模型F靠拢。
其次，对本申请实施例中涉及的实施环境进行说明，示意性的，请参考图2，该实施环境中涉及终端210、服务器220，终端210和服务器220之间通过通信网络230连接。
在一些实施例中，终端210包括第一终端211和第二终端212。
第一终端211用于向服务器220发送医学影像。示意性的，第一终端211为医生应用的终端，医生在通过医学影像对罕见病进行诊断的过程中，通过分类模型进行辅助诊断，从而提高诊断准确率；或者，第一终端211为用户应用的终端，如：问诊人本人，或者问诊人的亲属等，用户将医学影像发送至服务器，从而获取参考诊断结果；或者，第一终端211为医学影像扫描设备所连接的终端，医学影像扫描设备在扫描得到医学影像后传输至第一终端211，第一终端211在接收到医学影像后，将医学影像转发至服务器220进行辅助诊断。
服务器220通过上述图1所示的方式进行罕见病分类模型221的训练，得到罕见病分类模型221后，接收第一终端211上传的医学影像，并通过罕见病分类模型221对医学影像进行分类识别，得到医学影像在罕见病分类集中的分类诊断结果。服务器220将分类诊断结果反馈至第一终端211或者将分类诊断结果发送至第二终端212。
其中，当第一终端211实现为与医学影像扫描设备连接的终端时，服务器220将分类诊断结果发送至第二终端212，第二终端212实现为医生应用的终端或者用户应用的终端。
上述终端可以是手机、平板电脑、台式电脑、便携式笔记本电脑等多种形式的终端设备，本申请实施例对此不加以限定。
值得注意的是，上述服务器可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、内容分发网络(Content Delivery Network，CDN)、以及大数据和人工智能平台等基础云计算服务的云服务器。
其中，云技术(Cloud technology)是指在广域网或局域网内将硬件、软件、网络等系列资源统一起来，实现数据的计算、储存、处理和共享的一种托管技术。
在一些实施例中，上述服务器还可以实现为区块链系统中的节点。
需要说明的是，本申请所涉及的信息(包括但不限于用户设备信息、用户个人信息等)、数据(包括但不限于用于分析的数据、存储的数据、展示的数据等)以及信号，均为经用户单独授权或者经过各方充分授权的，且相关数据的收集、使用和处理需要遵守相关国家和地区的相关法律法规和标准。例如，本申请中涉及到的医学影像都是在充分授权的情况下获取 的。
另外，需要说明的是，本申请涉及到的医学影像在获取授权时，会充分表明医学影像的获取用途，并获得影像所有人的授权。
结合上述名词简介，对本申请实施例中涉及的应用场景进行举例说明。
第一，医生通过分类模型进行辅助诊断的场景。
也即，医生通过终端将医学影像发送至服务器，服务器通过训练好的分类模型对医学影像进行分类识别，得到与医学影像对应的分类诊断结果，并将分类诊断结果反馈至医生所应用的终端进行展示，从而医生通过分类诊断结果进行辅助诊断，并得出最终的诊断结果。
第二，用户通过分类模型进行预诊断。
用户(问诊人或者问诊人的亲友)将医学影像发送至服务器，服务器通过训练好的分类模型对医学影像进行分类识别，得到与医学影像对应的分类诊断结果，并将分类诊断结果反馈至用户应用的终端进行展示，用户根据分类诊断结果首先对异常生命状态进行初步了解，继而通过医生诊断得到详细诊断结果。
第三，分类模型还可以应用于其他分类场景。
示意性的，该分类模型还可以应用于物体识别场景、语音识别场景、笔迹识别场景等，本申请实施例对此不加以限定。
结合上述名词简介和应用场景，对本申请提供的数据分类识别方法进行说明，以该方法应用于服务器中为例，如图3所示，该方法包括如下步骤。
步骤301，获取第一数据集和第二数据集。
其中，第一数据集中包括第一数据，第二数据集中包括标注有样本标签的第二数据。
在一些实施例中，第一数据集中的第一数据为未标注有标签的数据，而第二数据集中的第二数据为标注有样本标签的数据；或者，第一数据集中的第一数据标注有标签，但第一数据标注的标签在本实施例中不作应用。
可选地，第一数据属于第一分类集，第二数据属于目标分类集，也即第一数据和第二数据属于不同分类集对应的数据。其中，分类集用于表示包含多个下属分类的集合，即，分类集属于分类的上层归纳概念，在上层归纳中包括属于该归纳特点的下层分类。示意性的，第一级归纳概念为疾病，属于疾病概念的分为两个下层分类，分别为常见病和罕见病，则第一分类集对应为常见病分类集，以眼部疾病为例，如：第一分类集中包括近视、远视、结膜炎等常见眼疾类型；目标分类集对应为罕见病分类集，以眼部疾病为例，如：目标分类集中包括干眼症、视雪症、遗传性视神经病变等罕见眼疾类型。
本申请实施例中，常见病和罕见病是针对同一器官或者同一身体部分对应的病症，或者，常见病和罕见病属于同一病症类型，如：身体疼痛类型、呕吐类型。
在一些实施例中，第一数据集中包括的第一数据为与常见病对应的医学影像，如：电子计算机断层扫描(Computed Tomography，CT)图像、X光图像、超声波图像等形式的影像；第二数据集中包括的第二数据为与和罕见病对应的医学影像，如：CT图像、X光图像、超声波图像等形式的影像。
值得注意的是，上述医学影像仅为示意性的举例，本申请实施例中的第一数据和第二数据还可以实现为其他类型的数据，如：语音数据、文本数据等，本申请实施例对此不加以限定。
可选地，第一数据集中第一数据的数据量(即医学影像的数量)大于第二数据集中第二数据的数据量(即医学影像的数量)，可选地，第二数据集中第二数据的数量在要求数量范围内，如：小于预设数量。
可选地，第一数据集中的第一数据是从基础数据集中随机采样的数据，基础数据集中包括常见病数据；第二数据集中的第二数据是从罕见病数据集中随机采样的数据，罕见病数据 集中包括罕见病数据，第二数据标注有罕见病信息，也即每个医学影像所对应的罕见病类型。
步骤302，通过第一数据以无监督训练模式，以及通过第二数据以监督训练模式对候选分类模型进行训练，得到第一分类模型。
在一些实施例中，基于第一数据对候选分类模型中的特征提取网络进行无监督训练，将候选分类模型中的分类回归网络与经过无监督训练的特征提取网络结合，得到分类模型，其中，分类回归网络用于进行数据分类，可选地，分类回归网络用于在目标分类集(也即第二数据所处的分类集)中进行数据分类。通过第二数据集中的第二数据和样本标签对分类模型进行监督训练，得到第一分类模型。
由于第一数据集中的第一数据为不存在对应标注的标签的数据，故，第一数据仅能够用于对特征提取网络进行无监督训练。而第二数据集中的第二数据存在对应标注的样本标签，故，第二数据能够用于对分类模型进行监督训练。
步骤303，获取第二分类模型。
第二分类模型为模型参数待调整的分类模型。
可选的，第二分类模型为随机初始化的分类模型，第二分类模型中包括模型参数，第二分类模型用于根据第一分类模型输出的知识进行蒸馏训练。其中，知识蒸馏是指将教师模型输出的监督信息作为知识，由学生模型对迁移自教师模型的监督信息进行学习的过程，利用教师模型进行监督训练来达到蒸馏的目的，最终得到更高性能和精度的学生模型。
步骤304，以第一分类模型对第一数据的第一预测结果为基准，基于第二分类模型对第一数据的第二预测结果对第二分类模型的模型参数进行调整，得到数据分类模型。
可选地，通过第一分类模型对第一数据进行分类预测，得到第一预测结果，可选地，将该第一预测结果作为第一数据对应目标分类集中的类别伪标签；通过第二分类模型对第一数据进行分类预测，得到第二预测结果，基于第一预测结果与第二预测结果之间的差异对第二分类模型的模型参数进行调整，得到数据分类模型。其中，第一分类模型和第二分类模型针对相同的第一数据进行预测，得到第一分类模型输出的第一预测结果，以及得到第二分类模型输出的第二预测结果。示意性的，第一数据集中包括数据A、数据B和数据C，则首先通过第一分类模型对数据A进行预测，得到第一预测结果a，通过第二分类模型对数据A进行预测，得到第二预测结果b，根据第一预测结果a和第二预测结果b之间的差异，对第二分类模型的模型参数进行调整。
也即，将第一分类模型对第一数据进行分类预测后输出的伪标签作为知识，由第二分类模型迁移该伪标签进行蒸馏，从而实现第二分类模型的知识蒸馏训练。其中，第二分类模型迁移该伪标签进行蒸馏是指，第二分类模型在对第一数据进行分类预测后，以该伪标签的基准，对预测结果进行评估，从而对模型参数进行调整，也即，将第一分类模型所识别得到的伪标签作为数据标签，指引第二分类模型的分类预测结果向伪标签靠齐，从而提高第二分类模型的分类预测准确率。
步骤305，通过数据分类模型对目标数据进行分类预测，得到目标数据的分类结果。
在一些实施例中，通过数据分类模型对目标数据进行分类预测，得到目标数据在目标分类集中所属的分类结果，其中，目标分类集为标注有样本标签的第二数据所属的分类集。
在第二分类模型经过训练后，得到数据分类模型，通过数据分类模型对目标数据进行分类，即可得到目标数据在目标分类集中的分类结果。其中，目标数据可以是实际应用时的数据，如：实际应用时的医学影像；或者，目标数据也可以是测试集中用于对数据分类模型进行测试的数据。
综上所述，本实施例提供的数据分类识别方法，在通过无标签的第一数据进行无监督训练以及有标签的第二数据进行监督训练后，得到第一分类模型，从而在第一分类模型的基础上，创建第二分类模型进行知识蒸馏训练，利用教师模型进行监督训练来达到蒸馏的目的，最终得到更高性能和精度的学生模型，训练主要依赖大量的第一数据，而对有标签的第二数据的数据量要求较小，避免了对样本数据进行大量标注的繁琐过程，提高了数据分类模型的 训练效率以及准确率。
在一些实施例中，通过第一分类模型对第二分类模型进行蒸馏训练的过程中，需要通过第一分类模型识别得到的伪标签作为知识，图4是本申请另一个示例性实施例提供的数据分类识别方法的流程图，以该方法应用于服务器中为例进行说明，如图4所示，该方法包括如下步骤。
步骤401，获取第一数据集和第二数据集。
其中，第一数据集中包括第一数据，第二数据集中包括标注有样本标签的第二数据。
在一些实施例中，第一数据集中的第一数据为未标注有标签的数据，而第二数据集中的第二数据为标注有样本标签的数据。
可选地，第一数据属于第一分类集，第二数据属于目标分类集，也即第一数据和第二数据属于不同分类集对应的数据。示意性的，第一分类集对应为常见病分类集；目标分类集对应为罕见病分类集。
步骤402，通过第一数据以无监督训练模式，以及通过第二数据以监督训练模式对候选分类模型进行训练，得到第一分类模型。
在一些实施例中，基于第一数据集中的第一数据对候选分类模型中的特征提取网络进行无监督训练，并将候选分类模型中的分类回归网络与经过无监督训练的特征提取网络结合，得到分类模型，其中，分类回归网络用于在目标分类集中进行数据分类，通过第二数据集中的第二数据和样本标签对分类模型进行监督训练，得到第一分类模型。
第一分类模型具有较好的分类性能，但在表征学习的过程中，忽略了与目标分类集相关的知识，故，本申请实施例中，将第一分类模型作为基准模型，通过第一分类模型输出的知识对第二分类模型进行蒸馏训练。其中，第二分类模型为模型参数待调整的用于在目标分类集进行分类的模型。
步骤403，通过第一分类模型对第一数据进行分类预测，得到第一预测结果。
由于即使第一数据集和第二数据集所包含的数据类别不同，但数据具有相似的特征，示意性的，第一数据集为常见病的医学影像，第二数据集为罕见病的医学影像，则第一数据集和第二数据集的数据在颜色、纹理或者形状上具有相似的特征。因此，采用第一分类模型作为基准模型预测第一数据集中的图像属于目标分类集中各个分类的概率。
在一些实施例中，通过第一分类模型对第一数据集中的第一数据进行分类预测，得到第一数据对应目标分类集中分类的概率值，并基于概率值从目标分类集中确定第一数据对应的伪标签作为第一预测结果。
可选地，首先通过第一分类模型确定第一数据对应目标分类集中类别的软标签，也即对应目标分类集中类别的概率：p＝F(x)＝[p 1，…，p n] T，其中，p表示概率，n表示目标分类集中类别数量，F(x)表示对第一数据x采用第一分类模型F进行分类预测，n个类别的概率和为1。根据上述确定的软标签进一步确定第一数据对应目标分类集中类别的硬标签。示意性的，根据上述概率p，将概率最大的类别标注为第一数据对应的伪标签，也即，概率最大的类别的硬标签为1，其他类别硬标签为0，而硬标签为1的类别即为第一数据对应的第一预测结果，也即为第一数据对应的伪标签。其中，针对一个第一数据x通过分类预测得到一个第一预测结果，也即，针对m个第一数据通过分类预测得到m个第一预测结果，m为正整数，每个第一数据能够通过分类预测得到一个第一预测结果。
步骤404，获取第二分类模型，并通过第二分类模型对第一数据进行分类预测，得到与第一数据对应的第二预测结果。
第二分类模型为模型参数待调整的模型，第二分类模型用于对应目标分类集对数据进行分类。
其中，通过第一分类模型和第二分类模型对同一个第一数据进行分类预测后，分别得到第一预测结果和第二预测结果，则对应同一第一数据的第一预测结果和第二预测结果之间存 在比对意义。
步骤405，基于第一预测结果与第二预测结果之间的差异对第二分类模型的模型参数进行调整，得到数据分类模型。
可选地，第二分类模型中包括第一查询编码器和第一键值编码器，则通过第一查询编码器对第一数据进行编码，得到第一编码结果，通过第一键值编码器对第一数据和第一预设动态字典中的数据进行编码的第二编码结果，基于第一编码结果和第二编码结果的差异对第二分类模型进行训练，得到数据分类模型。
在一些实施例中，结合伪标签监督方法与对比判别方法进行混合蒸馏损失的确定，其中伪标签监督方法即为基于第一预测结果与第二预测结果之间的差异对第二分类模型的模型参数进行调整，对比判别方法即为通过第一查询编码器与第一键值编码器对第二分类模型进行训练。可选地，采用随机初始化学生模型的策略，其中第二分类模型F’＝f’ c(f’ q)，f’ q对应为第一查询编码器，具有模型参数θ’ q，f’ c对应为回归分类网络，具有模型参数θ’ c，确定混合损失L dis，计算公式如下公式一所示。
公式一：L dis＝L con(x；θ’ q，θ’ k)+L cls(y，F’(x；θ’ q，θ’ c))
其中，x为第一数据集中的第一数据，θ’ k为第一键值编码器f’ k的参数，并随着θ’ q进行动量更新，第一键值编码器f’ k对应有第一预设动态词典，将第一查询编码器f’ q提取的第一编码结果与第一键值编码器对第一预设动态词典中的数据编码的第二编码结果进行比对，得到对比判别方法中的损失L con。y表示第一分类模型识别得到的第一预测结果，将第一预测结果与第二分类模型的第二预测结果进行比对，得到伪标签监督方法对应的损失L cls。从而将两部分损失相加，对模型参数θ’ q进行更新。
在一些实施例中，与基准模型不同的是，f’ c采用全连接层(后接softmax操作)实现，便于进行端到端的模型训练。
在训练中，由于罕见病对应的第二数据集中数据量较少以及其产生的噪声和偏差，第一分类模型生成的上述第一预测结果不是完全可用的并且可能对第二分类模型的训练造成不利影响。故，本申请实施例中，第一预测结果还对应有置信度参数，获取第一预测结果的置信度参数，确定第一预测结果在置信度参数下与第二预测结果之间的差异，并基于差异对第二分类模型的模型参数进行调整，得到数据分类模型。
示意性的，本实施例中，将第二分类模型的预测值p’与伪标签y(也即上述第一预测结果y)结合作为训练目标，请参考如下公式二。
公式二：y adpt＝(1-α)×y+α×p’
其中，α为置信度参数，控制第一分类模型生成的伪标签y所占训练目标的比例。通常α为一个固定值，然而，在训练的初始阶段，学生模型所产生的预测值的可信度较低。因此本申请采用线性增长方法，在第t个训练回合的α为：α t＝α T×(t/T)。其中，α T为最后一个训练回合的置信度参数值，示意性的，α T设置为0.7。T为总训练回合数。最后，用y adpt替代上述公式一中的y作为最终的损失函数。
步骤406，通过数据分类模型对目标数据进行分类预测，得到目标数据在目标分类集中所属的分类结果。
在第二分类模型经过训练后，得到数据分类模型，通过数据分类模型对目标数据进行分类，即可得到目标数据在目标分类集中的分类结果。
在一些实施例中，获取测试数据集，测试数据集中的测试数据用于对数据分类模型的训练效果进行测试，从测试数据集中获取目标数据，目标数据标注有参考分类信息，通过数据分类模型对目标数据进行分类预测得到分类结果后，基于参考分类信息和分类结果获取数据分类模型的训练效果数据。示意性的，获取测试数据集中的多个目标数据，分别进行分类预测，并与参考分类信息进行比对，根据比对结果正确的目标数据占被测试的目标数据总数的比例，确定训练效果，也即确定数据分类模型的预测准确率。
综上所述，本实施例提供的数据分类识别方法，在通过无标签的第一数据进行无监督训 练以及有标签的第二数据进行监督训练后，得到第一分类模型，从而在第一分类模型的基础上，创建第二分类模型进行知识蒸馏训练，利用教师模型进行监督训练来达到蒸馏的目的，最终得到更高性能和精度的学生模型，训练主要依赖大量的第一数据，而对有标签的第二数据的数据量要求较小，避免了对样本数据进行大量标注的繁琐过程，提高了数据分类模型的训练效率以及准确率。
本实施例提供的方法，结合伪标签监督方法与对比判别方法进行混合蒸馏损失的确定，在通过第一分类模型对第二分类模型进行蒸馏训练的同时，避免第二分类模型对数据的特征提取被蒸馏训练过程影响，提高了第二分类模型的训练效率和准确率。
本实施例提供的方法，在确定损失值时，由于第二数据集中数据量较少以及其产生的噪声和偏差，导致第一分类模型本身的第一预测结果存在一定偏差，故引入了第一预测结果的置信度参数，避免第一预测结果在准确率上的偏差影响第二分类模型的训练效果，提高了分类模型模型的预测准确率。
本实施例提供的方法，真对置信度参数引入线性增长方法，逐步调整置信度参数对第一预测结果的调整情况，避免置信度参数由于人为主观设置而导致对第一预测结果的干预过高或者过低，提高了损失值的计算准确率。
在一些实施例中，第一分类模型是通过第一数据的无监督训练和第二数据的监督训练得到的。图5是本申请另一个示例性实施例提供的数据分类识别方法的流程图，如图5所示，以该方法应用于服务器中为例，该方法包括如下步骤。
步骤501，获取第一数据集和第二数据集。
其中，第一数据集中包括第一数据，第二数据集中包括标注有样本标签的第二数据，第二数据属于目标分类集。
可选地，第一数据属于第一分类集，第二数据属于目标分类集，也即第一数据和第二数据属于不同分类集对应的数据。
步骤502，基于第一数据对候选分类模型中的特征提取网络进行无监督训练。
在一些实施例中，特征提取网络中包括第二查询编码器和第二键值编码器则通过第二查询编码器对第一数据进行编码，得到第三编码结果，获取第二键值编码器对第二预设动态字典中的数据进行编码的第四编码结果，基于第三编码结果和第四编码结果的差异对特征提取网络进行无监督训练。
无监督表征学习能够在无标注数据的情况下训练一个较好的特征提取模型，故，本申请实施例中，采用对比损失作为特征提取网络的优化函数。
可选地，在通过特征提取网络对第一数据进行特征提取时，将第一数据进行数据增强，以第一数据为医学影像为例，则对第一数据集中的医学影像进行图像增强。图像增强包括对比度增强、亮度增强、锐化增强等增强处理方式中的至少一种。其中，图像增强的次数两次，从而分别输入第二查询编码器和第二键值编码器，其中，两次图像增强是针对同一个增强方向进行两次不同程度的增强；或者，两次图像增强是针对不同增强方向进行两次相同或者不同程度的增强；或者，两次图像增强是针对同一增强方向进行两次叠加的相同或不同程度的增强。示意性的，对第一数据集中的每张图像进行两次图像增强，得到
Figure PCTCN2022090902-appb-000001
和
Figure PCTCN2022090902-appb-000002
其中，
Figure PCTCN2022090902-appb-000003
是经过第一次图像增强后得到的图像，
Figure PCTCN2022090902-appb-000004
是经过第二次图像增强后得到的图像。分别通过第二查询编码器
Figure PCTCN2022090902-appb-000005
和第二键值编码器
Figure PCTCN2022090902-appb-000006
进行特征提取，得到相应的特征，其中，f q和f k分别为由参数θ q和θ k组成的第二查询编码器和第二键值编码器，则对比损失的计算方式请参考如下公式三。
Figure PCTCN2022090902-appb-000007
其中，x i为第一数据集中的第一数据，l为存储在第二预设动态字典中的键值图像的数量， τ为平滑标签的超参数。通过对比损失的训练，模型能够区分图像x i与存储在第二预设动态字典中的键值图像，并根据图像x i与存储在第二预设动态字典中的键值图像的差异通过反向传播更新参数θ q，而θ k通过θ q进行动量更新：θ k←mθ k+(1-m)θ q，其中，m∈[0，1)。
通过第一数据集对特征提取网络完成无监督训练后，冻结参数θ q。
步骤503，将候选分类模型中的分类回归网络与经过无监督训练的特征提取网络结合，得到分类模型。
在一些实施例中，分类回归网络用于在目标分类集中进行数据分类。
可选地，由于上述特征提取网络对应有第二查询编码器和第二键值编码器，在将分类回归网络与特征提取网络结合时，本申请实施例中，将分类回归网络与经过无监督训练的第二查询编码器连接，得到分类模型。
步骤504，通过第二数据集中的第二数据和样本标签对分类模型进行监督训练，得到第一分类模型。
在一些实施例中，通过第二数据对分类模型进行监督训练时，将第二数据输入分类模型进行分类预测，得到预测结果，而第二数据本身标注有样本标签，用于指示第二数据的实际分类，从而根据样本标签与预测结果之间的差异反向对分类模型的模型参数进行调整，得到第一分类模型。
可选地，根据样本标签与预测结果计算该预测结果的损失值，从而根据损失值反向对分类模型的模型参数进行调整，直至预测结果对应的损失值收敛，如：预测结果对应的损失值小于预设阈值；或者，第q次迭代训练中预测结果的损失值，与第q-1次迭代训练中预测结果的损失值之差小于预设差值阈值，q为大于1的整数。
步骤505，获取第二分类模型。
第二分类模型为模型参数待调整的分类模型。
可选的，第二分类模型为随机初始化的分类模型，第二分类模型中包括模型参数，在随机初始化时，第二分类模型的初始模型参数为随机获取的。第二分类模型用于根据第一分类模型输出的知识进行蒸馏训练。其中，知识蒸馏是指将第一分类模型输出的监督信息，也即第一分类模型输出的第一预测结果作为知识，由第二分类模型学习迁移自第一分类模型的监督信息作为蒸馏过程，利用第一分类模型进行监督训练来达到蒸馏的目的，最终得到更高性能和精度的学生模型。
步骤506，以第一分类模型对第一数据的第一预测结果为基准，基于第二分类模型对第一数据的第二预测结果对第二分类模型的模型参数进行调整，得到数据分类模型。
可选地，通过第一分类模型对第一数据集中的第一数据进行分类预测，得到与第一数据对应目标分类集中类别的第一预测结果；通过第二分类模型对第一数据集中的第一数据进行分类预测，得到与第一数据对应的第二预测结果，基于第一预测结果与第二预测结果之间的差异对第二分类模型的模型参数进行调整，得到数据分类模型。
也即，将第一分类模型对第一数据进行分类预测后输出的伪标签作为知识，由第二分类模型迁移该伪标签进行蒸馏，从而实现第二分类模型的蒸馏训练。
步骤507，通过数据分类模型对目标数据进行分类预测，得到目标数据的分类结果。
在第二分类模型经过训练后，得到数据分类模型，通过数据分类模型对目标数据进行分类，即可得到目标数据在目标分类集中的分类结果。其中，目标数据可以是实际应用时的数据，如：实际应用时的医学影像；或者，目标数据也可以是测试集中用于对数据分类模型进行测试的数据。
综上所述，本实施例提供的数据分类识别方法，在通过无标签的第一数据进行无监督训练以及有标签的第二数据进行监督训练后，得到第一分类模型，从而在第一分类模型的基础上，创建第二分类模型进行知识蒸馏训练，利用教师模型进行监督训练来达到蒸馏的目的，最终得到更高性能和精度的学生模型，训练主要依赖大量的第一数据，而对有标签的第二数据的数据量要求较小，避免了对样本数据进行大量标注的繁琐过程，提高了数据分类模型的 训练效率以及准确率。
本实施例提供的方法，通过第一数据集中无标签的第一数据对特征提取网络进行无监督训练，从而通过第二数据集中有标签的第二数据对分类模型进行监督训练，从而在第二数据的采集过程较为繁琐，或者第二数据的收集难度较大时，仅需要少量采集第二数据，即可实现对第一分类模型的有效训练，提高了模型的训练效率。
本实施例提供的方法，在训练得到数据分类模型后，通过测试数据集中的测试数据对数据分类模型进行测试后，确定数据分类模型的训练效果，从而辅助数据分类模型的进一步训练或者应用，提高了数据分类模型的分类准确率。
结合上述内容，以上述第一数据集中的第一数据为常见病的医学影像，第二数据集中的第二数据为罕见病的医学影像为例，进行示意性的说明，图6是本申请一个示例性实施例提供的罕见病分类识别模型的训练过程整体示意图。
如图6所示，该过程中包括无监督训练阶段610、监督训练阶段620、伪标签生成阶段630以及第二分类模型的训练阶段640。
其中，在无监督训练阶段610中，将无标签标注的常见病医学影像611进行两次图像增强得到x q和x k，并通过查询编码器612和键值编码器613进行损失值的确定，从而完成对查询编码器612的训练，冻结查询编码器612的参数，并将无监督训练后的查询编码器612在监督训练阶段620应用于与分类回归模型621的连接。
在监督训练阶段620中，当查询编码器612与分类回归模型621连接后，得到待训练的第一分类模型622，通过标注有样本标签的罕见病的医学影像623对第一分类模型622进行监督训练时，根据罕见病的医学影像623对应标注的样本标签以及第一分类模型622的分类结果确定损失值，并实现对第一分类模型622的监督训练。
在第一分类模型622训练完毕后，在伪标签生成阶段630，通过第一分类模型622对常见病的医学影像611进行分类识别，得到常见病的医学影像611对应的伪标签。
在第二分类模型的训练阶段640，根据常见病的医学影像611对应的伪标签，以及第二分类模型641的预测结果得到第一损失值，根据第二分类模型641中查询编码器642和键值编码器643的编码结果得到第二损失值，从而根据第一损失值和第二损失值确定总的损失值对第二分类模型641进行训练，得到罕见病分类识别模型。
表一给出了本申请的技术方案在皮肤病变分类数据集上的结果对比。此数据集包含7个类别，将病例数量最多的四个类别的数据集作为第一数据集，剩余三个类别的数据集作为第二数据集。评价指标选择了准确率(Accuracy)、统计学中用来衡量二分类模型精确度的指标F1score。
表一
Figure PCTCN2022090902-appb-000008
表一中，N代表测试类别数，K代表每个测试类别提供的有标签的图片数量，本技术方案分别对比了K为1，3，5的结果。将罕见病数据集中剩余的图像组成Q作为测试集用于性 能评估。
由表一可见，本技术方案的分类指标优于全部相关技术。本技术方案在基准模型的基础上加入自蒸馏，提升了准确率约1-2％，F1score约3-5％。从表一中可以观察到在K＝5时，本技术方案无需任何常见病数据集的标注，准确率即可达到81.16％。此结果验证了本方法的假设：通过将伪标签监督信息注入到表征学习过程中并充分利用大量无标注数据集学习能够更好地学习罕见疾病数据的表征及其分类器。
图7是本申请一个示例性实施例提供的数据分类识别装置的结构示意图，如图7所示，该装置包括如下部分：
获取模块710，用于获取第一数据集和第二数据集，所述第一数据集中包括第一数据，所述第二数据集中包括标注有样本标签的第二数据；
训练模块720，用于通过所述第一数据以无监督训练模式，以及通过所述第二数据以监督训练模式对候选分类模型进行训练，得到第一分类模型；
所述获取模块710，还用于获取第二分类模型，所述第二分类模型为模型参数待调整的分类模型；
所述训练模块720，还用于以所述第一分类模型对所述第一数据的第一预测结果为基准，基于所述第二分类模型对所述第一数据的第二预测结果对所述第二分类模型的所述模型参数进行调整，得到数据分类模型；
预测模块730，用于通过所述数据分类模型对目标数据进行分类预测，得到所述目标数据的分类结果。
在一个可选的实施例中，所述预测模块730，还用于通过所述第一分类模型对所述第一数据进行分类预测，得到第一预测结果；
所述预测模块730，还用于通过所述第二分类模型对第一数据进行分类预测，得到第二预测结果；
如图8所示，训练模块720，还包括：
调整单元721，用于基于第一预测结果与所述第二预测结果之间的差异对所述第二分类模型的所述模型参数进行调整。
在一个可选的实施例中，所述获取模块710，还用于获取所述第一预测结果的置信度参数；
所述调整单元721，还用于确定所述第一预测结果在所述置信度参数下与所述第二预测结果之间的差异，并基于所述差异对所述第二分类模型的所述模型参数进行调整。
在一个可选的实施例中，所述预测模块730，还用于通过所述第一分类模型对所述第一数据进行分类预测，得到所述第一数据对应目标分类集中分类的概率值；基于所述概率值从所述目标分类集中确定所述第一预测结果。
在一个可选的实施例中，所述第二分类模型中包括第一查询编码器和第一键值编码器；
所述装置还包括：
编码模块740，用于通过所述第一查询编码器对所述第一数据进行编码，得到第一编码结果；
所述获取模块710，还用于获取所述第一键值编码器对所述第一数据和第一预设动态字典中的数据进行编码的第二编码结果；
所述训练模块720，还用于基于所述第一编码结果与所述第二编码结果的差异对所述第二分类模型进行训练。
在一个可选的实施例中，所述训练模块720，还用于基于第一数据对候选分类模型中的特征提取网络进行无监督训练；将候选分类模型中的分类回归网络与经过无监督训练的所述特征提取网络结合，得到分类模型，所述分类回归网络用于在所述目标分类集中进行数据分类；
所述训练模块720，还用于通过所述第二数据集中的所述第二数据和所述样本标签对所述分类模型进行监督训练，得到所述第一分类模型。
在一个可选的实施例中，所述特征提取网络中包括第二查询编码器和第二键值编码器；
所述装置还包括：
编码模块740，用于通过所述第二查询编码器对所述第一数据进行编码，得到第三编码结果；
所述获取模块710，还用于获取所述第二键值编码器对所述第一数据和第二预设动态字典中的数据进行编码的第四编码结果；
所述训练模块720，还用于基于所述第三编码结果与所述第四编码结果的差异对所述特征提取网络进行无监督训练。
在一个可选的实施例中，所述训练模块720，还用于将所述分类回归网络与经过无监督训练的所述第二查询编码器连接，得到所述分类模型。
在一个可选的实施例中，所述获取模块710，还用于获取测试数据集，所述测试数据集中的测试数据用于对所述数据分类模型的训练效果进行测试；从所述测试数据集中获取所述目标数据，所述目标数据标注有参考分类信息；
所述预测模块730，还用于通过所述数据分类模型对目标数据进行分类预测，得到所述分类结果；
所述获取模块710，还用于基于所述参考分类信息和所述分类结果获取所述数据分类模型的训练效果数据。
综上所述，本实施例提供的数据分类识别装置，在通过无标签的第一数据进行无监督训练以及有标签的第二数据进行监督训练后，得到第一分类模型，从而在第一分类模型的基础上，创建第二分类模型进行知识蒸馏训练，利用教师模型进行监督训练来达到蒸馏的目的，最终得到更高性能和精度的学生模型，训练主要依赖大量的第一数据，而对有标签的第二数据的数据量要求较小，避免了对样本数据进行大量标注的繁琐过程，提高了数据分类模型的训练效率以及准确率。
需要说明的是：上述实施例提供的数据分类识别装置，仅以上述各功能模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能模块完成，即将设备的内部结构划分成不同的功能模块，以完成以上描述的全部或者部分功能。另外，上述实施例提供的数据分类识别装置与数据分类识别方法实施例属于同一构思，其具体实现过程详见方法实施例，这里不再赘述。
图9示出了本申请一个示例性实施例提供的服务器的结构示意图。
具体来讲：服务器900包括中央处理单元(Central Processing Unit，CPU)901、包括随机存取存储器(Random Access Memory，RAM)902和只读存储器(Read Only Memory，ROM)903的系统存储器904，以及连接系统存储器904和中央处理单元901的系统总线905。服务器900还包括用于存储操作系统913、应用程序914和其他程序模块915的大容量存储设备906。
大容量存储设备906通过连接到系统总线905的大容量存储控制器(未示出)连接到中央处理单元901。大容量存储设备906及其相关联的计算机可读介质为服务器900提供非易失性存储。也就是说，大容量存储设备906可以包括诸如硬盘或者紧凑型光盘只读存储器(Compact Disc Read Only Memory，CD-ROM)驱动器之类的计算机可读介质(未示出)。
不失一般性，计算机可读介质可以包括计算机存储介质和通信介质。计算机存储介质包括以用于存储诸如计算机可读指令、数据结构、程序模块或其他数据等信息的任何方法或技术实现的易失性和非易失性、可移动和不可移动介质。计算机存储介质包括RAM、ROM、可擦除可编程只读存储器(Erasable Programmable Read Only Memory，EPROM)、带电可擦 可编程只读存储器(Electrically Erasable Programmable Read Only Memory，EEPROM)、闪存或其他固态存储其技术，CD-ROM、数字通用光盘(Digital Versatile Disc，DVD)或其他光学存储、磁带盒、磁带、磁盘存储或其他磁性存储设备。当然，本领域技术人员可知计算机存储介质不局限于上述几种。上述的系统存储器904和大容量存储设备906可以统称为存储器。
根据本申请的各种实施例，服务器900还可以通过诸如因特网等网络连接到网络上的远程计算机运行。也即服务器900可以通过连接在系统总线905上的网络接口单元911连接到网络912，或者说，也可以使用网络接口单元911来连接到其他类型的网络或远程计算机系统(未示出)。
上述存储器还包括一个或者一个以上的程序，一个或者一个以上程序存储于存储器中，被配置由CPU执行。
本申请的实施例还提供了一种计算机设备，该计算机设备包括处理器和存储器，该存储器中存储有至少一条指令、至少一段程序、代码集或指令集，至少一条指令、至少一段程序、代码集或指令集由处理器加载并执行以实现上述各方法实施例提供的数据分类识别方法。
本申请的实施例还提供了一种计算机可读存储介质，该计算机可读存储介质上存储有至少一条指令、至少一段程序、代码集或指令集，至少一条指令、至少一段程序、代码集或指令集由处理器加载并执行，以实现上述各方法实施例提供的数据分类识别方法。
可选地，该计算机可读存储介质可以包括：只读存储器(ROM，Read Only Memory)、随机存取记忆体(RAM，Random Access Memory)、固态硬盘(SSD，Solid State Drives)或光盘等。其中，随机存取记忆体可以包括电阻式随机存取记忆体(ReRAM，Resistance Random Access Memory)和动态随机存取存储器(DRAM，Dynamic Random Access Memory)。上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。
本申请的实施例还提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述实施例中任一所述的数据分类识别方法。

Claims (16)

  1. 一种数据分类识别方法，其特征在于，应用于计算机设备，所述方法包括：
    获取第一数据集和第二数据集，所述第一数据集中包括第一数据，所述第二数据集中包括标注有样本标签的第二数据；
    通过所述第一数据以无监督训练模式，以及通过所述第二数据以监督训练模式对候选分类模型进行训练，得到第一分类模型；
    获取第二分类模型，所述第二分类模型为模型参数待调整的分类模型；
    以所述第一分类模型对所述第一数据的第一预测结果为基准，基于所述第二分类模型对所述第一数据的第二预测结果对所述第二分类模型的所述模型参数进行调整，得到数据分类模型；
    通过所述数据分类模型对目标数据进行分类预测，得到所述目标数据的分类结果。
  2. 根据权利要求1所述的方法，其特征在于，所述以所述第一分类模型对所述第一数据的预测结果为基准，基于所述第二分类模型对所述第一数据的预测结果对所述第二分类模型的所述模型参数进行调整，包括：
    通过所述第一分类模型对所述第一数据进行分类预测，得到所述第一预测结果；
    通过所述第二分类模型对所述第一数据进行分类预测，得到所述第二预测结果；
    基于所述第一预测结果与所述第二预测结果之间的差异对所述第二分类模型的所述模型参数进行调整。
  3. 根据权利要求2所述的方法，其特征在于，所述基于所述第一预测结果与所述第二预测结果之间的差异对所述第二分类模型的所述模型参数进行调整，包括：
    获取所述第一预测结果的置信度参数；
    确定所述第一预测结果在所述置信度参数下与所述第二预测结果之间的差异，并基于所述差异对所述第二分类模型的所述模型参数进行调整。
  4. 根据权利要求2所述的方法，其特征在于，所述通过所述第一分类模型对所述第一数据进行分类预测，得到所述第一预测结果，包括：
    通过所述第一分类模型对所述第一数据进行分类预测，得到所述第一数据对应目标分类集中分类的概率值；
    基于所述概率值从所述目标分类集中确定所述第一预测结果。
  5. 根据权利要求2所述的方法，其特征在于，所述第二分类模型中包括第一查询编码器和第一键值编码器；
    所述方法还包括：
    通过所述第一查询编码器对所述第一数据进行编码，得到第一编码结果；
    获取所述第一键值编码器对所述第一数据和第一预设动态字典中的数据进行编码的第二编码结果；
    基于所述第一编码结果与所述第二编码结果的差异对所述第二分类模型进行训练。
  6. 根据权利要求1至5任一所述的方法，其特征在于，所述通过所述第一数据以无监督训练模式，以及通过所述第二数据以监督训练模式对候选分类模型进行训练，得到第一分类模型，包括：
    基于所述第一数据对所述候选分类模型中的特征提取网络进行无监督训练；
    将所述候选分类模型中的分类回归网络与经过无监督训练的所述特征提取网络结合，得到分类模型，所述分类回归网络用于在所述目标分类集中进行数据分类；
    通过所述第二数据集中的所述第二数据和所述样本标签对所述分类模型进行监督训练，得到所述第一分类模型。
  7. 根据权利要求6所述的方法，其特征在于，所述特征提取网络中包括第二查询编码器和第二键值编码器；
    所述基于所述第一数据对所述候选分类模型中的特征提取网络进行无监督训练，包括：
    通过所述第二查询编码器对所述第一数据进行编码，得到第三编码结果；
    获取所述第二键值编码器对所述第一数据和第二预设动态字典中的数据进行编码的第四编码结果；
    基于所述第三编码结果与所述第四编码结果的差异对所述特征提取网络进行无监督训练。
  8. 根据权利要求7所述的方法，其特征在于，所述将所述候选分类模型中的分类回归网络与经过无监督训练的所述特征提取网络结合，包括：
    将所述分类回归网络与经过无监督训练的所述第二查询编码器连接，得到所述分类模型。
  9. 根据权利要求1至5任一所述的方法，其特征在于，所述通过所述数据分类模型对目标数据进行分类预测，得到所述目标数据的分类结果，包括：
    获取测试数据集，所述测试数据集中的测试数据用于对所述数据分类模型的训练效果进行测试；
    从所述测试数据集中获取所述目标数据，所述目标数据标注有参考分类信息；
    通过所述数据分类模型对目标数据进行分类预测，得到所述分类结果；
    基于所述参考分类信息和所述分类结果获取所述数据分类模型的训练效果数据。
  10. 一种数据分类识别装置，其特征在于，所述装置包括：
    获取模块，用于获取第一数据集和第二数据集，所述第一数据集中包括第一数据，所述第二数据集中包括标注有样本标签的第二数据；
    训练模块，用于通过所述第一数据以无监督训练模式，以及通过所述第二数据以监督训练模式对候选分类模型进行训练，得到第一分类模型；
    所述获取模块，还用于获取第二分类模型，所述第二分类模型为模型参数待调整的分类模型；
    所述训练模块，还用于以所述第一分类模型对所述第一数据的第一预测结果为基准，基于所述第二分类模型对所述第一数据的第二预测结果对所述第二分类模型的所述模型参数进行调整，得到数据分类模型；
    预测模块，用于通过所述数据分类模型对目标数据进行分类预测，得到所述目标数据的分类结果。
  11. 根据权利要求10所述的装置，其特征在于，所述预测模块，还用于通过所述第一分类模型对所述第一数据进行分类预测，得到所述第一预测结果；
    所述预测模块，还用于通过所述第二分类模型对所述第一数据进行分类预测，得到所述第二预测结果；
    所述训练模块，还包括：
    调整单元，用于基于所述第一预测结果与所述第二预测结果之间的差异对所述第二分类模型的所述模型参数进行调整。
  12. 根据权利要求11所述的装置，其特征在于，所述获取模块，还用于获取所述第一预测结果的置信度参数；
    所述调整单元，还用于确定所述第一预测结果在所述置信度参数下与所述第二预测结果之间的差异，并基于所述差异对所述第二分类模型的所述模型参数进行调整。
  13. 根据权利要求11所述的装置，其特征在于，所述预测模块，还用于通过所述第一分类模型对所述第一数据进行分类预测，得到所述第一数据对应目标分类集中分类的概率值；基于所述概率值从所述目标分类集中确定所述第一预测结果。
  14. 一种计算机设备，其特征在于，所述计算机设备包括处理器和存储器，所述存储器中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由所述处理器加载并执行以实现如权利要求1至9任一所述的数据分类识别方法。
  15. 一种计算机可读存储介质，其特征在于，所述存储介质中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由处理器加载并执行以实现如权利要求1至9任一所述的数据分类识别方法。
  16. 一种计算机程序产品，其特征在于，包括计算机指令，所述计算机指令被处理器执行时实现如权利要求1至9任一所述的数据分类识别方法。
PCT/CN2022/090902 2021-05-17 2022-05-05 数据分类识别方法、装置、设备、介质及程序产品 WO2022242459A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/077,709 US20230105590A1 (en) 2021-05-17 2022-12-08 Data classification and recognition method and apparatus, device, and medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110532246.3A CN112949786B (zh) 2021-05-17 2021-05-17 数据分类识别方法、装置、设备及可读存储介质
CN202110532246.3 2021-05-17

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/077,709 Continuation US20230105590A1 (en) 2021-05-17 2022-12-08 Data classification and recognition method and apparatus, device, and medium

Publications (1)

Publication Number Publication Date
WO2022242459A1 true WO2022242459A1 (zh) 2022-11-24

Family

ID=76233883

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/090902 WO2022242459A1 (zh) 2021-05-17 2022-05-05 数据分类识别方法、装置、设备、介质及程序产品

Country Status (3)

Country Link
US (1) US20230105590A1 (zh)
CN (1) CN112949786B (zh)
WO (1) WO2022242459A1 (zh)

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112949786B (zh) * 2021-05-17 2021-08-06 腾讯科技（深圳）有限公司 数据分类识别方法、装置、设备及可读存储介质
CN113488023B (zh) * 2021-07-07 2022-06-14 合肥讯飞数码科技有限公司 一种语种识别模型构建方法、语种识别方法
CN113610111B (zh) * 2021-07-08 2023-11-03 中南民族大学 分布式多源数据的融合方法、装置、设备及存储介质
CN113645063B (zh) * 2021-07-16 2024-03-19 上海德衡数据科技有限公司 基于边缘计算的智能集成数据的方法及系统
CN113822339B (zh) * 2021-08-27 2024-05-31 北京工业大学 一种自知识蒸馏和无监督方法相结合的自然图像分类方法
CN114090770B (zh) * 2021-10-19 2022-11-04 杭州电子科技大学 一种多阶段的无监督域适应因果关系识别方法
CN114169392A (zh) * 2021-10-29 2022-03-11 阿里巴巴（中国）有限公司 模型训练方法及装置、任务处理方法、存储介质和处理器
CN113919499A (zh) * 2021-11-24 2022-01-11 威盛电子股份有限公司 模型训练方法与模型训练系统
CN114186097A (zh) * 2021-12-10 2022-03-15 北京百度网讯科技有限公司 用于训练模型的方法和装置
US20230245450A1 (en) * 2022-02-03 2023-08-03 Robert Bosch Gmbh Learning semantic segmentation models in the absence of a portion of class labels
CN114528937A (zh) * 2022-02-18 2022-05-24 支付宝(杭州)信息技术有限公司 模型训练方法、装置、设备及系统
CN114626520B (zh) * 2022-03-01 2024-05-10 腾讯科技（深圳）有限公司 训练模型的方法、装置、设备以及存储介质
CN115331088B (zh) * 2022-10-13 2023-01-03 南京航空航天大学 基于带有噪声和不平衡的类标签的鲁棒学习方法
CN116132527B (zh) * 2023-04-13 2023-06-16 深圳柯赛标识智能科技有限公司 管理指示牌的系统、方法及数据处理服务器
CN116934709B (zh) * 2023-07-20 2024-04-02 北京长木谷医疗科技股份有限公司 一种基于弱监督学习的脊柱滑脱智能识别方法及装置
CN116681123B (zh) * 2023-07-31 2023-11-14 福思(杭州)智能科技有限公司 感知模型训练方法、装置、计算机设备和存储介质
CN117195951B (zh) * 2023-09-22 2024-04-16 东南大学 一种基于架构搜索和自知识蒸馏的学习基因继承方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200110596A1 (en) * 2018-10-09 2020-04-09 Here Global B.V. Method and apparatus for identifying abandoned applications and services
CN111522958A (zh) * 2020-05-28 2020-08-11 泰康保险集团股份有限公司 文本分类方法和装置
CN112184508A (zh) * 2020-10-13 2021-01-05 上海依图网络科技有限公司 一种用于图像处理的学生模型的训练方法及装置
CN112347261A (zh) * 2020-12-07 2021-02-09 携程计算机技术（上海）有限公司 分类模型训练方法、系统、设备及存储介质
CN112686046A (zh) * 2021-01-06 2021-04-20 上海明略人工智能(集团)有限公司 模型训练方法、装置、设备及计算机可读介质
CN112949786A (zh) * 2021-05-17 2021-06-11 腾讯科技（深圳）有限公司 数据分类识别方法、装置、设备及可读存储介质

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP7022195B2 (ja) * 2018-02-28 2022-02-17 富士フイルム株式会社 機械学習装置、方法およびプログラム並びに記録媒体
US10685265B2 (en) * 2018-04-11 2020-06-16 International Business Machines Corporation Cognitive analysis and classification of apparel images
CN110852426B (zh) * 2019-11-19 2023-03-24 成都晓多科技有限公司 基于知识蒸馏的预训练模型集成加速方法及装置
CN111160553B (zh) * 2019-12-23 2022-10-25 中国人民解放军军事科学院国防科技创新研究院 一种新的领域自适应学习方法
CN111950638B (zh) * 2020-08-14 2024-02-06 厦门美图之家科技有限公司 基于模型蒸馏的图像分类方法、装置和电子设备
CN112232397A (zh) * 2020-09-30 2021-01-15 上海眼控科技股份有限公司 图像分类模型的知识蒸馏方法、装置和计算机设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200110596A1 (en) * 2018-10-09 2020-04-09 Here Global B.V. Method and apparatus for identifying abandoned applications and services
CN111522958A (zh) * 2020-05-28 2020-08-11 泰康保险集团股份有限公司 文本分类方法和装置
CN112184508A (zh) * 2020-10-13 2021-01-05 上海依图网络科技有限公司 一种用于图像处理的学生模型的训练方法及装置
CN112347261A (zh) * 2020-12-07 2021-02-09 携程计算机技术（上海）有限公司 分类模型训练方法、系统、设备及存储介质
CN112686046A (zh) * 2021-01-06 2021-04-20 上海明略人工智能(集团)有限公司 模型训练方法、装置、设备及计算机可读介质
CN112949786A (zh) * 2021-05-17 2021-06-11 腾讯科技（深圳）有限公司 数据分类识别方法、装置、设备及可读存储介质

Also Published As

Publication number Publication date
CN112949786A (zh) 2021-06-11
CN112949786B (zh) 2021-08-06
US20230105590A1 (en) 2023-04-06

Similar Documents

Publication Publication Date Title
WO2022242459A1 (zh) 数据分类识别方法、装置、设备、介质及程序产品
CN111126574B (zh) 基于内镜图像对机器学习模型进行训练的方法、装置和存储介质
Chen et al. Deep feature learning for medical image analysis with convolutional autoencoder neural network
CN109583332B (zh) 人脸识别方法、人脸识别系统、介质及电子设备
US10671895B2 (en) Automated selection of subjectively best image frames from burst captured image sequences
WO2020048389A1 (zh) 神经网络模型压缩方法、装置和计算机设备
CN110797101B (zh) 医学数据处理方法、装置、可读存储介质和计算机设备
US10984024B2 (en) Automatic processing of ambiguously labeled data
CN112784801A (zh) 基于文本和图片的双模态胃部疾病分类方法及装置
WO2020238353A1 (zh) 数据处理方法和装置、存储介质及电子装置
CN113516181B (zh) 一种数字病理图像的表征学习方法
CN113221983B (zh) 迁移学习模型的训练方法及装置、图像处理方法及装置
CN113707299A (zh) 基于问诊会话的辅助诊断方法、装置及计算机设备
CN114330499A (zh) 分类模型的训练方法、装置、设备、存储介质及程序产品
CN117036834B (zh) 基于人工智能的数据分类方法、装置及电子设备
CN114330482A (zh) 一种数据处理方法、装置及计算机可读存储介质
CN113920379A (zh) 一种基于知识辅助的零样本图像分类方法
CN113963793A (zh) 一种问诊方法及其相关设备
Viscaino et al. Computer-aided ear diagnosis system based on CNN-LSTM hybrid learning framework for video otoscopy examination
CN114764865A (zh) 数据分类模型训练方法、数据分类方法和装置
US20230325373A1 (en) Machine-learning based automated document integration into genealogical trees
CN116719840A (zh) 一种基于病历后结构化处理的医疗信息推送方法
CN115100723A (zh) 面色分类方法、装置、计算机可读程序介质及电子设备
CN116596836A (zh) 基于多视图邻域证据熵的肺炎ct影像属性约简方法
CN111582404B (zh) 内容分类方法、装置及可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22803783

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE